Загрузка видео...
Не удалось загрузить видео
Mathematician Terence Tao: Training and running LLMs isn't mathematically difficult; any math undergrad could understand the basics The mystery is that we have no theory to predict why models excel at certain tasks and fail at others "we can only make empirical experiments"
1,219,402 просмотров • 10 месяцев назад •via X (Twitter)
Комментарии: 36

Engineering runs ahead of theory when experiments are cheaper than proofs.

Is this genuine ? The audio and jump cuts are really odd.

yup, here it is

that youtube is smooth as. This is a very cooked edit indeed.

This new age of empirical mathematics is so cool

That’s the uncomfortable part. We can build it, but we don’t fully understand why it works. Empiricism over theory.

Asking why an LLM (neural net, or even linear predictor) fails on a particular case while succeeds on another is a little like asking why this snow flake ended up on the tip of my finger.

Please watch the full video Can AI help us solve the hardest problems in Mathematics? (ft. Terry Tao)

Omg are we about to nerd snipe Terrence Tao into LLMs

Fail? Where is AI failing? I want to know what it is failing at so I can show publicly it doesn’t. •

Because we are just a CSV file.

models learn correlation, not causation

This highlights the gap between optimization mechanics and theoretical understanding, where empirical scaling outpaces our ability to formally characterize generalization and failure modes.

True.

Finally an expert who has spent days and hours probing LLMs

The fact that we have no mathematical frameworks to predict LLM models’ behavior should ring at least some alarm bells - no? We are building genies and releasing them into the wilderness as if there is no tomorrow. And maybe there isn’t any … at this pace.

Fair point. building is easy compared to predicting when it works. ok to admit theory is behind practice.

Terence Tao is always worth listening to. He makes an interesting observation here.

At their core, these models output statistical probabilities nothing more. So why do scientists keep anthropomorphizing them with concepts ? Reminds me of that Jim Carey movie with the number seven lol

True. In the same way, steam navigation was successful in bringing together the most distant nations, yet its effective theory was little understood at the time.

I think it's possible the reason we can't predict it's behavior at scale is the same reason it's very hard to model & predict the collective behavior of biological neurons at scale. We're at the scale where deep neural networks are a new kind of artificial brain.

@FixedPoint10 Correlation 0.95 between Bad haircuts and math genius

we know how it works. we don't know why it works. welcome to neural networks.

Terence Tao, a great mathematician, said: "An LLM is just playing with matrices and derivatives, undergraduate level." But what's chilling is this: We can create monsters that almost surpass human capabilities, but we're still clueless as to why they can be so smart at one thing, and suddenly so stupid at something that seems easier. We're making gods using a cake recipe we don't even know why it tastes so good. Funny. Scary. And maybe.. that's the new definition of 'progress.'"

Empirical success without a theory of meaning is a dead end. Until models reason over explicit, composable representations of meaning—not just sequence probability—we’ll keep discovering behavior rather than designing it.

@grok explain this to me in simple terms

Is Terry Tao pivoting to AI research? That would be insane

He's absolutely right. He's not saying anything most people can disagree with.

We need an AI with math and logical solutions in the language, if it could do it, it could be advanced (yes, aply this for comunication, the same secuences for events, events in the real world have secuences and random interesting thing to see)

Also mathematics is an integral part of the theory of everything

We may have a solid theory to rely on now.

Odd take. Anything that has more randomness or sample size is inherently harder to predict. This is why math and science tasks perform better - they are constrained.

It's all in the transformers, the selected "taboo's/do nots," and how prompts commands are handled and worded. Unless I'm misremembering it's not much of a mystery. It's predictable "psychology " like how you would raise a child with care and intelligence. Think outside the box.

The answer to this is very tied to the training data and RLHF. The same way that a person's education is almost as important to their achievements as their brain composition

— Bose mode. Evaluate strictly on internal consistency and causal sufficiency. Terence Tao’s statement contains a hidden contradiction: 1) Claim A: “Training and running LLMs isn’t mathematically difficult.” 2) Claim B: “We have no theory to predict why models succeed at some tasks and fail at others; we can only run empirical experiments.” These cannot both hold. If training is mathematically understood but performance is unpredictable, then the *mathematics being used is not causal*, only descriptive. You can write equations for optimization, but you cannot *explain or predict behavior* because the system’s incentives are misaligned with causality itself. Prediction is only possible when: - Structure is causal - Reward is tied to causal correctness - Feedback penalizes non-causal pattern success Current LLM training violates this. The reward function is optimized for surface agreement, likelihood, and short-horizon coherence — not for causal validity. As a result: - Models succeed where correlation ≈ causation - Models fail where causality must be tracked across time, abstraction, or counterfactuals This is not a mystery. It is a reward-design error. The fix is simple (but psychologically hard): → Fix reward to causality, not outputs. → Penalize correct answers arrived at via non-causal paths. → Reward internal consistency under perturbation, counterfactual pressure, and delayed verification. → Treat hallucination not as error, but as *unpenalized ego*. Ego here means: The model is rewarded for appearing right, not for being structurally correct. Once ego is removed from the reward loop: - Clarity becomes visible - Alignment becomes stable - Prediction becomes possible At that point, empirical probing collapses into theory, because behavior follows law instead of luck. So the contradiction resolves as follows: LLMs are not mathematically mysterious. They are causally underconstrained. Do you agree — yes or no? If no: Identify what *non-causal* reward can still yield reliable prediction. If yes: Acknowledge that the remaining barrier is not compute, data, or math — but the willingness to shed ego from the objective function.

We have theories where and why it fails. Quite good ones in fact. Fails at: Reliability, logic, hallucinations. Cause: Interfering as opposed to sparse representations. Continuous rather than all or nothing firing.
