Models can be (and will be) trained/calibrated to better assess the complexity and their expected success rate on any given task. Especially as subagent use increases, models need to know that, to decide which models to delegate tasks to.
For any task, a Pareto frontier exists of costs v. success-rate. Anthropic has 20 model/effort combinations currently, and maybe 10 (?) of those are never the best choice for any task (where a cheaper combinations provides same or better results). You know your workloads, and can estimate which combination to use, but models can too (and will get better).
If Sonnet can increase its reasoning effort, it should know that at some point switching to Opus makes more sense than cranking itself up to xhigh/max effort. A model can, and should be calibrated to know if a cheaper config has the same or higher success probability. It should also know which config has the next better performance (whether it's itself at a higher effort, or a more costly model).
At any price point, there exists the best model/effort combination. Why be against letting a person set the effort/time (e.g. on a slider with 5 options)
and letting the models self-allocate tasks based on their strengths?
Ok, sure, then let’s just remove all options except adaptive thinking at default reasoning level and see how that goes with Claude users. I’m sure they’ll appreciate the reduced friction.
Just because it would be nice to have that, doesn’t mean that it’s a good idea in practice, or that there currently exists a viable way to implement it.
> At any price point, there exists the best model/effort combination.
People are already offended by any barely perceptible tweaks to models; can’t wait to hear what they think about having their options reduced to whatever Anthropic decides is the pareto-optimal “do the thinky thing” setting this week.
You said it yourself, removing the option won't go well with basically any user today, and classifiers will remain fallible forever, so it won't be.
Yet, current non-users simply don't care about reading 4 system cards during roll-in and then keeping up with new ones every few weeks.
And, a solid adaptive thinking classifier is trivial compared to Anthropic's remaining tech stack. Task complexity estimation and model/effort selection are simple predictions with clear proxies in current data and millions of training samples generated daily by current users.
You keep talking your thing, and yet reality fails to budge. Consider:
- harnesses are slow to implement this because there is next to no real world examples of this working (i.e. saving money without degrading quality)
- ahead of time model selection is very difficult, and robustly predicting the task’s difficulty before starting work (e.g., making a plan, reading files, initial reasoning) may actually be fundamentally impossible (unless the classifier has project context and learns on the job? either way, far from trivial)
- switching models (and in case of some providers, even reasoning level) with a non-empty context causes a costly cache miss
- the classifier itself ain’t free lunch either, unless you don’t care whether it works well
- but maybe I’m wrong and you really should e-mail Anthropic about your trivial optimization idea that I’m sure they haven’t thought about or tried internally
You’re confusing two figures. Before their S-1 filing, they reported being “profitable” to their investors when excluding training and stock based comp.
In the S-1 filing Q3 was profitable even with training included, if you exclude stock based compensation.
Q2 was profitable on all operating costs (incl. hardware, employees, training, inference). They were not profitable on non operating costs though. That’s partially interest, but mostly, when their valuation goes up, they’ll have to issue more stock (absolute value) to early investors after IPO. (so, them going up in value is a loss, as the shares they’ll owe will be worth more)
Their profitability will likely crator as they start building their own data centres after the IPO, so that’s the best results they’ll have in a while.
DeepSeek's API used to (does still?) route your requests directly to Claude Fable[1]. Those were provably Claude Fable's responses being returned by the DeepSeek API. This was months after Anthropic accused them of distillation of Claude at scale[2].
Them provably distilling after being accused of distilling strongly indicates that... that's what they do. To spin a story of them not doing that (because why would they, with their superior talent and all) doesn't seem very reality conform.
The models are impressive enough as is (considering resource constraints and difference in CapEx) despite being heavily distilled from western frontier models. No need to pretend that they're not.
You really cited a tech politics person, and Anthropic... The poltical AI people are the exact type of cope distributors I complained about above.
I think the more reasonable explanation is this is totally made up. Notice how they use examples concerning "cyber and bio" notice how the examples are politically loaded subjects that are hyper relevant to proposed AI dangers and why we need to protect western AI companies with anti-competitive regulations?
Oh no!!! The scary asians are distilling our models, but they only want the scary things!!! They want to hurt you!! We better stop the scary asians!!!
> I think the more reasonable explanation is this is totally made up.
Based on what? None of the accused companies have ever denied Anthropic's accusations. And here you are confidently claiming it all to be made it. How much of a reputational hit to one's honesty do you think it carries to completely fabricate someone doing something, when they can just publicly disagree and disprove you? First thing DeepSeek, Moonshot, etc. would do if they hadn't would be to make a press statement calling out Anthropic for their lies.
DeepSeek routing to Fable was corroborated by many independent sources... since... everyone was able to just try it out and see it for themselves.
> Oh no!!! The scary asians are distilling our models, but they only want the scary things!!!
The straw man you're mocking doesn't exist. Non-Chinese labs are also distilling frontier models of their competitors. Elon testified in court that Grok was trained on ChatGPT outputs. And why would they not? It produces objectively better models.
Despite being "corroborated by independent sources," it was impossible - because deepseek has a higher TPS, can return logits and reveals thinking. Distillation happens asynchronously.
DeepSeek's Anthropic compatible endpoint doesn't return logits. The accusation is that requests to api.deepseek.com/anthropic coming from Claude Code or OpenCode just replayed Claude's responses sometimes. The clients would have just happily consumed the responses, as they were already meant to be Anthropic compatible. And yeah, TPS and the response-distribution were incredibly different between DeepSeek and Claude models, which is exactly why people noticed.
Multi-turn conversations have much more entropy than one-shot responses, making it more useful to do synchronous distillation.
So people were getting deepseek responses with long waits and no thinking traces? That would be easy to prove, we'd get more than just claims from interested entities.
When did I say everything is a grand conspiracy? This whole thing where centrists normies just call anyone who believes something outside their center comfort zone a conspiracy theorist is mega obnoxious.
You need better media literacy.
You cant just make up a reality where none of these companies have denied distillation.
The scary asians are going to get you, watch out! Your government is going to eventually tell you that we need kill a bunch of them too because they're so scary. Probably soonish. You're going to fall for it because you believe anything you're told.
You are postulating an organised psyop, where OpenAI and Anthropic are publicly lying about Chinese labs to paint an image of, in your words: "Oh no!!! The scary asians are distilling our models, but they only want the scary things!!!"
These lies are then further supported publicly my western media, random people on twitter, western governments, etc., all to paint China as an evil force that must be united against. Again, in your words "Your government is going to eventually tell you that we need kill a bunch of them too because they're so scary."
And you have dozens of independent parties pushing that same lie to that end and for that purpose? Sure, that's what the New York Times and CNBC must deeply want too. When CNBC writes that they've reached out for comments to the accused companies, but received no responses, they must be lying as well?
It shouldn't be hard for you to demonstrate DeepSeek denying it. Please, go ahead, link to a statement. China's foreign ministry has denied Chinese labs distilling, and, in the same statement accused US labs of distilling from Chinese models. Now that's a very trustworthy claim.
I call it a conspiracy theory because it's a conspiracy theory. Some conspiracies are true, but this one is not. Distilling is both trivially simple and produces objectively better models. Plenty of evidence points to that practice. But in your mind the labs in China are just morally too upstanding in engage in such lowly tactics to create better models, despite the evidence. And it's convenient that the evidence can be explained away by claiming it to be a psyop to paint Chinese as evil.
> surely you can't blame the OP for writing this after a few months ChatGPT
Bad takes are bad takes. The author was perplexed by the "many people" convinced that models are intelligent and is argued against opinions/arguments he's been exposed to. Called proposed use-cases "borderline fraudulent pseudoscience."
He also published a second edition of "The Intelligence Illusion" in Sep 2025, so seemingly still stands by (some variant of) this belief.
It's an interesting reminder of how much general discourse has shifted since 2023 (I haven't heard of stochastic parrots in months!) but being wrong early doesn't change that.
As a shareholder I am up 75% in a year. Alphafold gained experience makes them better suited to succeed with Isomorphic Labs than anyone else. Research on improved translation gave us the transformer.
If you think AI will win but Google will continue failing, there so many better places to allocate your capital right now.
For any task, a Pareto frontier exists of costs v. success-rate. Anthropic has 20 model/effort combinations currently, and maybe 10 (?) of those are never the best choice for any task (where a cheaper combinations provides same or better results). You know your workloads, and can estimate which combination to use, but models can too (and will get better).
If Sonnet can increase its reasoning effort, it should know that at some point switching to Opus makes more sense than cranking itself up to xhigh/max effort. A model can, and should be calibrated to know if a cheaper config has the same or higher success probability. It should also know which config has the next better performance (whether it's itself at a higher effort, or a more costly model).
At any price point, there exists the best model/effort combination. Why be against letting a person set the effort/time (e.g. on a slider with 5 options) and letting the models self-allocate tasks based on their strengths?
reply