Latest news from renowned machine learning researcher Ilya Sutskever:
Below is another overview that was just released interview His work that is making waves is a little more technical. Essentially, Sutskever is saying that scaling (achieving AI improvements through more chips and more data) is reaching a plateau and new technologies are needed. he is even more open neurosymbolic technology and talent. It is clear that he does not predict a bright future for pure large-scale language models.
Sutskever also said this.What I think is most fundamental is that these models just somehow generalize dramatically worse than humans. And it’s very obvious. It seems so basic.”
Some of this may be news to many in the machine learning community. Sutskever, a deep learning icon, may be surprising. He specifically worked on an important 2012 paper showing how much GPUs can actually improve deep learning, the basis of LLM. He was also a co-founder of OpenAI and was thought by many to be the company’s lead researcher until he left after failing to oust Sam Altman.
But none of what Sutskever says should come as a surprise, especially to readers here at Substack and those who have followed me over the years. essentially all It was in my article before GPT 2018.Deep learning: A critical evaluation” advocated a neurosymbolic approach to complement neural networks (like current Satskeva). congenital (i.e., built-in rather than learned) constraints (what Sutskever calls “new inductive constraints”) and/or my “Deep learning is hitting a wall” The LLM evaluation clearly argues that Kaplan’s scaling law eventually reaches a point of diminishing returns (as Sutskever did) and that even as the model scales, problems with illusions, truth, generalization, and inference remain, many of which Sutskever just acknowledged.
Subbarao Kambanpati, on the other hand, has been arguing for years that: Limitations of planning with LLM. Emily Bender has long said that focusing too much on LLM “sucks the oxygen out of the room” compared to other research approaches. of Reasons why Apple was unfairly rejected I uncovered a generalization problem. Another paper calledIs LLM Thought Chain Reasoning a Mirage? Data Delivery Lens” is yet another nail in the coffin of LLM reasoning and generalization.
none Sutskever’s statement may come as a surprise. Alexia Jolicoeur Martineau, a machine learning researcher at Samsung, summed up the situation nicely on Tuesday, following the publication of Satskever’s interview.
§
Of course, it’s not over until it’s over. Probably pure scaling (adding data and compute without changing the underlying architecture) intention Somehow, magically, it solves what researchers like Sutskever, LeCun, Sutton, Chollet, and myself thought was no longer possible.
And investors may be reluctant to kick the habit. As Phil Libin presciently stated last year, what investors know best is scaling, not generating new ideas.
And not only do venture capitalists know more about scaling a business than inventing new ideas, but for venture capitalists who have led many fields, scaling, even when they fail, has been a great accomplishment. For a 2% management fee, they take other people’s money and invest it in really big, plausible bets that make them rich no matter the outcome. Indeed, successful investments make VCs richer. But they’re covered either way. Even if everything were to fail, the venture capitalists themselves would be wealthy just from operating costs. (Customers such as pension funds will be hurt). So venture capitalists may continue to support LLM mania, at least for a while.
But for the sake of argument, let’s assume that Sutskever and the rest of us are right, that AGI will never emerge directly from LLM, that we are on track to some extent, and that we actually need new ideas.
So the question is: What has it done to the field and society that it has taken the machine learning mainstream so long to understand what some of us (including virtually the entire neurosemiotic AI community) have been saying for years?
§
The first and most obvious answer is money. I estimate it to be (roughly) $1 trillion, with much of it going to Nvidia chips and huge salaries. (Zuckerberg apparently hired a machine learning expert for $100 million a year).
According to Ed Zitron’s calculations:Big tech companies need $2 trillion in AI revenue by 2030 or waste capital investmentIf Sutskever and I are right about the limitations of LLMs, the only way to reach that $2 trillion is by inventing new ideas.
If doing the same thing over and over and expecting different results is the definition of insanity, investing trillions of dollars in ever more expensive experiments aimed at achieving AGI may be delusion of the highest order.
First, in approximation, all the big tech companies, from OpenAI, Google, Meta, xAI, Anthropic, and even some Chinese companies, are repeating the same experiment over and over again, building ever-larger LLMs in hopes of reaching AGI.
It never worked. Each new model, larger and more expensive, has brought measurable improvements, but revenue appears to be decreasing (that’s what Sutskever is saying). Kaplan method) and, as Sutskever now recognizes, none of these experiments solve the core problems of hallucination, generalization, planning, and reasoning.
But it’s not just that more than $1 trillion could be wasted, it could also cause considerable collateral damage to society as a whole, both economically and otherwise (e.g., in terms of how). LLMs are undermining university education). As Roger Calma stated in a recent article in The Atlantic:The overall U.S. economy is supported by the promise of increased productivity, but that seems far from being realized.”
To be fair, no one knows what the explosion radius will be. If LLM-powered AI fails to meet expectations and its value declines, who will suffer the consequences? Will limited partners like pension funds be the only ones who entrust their funds to VC firms? Or could the impact be even more far-reaching? Could banks also go bankrupt in a 2008-style liquidity crisis, perhaps forcing taxpayers to bail them out? In the worst case scenario, the impact of the deflation of the AI bubble could be enormous. (Much of consumer spending is supported by the wealthy, which could hurt the stock market, but this could also fall, contributing to a recession.)
The White House has also acknowledged concerns about this. As White House AI and crypto czar David Sachs himself said earlier this week, citing a Wall Street Journal analysis, “Arles-related investments account for half of GDP growth. It’s a reversal.” [in that] The economy will be at risk of recession. ”
From Karma’s article in The Atlantic:
its prosperity [that GenAI was supposed to deliver] Other than stock prices, little has yet materialized. (The exception is Nvidia, which provides a key input that the rest of the Magnificent Seven are buying: advanced chips.) wall street journal According to reports, Alphabet, Amazon, Meta, and Microsoft. free cash flow It has decreased by 30% over the past two years. one by one estimateBy the end of this year, Meta, Amazon, Microsoft, Google, and Tesla will have spent a total of $560 billion in AI-related capital spending since the beginning of 2024, with only $35 billion in AI-related revenue. OpenAI and Anthropic bring Lots of revenue and fast growth, but still nowhere near make a profit. Their evaluation is roughly $300 billion and $183 billionrespectively, and rising— many times higher than current earnings. (OpenAI project This year’s revenue is about $13 billion. human$2 billion to $4 billion) Investors are betting big on the prospect that all this spending will soon yield record profits. But if that belief falters, investors could start selling en masse and the market could suffer a major and painful correction.
…
The dot-com crash was bad, but it didn’t cause a crisis. It might be different if the AI bubble bursts. AI-related investments are already being made exceeded The level that telecommunications’ share of the economy reached at the peak of the dot-com boom. In the first half of this year, corporate spending on AI exceeded overall consumer spending. combined. Many experts believe that the main reason the U.S. economy was able to weather tariffs and mass deportations without causing a recession is because all of this AI spending is working. words One economist described it as a “massive private sector stimulus program”. An AI crash could lead to widespread cuts in spending, fewer jobs, and slower growth, dragging the economy into recession. economist noah smith claim It could even lead to a financial crisis if the unregulated “private credit” loans that financed the industry’s expansion collapsed all at once.
The whole thing looks incredibly fragile.
§
Frankly, the world is “all in” on LLMs, but as Sutskever’s interview highlights, there are many reasons to doubt whether LLMs will deliver the rewards many people expected.
Sadly, most of the reason is that it has been known for a very long time, although it is not widely accepted. It all could have been avoided. However, the machine learning community has arrogantly excluded other voices and even other fields, such as cognitive science. And we may all be paying the price now.
The old saying about this kind of stupidity is, “Six months in a lab will save you an afternoon in the library.” Here we may have wasted a trillion dollars and years rediscovering what cognitive science already knew.
$1 trillion is probably a terrible amount of money to waste. If the explosion area is wide, it can spread even further. It’s all starting to feel like a story straight out of Greek tragedy, an avoidable mixture of arrogance and power that can ultimately destroy an economy.
