janalsncm

Checks out. China has 10,000 TWh, the US only has 4,000.

Your question is a little like asking why we arrest arsonists when there are fires to put out. The law does not need to choose.

If you are training an LLM, you need to tokenize the text before it’s trained on. A lot of time this can be done in parallel with the GPU though.

I have spent way too much time waiting 10-15 minutes tokenizing my training dataset only for the run to crash over some minor bug after that. (If I was smarter, I’d test on a smaller batch first.)

It is guaranteed to work worse than top_k=1, that’s for sure.

China can keep up because it's cheaper to run a frontier lab there

Not sure if this is what you meant, but their training runs are significantly cheaper. This was one of the big shockers from the Deepseek R1 paper. US foreign policy has helped to ensure that the Chinese are compute constrained, so they literally cannot buy the most expensive and powerful training rigs.

This has led to a steady drumbeat of innovations which are not revolutionary on their own but stack together to make things much more efficient.

Yes, the human actors in your scenario were the ones who built the autonomous cannon and turned it on while knowing that 1) a good neighbor does not destroy their neighbor’s property 2) cannons can destroy property.

Also OpenAI specifically turned off their own cybersecurity guardrails to run this experiment. In other words it was able to escape the lab specifically because they turned them off. A human made the choice to turn off the guardrails.

It would be a pretty big plot twist if we found out that Shai Halud was a worm created by GPT during testing.

Absolutely bewildering. If I am building a giant cannon and blow a hole straight through my neighbor’s house, I’m not going to say “we are working with our neighbors to improve their giant cannon defenses”.

OpenAI brought this weapon and as far as I’m concerned they used it on another party. Morally it probably matters that this happens because they don’t know how their weapon works. Legally I always thought it was ill-advised to accidentally hack people too.

By Anthropic. The government blocked it after it was already released.

Every LLM product goes through testing and alignment after training, maybe even some quick improvements here and there. Kimi probably did something similar.

Put another way, if Google says they have the best model in the world but won’t release it in December I will start caring in December, not before.

I’m glad you only find crime “annoying”. Others find it oppressive. For example the murder rate in New York City (a similarly dense urban area) is around 35-50x the rate of Singapore. This means your chance of being murdered within the first two weeks of living in NYC is equivalent to a whole year in Singapore.

You don’t have to walk around wondering if you will get mugged or shot. You don’t worry that your car or apartment will be broken into. Crime solve rates are very high because police have fewer crimes to deal with.

That and/or the business case isn’t as clear when serving enormous models? You’re constantly stuck in a red queen’s race where your profitability window is increasingly measured in weeks because the Chinese are right behind you.

For small models (which are probably distilled from their big ones) you can serve them economically all the time and not hemorrhage money.

There are two separate types of events with associated probabilities: 1) the probability of being harmed by criminals, and 2) the probability of being harmed by the government.

For most people in most places, police are ineffective and therefore the first is orders of magnitude more of a problem. For example in the Bay Area, car break-ins are incredibly common and police are incapable of solving it. This would be unthinkable in Singapore.

unless you happen to express disagreement with any part of it

On the other hand, the government in Singapore is genuinely popular. A majority eligible of voters chose to re-elect the incumbent. For reference, only 32% of eligible American voters chose the current US president.

He was their “chief AI scientist” (what an abomination of a title) and led FAIR. Maybe it’s true he was not working on LLMs but I can’t even imagine what more important thing he could be doing at Meta.

Which brings me to my other point. If you want to compete with serious labs (OpenAI, Anthropic, Deepseek, Alibaba) you need to be smart and focused, not messing around.

Right, but retail investors weren’t supposed to find that out until after the IPO.

1) No one is asking Anthropic to give tokens for free, but at market rates.

2) Any professor who tried to ban students from posting lecture notes online would be immediately mocked.

Probably a little of both.

Chinese labs have come up with a bunch of genuine innovations: GRPO, auxiliary loss free MoE load balancing, MLA, muon optimizer, and a bunch of other ones. The Deepseek papers are really well written, this isn’t just sneaking a peek at a peer.

The problems are inherently harder now too, partially because they take longer, so your training pipeline is waiting for long completions.

Also there probably is some “distillation” (technically pseudo-labeling, which is common in ML). But I wouldn’t put too much weight on it because that was true 18 months ago as well.

Why are you using the product release date for Kimi K3 and the training date for Fable? Either use the release date for both (6 weeks apart) or if you have it the training date for both.

I would argue your logic is exactly backwards. For most Americans, sharing data with the Chinese government is irrelevant. The CCP cannot put you in jail.

However, American providers are going to be subject to secret national security letters, FISA court warrants, and regular court orders.

I am no fan of Wang but he came after Llama got caught benchmaxxing Llama 4 rather than training a good model. My read is that Zuckerberg tried to buy his way out of the problem like he always does, and he ended up overpaying for a lemon.

At the time the whole thing was led by Yann LeCun who seemed to spend more time arguing with people on Twitter than figuring out new techniques to make Llama the best. Meanwhile Deepseek was figuring out large scale RL on kneecapped hardware like H800s and how to scale architectures an order of magnitude bigger with MoE.

China wants the US economy to flounder

It’s not that simple. If the US economy goes into a recession, it will take large sectors of the weak Chinese economy with it either directly or indirectly.

It’s probably more accurate to say they don’t want American LLMs to become dominant. The huge US data center build out doesn’t depend on Anthropic and OpenAI anyways. Those data centers can just as easily serve Qwen or GLM models.

“Bank error in your favor, collect $100”

It is a huge improvement to PPO because you don’t need a separate critic model which cuts memory costs in half and stabilizes training.

I wonder what Sutton thinks about some of the more recent innovations in RL like GRPO. In some ways it’s new, in other ways it’s an echo of RLOO.

I’m not reacting to the idea of using an LLM as a judge in general. That’s a proven path.

I’m specifically reacting to using it to reward the chain of thought during RL training because models love to hack their rewards, learning any possible shortcut rather than the task we want them to.

I really feel out of my depth because 2 out of the 3 methods here seem like they shouldn’t work?

To evaluate comprehensibility quantitatively, we employ an LLM-as-a-Judge framework

This isn’t the worst idea, but it’s still a bit incestuous. Adding an LLM judge to check for hallucinations creates two new kinds of problems: false positives, where your judge hallucinates an incorrect fact, and false negatives, where the judge lets a hallucination slip by.

We measure reproducibility through knowledge distillation. By fine-tuning a weaker model on the generated CoT traces, we use the downstream performance gain of the student as a proxy.

And my problem here, as a member of the GPU proletariat, is that this just seems incredibly inefficient. In other words, you’re going to generate a bunch of rollouts from your model then wait for the student to train? I guess if you have the compute to train a trillion params then maybe you don’t care.

1) Mixture of experts as an LLM architecture is not the same thing. In MoE each “expert” can activate for any given token.

2) Bitter lesson is misunderstood. Specialization and inductive biases still matter. ChatGPT isn’t the best chess player in the world just because it’s seen more math problems or read more Japanese poetry. Stockfish is, because it bakes in useful inductive biases like minimax.

software is solved with AI

Presumably every car manufacturer can use AI. Yet there are still bugs. If all bugs are solved with AI, and therefore every car manufacturer with access to AI writes bug-free software, the only remaining conclusion is that some car manufacturers don’t have AI yet.

3 of the 4 agents used Wan, the other one used Seedance.

I do mean losing that specific sale. I’m also referring to the economists who saw an increase in sales.

This means there are marginal readers who previously didn’t buy any book but now buy AI slop books, or buy slop books in addition to human ones.

You definitely can rule out the general case a priori. If the problem were possible, for every text there would be a unique provenance label “human” or “ai”. But since humans and machines have both written many texts, it is not possible.

As an example, you could imagine a giant lookup table that deterministically mapped every text ever written to “human” or “AI”. You would very quickly run into situations where the labels conflict for the same piece of text.

The data is statistically inseparable which makes it impossible to classify from text alone.