[dead]

the issue is the return is amorphous and unstructured

there's no contract. you send a bunch of text in (context etc) and it gives you some freeform text out.

Me too, but it was obviously wildly unsustainable. I was telling friends at xmas to enjoy all the subsidized and free compute funded by VC dollars while they can because it'll be gone soon.

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.

so it's also a skinner box

Whoops haha. Surely that can't be how black boxes normally work right?

And now it isn't. Pray they don't alter the deal any further.

its a drug. that is how it works. they ration it before the new stuff. seeing legends of programming shilling it pains me the most. so far there are a few decent non insane public people talking about it :Mitchel Hashimoto, Jeremy Howard, Casei Muratori. hell even DHH drank the coolaid while most of his interviews in the past years was how he went away from AWS and reduced the bill from 3 million to 1millions by basically loosing 9s, resiliency and availability. but it seems he is fine with loosing what makes his business work(programming) to a company that sells Overpowered stack overflow slot machines.

> It’s the official communication that sucks. It’s one thing for the product to be a black box if you can trust the company.

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.

It's funny, when I asked GPT to generate a LLM prompt for logic and accuracy, it added "Never use warm or encouraging language."

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).

Do you think the typos are helping or hurting output quality?

My settings are pretty standard:

% 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.

To me Claude Opus 4.6 seems even more confused.

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.

There is a certain amount of it which is the randomness of an LLM. You really want to ask most questions like this several times.

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've tried these with Claude various times and never get the wrong answer. I don't know why, but I am leaning they have stuff like "memory" turned on and possibly reusing sessions for everything? Only thing I think explains it to me.

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".

Claude Opus 4.7 responds with walk for me with and without adaptive thinking, but neither the basic model used when you Google search or GPT 5.4 do.

Idk but ironically, I had to re-read the first part of GP's comment three times, wondering WTF they're implying a mistake, before I noticed it's the car wash, not the car, that's 50 meters away.

I'd say it's a very human mistake to make.

That's an example the LLM came up with itself while analyzing its failed car wash walk/drive answer, it's not OP's question.

I've run several local models that get this right. Qwen 3.5 122B-A10B gets this right, as does Gemma 4 31B. These are local models I'm running on my laptop GPU (Strix Halo, 128 GiB of unified RAM).

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.

If you ask claude in chinese it thinks its deepseek.

> I'm imagining you're doing it because that's how Anthropic prefers to frame it

Correct.

> would it be valid to interpret that as an attack as well?

Yup.

Firehosing Anthropic to exfiltrate their model seems materially different than Anthropic downloading all of the Internet to create the model in the first place to me. But maybe that's just me?

Definitely Anthropic playing the victim after distilling the whole internet.

All extant human knowledge SO FAR. Remember, by the nature of the beast, the companies will always be operating in hindsight with outdated human knowledge.

Yep, that is exactly what happens. It's a disgrace that their models aren't open, after training on everything humanity has preserved.

They should at least release the weights of their old/deprecated models, but no, that would be losing money.

The usefulness of thinking tokens in my case might come down to the conditions I have claude working in.

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".

This can result in some funny interactions. I don't know if Claude will say anything, but I've had some models act "surprised" when I commented on something in their thinking, or even deny saying anything about it until I insisted that I can see their reasoning output.

Here is a 3blue1brown short about the relationship between words in a 3 dimensional vector space. [0] In order to show this conceptually to a human it requires reducing the dimensions from 10,000 or 20,000 to 3.

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.

[0] https://www.youtube.com/shorts/FJtFZwbvkI4

no, he's saying that in amongst whatever else is there, you can often see how you could refine your prompt to guide it better in the firtst place, helping it to avoid bad thinking threads to begin with.

I would expect to see a significant wall clock improvement if that was the case - Meta's Coconut paper was ~3x faster than tokenspace chain-of-thought because latents contain a lot more information than individual tokens.

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-...

If that's true, then we're following the timeline of https://ai-2027.com/

Don't.

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.

No, there is research in that direction and it shows some promise but that’s not what’s happening here.

Most likely, would be cool yes see a open source Nivel use diffusion for thinking.

Don't. thinking right now is just text. Chain of though, but just regular tokens and text being output by the model.

With the sunglasses, the long flamingo neck and the "joint", I immediately thought of the poster for Fear And Loathing In Las Vegas:

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!

> When was the last time they declined your subscription because they have no compute?

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.

Most weekdays.

https://status.claude.com/

Just last week. They cut off openclaw. And they added a price increased fast mode. And they announced today new features that are not included with max subscriptions.

They are short 5GW roughly and scrambling to add it.

This is low effort thinking, and a low effort comment. They have a lot of cash. They do not think they have achieved a "city of geniuses" in a datacenter yet. They are racing against two high quality frontier model teams, with meta in the wings. They have billions of dollars in cash that they are currently trying to spend to increase their datacenter capacity.

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?