Below is another summary of the recent release Interview The one that’s making waves is a little more technical. Basically Sutskever is saying that scaling (achieving improvements in AI through more chips and more data) is leveling off, and we need new technologies; he’s open to that too neurosymbolic Techniques, and spontaneity. He is clearly not predicting a bright future for pure large language models.
Sutskever also said that “What I find most fundamental is that these models somehow generalize dramatically worse than people. And this is very clear. This seems very basic.,
Some of this may come as news to the machine learning community; This may be surprising given that Sutskever, an icon of deep learning, has inter alia worked on an important 2012 paper showing how much GPUs can improve deep learning, the foundation of LLM, in practice. He is also the co-founder of OpenAI, considered by many to be its leading researcher until he left after a failed attempt to oust Sam Altman.
But none of what Sutskever said should really be a surprise, especially to readers of this Substack, or to anyone who has followed me over the years. necessarily All This was detailed in my pre-GPT 2018 article “Deep learning: a critical assessment”, which argued for a neurosymbolic approach to complement neural networks (as Sutskever now has), and for more congenital (i.e., inherent rather than learned) constraints (what Sutskever calls “new inductive constraints”) and/or “in my 2022”Deep learning is hitting a wallEvaluation of LLM, which explicitly argued that Kaplan scaling laws would eventually reach a point of diminishing returns (as Sutskever had just done), and that problems with hallucinations, truthiness, generalization, and logic would persist even if models scaled, most of which Sutskever had just acknowledged.
Meanwhile, Subbarao Kambhampati has been debating about this for years Limitations of planning with LLMEmily Bender has been saying for ages that the excessive focus on LLMs is “sucking the oxygen out of the room” compared to other research approaches, Apple reasoning paper unfairly rejected highlighted issues of generalizability; Another paper named “Is the chain-of-thought reasoning of LLM a mirage? a data delivery lens“LLM logic and generalization hammered another nail into the coffin.
nobody What Sutskever said should come as a surprise. Alexia Jolicoeur-Martineau, a machine learning researcher at Samsung, summarized the situation well in X on Tuesday, following the release of Sutskever’s interview:
,
Of course it’s not over until it’s over. Perhaps pure scaling (adding more data and compute without fundamental architectural changes) Desire Somehow magically it has still been solved in a way that researchers like Sutskever, LeCun, Sutton, Chollet and me no longer think can be.
And investors would not like to give up this habit. As Phil Libin eloquently put it last year, scaling—not the generation of new ideas—is what investors know best.
And it’s not just that venture capitalists know more about scaling businesses than inventing new ideas, it’s that for venture capitalists who have pioneered much of the field, scaling, even if it fails, is having a great run: It’s a way to take your 2% management fee by investing someone else’s money on credible-ish-sounding bets that were actually bigger, which makes them rich no matter how things turn out. To be sure, if investment increases, VCs become even richer. But they are covered up somehow; Even if all this falls apart, venture capitalists will become rich from management fees alone. (It is their clients, such as pension funds, who will be affected). So venture capitalists may continue to support the LLM mania, at least for some time.
But let’s assume for the sake of argument that Sutskever and the rest of us are right, and that AGI will never come straight out of LLM, and to some extent they have run their course, and we really need new ideas.
The question is, what was the cost to the field and to society that it took so long for the mainstream of machine learning to figure out what some of us, including almost the entire neurosymbolic AI community, had been saying for years?
,
The first and most obvious answer is money, which I estimate is (about) a trillion dollars behind the envelope, most of it on Nvidia chips and huge salaries. (Zuckerberg apparently employs some machine learning experts at salaries of $100,000,000 a year).
According to Ed Zitron’s calculations, “Big Tech needs $2 trillion in AI revenue by 2030 or they waste their capex“. If Sutskever and I are right about the limitations of LLM, the only way to reach that $2T is to invent new ideas.
If the definition of insanity is doing the same thing over and over again and expecting different results, then investing trillions of dollars in increasingly expensive experiments aimed at reaching AGI may be delusional to the highest degree.
At a first estimate, all the big tech companies, from OpenAI to Google, Meta, XAI, Anthropic to many Chinese companies, keep doing the same experiment over and over again: building big LLMs in the hope of reaching AGI.
It never worked. Each new larger, more expensive model brings measurable improvements, but there seem to be diminishing returns (that’s what Suitskever is saying). kaplan law) and none of these experiments resolved the main issues related to hallucinations, generalization, planning, and reasoning, as even Sutskever now believes.
But the point is not just that a trillion dollars or more could be wasted, but also that there could be considerable collateral damage to the rest of society, both economic and otherwise (for example, in terms of how LLM has weakened college educationAs Rose Karma said in a recent article in The Atlantic, “The entire US economy is being fueled by the promise of productivity gains that appear far from being realized.,
To be fair, no one knows for sure what the scope of the explosion would be. If LLM-powered AI does not live up to expectations and its value diminishes, who will bear the brunt? Will it be only “limited partners” like pension funds that entrust their money to VC firms? Or could its consequences be much more widespread? Could banks collapse in a 2008-style liquidity crisis, potentially forcing taxpayers to bail them out? In a worst-case scenario, the impact of an inflated AI bubble could be huge. (Consumer spending, much of it by the wealthy, could also decline, which could weigh on the stock market, a recipe for recession.)
Even the White House has acknowledged concerns about this. As White House AI and crypto czar David Sachs himself said earlier this week, citing Wall Street Journal analysis, “Al-related investments account for half of GDP growth. A reversal.” [in that] There will be a risk of recession.
Quote from Karma’s article in The Atlantic:
that prosperity [that GenAI was supposed to deliver] Apart from their share prices, this has not yet been implemented anywhere. (The exception is Nvidia, which provides the critical input – advanced chips – that the rest of the Magnificent Seven are buying.) wall street journal Reports, Alphabet, Amazon, Meta and Microsoft have seen free cash flow There has been a 30 percent decline in the last two years. one by one guessMeta, Amazon, Microsoft, Google and Tesla will have collectively spent $560 billion on AI-related capital expenditures from the beginning of 2024 to the end of this year and brought in only $35 billion in AI-related revenues. OpenAI and Anthropic are Bring in are in a lot of revenue and growing rapidly, but they are still Nowhere near Profitable. His assessment – broadly $300 billion And $183 billionrespectively, and is increasing-Many times their current revenue. (OpenAI projects Nearly $13 billion in revenue was generated this year; anthropic$2 billion to $4 billion.) Investors are betting heavily on the possibility that all this spending will soon generate record-breaking profits. However, if that confidence is broken, investors could begin selling en masse, causing the market to experience a large and painful correction.
,
The dot-com crash was bad, but it didn’t cause a crisis. AI-bubble crash could be different. AI-related investments have already taken place across The level that it was at the height of the dot-com boom as a share of the economy. In the first half of this year, business spending on AI contributed more to GDP growth than all consumer spending combined JointMany experts believe that a big reason why the US economy has been able to withstand tariffs and mass deportations without a recession is that all this AI spending is working, Word What one economist described as a “massive private sector stimulus program”. An AI crash could lead to broadly lower spending, fewer jobs, and slower growth, potentially dragging the economy into a recession. economist noah smith logic It could also lead to a financial crisis if the unregulated “private credit” loans that financed much of the industry’s expansion were to stop all at once.
The whole thing looks incredibly delicate.
,
To put it bluntly, the world has “gone all the way” on the LLM, but, as Sutskever’s interview highlights, there are many reasons to doubt that the LLM will ever deliver the rewards that many had hoped for.
The sad thing is that most of the causes have been known for a very long time – although not widely accepted. All this could have been avoided. But the machine learning community has arrogantly excluded other voices and indeed all other fields such as cognitive science. And we are all going to pay the price now.
There’s an old saying about such idiosyncrasies that “six months in the laboratory can save you an afternoon in the library”; Here we would have wasted a trillion dollars and many years trying to rediscover what cognitive science already knew.
A trillion dollars is an awful lot of money to potentially waste. This could be much higher if the area of the explosion is wide. It’s all starting to feel like a story straight out of Greek tragedy, an avoidable mix of ego and power that could ultimately ruin the economy.
<a href