AI's $600B Question
sequoiacap.com
sequoiacap.com
Meta has about 350,000 of these GPUs and a whole bunch of A100s. This means the ability to train 50 GPT-4 scale models every 90 days or 200 such models per year.
This level of overkill suggests to me that the core models will be commoditized to oblivion, making the actual profit margins from AI-centric companies close to 0, especially if Microsoft and Meta keep giving away these models for free.
This is actually terrible for investors, but amazing for builders (ironically).
The real value methinks is actually over the control of proprietary data used for training which is the single most important factor for model output quality. And this is actually as much an issue for copyright lawyers rather than software engineers once the big regulatory hammers start dropping to protect American workers.
If there's dedicated inferencing silicon (like say the thing created by Groq), all those GPUs will be power sucking liabilities, and then the REAL singularity superintelligence level training can begin.
The precisions and mantissa/exponent ratios you want for inference are just different to a mixed-precision, fault tolerant, model and data parallel pipeline.
Hopper is for training mega-huge attention decoders: TF32, bfloat16, hot paths to the SRAM end of the cache hierarchy with cache coherency semantics that you can reason about. Parity gear for fault tolerance, it’s just a different game.
And then someone else starts giving away shovels for free.
And their business model is shovel-fleet logistics and maintenance... :p
Ah, I see -- it's more like a "level 2 gold rush".
So a level 1 gold rush is: There's some gold in the ground, nobody knows where it is, so loads of people buy random bits of land for the chance to get rich. Most people lose, a handful of people win big. But the retailers buying shovels at wholesale and selling them at a premium make a safe, tidy profit.
But now that so many people know the maxim, "In a gold rush, sell shovels", there's now a level 2 gold rush: A rush to serve the miners rushing to find the gold. So loads of retailers buy loads and loads of shovels and set up shop in various places, hoping the miners will come. Probably some miners will come, and perhaps those retailers will make a profit; but not nearly as much as they expect, because there's guaranteed to be competition. But the company making the shovels and selling them at a premium makes a tidy profit.
So NVIDIA in this story is the manufacturer selling shovels to retailers; and all the companies building out massive GPU clouds are the retailers rushing to serve miners. NVIDIA is guaranteed to make a healthy profit off the GPU cloud rush as long as they play their cards right (and they've always done a pretty decent job of that in the past); but the vast majority of those rushing to build GPU clouds are going to lose their shirts.
https://hai.stanford.edu/news/ai-trial-legal-models-hallucin...
Largely depends on how much money the client has.
Human imperfections are a family of failure-modes which have a gajillion years of experience in detecting, analyzing, preventing, and repairing. Quirks in ML models... not so much.
A quick thought-experiment to illustrate the difference: Imagine there's a self-driving car that is exactly half as likely to cause death or injury than a human driver. That's a good failure rate. The twist is that its major failure mode is totally alien, where units attempt to inexplicably chase-murder random pedestrians. It would be difficult to get people to accept that tradeoff.
It's one thing if a human makes a wrong financial decision or a wrong driving decision, it's another thing if a model distributed to ten million computers in the world makes that decision five million times in one second before you can notice it's happening.
It's why if your coworker makes a weird noise you ask what's wrong, if the industrial furnace you stand next to makes a weird noise you take a few steps back.
Human lawyers fail by not being very zealous and most of them being very average, not having enough time to spend on any filings, and not having sufficient research skills. So really, depth-of-knowledge and talent. They generally won't get things wrong per se, but just won't find a good answer.
AI gets it wrong by just making up whole cases that it wishes existed to match the arguments it came up with, or that you are hinting that you want, perhaps subconsciously. AI just wants to "please you" and creates something to fit. Its depth-of-knowledge is unreal, its "talent" is unreal, but it has to be checked over.
It's the same arguments with AI computer code. I had AI create some amazing functions last night but it kept hallucinating the name of a method call that didn't exist. Luckily with code it's more obvious to spot an error like that because it simply won't compile, and in this case I got luckier than usual, in that the correct function did exist under another name.
A good real estate agent can guide people through this process while advising them on selling at the right price while avoiding the most stress often during an extremely difficult time in their life, such as going through divorce of breakup. They of course also help keep buyers interested while the seller is making up their mind about the correct offer to take.
I find your comment ignorant in so many ways. Maybe have some respect?
It takes a long time for cultures to shift and for people to start to trust information systems to entirely replace high touch stuff like that. And at some level there will always be some white glove service on top for special cases.
but for long time in the US you were "forced" to hire a real estate agent, if you wanted to get the market price.
Refer to the NAR settlement that pretty much admits to this.
https://www.realestatecommissionlitigation.com/
This is not to say that real estate agents cannot add value to a process; it is just that they were a cartel with anticompetitive practices.
The mandated and fixed 6% on each sale was and is ridiculous, when the median sell price is 400K in the US ... that is 24K commission
simply put the cost of selling a home should not be linearly related to the cost of the house,
and especially should not be a fixed constant across the entire country
When NAR settled the price collusion charge? Thus cartel or not, times do change.
The same thinking stopped many legacy tech companies from becoming a “cloud” company ~20 years ago.
Fast forward to today and the margin for cloud compute is still obscene. And they all wish in hindsight they got into the cloud business way sooner than they ultimately did.
> Plus you have to pay for your own control plane when that’s already baked into the cloud provider’s charge model.
There is a hypothetical "but what if we honestly actually really really do", but that's such a waste of engineering time when there are so many other problems to be solved that it's implausible. The only time multi-cloud makes sense is when you have to meet customers where they're at, and have resources in whichever cloud your customers are using. Or if you're running arbitrage between the clouds and are reselling compute.
What ended up happening was Amazon was better at scale and lockin than everyone else. They gave Netflix a sweet deal and used it as a massive advertisement. It ended up being a rock rolling down a hill and all the competitors except ones with very deep pockets and the ability to cross-subsidize from other businesses (MSFT and Google) got crushed.
I thought Nvidia recently took that crown recently though.
And then you come to companies that managed to streamline both and ran out of floor space in their data center because they had to hold onto assets for 3-5 years. At one previous employer, the smallest orderable unit of compute was a 44U rack. They eventually filled the primary data center and then it took them 2 years to Tetris their way out of it.
Have we been living in the same universe the last 10 years? I don't see this ever happening. Related recent news (literally posted yesterday) https://www.axios.com/2024/07/02/chevron-scotus-biden-cyber-...
But there's bigger fish to fry for American politics and worker obsolescence is not really top of mind for anyone.
Red state blue collar workers got their candidate to pass tariffs. What happens when both blue state white collar workers and red state blue collar workers need to contest with AI. Perhaps not within the next 10 years, but certainly within 20 years!
And if you think 20 years is a long time... 2004 was when Halo 2 came out
But agreed, between the unions with political pull and "AI safety" grifters I suspect there could be some level of regulatory risk, particularly for the megacorps in California. I doubt it will be some national thing in the US absent a major political upheaval. Definitely possible in the EU which will probably just be a price passed on to customers or reduced access, but that's nothing new for them.
Tucker Carlson at one point said if FSD was going to take away trucking jobs we should stop that with regulation.
But in the general sense, I think it's tautologically correct to say better models always lead to better predictions, which always give an edge in competitions on an individual or societal level. So long term I do believe learning trumps ignorance, not in all cases but on average.
I don't know what power you imagine SWEs and PhDs posses, but the last time their employers flexed their power by firing them in droves (despite record profits); the employees sure seemed powerless, and society shrugged it off and/or expressed barely-concealed schadenfreude.
That settlement favored Apple, Google and the other conspirators because they only paid out a fraction of what they would have paid in salaries absent the collusion - so the settlement was not exactly a show of force by the engineers. Additionally, this was after a judge had thrown out a lower settlement amount the lawyers representing the class had agreed to.
> What happens when both blue state white collar workers and red state blue collar workers need to contest with AI. Perhaps not within the next 10 years, but certainly within 20 years!
Populism. Probably the fascist right-wing kind, but I expect some form of populism. Related, if we're talking about a 20 year time horizon, I'm genuinely unsure if society will still exist in any recognizable fashion at the rate we're going...
> Think of how cancer could have been cured a decade ago if information was allowed to flow freely from the 50's forward
might be a bit fanciful? Unless you're referring to something particular I'm unaware of.
The people best equipped and trained to deliver a cure for cancer (and then some, since it tends not to be particularly field-restricred) do have access.
I think the loss is more likely in engineering (to the publication's science), cheaper methods, more reliably manufacturable versions of lab prototypes, etc.
I doubt there are many people capable of cancer research breakthroughs who don't have access to cancer research, personally.
(And to be clear: I'm not capable of it.)
The schools I’ve worked with have access to everything I’ve needed. They didn’t advertise it but it’s also free for students.
What IS a huge problem is the almost complete lack of systematically acquired quantitative data on human health (and diseases) for a very large number (1 million subjects) of diverse humans WITH multiple deep-tissue biopsies (yes, essentially impossible) that srr suitable for multiomics at many ages/stages and across many environments. (Note, we can do this using mice.)
Some specific examples/questions to drive this point home: What is the largest study of mRNA expression in humans? ANSWER: The small but very expensive NUH GTEx study (n max of about 1000 Americans). This study acquired postmortem biopsies for just over 50 tissues. And what is the largest study of protein expression in humans across tissues? Oh sorry, this has never been done although we know proteins are the work-horses of life. What about lipids, metabolites, metagenomics, epigenomics? Sorry again, there is no systematically acquired data at all.
What we have instead is a very large cottage-industry of lab-level studies that are structurally incoherent.
Some brag about the massive biomedical data we have, but it is truly a ghost and most real data evaporates with a few years.
Here is my rant on fundamental data design flaws and fundamental data integration flaws in biomedical research:
Herding Cats: The Sociology of Data Integration https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2751652/
But I also think the GP's claim and yours are not incompatible. I wonder how much survivorship bias this has since it only considers those that are able to do research, and not those that would have but ended up doing continuing with another STEM job. We could be asking the counterfactual that I think the GP is implying: would more people have been interested in becoming cancer researchers if publications were open?
We can sort of see the effect because we have scihub now, which basically unlocks journal access for those that are comfortable with it, and I consider it plausibly having a significant effect for the population that have a research background without an academic affiliation. I've met a few biotech startup founders that switched from tech to bio and did self study+scihub outside of the university. The impetus for change I've heard a few times is a loved one got X disease, and I studied it, quit my less impactful tech job to work on bio stuff.
2) There may be a few researchers who don't have unfettered access. Perhaps they paid $40 for a copy of a paper. Given the high cost of other parts of research labs, I find it hard to believe that any real possibility of curing cancer was halted because someone had to pay $40.
3) It's possible to imagine the opposite being the case. Perhaps someone had a key insight in a clever paper and decided to distribute it for free out of some info-anarchistic impulses. There it would sit in some FTP directory uncrated, unindexed and uncared for. Perhaps the right eyes would find it. Perhaps they wouldn't. Perhaps the cancer researcher would be able to handle all of the LaTeX and FTP chores without slowing down research. Perhaps they would be distracted by sys admin headaches and never make a crucial follow up discovery.
The copyrighted journal system provides curation and organization. Is it wonderful? Nah. Is it better than some ad hoc collection of FTP directories? Yes!
Your opinion may be that this scenario would never happen. In my opinion, this is more likely than your vision.
[0] https://en.wikipedia.org/wiki/Slavery_Abolition_Act_1833
I agree though that the returns on hardware rapidly diminish.
> I think it was because we were working on Reels. We always want to have enough capacity to build something that we can't quite see on the horizon yet. ... So let's order enough GPUs to do what we need to do on Reels and ranking content and feed. But let's also double that.
So there's an immense capacity inside Meta, but the _whole_ fleet isn't available for LLM training.
[0]: https://www.dwarkeshpatel.com/p/mark-zuckerberg?open=false#§...
Not anywhere close to that.
Those 350k GPUs you talk about aren't linked together. They also definitely aren't all H100s.
To train a GPT-4 scale model you need a single cluster, where all the GPUs are tightly linked together. At the scale of 20k+ GPUs, the price you pay in networking to link those GPUs is basically almost the same as the price of those GPUs themselves. It's really hard and expensive to do.
FB has maybe 2 such clusters, not more than that. And I'm somewhat confident one of those cluster is an A100 cluster.
So they can train maybe 6 GPT-4 every 90 days.
340,000 H100s 600,000 H100 equivalents (perhaps AMD Instinct cards?) On top of the hundreds of thousands of legacy A100s.
And I'm certain the order for B100s will be big. Very big.
Even the philanthropic org Chan-Zuckerberg institute current rocks 1000 H100s, probably none used for inference.
They are going ALL OUT
What do they use them for?
Just like they did for their metaverse play, and that didn't work out very well.
This was even true in Star Trek. People could do literally anything on a holodeck and the writers still had them going to Risa for a holiday.
There is no chance of VR going mainstream until someone solves the fundamental human problem of people preferring to do things in real life.
I don't think that's much of a problem? People already watch TV and play computers games and read novels, instead of real life.
I agree that VR has _some_ problem, but I don't think it's that people prefer real life.
Now, there may also be a physical laziness factor to overcome, but there are enough people that enjoy moving their bodies to really explode the industry even if all the lazy folks stay 2D.
You can't interact with people around you the same way you do if you play with, say, a console controller. VR is an all encompassing activity that you have to dedicated time for, instead of having it casually just exist around you. Then we have the cost. Only some people can have it, so it will be a lonely activity most of the time when it could be so much more.
I can afford it, but every time I am in front of a new set, I consider my life with it and say "maybe next time". Finally, I have not really explored them, but I have a feeling the experience is limited by the content that exists.
I dream of a VR experience where suddenly all content I currently enjoy on flat screens will automagically be VRified. But I am pretty sure that will not be the case. Only a very limited collection will be VR native.
But I want it all to be, or almost all, before I go all in.
It would be a bit better suited to a civilization that wasn't undergoing a catastrophic urban housing shortage crisis with demographic & economic effects for upcoming generations that are comparable to a world war or the Black Death. We are building huge exurban houses which nonetheless do not have VR-appropriate rooms, and tiny 1-bedroom apartments, and not much else. https://www.youtube.com/watch?v=4ZxzBcxB7Zc
The question is whether this is a chicken/egg problem that prevents us from launching next-generation VR plays.
But it’s too expensive and still too heavy on your face.
I bought the somewhat dated Quest 2 sometime in the last year, because I could get it for a good price. There's a mobile app I think I used for setup, you can also connect to the PC for Oculus Link, SteamVR, Virtual Desktop or any other number of OpenXR apps or games, but as for the device itself... there was basically nothing aside from logging in and downloading what I want from the store, if I wanted to run things directly on the headset. The controllers and tracking just works, you define the area you want to get warned about getting close to the borders of by just drawing in the room around you.
Actually, the only problems I've had have been in a PCVR use case after I got an Intel Arc - Oculus Link and SteamVR both don't support it natively (an allowlist in the case of the former and support only for NVENC I think in the case of the latter), whereas Virtual Desktop worked with AV1 and Intel QSV out of the box, while also allowing me to launch SteamVR through it.
There are warts and all (especially software like Immersed removing support for physical monitors, what were they thinking), but in general the hardware and everything around it, even hand tracking, are pretty well streamlined, surprisingly so.
> VR is an all encompassing activity that you have to dedicated time for, instead of having it casually just exist around you.
This kind of killed it for me, to be honest. There's more friction than just launching a game on the PC directly (in the case of PCVR: putting the headset on, connecting to the PC, then launching it on the PC, finally accessing it on the headset) in addition to needing to sometimes use the keyboard being especially annoying, since the on screen keyboard is just more annoying to use and having to find your regular keyboard taking a step or two, if you're standing instead of sitting while playing.
That said, VR in general still feels cool, even if it's a bit early.
If anything, that was a failure of imagination on writers' part, somewhat rectified over time and subsequent shows. Even in the core shows (TNG, DS9, VOY), we've seen the holodeck used for recreation, dating, study, simulation, brainstorming, physical training, hand-to-hand combat training, marksmanship training, gaming out scenarios for dangerous missions, extreme sports, engineering, accident investigation, crime scene reconstruction, and a bunch of other things. Still, the show was about boldly going where no one has gone before - not about boldly staying glued to a virtual reality display - so this affected what was or wasn't shown.
Plus, it's not either/or. People went to Risa to have real sex (er, jamaharon) with real people, and lots of (both it and them). This wasn't shown on-screen, just heavily implied, as this is where Roddenberry's vision of liberated humanity clashed with what could be shown on daytime TV. Holo-sex was a thing too, but it was shown to be treated more like pornography today - people do it, don't talk much about it, but if you try to do it too often and/or with facsimile of people you know, everyone gets seriously creeped out.
There was a similar episode of DS9 where Nog gets addicted to the holodeck due to war trauma.
The central take of the show is that real life is better for these people in this future communist space utopia and the only reason why you'd go to the holodeck is light entertainment, physical training, or if there's something wrong with your life that needs fixed.
Ready Player One had a pretty good answer to this: dystopia. Once real life is miserable enough, VR's time will have arrived.
Three counterpoints: Online gaming, social media, smartphones. All of these favor "virtual" over "real life", and have become massively popular over the last decades. Especially among the young, so the trend is likely to continue.
We’ve joked multiple times that whenever the “co-op Skyrim VR” of gaming comes out, we will never be heard from again lol.
Apple is SO CLOSE to this with its Vision Pro hardware, and yet so far… (no “co-op space” implementation, too expensive, too heavy on face)
Imagine seeing a live soccer match in 3D from incredible camera angles like just above the goals, but your buddy who is 3000 miles away is actually also sitting right next to you in that space, and you can see and hear each other…
Sounds like more unsubstantiated hype from a company desperate to sell a product that was very expensive to build. I guess we'll see, but I'm not optimistic for them.
Maybe. But we've barely scratched the surface of being more economical with data.
I remember back in the old days, there was lots of work on eg dropout and data augmentation etc. We haven't seen too much of that with the like of ChatGPT yet.
I'm also curious to see what the future of multimodal models holds: you can create almost arbitrarily amounts of extra data by pointing a webcam at the world, especially when combined with a robot, or letting your models also play StarCraft or Diplomacy against each other.
The US Supreme Court seems determined to make sure that big regulatory hammers are not going to be dropping, from what I can tell.
Could you expand on this? Who are "the builders" here? You mean the model developers? I don't see how this situation can be "amazing" for the builders - developers will just get a wage out of their work.
What it actually means is that they are training next gen models that are 50X larger.
And, considering MS and OpenAI are planning to build a $100 billion AI training computer, these 350K GPUs is just a tiny portion of what they are planning.
This isn't an overkill. This is the current plan: throw as much compute as possible and hope intelligence scales with compute.
I put this as equivalent to investing in Sun Microsystems and Netscape in the late 90s. We knew the internet was going to change the world, and we were right, but we were completely wrong as to how, and where the money would flow.
Sun Microsystems sold to Oracle for $7B, and Netscape was acquired by AOL for $10B.
So will Nvidia be worth $5 trillion AFTER the AI bust?
Cisco too.
Highly debatable.
When we look back during the internet and mobile waves it is overwhelmingly the companies that came in after the hype cycle had died that have been enduring.
Microsoft Office: wasn't close to the first office editing suite
Google: Wasn't close to the first search engine
Facebook: Wasn't close to the first social media website
Apple: ~~First "smart phone"~~ but not the first personal computer. Comments reminded me that it wasn't the first smartphone
Netflix: Wasn't close to the first video rental service.
Amazon: Wasn't close to the first web store
None of the big five were first in their dominate categories. They were first to offer some gimmick (i.e., google was fast, netflix was by mail, no late fees), but not first categorically.
Though they certainly did benefit from learnings of those that came before them.
Was it the first smartphone? I would call phones like the Palm Treo and later BlackBerries smartphones. There were even apps, but everything was lot more locked down and a lot more expensive.
And just plain... bad. The entire experience didn't have that "feel" that Apple turned into reality. It's comparable to today's AI landscape—the technology is pretty neat, but using it is a complete slog.
Edit: There were also the limitations of that era that held devices back in general. WAP internet[1] was awful, but most mobile services were too slow for much else.
[1] https://en.wikipedia.org/wiki/Wireless_Application_Protocol
The entire device was a regular Linux machine.
It's not just you; at the time these products were available, _everyone_ called them smartphones. Emphatically, Apple did not bring the first smartphone to market, not even close. They were, however, the first to popularize it beyond the field of nerds into the general public.
>it became a huge sales success for Nokia ... It managed to outsell rivals such as LG Viewty and iPhone.
However the iPhone got better.
The 770 was released in Q4 '05.
They definitely fell within the smartphone category, but oddly the first few iterations lacked GSM radio.
"key differentiator" and not necessarily easy to pull off or pay for
Just to quibble with this - that was not even close to the reason Google got popular. It was because Google was much, much better at finding what you actually wanted. It was just a far better product.
You can debate why this is exactly, Joel Spolsky pointed out many years ago that it was because Google got that what matters to users most isn't "finding all pages related to X" but rather "ranking" those pages, a take I agree with.
"Several studies have shown that pioneers have long-lived market share advantages and are likely to be market leaders in their product categories. However, that research has potential limitations: the reliance on a few established databases, the exclusion of nonsurvivors, and the use of single-informant self-reports for data collection. The authors of this study use an alternate method, historical analysis, to avoid these limitations. Approximately 500 brands in 50 product categories are analyzed. The results show that almost half of market pioneers fail and their mean market share is much lower than that found in other studies. Also, early market leaders have much greater long-term success and enter an average of 13 years after pioneers."
PDF available here:
https://people.duke.edu/~moorman/Marketing-Strategy-Seminar-...
> Founders and company builders will continue to build in AI—and they will be more likely to succeed, because they will benefit both from lower costs and from learnings accrued during this period of experimentation
This still lines up with the 2nd wave benefiting more. The first movers helped established the large scale AI hardware industry, got a bunch of smart kids trained on how to make AI, a bunch of people will fail and learn, etc and this experimentation stage sets the groundwork for OpenAI 2.0.
We could very well just in the Altavista vs Yahoo days of AI and an upstart takes over in 5yrs.
For the case where a company is using their own AI for their own cost reduction and productivity improvements, they can keep doing that but not offer to another party.
If they offer to another party, and that party is having benefits (like you have said), the price should be such that a part of the consumer benefit is shared with the producer resulting in benefits for the producer.
The real challenge here is because of price wars, i.e., too much competition already with producers willing to take a hit on profitability in anticipation that they will be able to do so later after creating a moat above and beyond competitors. Or they think that it will strenghen their overall bigger offering by adding an otherwise lossy feature.
In a nutshell, even if there's a lot of value for the consumers, it must result in a win-win for a new product to be sustainable in the market.
Well if there's value to you then how much did you pay for it and would it realistically cover operating cost once VC cash dries up? That's the only question.
> A huge amount of economic value is going to be created by AI. Company builders focused on delivering value to end users will be rewarded handsomely.
Such strong speculative predictions about the future, with no evidence. How can anyone be so certain about this? Do they have some kind of crystal ball? Later in the article they even admit that this is another one of tech's all-too-familiar "Speculative frenzies."
The whole AI thing just continues to baffle me. It's like everyone is in the same trance and simply assuming and chanting over and over that This Will Change Everything, just like previous technology hype cycles were surely going to Change Everything. I mean, we're seeing huge companies' entire product strategies changing overnight because We Must All Believe.
How can anyone speak definitively about what AI will do at this stage of the cycle?
I think doubt is OK, at least it is before any particular technology or product has actually proven itself.
Sometimes the future just gets here before we're ready for it.
Generative models are already changing how people live and work. Ignore the grifters, and ignore the entrepreneurs. Look at civilians, regular folks, and watch how it impacts them.
It doesn't have anything to do with the libertarian vision of BTC but it's the same technical concept.
It's like hiring someone to go to the gym for you, as far as I can see.
They are more intelligent than the average person I deal with on a daily basis.
The one thing us meat bags have going for us is that we have bodies and can do things.
No they aren't. They are differently intelligent. Which includes being more "intelligent" in some ways, and vastly less in others.
It is an all too common mistake to infer an LLM's capability based on what it would mean if a human could produce the same output. But it's not the same thing at all. A human that could produce the quality of many LLM outputs I've seen would be a person with superintelligence (and supercreativity, for that matter). But a human who is constrained to only the output that an LLM is capable of would be an untrustworthy idiot.
What's interesting to me is looking at the human+LLM combos that have recently entered the world. (As in, everyone who is regularly using good quality LLMs.) What are they capable of? Hopefully they'll be at least as intelligent as a regular human, though some of the articles on people blindly accepting LLM hallucinations do make me wonder.
Good luck with that genie
eBay, Amazon, Google, Yahoo etc were all around at the time and making serious money.
Not sure who those people were but it was very obvious to most that the internet was here to stay.
The problem with expert systems is that even if the tooling was perfect the people using them needed a rather nuanced and sophisticated understanding of ontologies. That just wasn’t going to happen. There is not enough of that kind of expertise to go around. Efforts to train people largely failed. I think the intentional undermining of developer salaries pushed a lot of smart people out of the software industry making the problem even worse.
That’s what makes AI special, the ability to deliver value even when used by unsophisticated operators. Many workflows can largely stay the same and AI can be sprinkled in where it makes the most sense. I use it for documentation writing and UI asset production and it’s better in that role than the people I used to pay.
The evidence is all around you. For anyone who has made any serious attempt to add AI to your current life and work process, you will fairly quickly notice that your productivity has doubled.
Now, do I as a random software engineer who is now producing higher quality code, twice as fast, know how to personally capture that value with a company? No. But the value is out there, for someone to capture.
> It's like everyone is in the same trance and simply assuming and repeating over and over that This Will Change Everything
It already is changing everything, in multiple fields. Go look up what happened to the online art commission market. It got obliterated over a year ago and is replaced by people getting images from midjourney/ect.
Furthermore, if you are a software engineer and you haven't included tools like github copilot, or cursor AI into your workflow yet, I simply don't consider you to be a serious engineer anymore. You've fallen behind.
And these facts are almost immediately obvious to anyone who has been paying attention in the startup space, at least.
Meanwhile, I have chatGPT open in background and go from unaware to informed for every new keyword I hear around me all day everyday. Not to mention annotating code, generating utlity functions, and tracing errors
I personally like co-pilot but I work across several languages and code bases where I seriously can’t remember how to do basic stuff. In those cases the automatic code generation from co-pilot speeds my efficiency, but it still can’t do anything actually useful aside from making me more productive.
I fully expect the tools to become “necessary” in making sure things like JSdoc and other domination is auto-updated when programmers alter something. Hell, if they become good enough at maintaining tests that would be amazing. So far there hasn’t been much improvement over the year we’ve used the tools though. Productivity isn’t even up across teams because too many developers put too much trust into what the LLMs tell them, which means we have far more cleanup to do than we did in the previous couple of years. I think we will handle this thing once we get our change management good enough at teaching people that LLMs aren’t necessarily more trustworthy than SO answers.
The best answer is for someone who has found ways to boost their own productivity, but also understands the caveats such as hallucinations and not pasting proprietary information into text boxes on the internet.
That sounds like you're fresh out of college. Copilot is great at scaffolding but doesn't do shit for bug fixing, design, or maintenance. How much scaffolding do you think a senior engineer does per week?
Sr. Eng adopted copilot and sung it's praises a lot faster then the jr engineers. Especially when working on codebases with less familiar languages.
Maybe take a look at tools like aider-chat with Claude 3.5 Sonnet. Or just have a discussion with gpt-4o about any programming area that you aren't particularly familiar with already.
Unless you literally decided you learned everything you need and don't try to solve new types of problems or use new (to you) platforms ever..
And yes cursor AI/copilot helps with bugs as well.
It works because when you have a bug/error message, instead of spending a bunch of time on Google/searching on stack overflow for the exact right answer, you can now do this:
"Hey AI. Here is my error message and stack trace. What part of the code could be causing it, and how should I fix it".
Even for debugging this is a massive speed up.
You can also ask the AI to just evaluate your code. Or explain it when you are trying to understand a new code base. Or lint it or format it. Or you can ask how it can be simplified or refactored or improved.
And every hour that you save not having to track down crazy bugs that might just be immediately solvable, is an hour that you can spend doing something else.
And that is without even getting into agents. I haven't figured out yet how to effectively use those yet, and even that is making me nervous/worried that I am missing some huge possible gains.
But sure, I'll agree that of all you are doing is making scaffolding, that is a fairly simply usecase.
That's not how I work since I stopped being a junior dev. I might google an error message/library combination if I don't understand it but in most cases, I just read the stacktrace and the docs, or maybe the code.
I don't doubt that LLMS can be quite useful when working with large, especially foreign, codebases. But I have yet to see the level of "if you don't use it you're not an engineer" some people like to throw around. To the contrary, I'd argue if you rely on an LLM to tell you what you should be doing, you aren't an engineer, you are a drone.
I just read the manual.
I once worked with someone who was brilliant, but fell apart when we tried to do pair-programming (acturial major who had moved into coding). The verbal communication overhead was too much for him.
I've always thought of software development as an inherently solo endeavor that happens entirely inside of one's own mind. When I'm faced with a software problem, I map out the data structures, data flows, algorithms and so on in my mind, and connect them together up there. Maybe taking some notes on a sheet of paper for very complex interactions. But I would not really think of sitting down with someone to "chat" about it. The act of articulating a question "What should this data structure look like and be composed of?" would take longer than it would take to simply build it and reason about it in my own brain. This idea that software is something we do in a group socially, with one or more people talking back and forth, is just not the way I operate.
Sure, when your software calls some other person's API, or when your system talks to someone else's system, or in general you are working on a team to build a large system, then you need to write documents and collaborate with them, and have this back-and-forth, but that's always kind of felt like a special case of programming to me.
The idea of asking ChatGPT to "write a method that performs a CRC32 on a block of data" seems silly to me, because it's just not how I would do it. I know how to write a CRC function, so I would just write it. The idea of asking ChatGPT to help write a program that shuffles a deck of cards and deals out hands of Poker is equally silly because before I even finished writing this sentence, I'm visualizing the proper data structures that will be used to represent the cards, the deck, and players' hands. I don't need someone (human or AI) to bounce ideas off of.
There's probably room for AI assistance for very, very junior programmers, who have not yet built up the capability of internally visualizing large systems. But for senior developers with more experience and capability, I'd expect the utility go down because we have already built out that skill.
Chat interface is annoying, though. Because it's natural language, I have to type a lot more, which is frustrating - but on the other hand, because it's natural language, I can just type my stream of thought and the LLM understands it. The two aspects cancel out each other, so in terms of efficiently, it's a wash.
We're nearing general quantum supremacy for practical jobs within the next 5 years. Autonomous cars are now rolling out in select geographies without safety drivers. And crypto is literally being pushed as an alternative for SWIFT among the BRICS, IMF, and BIS as a multi-national CBDC via the mBridge program.
I don't need a crystal ball for this. The impact is already evident for us early-adopters, it's just not evenly distributed yet.
That's not to say they're not OVER hyped - changing the entire company roadmap doesn't feel like a sensible path to me for most companies.
And/or, it's neither hard nor shameful to be True Believers, if what you believe in is plain fact.
Early adopters were using gpt-2 and telling us it was amazing.
I used it and it was completely shit and put me off openai for a good four years.
gpt-3 was nearly not shit, and 3.5 the same just a bit faster.
It wasn't until gpt-4 came out that I noticed that this AI thing should now be called AI because it was doing things that I didn't think I'd see in decades.
I started using GPT-3 via the playground UI for things like writing regular expressions. That's when this stuff began to get useful.
I've been using GPT-4 on an almost daily basis since it came out.
Turns out they were just grifters as the hilarious mess around Sam Altmans coup/counter coup/coup showed us.
We’ll talk counts when my grandma will be able to hey siri / okay google something like local hospital appointment or search for radish prices around her. It already is possible, just not integrated enough.
Coincidentally, I’m working on a tool at my job (unrelated to AI) that enables computer device automation on much higher level than playwright/etc. These two things combined will do miracles, for models good enough to use it.
I've just had GPT-4o write me a full-featured 2048 clone in ~6 hours of casual chat, in between of work, making dinner, and playing with kids; it cost me some $4 in OpenAI bills, and I didn't write a single line of code. I see non-tech people around me using ChatGPT for anything from comparison shopping to recipe adjustments. One person recently said to me that their dietitian is afraid for their career prospects because ChatGPT is already doing this job better than she is. This is a small fraction of cases in my family&friends circle; anyone who hasn't lived under the rock, or wasn't blinded by the memetic equivalent of looking at a nuclear weapon detonation, likely has a lot of similar things to say. And all of that is not will, it's is, right now.
You can't know this for certain until you look back on it in retrospect. We did not know mobile phones and e-commerce were going to be huge back in the 90s. We know now, of course, looking back, and the ones who guessed right back then can pat themselves on the back now.
Everyone is guessing. I'll admit it's totally possible LLMs and AI are going to be as earth shattering as its boosters claim it will be, but nobody can know this now with as much certainty as is being written.
It doesn't matter if it's general, what matters is that its useful. And if you don't find it useful just remember a lot of people in the 00s didn't find google useful either since they already had the yellow pages.
I strongly suggest paying for a subscription to either openai or anthropic and learning quickly.
Buy the subscription and use the turbo4 model.
After that api credits so you get access to the playground and change the system prompt. It makes a huge difference if you don't want to chat for 10 minutes before you get the result you want.
Learning what quickly exactly?
Correct, but the thing is, AI blown up much faster than phones - pretty much a decade in a single year, in comparison. Mobile phones weren't that useful early on, outside of niche cases. Generative AI is already spreading to every facet of peoples' lives, and has even greater bottom-up adoption among regular people, than top-down adoption in business.
I thought the same about adoption (across multiple audiences, not just tech workers and/or young people), was faced with surprising poor knowledge about GenAI when making surveys about it in my company. Maybe investors are asking the same questions right now.
https://www.pewresearch.org/short-reads/2024/03/26/americans...
23% said they'd used ChatGPT, 43% said they hadn't, 34% didn't know what it was.
> The Information recently reported that OpenAI’s revenue is now $3.4B, up from $1.6B in late 2023.
and links to
https://www.theinformation.com/articles/openais-annualized-r...
That's a lot of $20/month subscriptions. it's not all that but that's a lot of money, regardless.
There are a lot of companies building private GPTs and using their API.
What do you base that on though? Two years into the iPhone, Apple reported a $6.75b revenue on iPhone related sales. ChatGPT may reach or surpass that this year considering they're currently at $3.4b. That's not exactly what I would call growing faster than phones, however, and according to this article, very few people outside of nvidia and OpenAI are actually making big money on LLM's.
I do think it's silly to see this wave of AI to be referred to as the next blockchain, but I also think you may be hyping it a little beyond its current value. It being a fun and useful tool for a lot of things isn't necessarily the same thing at it being something that's actually worth the money investors are hoping it will be.
My childhood? I was a teen when mobile phones started to become widely used, and soon after pretty much necessary, in my part of the world. But, to reiterate:
> Two years into the iPhone, Apple reported a $6.75b revenue on iPhone related sales.
That's just an iteration, and not what I'm talking about. Smartphones were just different mobile phones. I'm talking about the adoption of a mobile phone as a personal device by general population.
> It being a fun and useful tool for a lot of things isn't necessarily the same thing at it being something that's actually worth the money investors are hoping it will be.
That's probably something which needs to be disentangled in these conversations. I personally don't care what investors think and do. AI may be hype for the VCs. It's not hype for regular Janes and Joes, who either already integrated ChatGPT into their daily lives, or see their friends doing so.
For a $399 device, Palm Pilot did well and had an excellent reputation for the time. Phones really took over the PDA market as a personal pocket-computer more-so than being used as ... a phone...
Really, I consider the modern smartphone a successor to the humble PDA. I grew up in that time too, and I remember the early Palm adopters having to explain why PDAs (and later Blackberries) were useful. That was already all figured out by the time iPhone took over.
And regular Janes and Joes are not using ChatGPT. Revenues would be 10-100x if that were the case.
3/4 of the people I know are actively using it are on free tier. And based on all the HN conversations in the last year, plenty of HNers commenting here are also using free tier. I'd never go back to GPT-3.5, but apparently most people find it useful enough to the point they're reluctant to pay that $20/month.
As for the rest, OpenAI is apparently the fastest-growing service of all time ever, so that says something.
A while back I used 3.5 to make a chat web page so I could get the better models as PAYG rather than subscription… and then OpenAI made it mostly pointless because they gave sufficient 4o access to the free tier to meet my needs.
Or they find it useless enough that they're unwilling to pay for the upgrade.
A phone on the go didn’t fundamentally alter anything except for making coordination while traveling easier. I went through both the cell phone adoption curve and the smartphone curve.
The latter was the massive impact that brought computing to the remaining 85% of the planet and upended targeting desktop operating systems for consumers by default.
Calling smartphones an iteration on cellphones is like calling ChatGPT an iteration on the neural networks we had 10 years ago.
They had tiny utility compared to modern smartphones but the first iPhone was a glorified iPod with a touchscreen and a cellular radio. It didn’t have an app store and the only thing it really did better than other mobile phones was web browsing, thanks to the touchscreen keyboard.
It wasn’t as revolutionary as hindsight now makes it seem. It was just an iteration on PalmPilots and Blackberries.
I had a blackberry before and it was just a glorified email and texting device.
It was immediately obvious how revolutionary the iPhone was. That’s why Android immediately pivoted hard to replicate the experience.
Isn't this a an odd take when you're discussing things on a VC website? In any case, if you like LLM's you probably should care considering it's the $10b Microsoft poured into OpenAI that's made the current landscape possible. Sure, most of those money were fuled directly into Azure because that's where OpenAI does all it's compute, but still.
> It's not hype for regular Janes and Joes, who either already integrated ChatGPT into their daily lives, or see their friends doing so.
Are they paying for it? If they aren't then will they pay for it? I think it's also interesting to view the numbers. ChatGPT had 1.6 billion visitors in January 2024, but it had 637 million in May 2024.
Again. I don't think it's all hype, I think it'll change the world, but maybe not as radically as some people expect. The way I currently view it is another tool in the automation tool-set. It's useful, but it's not decision making and because of the way it functions (which is essentially by being very good at being lucky) it can't be used for anything important. You really, really, wouldn't want your medical software to be written by a LLM's programmer. Which doesn't necessarily change the world too much because you really, really, didn't want it to be written by a search engine programmer either. On the flip-side, you can actually use ChatGPT to make a lot of things and be just fine. Because 90% (and this a number I've pulled out my ass, but from my anecdotal experience it's fairly accurate) of software doesn't actually require quality, fault tolerance or efficiency.
Even when cell phones started getting popular, often only one or two family members would get one. The transition time between “started becoming popular” and “everyone has one” was >5 years and even then it was relatively normal that people would just turn off their cell phone for a few days (to mixed reactions from friends and family).
>What do you base that on though? Two years into the iPhone, Apple reported a $6.75b revenue on iPhone related sales. ChatGPT may reach or surpass that this year considering they're currently at $3.4b.
But the iPhone was launched more than 10 years past mobile phones (in fact, more than 20, but that's stretching it). There were more than 1B mobile phones shipped in 2006, the year before the iPhone launched.
You're a long time HN contributor and I admit when I see your username, I stop and read the comment because it's always insightful, polite, and often makes me think about things in ways I never have before! But this discussion borders on religious fervor. "Every facet of peoples' lives?" Come on, man!
I'm aware of the HN bias, but in this case, I'm talking regular, non-tech, sportsball or TikTok watching crowd. Just within my closest circles, one person is using ChatGPT for recipes, and they're proficient at cooking so I was surprised when they told me the results are almost always good enough, even with dietary restrictions in place (such as modifying recipes without exceeding nutrient limits). Another person used it for comparison shopping of kids entertainment supplies. Another actually posted a car sale ad and used gen-AI to swap out background to something representative (no comment on ethics of that). Another is evaluating it for use in their medical practice.
(And I'm excluding a hundred random uses I have for it, like e.g. making colorbooks for my kids when they have very specific requests, like "dancing air conditioners" or whatever.)
He's intersted in meditation and mindfulness. He's not a native English speaker, so he's using AI to help him write content. He's then using AI text to voice to turn his scripts into YouTube videos. The videos have AI generated artwork too.
My dad is a retired welder in his late 60s. He's as "regular people" as it gets.
I'm a high school teacher and GPT has completely changed teaching. We're using it to help with report writing, lesson planning, resource creation, even for ourselves to get up to speed on new topics quickly.
I'm working on a tool for teachers that's only possible with GPT.
It's by far the single, most transformative technology I've ever encountered.
> work at FAANG, make $400K/yr
Truck drivers, construction crew, couriers, lawyers, sales people of all kinds, stock brokers, ... Most of the economy that isn't at a desk 996. Pretty big niche, bigger than the non-niche possibly.
You are in a bubble.
Eh? We did. The whole dot-com boom was predicated on that assumption. And it wasn't wrong. But most of the dot-com investments went sideways. In fact, they imploded hard enough to cause a recession.
In the same vein, even if we all agree that AI is fundamentally transformative, it doesn't mean that it's wise to invest money into it right now. It's possible that most or all of these early products and companies will go bust.
But hey, Nvidia is investing in companies.. to spend money on Nvidia .. infinite money glitch!
You don't need earth shattering though. The PC revolution was huge because every company got a bit more productive with things like word processors and printing and email.
The internet (and then later mobile) was big because every company got a revenue boost, from a small one with online presence to a a huge one for e-commerce to transformative with Netflix and streaming services.
Ignoring the more sci-fi claims of AGI or anything, if you just believe that AI is going to make every office worker 10% more productive, surely each company is goign to have to invest in AI, no? Anytime you have an industry that can appeal to every other company, it's going to be big.
Except AI is already being used by people (like myself) every day as part of their usual work flow - and it's a huge boost in productivity.
It's not IF it will make an impact - it IS currently making an impact. We're only just moving past early adopters and we're still in the early stages in terms of tooling.
I'm not saying that AI will become sentient and take over humanity, but to think that AI isn't making an impact is to really have your head in the sand at this point.
You're right. I pasted your comment to aider and it fixed it on the spot :).
EDIT: see https://git.sr.ht/~temporal/aider-2048/commit/9e24c20fc7145c....
A bit lazy approach, but also quite obvious. Pretty much what you'd get from a junior dev.
(Also if you're wondering about "// end of function ..." comments, I asked the AI to add those at some point, to serve as anchors, as the diffs generated by GPT-4o started becoming ambiguous and would add code in wrong places.)
(This kind of feedback driven generation is one of the things I do find very impressive about LLMs. But it's currently more or less the only thing.)
Yes, "many things are clones", but that just speaks to how uncreative we are all being. A 2048 clone, seriously? It was a mildly interesting game for about 3 minutes in 2014, and it only took the original author a weekend to build in the first place. Like how was that impactful that you were able to make another one yourself for $4?
It's been my "concentration ritual", an equivalent of doodling, for a few years in 2010s, so I have a soft spot for it. Tried getting back to it the other day, all my usual web and Android versions went through full enshittification. So that $4 and couple hours bought me a 2048 version that's lightweight, works on my phone, and doesn't surveil or monetize me. Scratched my own itch.
Of course, that's on top of gaining a lot of experience using aider-chat, by setting myself a goal of making a small, feature-complete app in a language I'm only moderately good at (and environment - the modern web - which I both hate and suck at), with extra constraint of not being allowed to write even a single line of code myself. I.e. a thing too boring for me to do, but easy enough to evaluate.
And no, the clone aspect wasn't really that important in this project. I could've asked it for something unique, and I expect it to work more-less the same way. In fact, this is what I'm trying right now, as I just added persistent state to the 2048 game (to work around Firefox Mobile aggressively unloading tabs you're not looking at, incidentally making PWAs mostly unusable) and I have my perfect distraction completely done.
EDIT:
BTW. did I ever tell you about the best voice assistant ever made, which is Home Assistant's voice assistant integrated with GPT-4o? I have a near-Star Trek experience at my home right now, being able to operate climate control and creature comforts by talking completely casually to my watch.
Try asking it something actually technologically hard or novel and see what answers you get.
In my experience, it repeatedly bails out with "this is hard and requires a lot of careful planning" regardless of how much I try to "convince" the model to live the life of a distributed systems engineering expert. Sure, it spits out some sample/toy code... that often doesn't/compile or has obvious flaws in it.
That said, the majority of my friends are doing relatively manual work (technician, restaurants, event gigs, sex work) and are neither threatened by LLMs nor find much use for them.
Even though I do think that almost any profession can potentially find use for LLMs in some shape or form. My opinion is LLMs can increase productivity and be a net positive the way the Internet/search engines are if used correctly.
To expand on my original comment: All that being said, I think the hype/media cycle overestimates the magnitude of the potential positive LLM effect. You’ll see numbers like 5x, 10x, 100x increase in productivity thrown around. If I have to bet, I would say the likely increase is going to be in the 1x-1.5x range but not much greater.
Most things in the world are not infinitely exponential, even if they initially seem to be.
(Not sure why the downvotes.)
Appreciated! But if we fret about a few downvotes, we're using the forum wrong. Some unpopular views need to be discussed - either because they hold some valuable truth that people are ignorant of, or because discussing them can shine a light on why the unpopular views are misguided. I suspect the downvotes are related to "Very few people are working on technologically hard or novel things" -- many HN users have been surrounded since elementary school by tons of people who do currently work on hard or novel problems, so they understandably think that >5-10% of people do that, when in fact it's closer to maybe 1-in-200. I've been part of social groups who went to high schools with absurd numbers of Rhodes' Scholars and peer groups where everyone in the group can trivially get through medical schools with top marks, receive faculty positions as professors at top-3 universities, found incredible startups through insane technical competence, and still all think they're stupid because they compare themselves to the true 1-in-a-million geniuses they grew up with who are doing research so advanced that it's far beyond their most remote chances of ever having even surface-level comprehension of that research. Their extended social group likely comprises >1% of all Americans working on "hard or novel problems", but since 75% of them are doing it, they have no idea that the real base rate is closer to 1-in-200, generously. They grossly underestimate their relative intelligence vs. the median and grossly overestimate the ability of average people (and explain away differences in outcome to personality issues like "laziness").
There are a surprising number of people from these peer groups on HN. These are the people who will never be threatened by LLM's -- they are capable of adapting to use any new tools and transcending any future paradigm, save war/disease/famine.
> To expand on my original comment: All that being said, I think the hype/media cycle overestimates the magnitude of the potential positive LLM effect. You’ll see numbers like 5x, 10x, 100x increase in productivity thrown around. If I have to bet, I would say the likely increase is going to be in the 1x-1.5x range but not much greater. Most things in the world are not infinitely exponential, even if they initially seem to be.
Yours is a very reasonable take that I wouldn't argue against. I also think it's reasonable that some people think it will be 5x-100x -- for the work some individuals are familiar with it very well might be already, or they might be more bullish on future advances in reinforcement learning / goal-seeking / "search" (iterative re-search to yield deep solutions).
> Even though I do think that almost any profession can potentially find use for LLMs in some shape or form.
I reactively feel this is stretching it for people who travel around just to load/unload boxes of equipment at events/concerts/etc. But the way you worded this is definitely not wrong - even manual laborers may find LLM's useful for determining whether they, their peers, and their bosses are following proper safety/health/HR regulations. More obviously, Sex workers will absolutely be using LLM's to screen potential customers for best mutual-fit and maintain engagement (as with lawyers who own small practices, a large number of non-billable hours goes towards client acquisition, as well as retention). LLM's are not "there" yet for transparent personalized client engagement which maintains the personality of the provider, but likely will be soon with some clever UX and RAG.
It's more than that when the LLMs go mutli-modal. A model (or an ensemble) that can hear you talking and see what you're showing it suddenly becomes very useful even for manual labor. Instructions, inspections, situational awareness, to think of few cases; this is a thoroughly unexplored space so far.
The mobile revolution needs three kinds of investment:
(A) The carrier has to build out a network
(B) You need to buy a handset
(C) Businesses need to invest in a mobile app.
The returns that anybody gets from investing in A, B or C depend on the investments that other people have made. For instance, why should I buy a handset if the network and the apps aren't there? Why should a business develop an app if the network and users aren't there? These concerns suppress the growth of mobile phones in the early phase.
ChatCPT depends on the existing network and existing clients for delivery so ChatGPT can make 100% of the investment required to bring their product to market which means they can avoid the two decades of waiting for the network and handsets to be there in order to motivate (C).
---
Note another thing that younger people might never have noticed was that the US was far behind the rest of the world in mobile adoption from maybe 1990 to 2005. When I changed apartments in the US in the 1990s I could get landline service turned on almost immediately by picking up the phone. When I was in Germany later I had no idea I could go into a store in most countries other than the US and walk out with a "handy" and be talking right away so I ended up waiting a month for DT to hook up my phone line.
Honestly I'm totally in the AI camp but 6 hours to make a 2048 clone?! And that's a good result? Come on.
Also, to be honest, it would've been much faster if GPT-4o didn't occasionally get confused by the braces, forcing me to figure out ways to coerce it into adding code in the right place. This is to say, there's still plenty of low-hanging fruits for improvement here.
But, tech being impactful doesn’t mean it will create and deliver value for others.
The closest I can think of would be the atom bomb, but even that arguably brought significant value in terms of relative geopolitical stability.
Very often technology advancements redistribute value, taking it away from many people and allowing companies to capture more profit and/or lower costs. Such as self-serve checkouts at supermarkets leading to less cashier jobs.
> A huge amount of economic value is going to be created by AI. Company builders focused on delivering value to end users will be rewarded handsomely.
Companies will acrue value, but 'common folk' will lose it.
I haven't analyzed its economics, but I would assume that by reducing the cashier jobs (e.g. allowing 1 operator to manage 8 self-serve checkout stations), supermarkets reduce their overall operating costs, and then use that to reduce their prices (and if they won't their competitors will), leading to a benefit to us 'common folk'.
As for cashier jobs, I don't think there's anything inherently good about them or inherently bad about them disappearing from the economy. As another example, I don't know many people who miss the Elevator Operators[0], and it makes perfect sense for us to press the buttons ourselves.
Unfortunately, what reduces the operating costs even more is giving the job of managing 8 self-serve checkouts to existing employee, on top of their current workload. With reliability of the checkout machines being as bad as it is, this results in a very frustrating shopping experience.
This is a pattern that we can see happening in everywhere across every sector in the market: the price for the buyer may get lower, but so does the quality. Often enough, it's impossible to retain quality and stay profitable, so the whole product or service class suddenly becomes worse. In the end, the delta between current satisfaction and the minimum satisfaction below which the customer would bounce, is a profit margin being extracted when competitive pressure is high - which, I dare say, sucks for the customer, even if they also get stuff cheaper.
The code ChatGPT generates is often bad in ways that are hard to detect. If you are not an experienced software engineer, the defects could be impossible to detect, until you/ChatGPT has gone and exposed all your customers to bad actors, or crash at runtime, or do something terribly incorrect.
As far as other thought work goes, I am not consulting ChatGPT over, say, a dietician or a doctor. The hallucination risk is too high. Producing an answer is the not the same as producing a correct answer.
- Are hard (or boring) to do, but easy to evaluate - for me, e.g. writing code, OCR, ideation; or
- Don't require a perfectly correct answer, but more of a starting point or map of the problem space; or
- Are very subjective, or creative, with there being no single correct answer,
is surprisingly large. It covers pretty much everything, but not everything for everyone at the same time.
Does it work though, yes it does. There are many human coders who write bad code and life goes.
I've seen both the good and the bad. I really like the good parts. Most recently, Claude Sonnet 3.5 fixed a math error in my code (I prompted it to check for it from a well-written bug report, and it did it fix it ever so perfectly).
These days, it is pretty much second nature for me to pull up a new file & prompt Copilot to complete writing the entire code from my comment trails. I don't think I've seen as much change in my coding behaviour since Borland Turbo C -> NetBeans.
My latest challenge is dealing with people that trust chatgp to be infallible, and just quote the garbage to make themselves look like they know what they are talking about.
LLMs are language model, it's crazy people expect them to be correct in anything beyond surface level language.
>I am not consulting ChatGPT over, say, a dietician or doctor
Do you know any doctors, by chance? You have way more faith in experts than I do.
I’ve had multiple friends get seriously ill before a doctor took their symptoms seriously, and this is a country with decent healthcare by all accounts.
Human biases are bad too.
So true. And it's hard to question a doctor's advice, because of their aura of authority, whereas it's easy to do further validation of an LLMs diagnosis.
I had to change doctor recently when moving towns. It was only when chancing on a good doctor that I realised how bad my old doctor was - a nice guy but cruising to retirement. And my experience with cardiologists has been the same.
Happy to get medical advice from an LLM though I'd certainly want prescriptions and action plans vetted by a human.
> It was only when chancing on a good doctor that I realised how bad my old doctor was
Wikipedia does not have dietary advice. It’s an encyclopedia.
If you can't have ChatGPT write testable code because of your architecture, you have other problems. People with bad process and bad architecture saying AI is bad because it doesn't work well with their dumpster fire systems, 100% facepalm.
There exist lots of reasons why code is hard to test automatically that have nothing to do with the architecture of the code, but with the domain for which the code is written and runs.
Sure I probably would have been able to do it without ChatGPT, but it was so much easier to have something to bounce ideas off-of. A safety net, if you will.
The hallucination risk was irrelevant: it did hallucinate a little early on. I told it it was a hallucinating, and we moved onto a different way of solving the problem. It was easy enough to verify it was working as expected.
But try to do something much more simple but has much fewer examples (a typical case is something which has bad documentation) in the data, and it falls apart. I even tried to use Perplexity to create a dead simple CLI command, and it hallucinated an answer (looking at the docs, it misused the parameter, and may have picked up on someone who gave an incorrect answer in the data.)
I keep hearing this, but it's incorrect. While I only know R, which is obviously a simple language, I would never type out all my code and go without testing to ensure it does what I intended before using it regularly.
So I can't imagine someone that knows a more complex language just typing out all of it before integrating it into business systems at their work or anything else before testing it.
Why would AI be any different?
Why the hell are AI skeptics acting like getting help from an LLM would involve not testing anything? Of course I test it! Why on earth wouldn't I? Just as I tested code made by freelancers I hired on commission before using the code I bought from them. Do AI skeptics really not test their own code? Are you all insane?
Take it from someone who started with R, R is 100% not a simple language. If you can write good R, you're probably a surprisingly good potential SE as R is kinda insane and inconsistent due to 50+ years of history (from S, to R etc).
It's an excellent language, I think, for many reasons. One is that you can work with data within hours because even before learning what packages or classes are, you got native objects for data storage, wrangling, and analysis. Even import my Excel data and rapidly learn the native function cheat sheet so fast that I was excited to learn what packages are because I couldn't wait to see what I could do.
That was my experience in like 2010, maybe, and after having C++ and Python go in and out my head during college multiple times. I view R as simple only because I actually felt more helpless to keep learning it than helpless to ever learn coding at all. Worth noting that I was a Stat/Probability tutor with a Finance degree and much Excel experience.
Ah yeah, makes sense. That's the happy path for learning R (know enough stats etc to decode the help pages).
That being said, R is an interesting language with lots of similarities to both C based languages and also Lisp (R was originally a scheme intepreter), so it's surprisingly good at lots of things (except string manipulation, it's terrible at that).
I wonder about this a lot, because there's a future here where a decent amount of software engineering is offloaded to these AIs and we reach a point, in the near future, where no one really knows or understands what's going on. That seems bad. Put another way, suppose that your primary care doctor is really just using MedAI to diagnose and recommend treatment for whatever it is you went in to see him about. Over time, these sorts of shortcuts metastasize and the doctor ends up not really knowing anything about you, or the other patients, or what he's really doing as a doctor ... it's just MedAI (with whatever wrongness rate is tolerable for the insurance adjusters). Again, seems bad. There's a palpable loss of human knowledge here that's enabled by a "tool" that's allegedly going to make us all better off.
As for what I think about them: I've been impressed with some aspects of code generation, but nothing else has really "wowed" me. Prose written with the various GPT models has an insincere quality that's impossible to overlook; AI-generated art tends to look glossy and overproduced in the same way that makes CGI-heavy movies hard to watch. I have not found that my Google Search experience was made better by their AI experiments; it made it harder, not easier, for me to find things.
While I absolutely agree that many movies over-use CGI, even with the relative decline in superhero movies, CGI-heavy movies still top the box office. Going over the list of highest-grossing movies each year [0], you have to go back about three decades to find a movie that isn't CGI-heavy, so apparently they're not that difficult for the general public to watch.
[0] https://en.wikipedia.org/wiki/List_of_highest-grossing_films
May we see it?
https://git.sr.ht/~temporal/aider-2048
There's a full transcript of the interactions with Aider in that repo (which I started doing manually before realizing Aider saves one of its own...).
Before anyone judges quality of the code - in my defense, I literally wrote 0 lines of it :).
Similar stuff will happen with a lot of other content, things that used to be costly will become very cheap. And then what? The amount of books people can consume doesn't scale into infinity. Their entertainment needs will be served by auto-generated AI content. Even the books themselves will be written by AI sooner or later.
Advertising industry might also start hurting badly, as while they will certainly try getting ads into AI content, users will have AI at home to filter it out. A lot of classic tricks and dark pattern to manipulate the user behavior will no longer work, since the user has a little AI helper to protect them from those tricks.
I don't doubt that the impact of AI will be gigantic, but a lot of AI produced content won't be worth anything, since it's so easy to create for everybody. And there isn't much of a moat either, since new models with better capabilities pop up all the time from different companies. Classic lock-in is also not really usable anymore, as AI can effortlessly translate between different APIs and user-interfaces.
Similar to a game; for many games players aren't intended to have exactly the same experience but there are still common things about the game platform to discuss.
I think it would be pretty cool to have a partially AI generated plot, it would be exciting to discuss what you got in the AI lottery with someone else who had also watched the show. Something like: [set plot point] [random catastrophic event] [set plot point] [etc]. "1 character must die in some meaningful way" turns to "omg, which one was killed when you watched? Adam got eaten by the giant spiders in mine" "oh man, Sarah tried to set a spider on fire in mine, but she ended up getting herself". A la https://en.wikipedia.org/wiki/Until_Dawn
As cool as this might be, what is the actual economic value of this? 2048 is free, you didn't even have to spend a dollar to get it.
Can ChatGPT materially and positively impact the code written by big companies? Can it do meaningful work in excel? Can it do meaningful PowerPoint work? Can it give effective advice on management?
Right now we don’t know the answer to those questions. LLM apps can still improve in many ways - better base models, better integration with common enterprise applications, agentic processes, verifiability and so on - so there is definitely hope that there will be significant value created. Companies and people are excited because there’s huge potential. But it is really just potential right now … current systems aren’t creating real enterprise value at this moment in time
> Right now we don’t know the answer to those questions.
I know the answer to the first three. Yes, yes, and yes. I've done them all, including all of them in the past few weeks.
(Which is how I learned that it's much better to ask ChatGPT to use Python evaluation mode and Pandoc and make you a PPTX, than trying to do anything with "Office 365 Copilot" in PowerPoint...)
As for the fourth question - well, ChatGPT can give you better advice than most advice on management/leadership articles, so I presume the answer here is "Yes" too - but I didn't verify it in practice.
> current systems aren’t creating real enterprise value at this moment in time
Yes, they are. They would be creating even more value if not for the copyright and exports uncertainty, which significantly slows enterprise adoption.
You say this but from a management perspective at a large enterprise software company I have not seen it.
Some of our developers use copilot and gpt and some don't and it is incredibly difficult to see any performance difference between the groups.
We aren't seeing higher overall levels of productivity.
We aren't seeing the developers who start using copilot/gpt rush ahead of their peers.
We aren't seeing any ability to cut back on developer spend.
We aren't seeing anything positive yet and many developers have been using copilot/gpt for >1 year.
In my opinion we are just regaining some of the economic value we lost when Google Search started degrading 5-10 years ago.
You can't measure productivity for shit, otherwise companies would look entirely differently. Starting from me not having to do my own finances or event planning or hundred other things that are not my job description, not my specialty, and which were done by dedicated staff just a few decades ago, before tech "improved office productivity".
> We aren't seeing the developers who start using copilot/gpt rush ahead of their peers.
That's because individual productivity is usually constrained by team productivity. Devs rushing ahead of their teammates makes the team dysfunctional.
> We aren't seeing any ability to cut back on developer spend.
Devs aren't stupid. They're not going to give you an opportunity if they can avoid it.
> We aren't seeing anything positive yet and many developers have been using copilot/gpt for >1 year.
My belief is that's because you aren't measuring the right things. But then, no one is. This is a problem well-known to be unsolved.
We can't measure small changes and we aren't great at comparing across orgs.
However, at the director level we can certainly see a 50% or 100% productivity improvement in our teams and with individuals in our teams.
We aren't seeing changes of this magnitude because they don't exist.
Perhaps developers are now slacking off.
Perhaps we have added more meetings because developers have more free time.
Or perhaps developers were never the bottleneck.
We can see large productivity improvements when we make simple changes like having product managers join the developers daily standup meetings. We can even measure productivity improvements from Slacks/Zooms auto-summary features. Yet gpt/copilot doesn't even register.
While not code generation, this auto-summary is powered by the same tech. I think using it to sift through and surface relevant information, as opposed to generation of new things, will have the biggest impact.
By far the greatest value I get out of LLMs is asking them to help me understand code written by others. I feel like this is an under-appreciated use. How long has this feature been in Copilot? Since February or so? Are people using it? I do not use Copilot.
Now that's dangerous thinking, but I think you are onto something.
Edit: and now I see chillfox made the same point.
For example:
- time between PRs being created and being picked up for review and merged
- time spent on releasing at end of sprint cycles
- time spent waiting for QA to review and approve
- extreme scrum practices like "you can only work on things in the sprint, even if all work is done"
How are you measuring developer productivity? Were those that adopted copilot and chatgpt now enabled to finally keep up with their faster peers (as opposed to outstrip them)? Is developer satisfaction improved, and therefore retention?
I guess we will see if smaller startups without many of our bottlenecks are suddenly able to be much more competitive.
> How are you measuring developer productivity?
We use a host of quantitative and qualitative measures. None of them show any positive improvements. These include the basics like roadmap reviews, demo sessions, feature cycle time, etc as well as fairly comprehensive business metrics.
In some teams every developer is using copilot and yet we can't see any correlation with it and improved business metrics.
At the same time we can measure the impact from changing the label on a button on our UI on these business metrics.
> Were those that adopted copilot and chatgpt now enabled to finally keep up with their faster peers
No.
> Is developer satisfaction improved, and therefore retention?
No.
Those are very high level. If there's no movement on those, I'd guess there are other things bottlenecking the teams. They can code as fast as possible and things still move at the same pace overall. Nice thing to know.
If you want to really test the hypothesis that Copilot and ChatGPT have no impact on coding speed, look at more granular metrics to do with just coding. The average time from the moment a developer picks up a work item to the time it gets merged (assuming code reviews happen in a timely fashion). Hopefully you have historical pre-AI data on that metric to compare to.
Edit: and average number of defects discovered from that work after merge
We do collect this data.
I personally don't put a lot of stock in these kinds of metrics because they depend far too much on the way specific teams operate.
For example perhaps Copilot helps developers understand the codebase better so they don't need to break up the tasks into such small units. Time to PR merge goes up but total coding time could easily go down.
Or perhaps Copilot works well with very small problem sizes (IMO it does) so developers start breaking the work into tiny chunks Copilot works well with. Time to PR merge goes way down but total code time for a feature stays the same.
For what it is worth I do not believe there have been any significant changes with these code level metrics either at the org level.
Very few of these are the superstars. But plenty are good solid senior developers.
As a dev, you can use the saved time to slow down and not be stressed, spend more time chatting with colleagues, learn new skills, maybe improve the quality of the code, etc. Or you can pass it on to management which will result in your workload being increased back to where you are stressed again and your slower colleagues will be let go, so now you get to feel bad about that and they won't be around to chat with.
I have never in my life seen workers actually get rewarded with pay raises for improved productivity, that is just a myth the foolish chase, like the pot of gold at the end of the rainbow.
I have also tried being the top performer on a team before (using automation tools to achieve it), and all I got was praise from management. That's nice, but I can't pay for my holidays with praise, so not worth it.
> We aren't seeing the developers who start using copilot/gpt rush ahead of their peers.
You think we are antsy worker bees, hastily rushing forwards to please the decision maker with his fancy car?
You are leadership. It's not hard. Cui bono, follow the money, etc. The incentives are clear.
If me and my peers were to receive a magic "do all my work for me" device I can assure you exactly zero percent of that knowledge will reach your position. Why would it? The company will give me a pat on the back. I cannot pay with pats on the back. Your Tesla cannot be financed with pats on the back. Surely you understand the nature of this issue.
Can you elaborate on what this saved over just making the ppt the old fashioned way?
Attach notes, paste, press Enter, wait half a minute, get back a PPTX you can build on, or just restyle[0].
Sure, it's faster to build the presentation yourself than to make ChatGPT make the whole thing for you. But the more time-consuming and boring parts, like making tables and diagrams and summaries from external data or notes, is something ChatGPT can do in a fraction of time, and can output directly into PPTX via Pandoc.
(There's a lot of fun things you can do with official ChatGPT and Python integration. The other day I made it design, write and train a multi-layer perceptron for playing tic-tac-toe, because why waste my own GPU-seconds :).)
--
[0] - In contrast, if you make the same request in PowerPoint's O365 Copilot, it'll barf. Last time I tried, it argued it has no capability to edit the document; the time before that, it made a new slide with text saying literally "data from the previous message".
The problem is that at the average medium sized company code looks like this - you have 1mln lines of code written over a decade by a few hundred people. A big portion of the code is redundant, some of it is incomplete, much of it is undocumented. Different companies have different coding styles, different testing approaches, different development dynamics. ChatGPT does not appreciate this context.
Excel has some similar problems. First of all Excel is 2 dimensional and LLMs really don’t think in 2 dimensions well. So you need to flatten the excel file for the LLM. A common approach to do this with LLMs is using pandas and then using the column and row names to index into the excel.
Unfortunately, excels at companies cannot be easily read using pandas. They are illogically structured, have tons of hardcoding, intersheet referencing is weird circular ways and so on. I spent some time in finance and sell side equity research models are written by highly trained financial analysts and are substantially better organized than the average excel model at a company. Even this subset of real world models is far from suitable for a direct pandas interpretation. Parsing sell side models requires a delicate and complex interpretation before being fed into an LLM.
It already has at a Fortune 100 company I contract with currently.
> Can it do meaningful work in excel?
We can quibble about what "meaningful" means, but it satisfactorily answered questions for two friends about how to build formulas for their datasets and is currently being used to summarize data insights from a database at a different large client (Excel =/= database, but the point stands).
> Can it do meaningful PowerPoint work?
I've used Midjourney multiple times a month to generate base imagery for various things in PowerPoint (usually requires modification in Photoshop, but saves me several hours each time compared to digital painting or photobashing from scratch).
> Can it give effective advice on management?
Again, what does "effective" mean in the context of management? I've seen VP-level individuals with hundreds of people in their orgs using AI tools for different things.
It really feels like a significant chunk of the HN crowd is living in a bubble with respect to AI in the real world right now. It's absolutely invading everything. As for how much revenue that will translate into long-term vs. the investment dollars being poured into it, that's a more interesting question to discuss.
My friends at McKinsey say that while it can’t fine-tune reports and presentations with quite enough nuance, it does a good job sifting through lots of shit to pick out important parts they should pay more attention to, highlighting data/talking points that contradict a working hypothesis, assisting in writing emails, and other time-consuming or very nit-picky tasks.
That said, no one I know has fed it real customer data, that would be a career-ending event. But self-hosted models like Gemma2 open up the possibility for using LLMs against real customer info.
> the lack of value of McKinsey
Yes, it absolutely can. I threw together a PowerPoint presentation with a script for a low-value, high visibility meeting a couple of weeks ago with ChatGPT 4.0 and a PowerPoint plugin. Everyone loved it.
> low-value, high visibility meeting
Yes it can.
But more importantly have you tried ChatGPT Data Analyst?: https://openai.com/index/improvements-to-data-analysis-in-ch...
It drops the barrier for "pretty good data analysis" to effectively zero.
> Can it do meaningful PowerPoint work?
Canva and Figma are both building this and they are pretty decent right now. Better than most PowerPoints I've seen.
The aforementioned Data Analyst does good presentations in a different way, too.
> Can it give effective advice on management?
Yes. Unfortunately can't talk about this except it is mindblowingly good.
And on top of that, it can do PowerPoint presentations too - magic keywords are "use Python and Pandoc".
That's actually a really good point. In the realm of programming, things that were previously not done because they were too expensive can now be done. Prior to ChatGPT, GP could have a) done it themselves, but the cost was too high/it wasn't worth their time, b) found enough time to write a spec, found someone on upwork/etc, paid them to make it, except that costs money they didn't want to spend, or c) just not do it. Now, GP can code this thing up while watching netflix with the kids or whatever. What programs do not exist that previously did not have the economic value to exist, but now can, thanks to programming time getting cheaper?
Now apply that to fields outside of programming. LLMs' ability to program is front and center here, since many of us can program, but they do other things as well.
This is not meant to be an offense, but you are in a bubble. The vast, vast majority of people do not use LLMs in their day-to-day life. That’s ok, we’re all in our own bubbles.
You should also post the 2048 clone as proof. Lots people saying they built X in Y minutes with AI. But, when it’s inspected, it’s revealed it very obviously doesn’t work right and needs more development.
I posted it twice already in this thread, but I guess third time's the charm: http://jacek.zlydach.pl/v/2048/ (code: https://git.sr.ht/~temporal/aider-2048).
It's definitely not 100% correct (I just spotted a syntactic issue in HTML, for example), and I bet a lot of people will find some visual issue on their browser/device configuration. I don't care. It works on my desktop, it works on my phone, it's even better than the now-enshittified web and Android versions I used to play. I'm content :).
>but in the process it breaks the CSS for larger screens.
So, no, it doesn’t fix it trivially. Also isn’t correctly sized on iPhone 11 Safari.
As mentioned above, I don't care. It's sized correctly for the devices I use to play it, and I'm not going to put any more work into this. I mean, even junior web devs get paid stupidly high salaries for doing Responsive Web Design; I ain't gonna work on this for free.
(But I will accept AI-generated patches, or human-generated advice I could paste into the prompt to get a correct solution :P.)
https://github.com/williamcotton/guish
The rest is Claude 3.5 (with a dash of GPT-4o) with a LOT of supervision!
I'd say I'm about 8 hours deep and that this would have taken me at least 30+ hours to get it to the current state of polish.
I used it to make some graphs at work today!
But seriously, do you have any thoughts or suggestions?
Boring.
Cursor with 3.5 sonnet has made all this way faster so that’s nice. Often LLMs are now featuring in these pipelines and I see libraries like data bonsai and instructor being helpful. But yeah idk. No bright ideas here but always on the lookout to optimise.
I have used chatgpt less and less, and bar copilot which is a useful autocomplete I just don't have much use for AI.
I know I'm not alone, and even though I've seen many people super excited by Dall-E first and chatgpt later they use very rarely both of them.
I still use GPT or Claude occasionally but I find switching over to prompting breaks my mental flow so it’s only a net win for certain kinds of tasks and even there it’s not a huge step up from searching Stack Overflow.
As for whether it's worth it, I argue this is the single most useful application of GPUs right now, both economically and in terms of non-monetary value delivered to users.
(And training them is, IMO, by far the most valuable and interesting part of almost all creative works available online, but that's another discussion.)
I've heard this asserted sometimes, and I just don't think it's true. ChatGPT's use cases as consumer software were discovered basically immediately after GPT 3 came out, and nothing new has really emerged since then. It's great for automating high school/undergrad B-quality writing and the occasional administrative email. Beyond that, it sometimes does better than 2024 Google(though probably still worse than 2019 Google) on knowledge questions.
ChatGPT is software. The barrier to entry is almost zero, and the tech industry has had decades of practice in enticing people into walled gardens and making sure they can never leave. If it's not completely taken over the world in the time it's had, I wouldn't bet on it doing so without a massive jump in capability or accuracy.
Not to mention all the proof-writing that will become simpler with this optimization/searcher now.
> We have essentially standardized a way to approximately solve optimization problems.
.. what does this mean? we had simplex solvers before. do you mean things like protein folding prediction?
We may not see universal adoption in people who are currently >30 but I think we will in the generations that are <25 now.
But it isn't translating into better across the board test results and at least in Australia we would be able to tell because we have yearly standardised testing.
And so schools are looking at it as more of a form of cheating and simply moving back to in-person, hand-written tests.
Why are you assuming they haven't? High school writing assignments and homework are the majority of the problems teenagers face daily, but they're also having fun with it, and why wouldn't they try it on new problems as they come along?
This is baffling to me, because these are two use cases I have tried and in which ChatGPT completely fails to produce any useful information.
I have a bunch of peers that haven’t used chatgpt at all and they are software developers. A bunch more tried it once, realized how terrible it was for anything you aren’t already knowledgeable in and then haven’t gone back to it.
Recipe adjustments has to be a joke unless they are really basic things like “cut it in half”. ChatGPT is terrible at changing recipes without fundamentally changing them and will offer multiple “substitutions” for an ingredient that have extremely different outcomes that it doesn’t warn or know about.
We should have seen massive revenue growth and raises in future quarter revenue forecasts in the most recent round of SaaS company earnings reports. I think tons of companies have hyped up AI as if it's just on the cusp of AGI. The lack of massive top line growth has proven we're not even close, but the enormous investor speculation these companies triggered is the main reason for this $600B gap.
I'm not at all saying AI won't be transformational because it definitely does bring revolutionary capabilities that weren't possible before.
I'm not saying AI is useless but it's certainly not the panacea that some say it is.
While I use AI quite often, none of my friends or family does. A few of them will use an image gen once or twice a year. And at work, only a few of my colleagues use AI.
So my impression is that current gen AI is too hard to use correctly, has too many rough edges and is not useful enough for most people.
Progress also seems to have stalled around the GPT-4 quality. Everything after GPT-4 (GPT-4 Turbo, GPT-4o, Claude 3 Opus, Claude 3.5 Sonnet) seems to be pretty much producing the same quality output, I have been using Open WebUI to bounce around between the different models and I can't really tell a difference in the quality of them, they are all roughly the same for my use case (programming/sysadmin stuff).
So the question of if a plateau has been hit or if scale can still improve quality is real to me.
EDIT : My bad, I see you posted the link elsewhere (link for posterity http://jacek.zlydach.pl/v/2048/ )
TBH 6 hours seems a lot longer than I would have expected.
Also: 6 hours is a lot if you sit down to it and know exactly what to write. Not when you half-remember how the game works, don't have uninterrupted focus time for it, and deal with executive function issues to boot.
I'm happy to wait for now and let the tech mature though.
This kind of example always confuses me. I don't see the value vs reading an article like this: https://www.freecodecamp.org/news/how-to-make-2048-game-in-r...
If I said I built a 2048 clone by following this tutorial, noone would be impressed. I just don't see how reading a similar tutorial via a chat interface is some groundbreaking advancement.
For the tutorial you linked to, there's a lot of prior knowledge assumed, which the author alludes to in the summary, which a chat interface would help with:
This time I decided to focus on the essence of the topic rather than building basic React and CSS, so I skipped those basic parts. I believe it makes this article easier to digest.
The perspective I take is the 15 year view: The iPhone 1 sucked objectively but by the iPhone 3 the trajectory was clear and 15 years later the world is a very different place.
You hear people very focused on specific shortfalls: "I asked it to write code and look it made a mistake". But there are very clear routes to fixing these and there are lots of people finding it useful despite these bugs.
I think AI is bigger than mobile. I'm nearly 50, and I remember the PC boom, the Internet boom, Social Networking boom, Mobile boom, SaaS boom - probably more that I forget.
I think the PC and Internet booms are the only ones that are as impactful as AI will be in 15 years.
Maybe mobile is as big, maybe not - depends if someone can build AI devices that replace the UX of phones sometime in the next 15 years.
Which technology are you talking about? ;-)
What I can clearly say is that I know no one from my social circles and extended social circles who used these AI chatbots for anything else than "simply trying out what is possible" (and ridiculing the results). A quote from a work colleague of me when I analyzed the output of some latest generation AI chatbot: "You shouldn't ask so complicated questions to the AI chat bots [with a huge smile on his face]. :-)"
I don't know the exact numbers, but I guess only maybe 5% of all investments in a given batch make any impact on the total return.
So for a VC, if there's a 10% chance that this whole AI thing will be a financial success, it's chance of success is already twice as high as average, so a pretty good bet.
In contrast, I've put in very little effort to use AI, but I'm noticing things.
I see high quality AI-generated images in blog posts. They look awesome.
I look over my coworker's shoulder and see vscode predict exactly the CSS properties and values he's looking for.
Another coworker uses AI to generate a working example of FPGA code that compiles and runs on a Xilinx datacenter device.
An AI assistant pops up in Facebook messenger. My girlfriend and I are immediately able to start sending each other high quality, ultra-specific inside joke AI generated memes. This has added real value to my life.
I'm starting to feel FOMO, a bit worried that if I don't go hard on learning this new tool I'm going to be left in the dust. To me at least, AI feels different.
Were regular people using it like ChatGpt?
After that it started really getting its scammy, investor-only reputation
Oh, it's really simple, you see if they don't get rewarded handsomely, that proves they didn't focus on delivering true [Scotsman] value. /s
this didn’t really happen the way you want it to. Fortune 50 companies never spent billions of dollars on crypto or NFTs like they are doing for AI. No NASDAQ listed companies got trillion-dollar valuations out of crypto.
There is buy-in happening this time, unlike previous times, because this time actually is different.
> The whole AI thing just continues to baffle me. It's like everyone is in the same trance and simply assuming and chanting over and over that This Will Change Everything
I mean, some people see a broad consensus forming and reactively assume everyone else must be stupid (not like ME!). That’s a reflection of your own personal contrarianism.
Instead, try to realize that a broad consensus forming means you actually hold heterodox opinions, and if you think you have a good basis for them that’s fine, but if the foundation for your point that everyone in the world is too stupid to see what’s REALLY going on then maybe your opinions aren’t as reasoned as you think they are. You need to at least understand the values differences that are leading you down the road to different conclusions before you just dismiss the whole thing as “everyone else is just too wrapped into the cult to see straight”.
Bitcoin was actually rebuttable on some easily-explicable grounds as to why nobody really needed it. Why do you think semantic embeddings, semantic indexes/generation, multimodal interfaces, and computationally-tractable optimization/approximation generators are not commercially useful ideas?
I haven't even formed much of an opinion either way, yet. Sure, I have doubt, but that's more of a default than something I reasoned myself into. I'm saying it's just way too early to make statements either way about the future of LLMs and AI that are anything beyond wild guesses. "This time it's different, it's fundamentally transformative and will obviously change the world" is a religious statement when made this early.
nvidia did very well out of crypto.
Got to imagine that IBM’s spending on their weird blockchain hobby was at least in the hundreds of millions.
And Facebook spent tens of billions of dollars on metaverse stuff, of course.
Crypto portended drastic and fundamental changes: programmable money, disintermediation, and the decentralization of the very foundations of our society (i.e. money, banking, commerce). Suffice to say that nothing close to this has happened, and probably will never happen.
So I can see how many people are equally skeptical that AI, as the next hyped transformative technology, will achieve anything near the many lofty predictions.
Online banking is great, and people knew it was coming since the dawn of the internet, but they mostly didn’t predict Stripe.
Given the accelerated invention/deployment cycles we're in, it's not hard to extrapolate GPT4o to $0 token cost and 0ms latency. Even assuming stagnation in context lengths and cognition, the extreme scope of impact on every computerized industry becomes self-evident.
It's not yet clear what (1) has to do with (2). Maybe it turns out that LLMs or similar can do (2). And maybe not.
I can understand being skeptical about the economic value of (1). But the economic value of (2) seems obviously enormous, almost certainly far more than all value created by humanity to date.
The question of "how do we make money from it" is a much harder to answer. Using every available computer to run quadratic time brute force on everything you can scrape from the internet is an unbounded resource sink that offers little practical return for almost everybody, but leveraging modest and practical use of generative machine learning where it works well will absolutely create some real value.
> Such strong speculative predictions about the future, with no evidence.
The speculation makes sense from a VC's perspective, but perhaps not from the perspective of society at large (i.e. human workers).
From the revenue-generating use cases of LLMs (== AI in the article) that I've seen so far, most seem to be about replacing human mental labor with LLMs.
The replacement of workers with AI-based machines will likely happen in mature industries whose market growth is basically capped. Productivity will stay mostly the same, but the returns will increase dramatically as the human workforce is hollowed out.
To the extent that AI instead empowers some workers to multiply their productivity with the same amount of effort, then it can create more economic value overall, and this may happen in industries with a long growth runway ahead.
On balance, it's not clear to me whether the growth (in productivity and employment) that comes from the latter will be enough to offset the employment losses from the former.
But in either scenario, the VCs investing in AI win, either from efficiency gains, or from accelerating growth in new industries.
All of society is so freaking leveraged at this point, something has to give.
can you elaborate on this? in what sense?
during WW2 it was also this high. the US economy is still growing. and inflation is around 3%. not the usual target 2, but it's crazy good considering the recent pandemic and land-war-in-europe "situation".
eventually spending has to decrease ... or taxation to increase (US tax to GDP ratio is around 27% while OECD average is around 34%) ... and the US demographic stats are looking much better than a lot of countries'.
the feeling of impending doom is definitely not great, but it's a (social) media phenomenon.
They already understand spoken English and can respond in kind.
This is Siri or Alexa on steroids. Just that alone is a “killer app” for everyone with a mobile phone or a home assistant device!
What’s the addressable market for that right now? Five billion customers? Six or seven maybe?
Computer RPG games are about to become completely different. You’ll be able to actually converse with characters.
Etc…
I’m a short-term pessimist but a long-term optimist.
This all reminds me of 3D graphics a in the early 1990s. The nerds were impressed but nobody else thought it was interesting. Now we have Pixar and a computer games industry bigger than Hollywood.
Siri and Alexa are bad. Very, very bad.
They pretend to understand spoken English, but they don't, because they're just a huge set of hard-coded rules written out one-at-a-time by enormous numbers of very expensive developers.
This is the 1990s approach to AI: Fuzzy logic, heuristics, solvers, dynamic programming, etc...
That approach has been thoroughly blown out of the water by Transformers, which does all of that and much more with a thousand lines of code that can be banged out in a couple of hours by one guy while talking in a YouTube video: https://www.youtube.com/watch?v=kCc8FmEb1nY
Transformers will revolutionise this entire space, and more that don't even exist yet as a market.
Take for example humanoid robots: Boston Dynamics has had the hardware working well enough for a decade, but not the software. You can't walk up to one of their robots, point at something, and tell the robot to complete a task. It can't understand what it is seeing, and can't understand English instructions. Programming that in with traditional AI methods would take man-millenia of effort, and might never work well enough.
If we could speed up GPT-4o (the one with vision) to just 10x or 50x its current speed, with some light fine-tuning it could control a humanoid robot right now with a level of understanding comparable to C3P0 from Star Wars!
The question is when exactly do we get to human level intelligence on all tasks (agi)? Is it going to be GPT5? GPT6? Or has the performance improvements saturated already? It makes a huge difference in terms of your investment and even career decisions whether you expect it to happen next year or 10 years from now.
Think about all human involved in producing unstructured document that people have to read like public councils, court of laws, teachers or other evaluators. Llm can be given a relevancy metric and flag content so that those in need of or waiting for a certain information aren't drowned in noise
Llm are unlocking digital transformation in sectors that have been historically resistant to it, and because they are goal driven they do it with minimal programming, just a few good prompts and a data pipeline.
And it's just the tip. They don't get tired, they don't forget nor omit, they are absolutely bonkers to find relevant products and services given a database a need and a set of preferences and they will transform how the next generation will make purchasing decisions, travel decision and all these opinion based choices.
And while llm are likely not the path to agi and they will cap out in capabilities at some point, they are coming down in price super fast, which will propel adoption even for those cases where other options would be more sensible, just because of the sheer convenience of just asking them to do stuff.
The biggest takeaway from this piece is the stark realism of this article (maybe a bit too bearish, imo) compared to the usual Sequoia VC-speak. Maybe FTX did teach them something, after all.
VCs did get burned by speculative investing in the 2019-21 period but FTX wasn’t like the others. At the time of its collapse, FTX was profitable and had $1+ bn in revenue, its doom had nothing to do with product market fit, revenue, margins, etc.
Quibi might be a more relevant example of a learning opportunity.
Comparing this article to the (now deleted) SBF profile is night and day.
> Maybe FTX did teach them something
The new tone is all about business fundamentals and the FTX collapse had literally nothing to do with business fundamentals. If this is the lesson they learned they learned the wrong lesson.
1. That all datacenter GPUs being purchased are feeding AI. You might be able to argue that some are or a lot are, but you don't know how many just looking at Nvidia sales numbers. I know of at least two projects deploying rows of cabinets in datacenters full of GPUs for non-AI workloads.
2. The assumption that pay-for-an-API is the only AI business model. What we now call "AI" has been driving Google's search and ad businesses for nearly a decade, sooo AI is already doing $300B/yr in revenue? There is no way for this guy to quantify how AI is solving problems that aren't SaaS.
David Chan, if you are reading this feel free to email me if you want a fact check for what will surely be the third installment in the series.
$600 billion is $434 per person, $36 per month per person, 1.2% percent of GDP.
If 75% of the spending goes to increasing productivity, I could see it. Get rid of $450 billion in labor costs (12 - 15 million work years). Cut worked hours in call centers, customer service, many services, menial programming jobs, ...
But I don't see it happening fast enough to pay for current investments.
So why aren't there more entrants in the CPU cloud area? The technology is a commodity. Google and Amazon don't make CPUs.
Because the market is saturated with players.
AWS, Azure, GCP, Alibaba, IBM Cloud, Digital Ocean, Tencent, Oracle Cloud, Huawei Cloud, that Dell/VMware thing, Linode/Akamai, HP, Scaleway, Vultr, GoDaddy, OVH, Hetzner, etc.
Source: I run a cloud company and have plenty of friends in the space.
It’s not shocking at all that margins are that high when a team pushes to pay AWS $40k a month for a static workload that could run on two racks of servers.
I would say “irrational when it comes to costs”. Not all developers ask questions such as:
Does paying 10K for convenience save me and my team more than 10k worth of time? Does this vendor make it easy to migrate out?
Of course I’m biased, because I like it better when the answers favor us, cloud vendors. But there are no catch all answers; as always, it depends.
I'm reminded of Bitcoin/crpyto, which in its early history was all operated on GPUs. And, then, almost overnight, the whole thing was run on ASICs.
Is there an intrinsic reason something similar couldn't happen with LLMs? If so, the idea of a bubble seems even more concerning.
I found a short discussion[2] you may find useful.
[2]: https://www.lesswrong.com/posts/qhpB9NjcCHjdNDsMG/new-fast-t...
Only if we can increase the efficiency of LLMs by 2-3 orders of magnitude, there are only some in lab examples of this and nothing really being publicly shown.
Even then the models are still going to require rather large amounts of memory, and any performance increases that could boost model efficiency would very likely increase performance on GPU hardware to the point we could get continuous learning models from multimodal input like video data and other sensors.
While people may say something is a Transformer that's more of a general description. It's not a specific algorithm; there are countless transformers and people are making progress on finding new ones.
Bitcoin runs a specific algorithm that never changes. That's for an ASIC. AI/ML runs a large class of models. GPUs are already finely tuned for his case.
What if these AI investment were only to protect or strengthen their current business? AI in Windows, macOS, Adobe, iPhone, Facebook, Instagram may not bring in any additional revenue. But it add additional value to their current product line, making competition harder, further hardening their moat.
Nvidia or Jensen is also smart to play the national security card. Does the European want their model to be all US based. Are the answer culturally correct? Just like how every single country invested in their own Telecom or Internet infrastructure, if this pitch were even half as successful, do these numbers we are looking at even matter when it is spread out across G7 or G20?
While I believe we are still far, or at least 10+ years away from AGI, the current form of AI still has a lot of improvement incoming and are already bringing in real world benefits and value to a lot users. The adoption curve will accelerate once it is integrated into Windows, Office and Mac. So even if we are in a bubble, I still think we are very early in the curve before it burst.
So we have this bearish piece and the previous bearish Goldman Sachs piece. While I agree with their analysis in this case, there is a lingering doubt that some banks might just want to tank Nvidia a little in order to go long. Or something like that.
But other than FOMO why would someone buy better chips when they don't actually know what to do with their old ones?
So while I initially thought customer-facing roles would be front and center in the "AI revolution." Today, I tend to think they'll be bringing up the rear, with entertainment/smut applications at the forefront along with a few unexpected applications where LLMs operate behind the scenes.
#1 Just like the intranet there are a lot of productivity gains that firms like Tesla, Meta, Google and Amazon can gain internally by optimizing their own workflows. That in itself should justify the investment. Granted some of these optimisations will use their own chips instead of Nvidia, but Nvidia will get a lion-share of this.
#2. Then there are other verticals - pharma, oil and gas, logistics who can optimize their internal workflows to gain productivity. It just helps them improve their margins. No end user benefit may be realized and that’s fine.
#3. Nation states are buying GPUs too. Ex: Falcon2 was trained on a cluster owned by a middle eastern country. Nation states see something larger at stake here than just releasing an app. This does not have to even be a profitable endeavour.
The returns are going to chip-makers and employers, including single founder startups, who don't have to hire a lot of people, and additionally get productivity they never thought possible. A surgeon I know uses AI every day - to translate, to explain, to figure out problems, to write. They wouldn't have paid someone to do that, but now they get that output in seconds. This is a time to solve all kinds of problems we didn't think possible - because AI has made the enterprising among us instantly smarter.
All the fodder about AGI being a next step is smoke and mirrors - for everyone using OpenAI knows they don't need any more niche tools as their one $20 subscription is doing more for them every day. AGI is here. Experts can correct AI generated mistakes, but those are getting less and less too. The real benchmark is: Name how many people you know who can out-do ChatGPT on a question. You won't bother to check LinkedIn for that.
The gains are aggregating towards Chips, Clouds and Entrepreneurs. The VCs, since A16Z's original AI blog post (all expense, little return, echoed this Sequoia post but did it but years ago), know they are not needed as much anymore. Fewer VCs will beat the market when founders can grow startups without raising too much money (they don't need to hire as many people). Hiring needs lead to PR waves which require VC funding. Valuation is not a big deal for founders making money either, so they may not even disclose how successful their companies are. Bragging about your gains only invites competition. So other than ponzi-type ventures where you need to attract the dinner to serve dinner, you won't hear much about the good ones.
A different era indeed. The tech giants are in for a lot of change as well. Those who have distribution may try to push their models to the masses to be the point of reference, but that can get expensive, especially for those who don't charge. AI will help improve AI performance as well and that means cheaper better performance with time.
What's most needed in this era are people who know what the world needs that hasn't been invented yet. They need to be inventing and monetizing it. Little stopping you now.
The holy grail of AI is still in future. Where it can interact with software tools like we do. Competent AI Agents will be a huge productivity unlock.
I am also using this to solve problems that are usually delegated to junior developers. I get faster results (that I would have to fix up anyways) for far less effort.
Not just that AI is a great tool that will over time have a significant impact on society. It’s the breathless hype e.g. we have AGI today, society will be instantly smarter because of it and every tool and employee will be impacted. The FOMO i.e. you must jump on board now or be left behind. And the complete lack of any data or evidence.
The people expressing themselves with image AI would never have gone to fiverr or upwork to begin with.
And only some of the existing clientele would change their behaviors too.
Additionally, I know artists that always wanted to express themselves differently than what they became known for, and actually create and sell AI generated work to their clientbase now.
>The Coffee Test (Wozniak)
>A machine is required to enter an average American home and figure out how to make coffee: find the coffee machine, find the coffee, add water, find a mug, and brew the coffee by pushing the proper buttons.
That is a step yet to come.
Are the gains aggregating to entrepreneurs? It seems like every just a matter of time before the function of the big models overtakes any entrepreneurial idea other than hardware and the models themselves.
In software the UX is the big burden AI will overcome. Not having to deal with all the typing and clicking frees up more tasks and energy for deciding what needs to be done. Brain energy.
Distraction industries, like sugar, become obviously tart when oversaturated and AI is quickly pushing that boundary. Once people realize they’ve been served a pacifier to the brain in every free minute, they will start thinking more and consuming less.
So now AI frees more human brain energy towards what is actually needed - what problems need solved for humans, not for overloaded interfaces or zombified content consumers.
Examples: Trustworthy, efficient and safe transport for kids to and from enrichment activities. Education of others on tasks that could earn them a living. A better healthcare experience, because 60% of the admin paperwork is AI automated and health workers can talk to patients instead of typing. Organizing people to do more things together to clear the social-media-generated loneliness epidemic. Making new, interactive experiences people would pay to do, especially with friends. People ask for more things to look forward to. Not more things to type. AI is a new hands-free take on UX and UI.
But in which sector will these extemely important companies be active? Adtech? Knowledge management / productivity tools? Some completely new category?
What is an undeniable fact is the drastic commodization of the hardware / software stack for certain classes of algorithms. How is this technological development going to be absorbed and internalized by the economy feels still rather uncertain.
I dont think anyone knows the answer to that question.
Modern web is full of examples of platforms that really don't make money...
Sell! Sell! Sell now before it's too late!
Most AI startups aren't building massive data centers so they're unaffected. Most money isn't spent on compute in most startups. Only a few companies spend big.
It's obviously a terrible idea to invest massive amounts into compute when Nvidia's profit margins are so astronomical if you need ROI in the long term. The massive corporations won't get their money back for these investments; but they don't have to.
Investors first need to ask what that ratio might look like in 10 years. 10-to-1? 100-to-1? Inference-to-Training
Assuming for each NVDA training GPU sold there are 100 open source / commodity GPUs doing inference, who owns and supplies those data centers and hardware?
This is one of the first time in the tech industry where the value was fully reaped by the hardware itself and not by the differentiated software that ran on top of it
Silicon is currently the most advanced tech humans have ever made and those GPUs are on the cutting edge of that
The author of this blog makes a great argument that there is a risk of AI investments not paying off because if a revenue gap: The author argues that the gap between the revenue expectations implied by the AI infrastructure build-out and actual revenue growth in the AI ecosystem has increased from $200B to $600B. This is due to factors such as the subsiding of the GPU supply shortage, growing GPU stockpiles, and the dominance of OpenAI in AI revenue. The author also notes that the $125B hole in AI revenue has now become a $500B hole.
However, my experience with previous AI winters is not relevant here because now is the first time in history that there is a possibility of what we used to call “real AI” and now call AGI. No investor wants to miss out completely on the creation of near limitless wealth.
Assuming startups like Etched (with its recent massive funding) could shrink CapEx quite a bit (and make it not such a large revenue shortfall)
Just two words to describe all the idiocy of the current wave of AI offerings.
> A huge amount of economic value is going to be created by AI. Company builders focused on delivering value to end users will be rewarded handsomely. We are living through what has the potential to be a generation-defining technology wave.
No. It's crap no matter how much money you throw at it. It will end in tears, because there is no standard, no operating system, no protocol, only a handful of APIs controlled by a couple of companies. They are not building networks, but hubs with a single point of failure--the API provider. The internet is such a world-changing force because it is built on top of TCP/IP, which allows the rest of the internet to survive even if a part of it goes down. When an AI API provider shuts down, all those bullshitters repackaging LLMs will be left holding the bag, or rather their investors will be. In a way, the coming AI bubble burst is going to be a self-fulfilling prophecy--AI will make its creators redundant.
> But we need to make sure not to believe in the delusion that has now spread from Silicon Valley to the rest of the country, and indeed the world. That delusion says that we’re all going to get rich quick, because AGI is coming tomorrow, and we all need to stockpile the only valuable resource, which is GPUs.
No, the delusion is that Gen AI is good for anything. It is not.
Soon, the $600B question is going to be, "where's the money gone?"
What do you get if you multiply six by seven