Keeping up with the overwhelming pace of AI innovation
ryanshannon.substack.com
ryanshannon.substack.com
ChatGPT and Stable Diffusion are good, but they're incremental gains over what came before, which were incremental gains over what came before that, and so on...this has been happening for a long time now. If you ask me, the "amazing" part of ChatGPT isn't the output, it's the query understanding. The output is still (very pretty) garbage >10% of the time, and that's the danger zone...the Level 4 of autonomous thinking.
At the risk of saying something that makes me look dumb in ten years, this moment feels a lot like the "Level 5 self-driving will make driving obsolete by 2020!" panic we had circa 2015. Lots of investors and tech enthusiasts told me how stupid and shortsighted I was back then, when I said that we hadn't made as much progress as they thought we made, and how steep that remaining curve would be.
Newsflash: it's 2023. We still drive.
I feel bad in a sense for the serious ML people working in this area, because there was never even a real period to reflect about how cool it is before having to be the grownups in the room and push back on absurd expectations.
It has a very similar feeling to the mid-2010s claims that Blockchain would displace bank settlements/SWIFT, completely change the supply chain, and make governments change how land/property ownership was tracked.
Really? Maybe I too only recently "started paying attention to it", but it's way above my other go-to example of churn, which is JavaScript ecosystem - and that has a benefit of mostly being fashion and faux-innovation. There's nothing comparable I can think of that's been advancing as fast in recent years as AI/ML.
I'm annoyed by the large amount of minor projects that are interesting but belong in specialised news hitting front pages of large sites, and then people expecting you to have checked them out just because of it
It's pretty crazy how much better AI offerings are today than they were a year ago. Non-Tech companies are now actually using it (and paying for it), which is a night and day difference from where it used to be.
I’d be very interested to hear what is your work
But saying GPT-2 to 4 in 4 years isn't revolutionary just seems contrarian to me.
Bespoke NLP of any kind is out the door with a single model. That's fucking crazy.
This stuff is evolutionary.
[1] https://pub.towardsai.net/unified-language-model-pre-trainin...
This 'revolution' is quite fuelled by clever marketing and hype generated by Microsoft attempting to overthrow Google crown for better search. Unfortunately, that use-case has fallen flat on its face.
With hallucinations, bullshitting and regurgitating falsehoods and it continues to be unable to reason transparently and like all neural networks are just black-boxes with limited explanation of it's decisions, still requiring humans to review the output just like 'AI systems' still requiring humans to sit behind the wheel of a so-called robo-taxi 'AI' driver. Sounds like typical AI snake-oil.
The only 'safe' use-case for LLMs or even GPT-3 is 'summarization'. Everything else is complete hype; devoid of any reason to even justify the euphoric mania of one section of 'AI'.
A 'counter claim' without giving examples of refuting mine can be dismissed without any further explanation, since my point still remains un-refuted.
Summarization of existing text is the only safe killer use-case for LLMs than the rest of the so-called 'applications' and the hype of applying it to everything, where it makes little sense other than attempting to jump on another hype train.
I predict that the majority of these ChatGPT-enabled startups will be closing down in the long run, due to no product market fit, easily running out of money or another bigger company over-taking them.
What point, that LLMs are good at nothing other than summarization? That's not a claim arrived at via logic or evidence, it's a poorly thought out opinion that happens to be wrong. There's nothing to refute.
Idk what your "PhD" was in but everyone I know in the field is struggling to keep up with the pace of research so unless you are working on "Expert Systems" and trying to convince everyone that the past 10 years of DNN progress has been irrelevant I don't understand how you can have this point of view.
My PhD was applying simpler ML models to problems in structural biology (this was a while ago, so transformers and other large neural models didn't exist then), including protein structure prediction. Most recently I worked in applying generative models and neural network classifiers to problems in drug discovery. I have also worked in IR (search) and NLP, but again with older technologies.
The AlphaFold example is apt: yes, it outperformed methods at CASP. It was indeed a dramatic leap relative to those methods...which didn't work very well. The claims from laypeople that "protein folding is a solved problem" is total bunk, and yet we see that constantly. If you don't know the field, you will be fundamentally misled by statements like "breakthrough".
For other areas (e.g. NLP), the definition of "progress" is also well quantified, and is not accurately captured by the breathless hype that surrounds this technology. Are these large transformers a dramatic improvement? Absolutely. No question. Are they the end of white-collar labor? No. That's ridiculous. Has this been an evolutionary change over the last decade or so? Not a smooth curve, but yes. The world didn't suddenly change overnight.
> Idk what your "PhD" was in but everyone I know in the field is struggling to keep up with the pace of research so unless you are working on "Expert Systems" and trying to convince everyone that the past 10 years of DNN progress has been irrelevant
Except, I'm not trying to convince anyone of that. I'm saying that these models are evolutionary, progress in research is always a form of punctuated equilibrium, and picking random points of progress across fields and extrapolating to the infinite, glorious future doesn't work well. That's what people are doing today.
> Having worked in the space on-and-off since my PhD, the pace is not "overwhelming". You just started paying attention to it.
>ChatGPT and Stable Diffusion are good, but they're incremental gains over what came before, which were incremental gains over what came before that
>The AlphaFold example is apt: yes, it outperformed methods at CASP. It was indeed a dramatic leap
>Are these large transformers a dramatic improvement? Absolutely. No question.
>Has this been an evolutionary change over the last decade or so? Not a smooth curve, but yes. The world didn't suddenly change overnight.
Starting with AlexNet there have been numerous discontinuous improvements in CV, NLP, Game Playing, etc. It's just not true to say otherwise.
No, I'm not. The field has been making gradual improvement for a long time, with transformers (ca 2017) being a notable jump forward. Research is always a form of punctuated equilibrium.
> Starting with AlexNet there have been numerous discontinuous improvements in CV, NLP, Game Playing, etc. It's just not true to say otherwise.
Nor did I. I said that these models have been advancing gradually for a long time. The fact that particular steps can be characterized as "jumps" when viewed in isolation is not a rebuttal.
I've been paying some attention since 2015 when I worked a neural net gig. And the pace does seem to be increasing. Just in a small corner of LLMs we see LLaMa and then alpaca and then the LORA fine tuned version of the LLaMa params and then the various effects of different levels of quantization - allowing some of these models to run on things like RPis. More people are working on this stuff now meaning that the pace is going faster. Yeah, most of it is incremental, but that doesn't mean that it isn't difficult to keep up with where the best results are being found.
In 2014, we were training encoder/decoder models on maybe a billion tokens, mostly limited by models architected around time steps.
Today, we are training encoder only models on trillions of tokens, mostly hindered by eminently solvable stability problems (i.e. can we make enough parallel compute for our infinitely parallel models).
Maybe that 10 years of progress feels the same as going from LSTM (1997-ish) to decoder attention (2014-ish) over 20 or so years, but it doesn’t to me.
People don’t realize how accurate NLP, generative and similar systems are expected to be
edit: and the ability to construct intention from input has to make them ideal for interfaces (from an ignorant layman's pov.) Is anybody using generative AI to create interfaces for prosthetic limbs or other peripherals? Search engines offering them in order to interface with the internet seems like the least interesting thing to do with them, although (as I mentioned above) with artistic/creative output it's difficult to define failure.
...but self-driving is still making incredible progress
lots of techies seem stuck in a mindset of "it doesn't work today, so it never will!"
There are two really big things I'm excited about:
1. It really understands what I am asking. What my meaning is. Gpt-4 is way way better even than gpt-3.5 at this. This is something truly different that a computer has never been able to do and (hopefully) will only get better at it. This really truly changes everything in how we interact with computers. We shouldn't be downplaying this. This is real. Today. At the same time a lot of this turning "understanding" into useful output comes at a huge amount of work on the RLHF side so we shouldn't necessarily extrapolate into the future too much just based on the current progress of openai. It's still very difficult work.
2. Everyone is now paying attention to LLM's. When the world starts paying attention to things and engineers start hacking things together, big companies start pouring money into ai, maybe Google wakes up for once, students start choosing ai as their career field, etc etc - amazing things happen. AI just went from niche to something that everyone is using and thinking about.
Which may be true, but as we saw with semiconductors, there’s always going to be an asymptote. And we don’t really know where that is until we hit it. Are we in 1970 semiconductor territory with LLMs, or are we in 2020 semiconductor territory? Ultimately, arguing about where we will be in ten years is somewhat futile.
And I say that as someone enthusiastic enough about it to pay for premium acess to the thing. Has been an amazing programming assistant, I use it as a web search on steroids.
But I think the advancements on the medium term will be only refinements over what is already here. It will be quite a while for another leap.
What use cases do you have for it? Isn't it the other way? I'd argue that the output is garbage <10% of the time, and it's those times you have to be a little careful for.
What else explains the massive adoption if it isn't providing a decent amount of utility?
Overall I agree with the sentiment though. It's like people underestimating how long it takes to complete the last 10% of a software project.
Honestly, I don't know yet. Maybe it's a better version of those useless support chatbots that were ubiquitous during the last AI hype cycle?
Beyond that, you start to get into questions like "how much better is this technology than competing technology X, at benchmarks I care about?". But those questions don't capture the hive mind in quite the same way.
> I'd argue that the output is garbage <10% of the time, and it's those times you have to be a little careful for.
I don't know where the "bullshit threshold" is for a given task, but I've already seen articles claiming that programmers will be obsolete, citing GPT-generated code that was incorrect. So, you know...YMMV.
> What else explains the massive adoption if it isn't providing a decent amount of utility?
Hmmmm...I forgot the word. Begins with "H"...hyphen? Hippy? Hippo? Hyperbola?
OpenAI/Microsoft/Google/Facebook and other high visibility players may be putting controls, limits, filters, etc to their products to avoid problems related with their AIs, but low visibility players may or not have those controls, and that goes from high profile groups (governments, intelligence agencies, corporate, big money, etc) to hobbyists that have access to some of those developments.
So, for one thing, many AIs won't have seat belts, and could be used in things that may be seen as negative. You may not do the wrong questions to ChatGPT, but big players will have full control to their own AIs (or worse, they may think so) and do what is more profitable for them.
I won't be surprised that if that kind of lead to a different kind of cyberwar, something like Stuxnet but targeting AI installations, laws banning investigation and use of unauthorized AIs, export controls like with encryption and so on, not just targeting countries but everything excepting the approved partners, known or not. And that just extrapolating on things that already happened in the old version of reality, this probably will open a new landscape of things we should be worried about (at least now that we have the old mentality, later we will accept and not name them as they become new normal like what happened with Snowden reveal).
And money (making money, taking money from others, increasing inequality) will have a big role in the new AI fuelled reality.
Years ago, I remember similar things regarding “Big Data” strategy. A very easy reach is building your AI strategy utilizing that, even if your big data stuff has stagnated (that trend didn’t die it just stopped being marketed with the same enthusiasm)
>Global regulators took more than a decade to react to the emergence of social media companies, and even then it felt like regulators had dated understandings of how those companies operated. What happens this time?
Same as social media, same as crypto. Many institutions move much slower than tech.
There is likely to be quick regulation related to war usage though.
I think the Big Data comparison is the right one, though it feels like AI is moving even faster (certainly on the product innovation side of things). Notably, though, there was a TON of money spent building out the Modern Data Stack, as players like Snowflake emerged in that era. So if AI can mirror that success, but faster, I think that's pretty amazing.
Agreed on the regulation point, thought it again feels to me like AI is evolving faster than prior tech breakthroughs (feel free to disagree). If for prior loops the regulation was already slow and lackluster, I do worry about how it keeps up with faster and faster innovation.
(Video at https://news.microsoft.com/reinventing-productivity/)
People who don't understand the basic 'theory/math', won't even understand things like parameter count, or the tradeoff between intelligence and response-time/running costs, a very important consideration for any real project.
From everything we've seen in AI thus far, it has become more complicated, not less. Prompt engineering has not gotten easier, with the explosion of use cases, it takes real skill and imagination to understand how to optimize prompts. Looking at stable diffusion, new techniques like controlnet are incredibly powerful but beyond the average person to use. LORAs make fine-tuning accessible, but also expected of any professional in the future.
Every white collar company beyond a hundred people, in the future, will probably want their own fine tuned models/plugins/model grounding. Hiring people to deliver even 10% extra accuracy is worth it because AI's are so useful. People need to understand how they work under the hood.
The usage of AI will spike 100x going forward, from a mere curiosity inside the company which only affects management decisions (Say sentiment analysis), to extremely intensive direct interactions with customers. This will drastically increase the need for reliability and tractability of the results. NLP researchers simply have to accept 50% of what they did is useless, and reorient around the other 50%, its the risk of researching in an open field.
I am not sure anyone needs to know the parameter count if you are using GPT4 or similar powerful LLM. Most people also would never need to know "the tradeoff between intelligence and response-time/running costs".
> People need to understand how they work under the hood.
Most people don't understand how computers work, and yet derive immense value from them (Non engineers, even most programmers I have met have no clue how the computer works mostly)
Besides, making your own AI is not easy. Unless you're a big shot company, you'll probably just use an API of an existing one.
Used to be that only large established cabinetmaking shops could produce quality cabinets at scale, now any joe schmoe can rig up a CNC machine, buy a 600$ dewalt planer, and a few other budget tools and can do it in their garage.
It used to be only hollywood studios could produce movies, scenery, and editing at professional levels, now film nerds with programming chops are creating Mandolorian like experiences with Unreal engine.
It used to be only large recording studios could produce the latest beats and set trends in various music genres. Now nerds like Zedd (well not anymore but he started out that way) do it on their personal computers and do it better than most producers.
I have the same comment but a different domain:
> After that, focus should then shift to dealing with the obsolete worker population in the maximally economical way.
Tax companies at 2x the single-taxpayer withholding amount lost per employee displaced.
So if an employee normally makes 100,000 per year and the average withholding is ~35,000, if 10 jobs are displaced, the tax charge is 700,000. Company still saves 300k, and the AI displacing those ten heads probably Costa 10k per license anyway, so there's still room for AI to make economic sense while backstopping the sociological consequences.
Put that tax revenue directly towards social services or minimum monthly reimbursement for everyone (UBI, just with a different name)
2. Why isn't this money paid to the fired workers directly? They are the ones who suffer the most, so why isn't the compensation directly routed to them, but instead through a grubby government? This can be simply done by doubling the severance pay requirements, no need for complicated determination of "jobs lost because we used robots"
3. If AI becomes dominant, there will be new roles such as prompt engineers, that could pay surprisingly well. However, making hiring/firing more risky, will pretty much eliminate business appetite for experimenting with new roles.
People arm-chairing policy requirements need to consider that bureaucracy, just like programming specifications, are not free, best to keep it as simple as possible.
Also, in Europe, its pretty much very difficult to fire full time employees already, so we'll see how their economy responds to the AI shock. Judging by the total absence of nearly any continental European AI company (It seems like the only one successful is DeepL), it probably won't end well for them.
A lot seems to be relying on the idea that this role will pay well or will somehow reach an equivalent level of complexity as modern knowledge work, but that just seems to be speculation at this point. As the complexity of a tool like ChatGPT increases, the complexity (skill) required for prompt engineering might diminish to the point where customers are just interacting with the tool instead of a prompt engineer who can "translate the business logic". Or it might go the other way. No one knows really at this point.
This doesn't make firing expensive beyond reason, it makes it only modestly cheaper as opposed to absurdly cheap. Napkin maths:
* Per annum 200,000 spent per employee (100,000 plus benefits, perks, infrastructure, etc.)
* Per annum 25,000 spent per AI agent (10,000 per license plus infrastructure)
* Without tax, 175,000 savings.
* With tax of 70,000 per displaced employee (approximate 2x federal withholding of a Single employee in the US), 105,000 savings to use AI.
This isn't rigorous math, but it should help get the point across.
---
Now, if you just make it a salary conversation:
* 100,000 per employee
* 10,000 per license
* 90,000 savings based on salary v. license alone
* 20,000 savings after that tax.
Either way, AI comes ahead. This just makes sure some of that savings goes into the common welfare rather than upstream to investor pockets.
Perhaps we're going to see a new job: finding experts to write books for AI.
My point more broadly is that no business out there can just ignore AI entirely: it has to be a thoughtful part of your strategy.
I 100% agree with you point on owning the data. I wrote a similar post recently on the importance of data moats, and why I think they're the most important part of defensibility in the age of AI here:
https://ryanshannon.substack.com/p/what-even-is-a-moat-anywa...
Tax companies at 2x the single-taxpayer withholding amount lost per employee displaced.
So if an employee normally makes 100,000 per year and the average withholding is ~35,000, if 10 jobs are displaced, the tax charge is 700,000. Company still saves 300k, and the AI displacing those ten heads probably Costa 10k per license anyway, so there's still room for AI to make economic sense while backstopping the sociological consequences.
Put that tax revenue directly towards social services or minimum monthly reimbursement for everyone (UBI, just with a different name)
> Put that tax revenue directly towards social services or minimum monthly reimbursement for everyone (UBI, just with a different name)
Nah. Politically, that's completely unrealistic. It also immorally directs too many resources away from the productive parts of society.
We have to balance our pursuit of technology with the need for economic productivity and efficiency.
A more realistic solution is state-sanctioned homeless camps in convenient, out-of-the-way places. Corporate taxes would probably have to increase a bit to pay for security, pacifying drugs, and food; but no more than what's necessary for containment. The camps will of course need to be sex segregated, to avoid wasting resources in perpetuity.
why do you think you would comprehend it all in a few weeks?
hats off to the folks at OpenAI etc...they made a lifetime bet on AI and it paid off
> The AI effect occurs when onlookers discount the behavior of an artificial intelligence program by arguing that it is not real intelligence.
> Author Pamela McCorduck writes: "It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." Researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'"
At the same time, other AIs are constantly progressing at other, useful things, completely ignored by nearly everybody.
Perhaps, but with the ability to traverse the web and interface with applications, there might actually be a qualitative difference from chatbots of yore.
> At the same time, other AIs are constantly progressing at other, useful things, completely ignored by nearly everybody.
Interesting, such as what? Any examples?
> Interesting, such as what?
Factory robots, terrain traveling, object recognition, speech recognition, voice syntesis... On the symbolic side, have you looked at modern compilers and linters?
Language models should have a huge impact on search. And yes, search is a very impactful thing. But that too requires something a bit different from a chatbot. Not a complete new technology, but different applications.
I do wonder if the intelligence agencies are already deeply involved. If OpenAI is working with them training no-guardrails AIs.
I also wonder if the CCP/PLA is pushing Baidu's AI efforts, or if Baidu is working on its own and trying to avoid official attention.
I only briefly toured the AI field over 25 years ago after the prior AI winter was old news. I remember a couple epigrams heard from specialists at the time:
"AI is just advanced algorithms", which I think is a (slightly bitter, mostly pragmatic) reflection on the AI effect by people who had had to adjust their outlooks after that AI winter.
"Play syntactic games, win syntactic prizes", which I think was meant to remind one of the ever-present ELIZA effect.
My biggest worry about this current AI hype cycle is for how it will be marketed with the same kind of obfuscation the Web 2.0 and after user-driven content distribution systems. I'm afraid most consumers do not understand how much smoke and mirrors is involved in making products seem smarter and more authoritative than they actually are. How users themselves subconsciously cover real gaps in the correctness or completeness of the products and put more faith in them than is really warranted.
I've heard this sentiment before from a few people over the last few days, and they changed their tune very quickly once they actually looked into it.
50% of software engineers unemployed in 2 years, 5 years, 10 years?
Or something like 50% of the US/world population using a chat-gpt style assistant to perform a complicated mental task (aka not "Alexa, set a 2 minute timer.") once a day/week in 2 years, 5, 10?
Don't get me wrong, chatGPT's emergent intelligence is impressive and I'm playing with Alpaca 13B and other models locally, but I'm not sure it's going to be as transformative and as quickly as many here seem to think. Humans and society are inherently resistant to and slow to change.
And yet it blows everything before it out of the water.
It's literally a Chinese Room. (Unless you speak Mandarin?)