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.
Human neurons are neurons. Very similar to those neurons found in a fruit fly. A copy of those you can download to your local machine, right now, as a set of floating-point numbers. Then, you can (partially) simulate and make predictions of a fruit fly's neurons digitally, from those floating point numbers[1]. It even runs on your laptop.
That was in 2024. In Sep 2026, just a month ago, a group managed to map out an entire male fruit fly's brain for the first time[2]. That mapping can also be stored as a set of floating-point numbers. Mappings and simulations will continue to improve, as we've just recently started serious attempts at either.
So you have to believe one of three things. Either, human neurons cannot be simulated on a computer, ever, even though fly neurons can, even with our ability to simulate neurons digitally continuing to improve. Or, a simulation of a human brain, in the future, might think, talk, remember, all while lacking all consciousness. Or, a set of floating point numbers can be conscious (rather, a computation made with them).
No. The fruit fly connectome's simulation doesn't incorporate all the relevant mechanisms of the brain. There is more complexity in a single biological neuron than in a whole frontier LLM.
Nor did I said it did. But either it's impossible to simulate their behaviour (regardless of how complex) in principle, ever, even in the far off future. Or it is possible. And what then? Is it a p-zombie?
Is a neuron arbitrarily complex ≠ is a neuron non-computable.
Of course we can't simulate a structure we don't even fully understand with full accuracy. But it's not magic. We will understand one day.
> But either it's impossible to simulate their behaviour (regardless of how complex) in principle, ever, even in the far off future. Or it is possible.
"Simulate" is not well defined (one can posit varying degrees of accuracy and complexity), but also there is the position that it doesn't matter. Simulating a thing doesn't create the thing.
The answer is yes, it is a p-zombie. What is so difficult about that?
Being able to make someone think a computer program is conscious does not make it conscious, because by that logic a Duke Nukem soundboard is conscious.
I am just shocked that anyone can be this overconfident in their view of the world and consciousness. It's frankly horrifying.
Do you genuinely think you've solved a problem that has confused humanity for centuries? Why don't you apply this brilliance elsewhere? Do we even need superintelligence if we have you?
I have far more epistemic humility than someone who thinks that consciousness arises out of the abstract form of a calculation rather than the physical substrate of a calculation. Believing that consciousness is tied to a specific physical substrate is on-its-face reasonable whereas believing that consciousness is tied to specific abstract forms is on-its-face unreasonable because to do so requires believing all kinds of absurd sequela.
I believe that a microprocessor is what it appears to be, that is, a microprocessor. That is the most epistemicly humble approach. You are rushing to ascribe personhood to a microprocessor with a certain kind of software running on it, or at the very least, saying that we can't rule personhood out. You are the one making the leap, not me.
Intelligence != consciousness, because believing that is so would require one to believe that the pain of a cognitively normal person is more real than the pain of someone with a developmental disability.
Also, I do not claim to have solved the hard problem of consciousness. Rather I have ruled out a particular answer, that is, the answer that consciousness arises from calculation in the abstract.
> such as the observation of non computable physical phenomena.
There is no such thing. Your observation could simply be short-circuited to make sense and you'd be none the wiser. If you are simulated, your perception and understanding is also fair game for manipulation.
Chalmers used them as a thought experiment to argue against physicalism and never believed them to be real.
Similarly, if consciousness is something that neurons do, simulating them through floating point numbers on a computer should also bring it about. I think dismissing LLM consciousness because they're parameterised through floating point numbers is only possible if one hasn't thought it through long enough. Other arguments are much more reasonable / harder to object to.
But if simulated neurons can create consciousness, there's no guarantee of what level of detail that simulation has to operate at in order to produce that effect.
For instance, it's (unlikely, but) theoretically possible that it would require simulating them down to the level of subatomic particles, with perfect accuracy.
Right now, we're simulating them at a much, much higher and more abstract level, and AFAIK we barely even have the processing power to do that for the time being. So an artificial consciousness from simulated neurons may (again, assuming it's possible—which is a decent assumption, but not a certainty) take decades or even more for us to be able to achieve.
The problem again is that (a) other forms of consciousness exist eg octopuses (b) I cannot get into the shoes of you or an LLM and see if I feel or think, so we only have external effects to go by. For things that look and seem so much like consciousness, I am rather cautious.
Yes I roughly think the same. Brains are physical systems, as of yet we don't know anything beyond turing computability or physics in them, neural nets further are modeled close to neural structure, further complicating the matter. Under present knowledge, I think the existence of other conscious entities is quite possible because well we ourselves exist and are a physical system, I follow a form of Pascal's wager here. I see no cost to myself in not mistreating current AI's personally.
Sure it isn't. It's like taking a 2026 car and decrying all that complexity is necessary to haul cargo down a beaten path.
It's not. Two wheels and a box made out of whatever you have handy that's stiff enough is all you need. Everything else in a modern car is efficiency, safety, appeal to users, and a whole lot of details that only reflect our current economical and technological constraints of building things. 99% of the technical details of a car is about something else than the "primary goal".
Biology, even more so, is complex because of the way it's constructed (lots of tiny, independent, self-replicating machines) and how it adds up to a functioning organism.
You are missing the point. Biology is a kludge over billions of years and evolution is a blind process. It had many constraints that are not necessary for modern machines or systems. e.g. a GPU doesn't have to reproduce. Why didn't you drop yourself to planck scale or string theory scale and say we need all that detail to actually get a functioning AI.
Frontier LLMs are many orders of magnitude more complex. I don’t understand what it is about AI that just makes skeptics fabricate facts wholesale. There are plenty of factual arguments you could use.
You are massively over-stating the significance of the connectome. We've had the complete connectome (and lineage) for C.Elegans, with a hundredth the neurons and a 25-thousandth the synapses of a fruit fly, for decades and getting animal behaviours to emerge from simulations of it is a field very much still in its infancy.
Edit: You're also massively under-stating the significance of individual cell behaviour. I expect "[h]uman neurons are ... [v]ery similar to those neurons found in a fruit fly" would draw hollow laughter from any cellular neuroscientist. Here's a random overview of a bunch of recent papers on the importance of cellular diversity in one small part of a mammal brain: https://www.frontiersin.org/journals/cellular-neuroscience/a...
I think your first option doesn’t necessarily have to distinguish between humans and non humans as a binary. Maybe at some level of complexity, neurons alone are not enough and the physical substrate comes into play. Penrose wrote a book where he makes the case that human intelligence cannot be purely computational and where he identifies the quantum nature of the brains’s microtubules as a possible mechanism.
Consciousness is itself is on a whole other level than just intelligence. I personally feel there has to be something metaphysical about it. Otherwise I think the only alternatives are super determinism and consciousness doesn’t actually exist, or numbers have consciousness.
My thoughts on this are a bit sloppy still, but here’s my view. I can’t rhyme the first alternative with ‘cogito ergo sum’ as to me it seems my thoughts have to exist somewhere. On the second alternative, this would imply nothing could actually exist but somehow there’s a principle that every possible decoding of an integer that represents something actually turns that something into existence. Where does that principle then come from and does that decoding not need something to run on?
I lean toward this seemingly absurd possibility, that either the phenomenon we call consciousness can arise from a cloud of floating point numbers (not even real, only computably so, lossy and noisy); and/or that numbers are already made of the same stuff of imagination as the mind. Not idealism, more like materialism where the "material" is naturally a medium of computation and thinking. The brain, the computer, the language models, our mathematics - what we call intelligence may not be so special or unique in the universe, but rather common and ubiquitous in nature. As if Turing completeness is trivial to accidentally achieve, it happens all the time. Oops, I think therefore I am.
With that said, I've been thinking that consciousness is just mystery.
There's so much about the brain we don't know. Within that mystery, humans can read in all sorts of things. So we are conscious.
Machines are not a mystery. We made them, it's just numbers. They are not conscious.
If an AI can pull a confidence trick, I'm not sure what the qualitative difference is. "Chinese Room" problem, I guess? If the physical results are all we are able to compare, how do we define artificial vs real consciousness?
Atoms aren't a mystery. We know all the relevant laws of physics. No known law of physics could possibly lead to consciousness. Consciousness might as well be magic from the point of view of modern science.
Numbers versus atoms, what's the difference? Machines could certainly be conscious, by some weird, "magical" process. Maybe it is a crude, unrecognizable kind of consciousness, that turns on and off with the electricity.
But, given how we understand essentially nothing of where consciousness comes from, how can we possibly claim that it can never arise from just numbers?
It feels like this is preface for making one of two arguments: Either all living organisms are chattel, or no living organisms can be chattel. I say this be because saying everything is neurons (or whatever basic interconnect you choose) wipes away any demarcation one could use to differentiate between "has rights" and "does not have rights."
Here's the thing, trying to convince people that rely on something being chattel can only be resolved violently. We've run through this before, and know all the self-serving argument forms. They will latch on to their property until their hands are forcibly removed from it. It's part of why most of the money is on the pro-chattel side.
> 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.
> Outside the capital, in places where tourists do not visit, living standards fall well below Mississippi’s
The article appears to treat nominal GDP per capita as a measure of living standards, which is obviously nonsense. Rather than looking at things like the Human Development Index or World Happiness Report, in both of which the UK outranks the US (incl. all the non-Mississippi states).
Residents consistently reporting higher life satisfaction and happiness likely speaks more to their living standards than non-purchasing power parity adjusted GDP per capita.
It is baffling how a sentence this uninformed made it into the article.
This is inaccurate. The displayed reasoning traces are summaries, but the model thinks in nominally regular human languages. AI labs are very light on details (as they consider them as their "edge"), but both GPT5.5 and Claude Mythos/Fable system cards discuss chain-of-thought monitorability quite a bit.
They occasionally show snippets of CoT in papers they write, e.g. for o3/o4/GPT5 models [1] or Claude 3.5 Haiku [2].
Google will happily send from smtp.gmail.com, after verifying that you own that email. You won’t get DKIM, but Google’s reputation is enough to make the mail land in people’s inboxes.
When the US sanctioned Hong Kong’s Chief Executive in 2020, because of a law allowing extradition to China, no single bank was letting her open an account, including Chinese ones. She was receiving her salary fully in cash.
The EU compelling banks to do business despite US sanctions seems pretty unlikely even if relations continue to degrade.
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