A small farm in the mountains.
Joined February 2013
- Tweets334
- Following6K
- Followers772
- Likes5.9K
OpsConfig retweeted
Narrator: he was not, in fact, allowed to say “I’m quite unhappy with much of what OpenAI does.”
I’m quite unhappy with much of what OpenAI does.
I am very happy that I’m allowed to say “I’m quite unhappy with much of what OpenAI does.”
Readers added context they thought people might want to know
Korbak was fired ~3 weeks later. He says for safety concerns/METR talks (his job); OpenAI says mishandling sensitive info, not for speaking out.
x.com/tomekkorbak/st…
x.com/OpenAINewsroom…
OpsConfig retweeted
Getting to watch these beagles from Marshall Bioresources run on the grass for the first time was pure joy!
#beagles #dogsofx #marshallfarms #labdogs
Based on the tokenizer data the new stealth model "space-bunny-alpha" appears to likely be part of MiniMax lineage. xitter.cf/OpsConfig/status/21029…
Back when they supported fine-tuning distillation was a celebrated feature of the API:
(granted they only ever supported distilling from one OpenAI model to another. )
OpsConfig retweeted
New Paper: Turns out that LLM think traces don't necessarily have any end-user semantics even on benchmarks specifically designed to showcase such semantics.. 🤷♂️
As some of y'all know, we have been looking at the question of end-user interpretable semantics for the reasoning models for a while (c.f. arxiv.org/abs/2504.09762 (#ICML2026),
arxiv.org/abs/2505.13792 (#ACL2026) &
arxiv.org/abs/2505.13775 (#TMLR➡️#NeurIPS2026))
One question I get asked is whether that lack of semantics just an artifact of the benchmarks *we* made.
What if we look at benchmarks that were specifically proposed to show off the semantics of thinking traces?
As it happens, earlier work by AllenZhu @ZeyuanAllenZhu has indeed created such a benchmark, iGSM, and made claims about intermediate token semantics based on it.
Since AllenZhu's work on "Physics of LLMs"--that is partly based on their analysis on iGSM was influential, we wondered whether they found conditions under which traces do indeed have semantics.
Turns out, (perhaps unfortunately for the "please let thinking traces have semantics" contingent), that iGSM traces don't have semantics either, and models trained with swapped/corrupted traces can do well too. The results seem quite in line with--rather than being contrary to--the position we have taken in our earlier work.
Read the details of this very crisp work, lead by @ratishsp and Pranabendu, in the following thread.👇 – at Tempe, AZ
Happy to share our paper, "Correct Answers, Invalid Traces: What Verifiable Grade-School Math Reveals About Chain-of-Thought Traces".
When a model gets the right answer, is the reasoning it produces also right? We test this on synthetic grade-school math where every step of a trace can be checked by a program.
Thread below.
Agreed, It is an admirable achievement, and the article presents a measured interpretation of it. I was gently poking fun at the headline which was more in response to how Space X presented timelines for full fledged datacenters in space as part of their IPO rather than how Google has represented their work in this area.
I would suspect much of the compute capacity sitting in space is on classified satellites, but outside of that StarCloud-1 put an H100 in space in 2025 which I believe at the time was the most GPU compute in space.
Comparing thinking tokens across models/providers is fraught, and of limited value.
Different models, different backend implementations, different tokenizers, etc.
Stronger models should be able to solve more complex problems using less thinking tokens/test time compute.
The older pre-adaptive thinking Opus models are weaker so even with higher thinking token allocation their performance on chess doesn't improve enough to catchup to latest OpenAI models.
With that said, I do think there is value in tracking reasoning/thinking token allocation to see how often low/zero thinking token allocation corresponds to poor performance.
While regardless of chess performance Anthropic models have continued to get stronger in other domains, my own anecdotal experience was that Opus 4.5 and Opus 4.6 were the last Opus models to feel consistent.
Ever since Opus 4.7 I feel like I am constantly having to correct and double-check the model's work.
One moment the model feels smart, the next it is apologizing to me for the umpteenth time for having made a mistake.
These days anytime I use an Anthropic model for any serious work I am constantly farming out adversarial reviews to codex/OpenAI models.
This change in anecdotal vibes for me coincided with Anthropic's shift to "adaptive thinking".
The older model's used the now deprecated "extended thinking" feature. If we add in data on the older model's we see a profile far closer to the OpenAI models as far as thinking token allocation.
Adaptive thinking definitely saves money on a per API call basis, but I do wonder to what extent that comes at the cost of quality.
8/
Another way of parsing thinking token data is to not look just at when zero tokens are used, but to bucket the distribution of thinking tokens across different reasoning effort levels.
7/
So at best it seems like at least in these cases the Anthropic model's are failing to correctly identify the complexity of the task they are facing.
While in chess this can manifest in the model making a bad move, in real-world tasks this can feel like the model is lazy or suddenly dumb.
When the thinking token allocation correctly matches the complexity of the task the model hums along perfectly.
But the moment a mismatch occurs on something that the model might otherwise be capable of solving correctly it creates the jagged experience where the model that feels smart one moment, suddenly faceplants.
6/
The absence of thinking tokens at low and medium reasoning effort was not necessarily surprising. But I was genuinely surprised to see that 11% of moves Opus 5 makes at high, and xhigh generate zero thinking tokens.
Previously in their docs Anthropic described zero thinking at high effort to be extremely rare. Clearly, this isn't always the case.
(Apologies to @TheStalwart for doubting you.)
If a model does this and outperforms its peers, I would just assume it is a stronger model. But when the lack of thinking tokens is coupled with lower performance, it is a correlation worth noting.
4/
One of the first things that caught my eye when poking though the data from hundreds of Anthropic model chess games was how these models allocate reasoning/thinking tokens.
In particular, how often an Anthropic model will respond without generating any thinking tokens when playing chess.
3/
OpenAI's models have consistently performed better at chess than Anthropic's models.
This doesn't tell us much, but it is interesting in isolation since as best I can tell neither lab is optimizing to be better at chess.
When an improvement happens the cause is indirect, and this can be useful to track.
2/
A few thoughts on why @AnthropicAI's Opus 5 sometimes felt lazy compared to @OpenAI 's models.
I like to use chess as a benchmark/eval for AI models.
It is a crude/toy benchmark, but it does a good job of giving a sense of when the frontier model's make leaps forward and helps temper some of the benchmaxxing.
1/
Out of curiosity, I ran it through my standard chess eval gauntlet on low, medium, high, and max. It lost every game except one.
Here is one of its losses, mate in 7 moves with stockfish at 1320.
All match replays have been added to the library:
codeandstream.com/chess/#g=s…