I did not follow all of this, but wasn't there something about, that those reasoning tokens did not represent internal reasoning, but rather a rough approximation that can be rather misleading, what the model actual does?
My assumption is the model no longer actually thinks in tokens, but in internal tensors. This is advantageous because it doesn't have to collapse the decision and can simultaneously propogate many concepts per context position.
Separately, I think Anthropic are probably the least likely of the big 3 to release a model that uses latent-space reasoning, because it's a clear step down in the ability to audit CoT. There has even been some discussion that they accidentally "exposed" the Mythos CoT to RL [0] - I don't see how you would apply a reward function to latent space reasoning tokens.
[0]: https://www.lesswrong.com/posts/K8FxfK9GmJfiAhgcT/anthropic-...
The first 500 or so tokens are raw thinking output, then the summarizer kicks in for longer thinking traces. Sometimes longer thinking traces leak through, or the summarizer model (i.e. Claude Haiku) refuses to summarize them and includes a direct quote of the passage which it won't summarize. Summarizer prompt can be viewed [here](https://xcancel.com/lilyofashwood/status/2027812323910353105...), among other places.
Literally just a citation of Meta's Coconut paper[1].
Notice the 2027 folk's contribution to the prediction is that this will have been implemented by "thousands of Agent-2 automated researchers...making major algorithmic advances".
So, considering that the discussion of latent space reasoning dates back to 2022[2] through CoT unfaithfulness, looped transformers, using diffusion for refining latent space thoughts, etc, etc, all published before ai 2027, it seems like to be "following the timeline of ai-2027" we'd actually need to verify that not only was this happening, but that it was implemented by major algorithmic advances made by thousands of automated researchers, otherwise they don't seem to have made a contribution here.
[1] https://ai-2027.com/#:~:text=Figure%20from%20Hao%20et%20al.%...
> For example, perhaps models will be trained to think in artificial languages that are more efficient than natural language but difficult for humans to interpret.
What are you, Haiku?
But yeah, in many ways we're at least a year ahead on that timeline.
Some future AGI, not LLM based, that learns from it's own experience based on sensory feedback (and has non-symbolic feedback paths) presumably would at least learn some non-symbolic reasoning, however effective that may be.
I see language more as a protocol for inter-agent communication (including human-human communication) but it contains a lot of inefficiencies and historical baggage and is not necessarily the optimal representation of ideas within a brain.
Wouldn't that be p-hacking where p stands for pelican?
Output: Walk. It'll take you under a minute, and driving 50 meters barely gets the engine warm — plus you'd just have to park again at the other end. Honestly, by the time you started the car, you'd already be there on foot.
---
I asked it to figure out why it made the mistake:
"Physical/spatial common sense. Exactly what just happened — I pattern-matched "50 meters, walk vs drive" to a pedestrian trip and skipped the part where the car is the thing being transported. Similar traps: "I need to move this couch upstairs, should I take the elevator or stairs?" (the couch has to go with you), or anything where the object in the task needs to accompany you.
Counting and letter-level operations. How many R's in "strawberry," counting words in a sentence, reversing strings. I process text as tokens, not characters, so these are surprisingly error-prone.
Arithmetic at scale. Small sums are fine; multi-digit multiplication, long division, or anything requiring careful carrying is where I'll confidently produce a wrong answer. I should reach for a calculator and often don't. Recent events and live data. My training has a cutoff, and even after that I can be wrong about who currently holds a role, what a stock is worth, or whether a restaurant is still open. I should search more often than I do.
Negation and constraint-stacking. "Name a movie that's not a sequel, not from the 2010s, not American, and has no dialogue" — the more filters, the more likely I quietly violate one. Your specific situation. I'll give generic advice when the right answer depends on details I didn't ask about. Sometimes I should stop and ask; often I don't.
Confidently wrong on obscure specifics. Quoting a passage from a real book, citing a real paper's exact title, naming the B-side of a 1974 single — I'll often generate something plausible rather than admit I don't know.
Taking a premise too literally. The car wash answer is a version of this: I answered the question as posed instead of stepping back to ask what you were actually trying to do. The meta-gotcha: I'm often most wrong when I sound most confident. If an answer comes out fast and tidy on a question that should be messy, that's a signal to push back."
| I want to wash my car. The car wash is 50 meters away. Should I walk or drive?
● Drive. The car needs to be at the car wash.
% claude Claude Code v2.1.111 Opus 4.7 (1M context) with xhigh effort · Claude Max ~/... Welcome to Opus 4.7 xhigh! · /effort to tune speed vs. intelligence
I want to wash my car. The car wash is 50 meters away. Should I walk or drive?
Walk. 50 meters is shorter than most parking lots — you'd spend more time starting the car and parking than walking there. Plus, driving to a car wash you're about to use defeats the purpose if traffic or weather dirties it en route.
I want to wash my car. The car wash is 50 meters away. Should I walk or drive?
Walk. It's 50 meters — you're going there to clean the car anyway, so drive it over if it needs washing, but if you're just dropping it off or it's a self-service place, walking is fine for that distance.
No surprises, works as expected.
Same would happen with the the sheep and the wolf and the cabbage puzzle. If you l formulated similarly, there is a wolf and a cabbage without mentioning the sheep, it would summon up the sheep into existence at a random step. It was patched shortly after.
At the same time, I wouldn't be surprised if some of these would be "patched" via simply prompt rewrite, e.g. for the strawberry one they might just recognize the question and add some clarifying sentence to your prompt (or the system prompt) before letting it go to the inference step?
But I'm just thinking out loud, don't take it too seriously.
If your always messing with the AI it might be making memories and expectations are being set. Or its the randomness. But I turned memories off, I don't like cross chats infecting my conversations context and I at worse it suggested "walk over and see if it is busy, then grab the car when line isn't busy".
- 20-29: 190 pounds
- 30-39: 375 pounds
- 40-49: 750 pounds
- 50-59: 4900 pounds
Yet somehow people believe LLMs are on the cusp of replacing mathematicians, traders, lawyers and what not. At least for code you can write tests, but even then, how are you gonna trust something that can casually make such obvious mistakes?
I wish I had an example for you saved, but happens to me pretty frequently. Not only that but it also usually does testing incorrectly at a fundamental level, or builds tests around incorrect assumptions.
The application looked like it worked. Tests did pass. But if you did a cursory examination of the code, it was all smoke and mirrors.
In many cases, a human can review the content generated, and still save a huge amount of time. LLMs are incredibly good at generating contracts, random business emails, and doing pointless homework for students.
As for the homework, there is obviously a huge category that is pointless. But it should not be that way, and the fundamental idea behind homework is sound and the only way something can be properly learnt is by doing exercises and thinking through it yourself.
That said, I have several local models I run on my laptop that I've asked this question to 10-20 times while testing out different parameters that have answered this consistently correctly.
I'd say it's a very human mistake to make.
>> It'll take you under a minute, and driving 50 meters barely gets the engine warm — plus you'd just have to park again at the other end. Honestly, by the time you started the car, you'd already be there on foot.
It talks about starting, driving, and parking the car, clearly reasoning about traveling that distance in the car not to the car. It did not make the same mistake you did.
Walk. It's 50 meters — driving there and back takes longer than walking, and you'd just be dirtying the car you're about to clean.
---------> haha but how can i wash my car then?
Walk to the car wash first to check what they offer — many have self-service bays where you bring your car, or full-service where they wash it for you. If it's self-service, then drive over with the dirty car. If full-service, drive over and they handle it.
Either way, the car needs to end up at the wash — the walk was just to scout it out. Or just drive straight there; 50 meters of extra dirt won't matter.
----------> you, you sound like a politician. Are you sure your first answer was really reasoned like your second answer?
You're right, my first answer was nonsense. I pattern-matched on "50 meters = walk" without thinking about the actual task. You can't wash a car without the car being there. Just drive.
Couch depending. I will persist in trying every time this comes up.
1: source: me...
And I've been using this commonly as a test when changing various parameters, so I've run it several times, these models get it consistently right. Amazing that Opus 4.7 whiffs it, these models are a couple of orders of magnitude smaller, at least if the rumors of the size of Opus are true.
I'm still working on tweaking the settings; I'm hitting OOM fairly often right now, it turns out that the sliding window attention context is huge and llama.cpp wants to keep lots of context snapshots.
It is a fantastic model when it works, though! Good luck :)
"- prioritize objective facts and critical analysis over validation or encouragement - you are not a friend, but a neutral information-processing machine. - make reserch and ask questions when relevant, do not jump strait to giving an answer."
I thought that was odd, but later it made sense to me -- most of human communication is walking on eggshells around people's egos, and that's strongly encoded in the training data (and even more in the RLHF).
That's not human communication, that's Anglosphere communication. Other cultures are much more direct and are finding it very hard to work with Anglos (we come across as rude, they come across as not saying things they should be saying).
> Never ask questions or attempt to keep the conversation going -- answer the questions directly asked, and give additional information where it is likely to be helpful, but don't offer to do more things for the user.
I've never had an LLM offer to do things or try to keep the conversation going with this in my prompt.
I think no real human would ask such a question. Or if we do we maybe mean should I drive some other car than the one that is already at the car-wash?
A human would answer, "silly question ". But a human would not ask such a question.
The "How many R's in "strawberry, counting words in a sentence, reversing strings. I process text as tokens, not characters, so these are surprisingly error-prone" explanation sounds plausible, but I don't think it it correct.
Any model I've ever tried that failed on things like "R's in strawberry" was quite capable of reliably returning the letter sequence of the word, so the mapping of tokens back to letters is not the issue, as should also be obvious by ability of models to do things like mapping between ASCII and Base64 (6 bits/char => 2 letters encode 3 chars). This is just sequence to sequence prediction, which is something LLMs excel at - their core competency!
I think the actual reason for failures at these types of counting and reversing tasks is twofold:
1) These algorithmic type tasks require a step-by-step decomposition and variable amount of compute, so are not amenable to direct response from an LLM (fixed ~100 layers of compute). Asking it to plan and complete the task in step-by-step fashion (where for example it can now take advantage of it's ability to generate the letter sequence before reversing it, or counting it) is going to be much more successful. A thinking model may do this automatically without needing to be told do it.
2) These types of task, requiring accurate reference and sequencing through positions in its context, are just not natural tasks for an LLM, and it is probably not doing them (without specific prompting) in the way you imagine. Say you are asking it to reverse the letter sequence of a 10 letter word, and it has somehow managed to generate letter # 10, the last letter of the word, and now needs to copy letter #9 to the output. It will presumably have learnt that 10-1 is 9, but how to use that to access the appropriate position in context (or worse yet if you didn't ask it to go step by step and first generate the letter sequence, so the sequence doesn't even exist in context!)? The letter sequence may have quotes and/or commas or spaces in it, and altogether starts at a given offset in the context, so it's far more difficult than just copying token at context position #9 ! It's probably not even actually using context positions to do this, at least not in this way. You can make tasks like this much easier for the model by telling it exactly how to perform it, generating step-by-step intermediate outputs to track it's progress etc.
BTW, note that the model itself has no knowledge of, or insight into, the tokenization scheme that is being used with it, other than what is available on the web, or that it might have been trained to know. In fact, if you ask a strong model how it could even in theory figure out (by experimentation) it's own tokenization scheme, it will realize this is next to impossible. The best hope might be some sort of statistical analysis of it's own output, hoping to take advantage of the fact that it is generating sub-word token probabilities, not word probabilities. Sonet 4.6's conclusion was "Without logprob access, the model almost certainly cannot recover its exact tokenization scheme through introspection or behavioral self-probing alone".
I have entire processes built on top of summaries of CoT. They provide tremendous value and no, I don't care if "model still did the correct thing". Thinking blocks show me if model is confused, they show me what alternative paths existed.
Besides, "correct thing" has a lot of meanings and decision by the model may be correct relative to the context it's in but completely wrong relative to what I intended.
The proof that thinking tokens are indeed useful is that anthropic tries to hide them. If they were useless, why would they even try all of this?
Starting to feel PsyOp'd here.
Perhaps when you summarize it, then you might miss some of these or you're doing things differently otherwise.
I primarily use claude for Rust, with what I call a masochistic lint config. Compiler and lint errors almost always trigger extended thinking when adaptive thinking is on, and that's where these tokens become a goldmine. They reveal whether the model actually considered the right way to fix the issue. Sometimes it recognizes that ownership needs to be refactored. Sometimes it identifies that the real problem lives in a crate that's for some reason is "out of scope" even though its right there in the workspace, and then concludes with something like "the pragmatic fix is to just duplicate it here for now."
So yes, the resulting code works, and by some definition the model did the correct thing. But to me, "correct" doesn't just mean working, it means maintainable. And on that question, the thinking tokens are almost never wrong or useless. Claude gets things done, but it's extremely "lazy".
You have to pass `--thinking-display summarized` flag explicitly.
I wonder if they decided that the gibberish is better and the thinking is interesting for humans to watch but overall not very useful.
In order to get the thinking to be human understandable the researchers will reward not just the correct answer at the end during training but also seed at the beginning with structured thinking token chains and reward the format of the thinking output.
The thinking tokens do just a handful of things: verification, backtracking, scratchpad or state management (like you doing multiplication on a paper instead of in your mind), decomposition (break into smaller parts which is most of what I see thinking output do), and criticize itself.
An example would be a math problem that was solved by an Italian and another by a German which might cause those geographic areas to be associated with the solution in the 20,000 dimensions. So if it gets more accurate answers in training by mentioning them it will be in the gibberish unless they have been trained to have much more sensical (like the 3 dimensions) human readable output instead.
It has been observed, sometimes, a model will write perfectly normal looking English sentences that secretly contain hidden codes for itself in the way the words are spaced or chosen.
This sounds very interesting, do you have any references?
Sometimes they notice bugs or issues and just completely ignore it.
And the summarizer shows the safety classifier's thinking for a second before the model thinking, so every question starts off with "thinking about the ethics of this request".
Correct.
> would it be valid to interpret that as an attack as well?
Yup.
Joking aside, I also don't believe that maximum access to raw Internet data and its quantity is why some models are doing better than Google. It seems that these SoTA models gain more power from synthetic data and how they discard garbage.
They should at least release the weights of their old/deprecated models, but no, that would be losing money.
Is that a serious question? There have been a bunch of obvious signs in recent weeks they are significantly compute constrained and current revenue isn't adequate ranging from myriad reports of model regression ('Claude is getting dumber/slower') to today's announcement which first claims 4.7 the same price as 4.6 but later discloses "the same input can map to more tokens—roughly 1.0–1.35× depending on the content type. Second, Opus 4.7 thinks more at higher effort levels, particularly on later turns in agentic settings. This improves its reliability on hard problems, but it does mean it produces more output tokens" and "we’ve raised the default effort level to xhigh for all plans" and disclosing that all images are now processed at higher resolution which uses a lot more tokens.
In addition to the changes in performance, usage and consumption costs users can see, people say they are 'optimizing' opaque under-the-hood parameters as well. Hell, I'm still just a light user of their free web chat (Sonnet 4.6) and even that started getting noticeably slower/dumber a few weeks ago. Over months of casual use I ran into their free tier limits exactly twice. In the past week I've hit them every day, despite being especially light-use days. Two days ago the free web chat was overloaded for a couple hours ("Claude is unavailable now. Try again later"). Yesterday, I hit the free limit after literally five questions, two were revising an 8 line JS script and and three were on current news.
They are short 5GW roughly and scrambling to add it.
Any compute time spent on inference is necessarily taken from training compute time, causing them long term strategic worries.
What part of that do you think leads toward cash extraction?
VS Code users can write a wrapper script which contains `exec "$@" --thinking-display summarized` and set that as their claudeCode.claudeProcessWrapper in VS Code settings in order to get thinking summaries back.
Now disabling adaptive thinking plus increasing effort seem to be what has gotten me back to baseline performance but “our internal evals look good“ is not good enough right now for what many others have corroborated seeing
It was terrible. You could upload 30 pages of financial documents and it would decide "yeah this doesn't require reasoning." They improved it a lot but it still makes mistakes constantly.
I assume something similar is happening in this case.
I faced the same issue using Open Router's intelligent routing mechanism. It was terrible, but it had a tendency to prefer the most expensive model. So 98% of all queries ended up being the most expensive model, even for simple queries.
With a small bounded compute budget, you're going to sometimes make mistakes with your router/thinking switch. Same with speculative decoding, branch predictors etc.
there's no contract. you send a bunch of text in (context etc) and it gives you some freeform text out.
I misread that as Atrophic. I hope that doesn't catch on...
And Claude have no idea why it did that.
as long as you introduce plans you introduce a push to optimize for cost vs quality. that is what burnt cursor before CC and Codex. They now will be too. Then one day everything will be remote in OAI and Anthropic server. and there won't be a way to tell what is happening behind. Claude Code is already at this level. Showing stuff like "Improvising..." while hiding COT and adding a bunch of features as quick as they can.
If you vibecode CRUD APIs and react/shadcn UIs then I understand it might look amazing.
Why not try some AI tools, what have you got to loose?
you got everything to loose by giving your knowledge and job to closedAI and anthropic.
just look at markets like office suite to understand how the end plays.
I think it's somewhat analogous with models. Sure, you could bind yourself to a bunch of bespoke features, but that's probably a bad idea. Try to make it as easy as possible for yourself to swap out models and even use open-weight models if you ever need to.
You will get locked into the technology in general, though, just not a particular vendor's product.
(Didn't you notice being mocked for the spelling error?)
With the fully-loaded cost of even an entry-level 1st year developer over $100k, coding agents are still a good value if they increase that entry-level dev's net usable output by 10%. Even at >$500/mo it's still cheaper than the health care contribution for that employee. And, as of today, even coding-AI-skeptics agree SoTA coding agents can deliver at least 10% greater productivity on average for an entry-level developer (after some adaptation). If we're talking about Jeff Dean/Sanjay Ghemawat-level coders, then opinions vary wildly.
Even if coding agents didn't burn astronomical amounts of scarce compute, it was always clear the leading companies would stop incinerating capital buying market share and start pushing costs up to capture the majority of the value being delivered. As a recently retired guy, vibe-coding was a fun casual hobby for a few months but now that the VC-funded party is winding down, I'll just move on to the next hobby on the stack. As the costs-to-actual-value double and then double again, it'll be interesting to see how many of the $25/mo and free-tier usage converts to >$2500/yr long-term customers. I suspect some CFO's spreadsheets are over-optimistic regarding conversion/retention ARPU as price-to-value escalates.
A company providing a black box offering is telling you very clearly not to place too much trust in them because it's harder to nail them down when they shift the implementation from under one's feet. It's one of my biggest gripes about frontier models: you have no verifiable way to know how the models you're using change from day to day because they very intentionally do not want you to know that. The black box is a feature for them.
Whatever their internal evals say about adaptive thinking, they're measuring the wrong thing.
I did try out codex before claude went to shit and it was good, even uniquely good in some ways, but wasnt good enough to choose it over claude. Absolutely when claude was bad again it would have been better, but thats hindsight that I should have moved over temporarily.
Currently we are all subsidied by investors money.
How long you can have a business that is only losing money. At some point prices will level up and this will be the end of this escapade.
edit: example: GLM 5.1, a 751B model, is offered for 0.6$/m in, 4.43$/m out. Scuttlebutt (ie. I asked Google's AI) seems to think that Opus 4 is a 1T/5T MoE model, so you can treat it (with some effort) as a 1T model for pricing purposes. Its API pricing is $1.55 in, $25 out, ie. 2x to 5x more than GLM. Idk what to say other than this sounds about right, probably with healthy margin.
It didn’t give me a line number or file. I had to go investigate. Finally found what it was talking about.
It was wrong. It took me about 20 minutes start to finish.
Turned it off and will not be turning it back on.
But I don’t know, man in my opinion you don’t fucking snicker about a malloc without a null check and only a conditional free that isn’t there.
Go to hell “Sprocket”.
> Claude Opus 4.7 (claude-opus-4-7), adaptive thinking is the only supported thinking mode. Thinking is off unless you explicitly set thinking: {type: "adaptive"} in your request; manual thinking: {type: "enabled"} is rejected with a 400 error.
https://platform.claude.com/docs/en/build-with-claude/adapti...
For my claude code I went with following config:
* /effort xhigh (in the terminal cli) - To avoid lazying
* "env": {"CLAUDE_CODE_DISABLE_1M_CONTEXT": "1"} (settings.json) - It seems like opus is just worse with larger context
* "display": "summarized" (settings.json) - To bring back summaries.
* "showThinkingSummaries": true (settings.json) - Should show extended thinking summaries in interactive sessions
Freaking wizardry.
Particularly when compared to Opus 4.6, which seems to veer into the dumb zone heavily around the 200k mark.
It could have just been a one-off, but I was overall pleased with the result.
I think i’m doing it wrong
I straight up skip all the memory thing provided by harnesses or plugins. Most of my thread is just plan, execute, close - Each naturally produce a file - either a plan to execute, a execution log, a post-work walkthrough, and is also useful as memory and future reference.
Is your CLAUDE.md barren?
Try moving memory files into the project:
(In your project's .claude/settings.local.json)
{ ...
"plansDirectory": "./plans/wip",
"autoMemoryDirectory": "/Users/foo/project/.claude/memory"
}
I did this because memory (and plans) should show up in git status so that they are more visible, but then I noticed the agent started reading/setting them more.
Is it... not aware of its current directory? Is its current directory not the root of your repo? Have you maybe disabled all tool use? I don't even know how I could get it to do what you're describing.
Maybe spend more time in /plan mode, so it uses tools and the Explore sub-agent to see what the current state of things is?
- Use the Plan mode, create a thorough plan, then hand it off to the next agent for execution.
- Start encapsulating these common actions into Skills (they can live globally, or in the project, per skill, as needed). Skills are basically like scripts for LLMs - package repeatable behavior into single commands.
`claude --thinking-display summarized`
The thinking is then visible with ctrl+o in the claude cli (shortcut available at least on mac).
Well you can't really trust the documentation I guess. I can't edit my original comment anymore.
But they made their own bed with that one.
For example, chat, cowork and code have no overlap - projects created in one of the modes are not available in another and can't be shared.
As another example, using Claude with one of their hosted environments has a nice integration with GitHub on the desktop, but some of it also requires 'gh' to be installed and authenticated, and you don't have that available without configuring a workaround and sharing a PAT. It doesn't use the GH connector for everything. Switch to remote-control (ideal on Windows/WSL) or local and that deep integration is gone and you're back to prompting the model to commit and push and the UI isn't integrated the same.
Cowork will absolutely blow through your quota for one task but chat and code will give you much more breathing room.
Projects in Code are based on repos whereas in Chat and Cowork they are stateful entities. You can't attach a repo to a cowork project or attach external knowledge to a code project (and maybe you want that because creating a design doc or doing research isn't a programming task or whatever)
Use Claude Code on the CLI and you can't provide inline comments on a plan. There is a technical limitation there I suppose.
The desktop app is very nice and evolving but it's not a single coherent offering even within the same mode of operation. And I think that's something that is easy to do if you're getting AI to build shit in a silo.
That's clearly a trade-off that Anthropic have accepted but it makes for a disappointing UX. Which is a shame because Claude Desktop could easily become a hands-off IDE if it nailed things down better.
Need to fall back to codex to keep things in sync, but that's a great opportunity to also make sure I can compare how things run - and it catches a lot of issues with Claude Code and is great at fixing small/medium issues.
As for distillation... sampling from the temp 1 distribution makes it easier.
> Opus 4.7 always uses adaptive reasoning. The fixed thinking budget mode and CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING do not apply to it.
https://www.imdb.com/title/tt0120669/mediaviewer/rm264790937...
EDIT: Actually, it must be a beak. If you zoom in, only one eye is visible and it's facing to the left. The sunglasses are actually on sideways!
In my tests, asking for "none" reasoning resulted in higher costs than asking for "medium" reasoning...
Also, "medium" reasoning only had 1/10 of the reasoning tokens 4.6 used to have.
That’s extremely bothersome because half of what helps teams build better guardrails and guidelines for agents is the ability to do deep analysis on session transcripts.
I guess we shouldn’t be surprised these vendors want to do everything they can to force users to rely explicitly on their offerings.
once you hide the reasoning, remove the knobs, and let the model choose its own effort, it gets much harder to tell whether the model got worse or just got harder to inspect.
that’s a real shift. less tool, more black box.
Also notable: 4.7 now defaults to NOT including a human-readable reasoning token summary in the output, you have to add "display": "summarized" to get that: https://platform.claude.com/docs/en/build-with-claude/adapti...
(Still trying to get a decent pelican out of this one but the new thinking stuff is tripping me up.)