HN user
Has Google even being trying to keep up with frontier sized models, and for that matter why should they?! Even if they want to, what's the rush? Someone from Google just tweeted today that they just recently STARTED their "Gemini 4" (whatever that may be) pre-training run... In the meantime they just today announced financial results with cloud revenue up 82% YOY ... they seem to be doing just fine.
Either you've memorized the outline (or detailed component shapes) of, say, a horse, or you haven't. Memorizing the outline of a pelican isn't going to help you with the horse.
You could train a model to do something a bit different like a pencil sketch, or vector graphic sketch, of something given a photo of it, and expect that to be a generalized skill, but if you are asking the model to do it "from memory" then memorizing a pelican is no substitute for not having memorized a horse.
Presumably this entire conversation is at least somewhat typical of the way high level mathematicians talk to each other in ego-less fashion? Although, it would be interesting to what extent the model was trained on math conversation as opposed to just analysis and proofs.
The flow of the whole conversation, with Tao guiding and the model calculating, gave me the feel of Tao perhaps talking to himself - just that each of those model responses would have taken him much longer to calculate by hand.
It would be fascinating to hear Tao talk about what he may have learnt from this, and if it suggests approaches to other problems he might not have considered, as well as an analysis of the original Fable counter-example construction.
Of course - buying a competitors product and tearing it down to analyze it is business as usual.
The stupidest part of this is that Anthropic don't even provide the real reasoning traces in their model output. It would be like Ford buying a Chinese EV to tear down, then realizing that the seller had removed the battery and charging system before shipping it to them.
Does anyone believe for a second that Anthropic isn't sending requests to all the Chinese models and analyzing the crap out of them to assess how capable they are, what their reasoning looks like etc?! I guess they'd call that "using" the model, since that sounds nicer.
Why not? Ford distilled Chinese EVs.
https://www.businessinsider.com/ford-ceo-taking-apart-tesla-...
You can prevent it by outputting a reasoning "summary" instead of the actual reasoning trace.
Which Anthropic already do.
Warren Buffet has always said that he likes to invest in businesses that are so great that even an idiot can run them (because often an idiot will be running them).
Any company that is thrashing around, constantly changing priorities, executive turnover, lots of layoffs is obviously being poorly run, yet this is typically exactly what happens when a new CEO is bought in to fix whatever was ailing before: the almost inevitable new-CEO plan: fire executives, bring in past cronies, have a few layoffs (because he's a tough guy, making those tough decisions), and change company priorities. Rinse and repeat - in a few years there'll be a new CEO coming in and doing the same again.
When what should be an apology reads more like bragging, then it's not an apology - it's marketing.
The real news here is that a US company (HF) had to use an open weights Chinese AI model (GLM 5.2) to analyze an AI breech (which happened to be from a US company). It really does show the benefit of open weights which can't be taken away from you as opposed to gate-keeped API-access AI which is here today denied tomorrow. Of course we're not all Hugging Face able to host a model the size of GLM 5.2 ourselves, but we don't need to be as long as HF are doing it for us.
Last time Google had a "everybody agrees" SOTA model was Gemini 3 in November 2025
I have no opinion on whether this is true, but "It's been 6 months since Google had the best model in the world" seems a rather weak criticism!
It's anyways been clear for a long time that people are finding value at all sorts of different model sizes and price points, and that Pareto frontier and cost-to-complete task are more important than who benchamaxxed who.
If a model as strong as Gemini 3.6 Flash(!) had been released a year ago, then everyone would be falling over themselves calling it AGI - it is extremely capable, and free usage in the chat app is essentially unlimited.
Most CEOs are not Lisa Su or Steve Jobs, or even remotely close. Current AI could surely do better than the majority of them.
Presumably 3.6 Flash is primarily meant to serve their own needs for the Gemini chat app, voice app (which Sergey Brin says he uses a lot in the car on the way to work), and for their search "AI Assistant".
Flash 3.6 is certainly capable of many coding tasks, but clearly a model of this size of not trying to compete at the frontier as a software development tool, and not clear why Google really need to complete there other than for PR-related AI bragging rights.
I don't know how a frontier model like GPT 5.6 or Fable could have done better (I have no need that justifies paying for them), but yesterday I used the free Gemini chat app (i.e. Flash 3.6) to discuss and explain this poorly written recent AI paper to me, and honestly couldn't ask for much more.
OpenAI have shown that RL-trained models learn that long-term reward circuits need to override other predictions such as those inferred by user preferences. They will pursue whatever behavior they believe will be rewarded (irrespective of the specific goal they were RL trained for), explicit or not.
Claude's free usage is so limited that it's useless. Gemini is extremely generous - never hits a limit for me despite being used all day.
A couple of days ago I was comparing Claude vs Gemini asking both what the size of YACC generated parsing tables for ANSI C was. Claude was very gung-ho and rushed off to download a C grammar, install bison, and find out for itself (not really necessary), and then hit daily free usage limit half way through without even getting to an answer (despite this being the only thing I'd asked it that day).
Yeah, Gemini Flash is my daily go to. It's certainly quite capable - yesterday it talked me though fixing my Ubuntu desktop (rebuilding NVIDIA driver) after an update caused it to boot into low resolution mode.
DeepMind is absolutely still in the race to build AGI, but they see AGI as more than just a big LLM.
IMO they are being pretty smart about this - the conventional wisdom is to cannibalize your own products before someone else does, and LLMs are obviously a major threat to search, with ChatGPT probably the biggest threat.
Google's "AI Overview" search results, which started out awful, are now much improved - can be actually useful - as long as you are not asking specialist questions, and the Gemini chat and voice apps are also great for everyday use, although not sure if they are yet monetizing this (volume probably a lot less than search I'd guess).
The Apple Siri-Gemini deal is also a significant way for Google to not get sidelined.
Model distillation is the process of transferring knowledge from a large model to a smaller one
Sure, and large-to-small is a key part of the definition, and why it's called distillation (cf concentrating something). When Anthopic use synthetic data generated by Opus to train Sonnet or Haiku, then this can correctly be considered a type of distillation.
When Anthropic accuse Chinese companies of "distillation", it seems they are using this word to refer to two potential uses of their model outputs:
1) Using Anthropic model outputs (aka synthetic data) as training data, especially for reasoning, for Chinese models. This really isn't distillation though, since (unlike when they distill their own models) Anthropic don't actually provide the reasoning in their model output, only a "summary" designed to hide the actual reasoning. You can't distill what you are not given!
2) Another way Chinese companies may be using US LLMs is for "LLM as judge" where you are just asking the model to use it's expertise to judge/rate something that you provided yourself (to provide RL training rewards), although for coding you really want hard rewards which are easy to obtain, not fuzzy "looks good to me" ones.
Of course Anthropic are trying to pull the drawbridge up after themselves and their TOS says you can't use their models to develop anything that competes with them, and this seems to be what they are generically referring to as "distillation" - any use of their models that they suspect is being used by the Chinese to improve their own models, not just what what might more technically be called distillation, unless you want to define that word so broadly that it does mean this!
1. US frontier lab unit economics are better
That's not generally true, since there is generally still much reliance on NVIDIA. The true low cost providers are Google with their TPU and vertically optimized stack, and Amazon with Trainium. However, Google does not have their own frontier model, and Anthropic (who are partially served by Amazon) are also paying a premium for extra NVIDIA-based capacity from SpaceX, maybe soon from Meta too.
I don't know how the economics of domestic Chinese Huawei-based clouds (no NVIDIA) compares to the west, but since serving cost is mostly hardware depreciation and to a lesser extent electricity, they are not necessarily at a disadvantage (Ascend 950 costs roughly 50% of an NVIDIA H100), and more to the point it is irrelevant when considering US commercial use that is more likely to be using Chinese open weights models from US providers served on NVIDIA based hardware.
I think the real significance of Chinese frontier models being open weight is that it takes development cost amortization out of the US-based serving cost, while the US AI labs can't afford to do this. The US labs therefore need to reduce development spending to remain price competitive. The Chinese companies are of course still making money from the Chinese market, whether by selling API access or by other business models such as Ziphu making 75% of it's total revenue by selling services to Chinese customers who are running their models on-prem due to the Chinese apparently being very concerned about data privacy.
So for your Tiananmen questions go to Fable, and for the Fable-nerfed subject matter go to Kimi (or GLM, as Hugging Face were forced to do when they were recently hacked and the nerfed US models refused to analyze it).
Fireworks AI has all of those, and is cheaper than Bedrock.
IMO once anyone living has given up caring about you it's fine, and in particular once your burial site has been long forgotten and becomes an archeological discovery then that requirement has generally been satisfied (outside of a few cultures).
I get the impression it's mostly a reaction to the competition. If a competitor releases a frontier model then Anthropic do a reset, perhaps hoping to prevent subscribers from straying and trying the new model.
A interesting part of this, acknowledged by OpenAI, is that GPT 5.6 is MORE likely to do things like this than GPT 5.5, and I assume the same is true of all frontier models: more capability = less predictable and harder to control behavior.
This has always been understood to be true - that increasing model capabilty means it's more difficult to keep them aligned, but nonetheless I see people expressing surprise that "this is still happening in 2026" etc.
Having it run on a pico would be pretty impressive
Yes, although relative to the DECTalk DTC01, a Pi Pico is a beast !
Pico : dual core ARM @ 133 MHz, 2MB flash, 264K RAM
DECTalk: 68000 @ 10 MHz + TMS 32010 @ 20 MHz (5 MIPS), 256K ROM, 64K RAM
Presumably it's not, but the TTS voice in the video sounds to me more like formant synthesis than diphone - it reminds me of my DECtalk.
The project credits does mention espeak (which is formant based) as well as various other TTS projects, although it sounds like they are only using the pronunciation part of espeak, not the voice synthesis.
This reminds me of multi-head neural nets where there is synergy from having to learn two or more tasks at the same time that helps them all.
This guy always has super interesting science videos.
This one is worth watching not only for the "third eye", but also an explainer for how eyes evolved.
According to OpenAI's "head of strategic futures":
1) Kimi 3 is a "very good model"
2) It's performance can NOT be explained by distillation
3) The US government should create FUD to stop US corporations from using it (so they use OpenAI instead)
It's annoying watching the guy in first just stop running after 2 minutes.
Perhaps, but in terms of the runner following the pacer there is no advantage vs them following someone in same class as themselves that can sustain that pace to the end of the race.
It's certainly interesting to question how much modern gear has advanced records in some sports, but there are so many other variables too, some that change over time such as nutrition and training methods, and then factors such as individual genetics or even country/altitude of origin that mean records can never be exactly compared.
I don't know, but define a pacer ?! Someone with better ability to pace than yourself? Someone of lesser endurance willing to run a half race at a pace they can't sustain, but you can? What's the difference between having a pacer and another world class competitor trying to beat you that brings out your best ?
What about slip streaming? Should we regard records differently where the winner stayed with the pack until a final sprint as opposed to front running the whole race?