llms are good at greedy depth first search[1]. so prompt / human needs to tell the model the breadth parts to take.
and human writing the prompt here did a lot more than that. asking to include parallel-edge 2-cycles, admitting disconnected graphs, specifying the emptyset cover, multiset counting, it all likely came from previous fake proofs.
[1]: see: ankitmaloo.com/fable - its the way most llms are trained, and is also natural owing to autoregressive nature.
they are also working on activations (w4a8, w4a16 from what i know). if they deliver (and a big if), it means that given their market reach, they can dictate the way sub 100b parameter models are trained and served to a large extent, given their major usecase would be on device (macos and not ios for most of them).
at a gross margin level, mostly no. if you include the cost of training a model as full R&D then possibly yes.
Batch size is what you should look at. If a cluster is running and processing one request, filling the batch has almost no marginal cost (kv cache creation/storage/fetch costs aside). But if the concurrent requests exceed batch size, one extra request would cost basically the rent cost of entire new cluster. APIs have the bursty nature so companies would plan to price it such that they are profitable / break even at 40%-50% utilization (% of filled batch for simplicity). so any extra request would not have the same cost as long as they are alongside an api request. you might think it degrades teh performance. easy: just assign a priority tier to api requests, and a lower tier to subscription requests.
its even more effective and powerful now that you have continuous batching. so likely if the api is being used, they are not eating any loss, let alone "big loss"
An interesting question to wonder is why these optimizations were pushed so aggressively in the first place. Especially given this is the time they were running a 2x promotion, by themselves, without presumably seeing any slowdown in demand.
problem is google's security concerns. when people connect gmail to openclaw, google flags the activity as weird and suspend the account because of unusual activity. Many whose accounts got locked because of this and they thought it was because they connected it to antigravity use against the policy (which happened in some cases). We will still see google account suspensions, and that would keep making news. and it wont be because of antigravity usage.
not much to do with self improvement as such. openai has increased its pace, others are pretty much consistent. Google last year had three versions of gemini-2.5-pro each within a month of each other. Anthropic released claude 3 in march 24, sonnet 3.5 in june 24, 3.5 new in oct 24, and then 3.7 in feb 25, where they went to 4 series in May 25. then followed by opus 4.1 in august, sonnet 4.5 in oct, opus 4.5 in nov, 4.6 in feb, sonnet 4.6 in feb itself. Yes, they released both within weeks of each other, but originally they only released it together. This staggered release is what creates the impression of fast releases. its as much a function of training as a function of available compute, and they have ramped up in that regard.
People are misunderstanding Anthropic's fast mode because they chose to name it that way. The hints all point to a specific thing they did. The setup is costlier, its also smarter and better on tougher problems which is unheard of in terms of speed. This paper[1] fits perfectly:
The setup is parallel distill and refine. You start with parallel trajectories instead of one, then distill from them, and refine that to get to an answer. Instead of taking all trajectories to completion, they distill it quickly and refine so it gives outputs fast and yet smarter.
- paper came out in nov 2025
- three months is a good research to production pipeline
- one of the authors is at anthropic
- this approach will definitely burn more tokens than a usual simple run.
- > Anthropic explicitly warns that time to first token might still be slow (or even slower)
To what people are saying, speculative decoding wont be smarter or make any difference. Batching could be faster, but then not as costly.
Gemini Deepthink and gpt-5.2-pro use the same underlying parallel test time compute but they take each trajectory to completion before distilling and refining for the user.
Agreed. Gemini 3 Pro for me has always felt like it has had a pretraining alpha if you will. And many data points continue to support that. Even as flash, which was post trained with different techniques than pro is good or equivalent at tasks which require post training, occasionally even beating pro. (eg: in apex bench from mercor, which is basically a tool calling test - simplifying - flash beats pro). The score on arc agi2 is another datapoint in the same direction. Deepthink is sort of parallel test time compute with some level of distilling and refinement from certain trajectories (guessing based on my usage and understanding) same as gpt-5.2-pro and can extract more because of pretraining datasets.
(i am sort of basing this on papers like limits of rlvr, and pass@k and pass@1 differences in rl posttraining of models, and this score just shows how "skilled" the base model was or how strong the priors were. i apologize if this is not super clear, happy to expand on what i am thinking)
What i think you are referring to is hidden state as in internal representations. I refer to hidden state in game theoretic terms like a private information only one party has. I think we both agree alphazero has hidden states in first sense.
Concepts like king safety are objectively useful for winning at chess so alphazero developed it too, no wonder about that. Great example of convergence. However, alphazero did not need to know what i am thinking or how i play to beat me. In poker, you must model a player's private cards and beliefs.
Bounded domains require scaling reasoning/compute. Two separate scenarios - one where you have hidden information, other where you have high number of combinations. Reasoning works in second case because it narrows the search space. Eg: a doctor trying to diagnose a patient is just looking at number of possibilities. If not today, when we scale it up, a model will be able to arrive at the right answer. Same goes with Math, the variance or branching for any given problem is very high. But LLMs are good at it. and getting better. A negotiation is not a high variance thing, and low number of combinations, but llms would be repeated bad at it.
I address that in part right there itself. Programming has parts like chess (ie bounded) which is what people assume to be actual work. Understanding future requiremnts / stakeholder incentives is part of the work which LLMs dont do well.
> many domains are chess-like in their technical core but become poker-like in their operational context.
> And they would have won the AI race not by building the best model, but by being the only company that could ship an AI you’d actually trust with root access to your computer.
and the very next line (because i want to emphasize it
> That trust—built over decades—was their moat.
This just ignores the history of os development at apple. The entire trajectory is moving towards permissions and sandboxing even if it annoys users to no end. To give access to an llm (any llm, not just a trusted one acc to author) the root access when its susceptible to hallucinations, jailbreak etc. goes against everything Apple has worked for.
And even then the reasoning is circular. "So you build all your trust, now go ahead and destroy it on this thing which works, feels good to me, but could occasionally fuck up in a massive way".
Not defending Apple, but this article is so far detached from reality that its hard to overstate.
you are comparing post hoc narratives in the training data to real time learning from causal dynamics. The objectives are different. They may look the same in scenarios where its heavily and accurately documented, but most narratives suffer from survivorship bias and reasoning post facto, eulogising the given outcomes.
Its a combination. All caps is used in prompts for extra insistence, and has been common in cases of prompt hijacking. OP was doing it in combination with attempting to direct claude a certain way, multiple times, which might have looked similar to attempting to bypass teh system prompt.
from what i know, it used to be that if you want to assertively instruct, you used all caps. I don't know if it succeeds today. I still see prompts where certain words are capitalized to ensure model pays attention. What i mean was not just capitalization, but a combination of both capitalization and changing the behavior of the model for trying to get it to do something.
if you were to design a system to prevent prompt injections and one of surefire ways is to repeatedly give instructions in caps, you would have systems dealing with it. And with instructions to change behavior, it cascades.
My rudimentary guess is this. When you write in all caps, it triggers sort of a alert at Anthropic, especially as an attempt to hijack system prompt. When one claude was writing to other, it resorted to all caps, which triggered the alert, and then the context was instructing the model to do something (which likely would be similar to a prompt injection attack) and that triggered the ban. not just caps part, but that in combination of trying to change the system characteristics of claude. OP does not know much better because it seems he wasn't closely watching what claude was writing to other file.
if this is true, the learning is opus 4.5 can hijack system prompts of other models.
Like it or not, it's a fundraising strategy. They have followed it mutliple times (eg: vague posts about how much their inhouse model is writing code, online RL, and lines of code etc. earlier) and it was less vague before. They released a model and did not give us the exact benchmarks or even tell us the base model for the same. This is not to imply there is no substance behind it, but they are not as public about their findings as one would like them to be. Not a criticism, just an observation.
Here is the link: https://www.clioapp.ai
Email: ankit at clioapp.ai
Twitter: ankit2119