Generative AI hype peaking?
bjornwestergard.com
bjornwestergard.com
Yeah, okay. I work each day with Copilot etc and the practical value is there, but there are so many steps missing for this statement to be true, that I highly doubt it.
My case is, wouldn't we already see the tools that are at least getting close to this goal? I can't believe that (or AGI in fact) to be a big bang release. It looks more like baby steps for now.
That cant even be considered a useful metric seeing as I spend similar time reviewing gpt written code
You might disagree with this assessment, but it doesn't show that he doesn't understand the field, since many of the people in the field also think this. At least to the definition of AGI he used - AI that can replace most of the economic work done by humans.
Have you ever interviewed people for software positions?
After being a part of the interview loop myself for a while, I am confident that “many of the people in the field” also don’t understand it. Not even talking about the leetcode gauntlet or complex systems design at all, just pure fundamentals and very basics. Not even talking about the larger scale picture of the business .
With that in mind, the fact that “many of the people in the field” he interviewed at “places with high amount of coding” agreed with him doesn’t say much.
And while it's true that many people don't understand much, I don't think this applies to many people working at frontier model or AI companies. Especially not the high level people Ezra said he talked with.
Maybe I am just morally corrupt enough to entertain this idea, but people working at AI companies overexaggerating and overhyping the impact of their work sounds like an obvious move.
Definitely not yet peaked IMO. However yea, I don’t see it fully replacing developers in the next 1-2 years — still gets caught in loops way too often and makes silly mistakes.
I am arguing the hype has peaked, and that there will likely be a pull back in investment in the next year. This is not to say the technology has "peaked", which I'm not sure one could even define precisely.
Important technologies emerged during each past "AI summer", and did not disappear during "AI winter". LISP is more popular than ever, despite the collapse of hype in symbolic reasoning AI decades ago.
As I mention in the OP, I think productivity enhancing tools for developers are one of the LLM applications that is here to stay. If I didn't think so, I wouldn't be concerned about the impact on skill development among developers.
The decline in NVDA stock price may also be due to newer models that require fewer GPUs, specifically from NVDA.
In other words, the demand may stay the same but if fewer GPUs in general or non-NVDA GPUs specifically get you to the same point performance-wise then the supply just went up.
Market may be pricing in possible takeover of Taiwan by China.
Not sure if this arrogant POV is very sustainable. This technology is replacing human coders as we speak, right now. Just not all of them yet, of course.
Amazon is especially vulnerable to tariffs. To be clear, I'm not really staking out a hard position on gen AI being the one true cause here, I just don't find "it's just part of the overall tariff driven market decline" to be very convincing.
It's unaffected by tariffs, but its insane valuation is driven by the narrative that Reddit posts can be used to train AI. Without that narrative, you have a semi-toxic collection of forums and the valuation would probably be somewhere in the millions at best, not the current $20 BB.
I think bootcamps will bloom again and companies will hire people from there. The bootcamp pipeline is way faster than 4 year degrees and easy to spin up if the industry decides the dev pipeline needs more juniors. Most businesses don't need CompSci degrees for the implementation work because it's mostly CRUD apps, so the degree is often a signal of intellect.
This model has a few advantages to employers (provided the bootcamps aren't being predatory) like ISAs and referrals. Bootcamp reputations probably need some work though.
What I think will go away is the bootstraps idea that you can self-teach and do projects by yourself and cold-apply to junior positions and expect an interview on merit alone. You'll need to network to get an 'in' at a company, but that can be slow. Or do visible open source work which is also slow.
That's not to say I never hire bootcamp candidates, it is that going to a bootcamp is not really a positive in my assessment.
And after this period, when companies start hiring juniors again, the amount of Stanford-like graduates may still be small because few wanted to go into CS. You have like a 2-4 year wait for people deciding to go into CS again.
If you are FAANG, you can throw money at the problem to get the best, but ordinary businesses probably won't be able to get Stanford grads during a junior-hiring boom.
Potential employers can (and do) verify your educational background such as a degree from an accredited university. Even if you had a certificate from a "legitimate code camp" (though I'd argue they're about as valuable as an online TEFL cert) - they have no way to verify it.
There is also a race to zero where the best AI models are getting cheaper and big tech is there attempting to kill your startup (again) by lowering prices until it is free for as long as they want it.
More YC startups accepted are so-called AI startups are just vehicles for OpenAI to copy the best one and for the rest of the 90% of them to die.
This is an obvious bubble waiting to burst. With Big Tech coming out stronger, AI frontier companies becoming a new elite group "Big AI" and the so-called other startups getting wiped out.
>100x growth ahead for sure.
Regardless of whether that distinction is useful, the author makes some fairly specific claims about inelasticity of demand, and seems only to lack confidence regarding the timing of Nvidia's fall not its inevitability.
I disagree with all of that.
But honestly, even chat apps are nowhere near their peak. Hallucinations and fine tuning issues are holding that segment back. There's a lot of growth potential there too as confidence and training help to increase adoption.
Huge amounts internet growth still happened after the 2000 crash, but networking gear and fiber optic networking became a commodity play, meaning the ROI shifted. The companies that survived ended up piggybacking the over-investment on the cheap, including Amazon and Google.
Even going way back, the real productive growth of the American railroads didn't happen until after the panic of 1873 after overbuilding was rationalized.
I hate to say "this time is different", but it really doesn't feel the same way, at least in public equities. nvidia has a high stock price, but they also have a P/E of ~36. Meanwhile, modern Cisco is ~27.
There might be some parallels though. OpenAI as modern Netscape might not be that far off.
I don't believe the previous 'summers' entailed quite the scale of [mal]investment that has occurred this time, so the impending winter will be correspondingly savage.
This isn't quite like the dot-com era, with tons of pure play investment options avaialable, regardless of whether you want to go long or short.
Yup.
2009 https://news.ycombinator.com/item?id=855684
2017 https://news.ycombinator.com/item?id=14790251
2018 https://news.ycombinator.com/item?id=17184054
2019 https://news.ycombinator.com/item?id=21613997
2022 https://news.ycombinator.com/item?id=30622300
It will get so much worse before it starts to fade.
Infecting every commercial, movie plot, and article that you read.
I can still here the Yahoo yodel in my head from radio and TV commercials.
I wanted to hear this again. Leaving it here for the next person: https://www.youtube.com/watch?v=Fm5FE0x9eY0
What a blunder by YC. What is this tool going to add to the conversation?
Hope he gets removed from speaker list.
I do think we're overbuilding on Nvidia and the CUDA moat isn't as big as people think, inference workloads will dominate, and purpose-built inference accelerators will be preferred in the next hardware-cycle.
On the one hand it has been two years of "x is cooked because this week y came out..." and on the other hand, people who seem to have formed their opinions based on ChatGPT 3.5 and have never checked in again on the state-of-the-art LLMs.
In the same time period, social media has done its thing of splitting people into camps on the matter. So, people – broadly speaking, no not you wise HN reader – are either in the "AI is theft and slop" camp or the "AI will bring about infinite prosperity" camp.
Reality is way more nuanced, as usual. There are incredible things you can do today with AI that would have seemed impossible twenty years ago. I can quickly make some python script that solves a real-world problem for me, by giving fuzzy instructions to a computer. I can bounce ideas off of an LLM and, even if it's not always 'correct', it's still a valuable rubber-ducky.
If you look at the pace of development – compare MidJourney images from a few years ago to the relatively stable generative video clips being created today – it's really hard to say with a straight face that things aren't progressing at a dizzying rate.
I can kind of stand in between these two extreme points of view, and paradigm-shift myself into them for a moment. It's not surprising that creative people who have been promised a wonderful world from technology are skeptical – lots of broken promises and regressions from big tech over the past couple of decades. Also unclear why suddenly society would become redistributive when nobody has to work anymore, when the trend has been a concentration of wealth in the hands of the people who own the algorithms.
On the other hand, there is a lot of drudgery in modern society. There's a lot of evolution in our brains that's biased to roaming around picking berries and playing music and dancing with our little bands. Sitting in traffic to go sit in a small phone both and review spreadsheets is something a lot of people would happily outsource to an AI.
The bottom line – if there is one – is that uncertainty and risk are also huge opportunities. But, it's really hard for anyone to say where all of this is actually headed.
I come back to the simultaneity of over-hyped/under-hyped.
The rest of society and our economy doesn't seem to be adjusting to hundreds of thousands or millions of people being "outsourced", so it's not likely there will be a lot of playing music and dancing for these people, though you may be more prescient than either of us are comfortable with, with the "berry picking" prediction...
Depends on your timescale, I suppose. But if AI is really about to take over everyone's work then we need to be having a much bigger discussion about what they imagine the billions of people on this planet will be doing in that scenario, and what kind of economic life they'll be living.
I had two defined types, both with the exact same field names. The only difference is one has field names written in snake_case, and the other had names written in camelCase. Otherwise the exact same
I wanted a function that would take an object of the snake_case type, and output an object of the camelCase type. The object only had about 10 fields
It missed about half of the fields, and inserted fields that didn't even exist on either object
You cannot convince me that AI is anywhere near to this level if it cannot even generate a function that can convert "is_enabled" to "isEnabled" inside an object
Every time I try this stuff I'm so disappointed with it. It makes me think anyone who is hyped about it is an absolute fraud that does not know at all what they are doing
If it cannot do that, then why is anyone saying it is a productivity booster?
The more code you ask it to generate, the higher the chances that it will introduce an issue. Even if the code compiles, subtle bugs can easily creep in. Not unlike a human programmer, you might say, but a human programmer wouldn't hallucinate APIs. LLMs make entirely unique types of errors, and do so confidently.
You really need to carefully review every single line, which sometimes takes more effort than just writing it yourself. I would be particularly wary of generated code for a dynamic language like Python.
It took a decade to reach LLMs. It will likely be another decade for agi. There is still clear trendline progress and we have clear real world targets of actual human level intelligence that exists so we know it can be done.
This person is just trying to get ahead of the game!
I propose an internet ban for anyone calling the generative ai top, and a public tar and feathering
Just think about the innovations over the last 100 years, how the world looked in 1925
I remember the 90s, and the time when translation programs were truly awful. They had pretty much no memory, not knowing what "it", or "she" might refer to, from the previous sentence. They tended to pick a translation based on the selected dictionary, and choked on typos and trademarks. A popular past time in my neck of the woods was feeding a translator the manual for a computer mouse, selecting the medical dictionary and giggling at all the incoherent talk about mouse testicles and their removal and cleaning procedures.
In light of that, ChatGPT is scifi magic. It translates very well, it deals with informal speech, it can deal with typos, and you can even feed it a random meme JPG and it'll tell you what's it about. It's really like living in the future.
That doesn’t mean I agree all jobs are going away, AGI is here, or all these other guesses people are making about the future.
That, of course, is the optimist's view. If you cynically see LLMs as a dead end that won't ever get that far even with Moore's law, then any day now we're going to come to our senses and give up on the whole thing, but looking at how we've come to our senses about crypto and Bitcoin is now worth $0, all I can say is that I'm along for the ride.
This much capital being poured into something and having very little to show for it is actually a bad sign, not a positive.
Putting it on the same shelf as the transistor, the jet engine, and the nuclear bomb is pretty funny. It's a probabilistic token generator. Relax.
Just like IoT, just like web3, just like blockchain, just like...
...but now it'll be exciting to let them bake. We need some time to really explore what we can do with them. We're still mostly operating in back-and-forth chats, I think there's going to be lots of experimentation with different modalities of interaction here.
It's like we've just gotten past the `Pets.com` era of GenAI and are getting ready to transition to the app era.
Perhaps as an extension of procedural generation, in interesting mechanics such as [1], or eventually even fully interactive NPCs.
PCs are starting to become more capable of running models locally, which can only make this tech more accessible. Like you say, we've barely begun exploring the possible use cases.
I also think it does the common-but-wrong thing of conflating between investment in big AI companies, and how useful GenAI is and will be. It's completely possible for the investments in OpenAI to end up worthless, and for it to collapse completely, while GenAI still ends up as big as most people clailm.
Lastly, I think this article severely downplays how useful LLMs are now.
> In my occupation of software development, querying ChatGPT and DeepSeek has largely replaced searching sites like StackOverflow. These chatbots generally save time with prompts like "write a TypeScript type declaration for objects that look like this", "convert this function from Python to Javascript", "provide me with a list of the scientific names of bird species mentioned directly or indirectly in this essay".
I mean, yes, they do that... but there are tools today that are starting to be able to look at a real codebase, get a prompt like "fix this bug" or "implement this feature", and actually do it. None of them are perfect yet, and they're all still limited... but I think you have to have zero imagination to think that they are going to stop exactly here.
I think even with no fundamental advances in the underlying tech, it is entirely possible we will be replacing most programming with prompting. I don't think that will make software devs obsolete, it might be the opposite - but "LLMs are a slightly better StackOverflow" is a huge understatement.
Maybe in the US.
I keep posting our work as an example and NO ONE here (Old HN is dead) has managed to point out any reasoning issues (we redacted the in-between thinking most recently like the thinking traces that people were treating as the final answer)
I dare you to tell me this is not useful when we are signing up customers daily for trial:
Https://labs.sunami.ai/feed
What is this?
Who is it for?
Why do you think this is demonstrates the OP is BS?
GOD
This wouldn’t be my biggest worry.