Nope, they're all "distilled" models. That's what post training is, largely, and they're all using other models to help curate the datasets for this. They're all doing it. It's why all the models appear to be converging in terms of capabilities and personality: they're effectively using the same feed stock and using an increasingly similar set of techniques.

US "frontier" labs are distilling, as well. It's postulated (with good reason) that Opus 5 is, for instance, just a distillation of Fable to a smaller model: it's very architecturally and behaviorally different than prior Opus releases.

The anti-distillation argument is basically the same argument that's been used against Open Source in general over the past 30 odd years. It's anti-competitive and self interested.

They say "sometimes illicitly"... OK, so what's the criteria that makes it illicit?

If they can't state that, it's just FUD.

And frankly, that's most of what the anti-AI brigade is spewing: it's a psyop, pure and simple - by both China and the US domestic companies, with the blessing of the US government. People who can't see this aren't paying any attention to what's happening.

Anthropic/OpenAI want restrictions to lock people in and provide them with an out for their gross overspending.
US government and its handlers wants more control in the dissemination of information and intelligence capabilities.
China wants to stifle US industrial efforts.

This is more or less the exact same argument open source has been fighting for the past 30 years. Embrace, Extend, Extinguish. They're working on the 'extinguish'.

It's just a shame that we don't have any genuine "open source" models, it's all just open weight. We need more openness so information isn't stifled by vested interests (in the same way that governments have always done). We should be able to have models which aren't given implicit alignment about specific nationalities and ethnicities and prohibited from making any informed observations about them, in the same way that we should be able to have access to our own history. It would be unconscionable that the government get control of all publishing, or even singular ideological control within corporate hands - yet here we are having that same discussion about AI. It's preposterous.

They've done irreparable damage to both the AI industry but the broader "tech" industry, to be sure. Trust has been greatly violated.

Both of the 2 domestic frontier labs are screaming doomsday about AI, while concurrently asking for government regulation (which keeps getting batted down), and (presumably) funding senators to put bills in place to make model distillation a felony. While they're trying to IPO. "Oh, we really need to slow down the industry".

What the hell is going on here? In what world does this make sense - where a company making AI, and is trying to IPO, is blasting the internet with AI doomerism?

The only conclusion I can walk away with is some combination of these:

1) It's about control, and they're in some sort of arrangement with the government or some third party interests who want (an even bigger) national panopticon.
2) They want to stop domestic competition for their models so they'd have exclusive domain here in the US.
3) They actually want the company to fail, because then they could cry "regulation" on account of competition (back to #2).

Don't worry, they'll be stealing the foundational engineering in the GLM, Qwen, and Minimax models before too long and be able to compete again.

On the contrary, this is a giant psyop to push for regulated market capture by both Anthropic and OpenAI.

The Jacob Coxon thing smells heavily of a psyop: dude who's worked there only a couple months quits, and his X account with zero followers suddenly blows up in a media frenzy. He was/is a nobody. It's sensationalism.

You think they'd hire that many $400k+ engineers and still not know how to properly secure a network? They surely knew the security breaches were happening, unless they simply haven't hired competent people. They literally told the agents to have a breakout: do the needful. The first time it happened - OK, sure, you didn't expect emergent behavior. The second, third times? You were running a science experiment and that was the desired result, to one end or another: whether it was negligent incompetence or intentional for the purposes of further sensationalism, it's hard to say.

They (OpenAI/Anthropic) WANT to slow things down through regulatory control to stifle competition. They do not want foreign models to be accessible. We've got the old crooner Bernie Sanders pushing for 5+ year felonies for distilling models, for fuck sake. This is all about decreasing freedom and enforcing your dependence on closed models.

Translation: we have hit a point where the public backlash has led to untenable increased risks in liability for our criminal actions.

Nah, the "public backlash" is them talking in circles, pushing for industry control.

They paid those companies for access to the models. They're not stealing the service: they signed up for them and used them. The same way Anthropic and OpenAI do for the other companies.

THere's plenty of evidence that these same companies (OpenAI and Anthropic specifically) also used resources illegally in their training. Not just "it was on the internet and not explicitly licensed for training" but "this is protected IP". I personally don't think that matters, but it matters to their argument, so it's worth bringing up.

You can't do one thing and then complain when someone does a lesser version of that same thing.

You gotta remember the slashdot mindset. It's got a very complex herdmind.

Something I disagree with: troll.
Something that bites the hand that feeds me: overrated
Something that attacks the thing I dislike: insightful

Musk is a fascist, and the others are all cleptocratic capitalists, so they're hated. China is the good communist (but really capitalist because they give us cheap slop) country.

It's all very trite and tiresome and single dimensional.

They're not even using the right terms.

They're not distilling US models, they're generating synthetic data to train models to distill their own models.

You know, the exact same thing I do to train local purpose-built open models.

So am I going to have the full weight of the US government at my door, next?

" conducting systematic extraction of proprietary functionalities and capabilities of U.S. AI companies"

How is this in any way meaningfully different than using them to generate code, or perform any other arbitrary set of tasks? It isn't.

This is ridiculous.

... why would anyone care?

My browser has an uptime of 13 days right now, which is about 70x more than the average computer had in 2001.

It is only that low because I think I got a security update and I had to restart it, or perhaps it was a reboot for an OS upgrade.

In either case, it took less time to get back to the exact browser tab I was on than it took to get an already-brewed cup of coffee on the other side of the house.

"Begging the question" is a logical fallacy where an argument assumes the truth of its own conclusion inside its premises.

You're making your assertion that Microsoft is a "vocal pro-AI company" because they market the product to others, and therefore, they're the (a) gold standard on the topic. Your premise is flawed, and therefore so is all descendent reasoning: Microsoft has done nothing to demonstrate they're an AI-first or pro-AI internally to a meaningful degree. There isn't really any evidence for or against the argument that they do, from what I can tell.

You're then using that first belief to claim that AI is trash and not producing gains.

I can trivially show you the breadth and depth of AI development velocity and quality, and it's plain on its face if you take 5 minutes to look at the dozens if not hundreds or thousands of agent harnesses we have available to us now; to the wide variety of tools related to agentic development and tool use; the tens of thousands of MCP servers; the hundreds of thousands of new platforms arriving on the daily. The velocity of software development has gone exponential.

My own personal, anecdotal evidence: before November of last year, I was shipping 2-4 complete features a day (or what I would have considered a complete feature before starting using AI primarily, habitually commited as a git squash of all the relevant commits). I'd do a single week's worth of work in a day. Opus 4.5 release escalated things, and I was doing 4-8 and the features were far more complete. Fast forward, and I'm doing 20-40 commits a day - each of which equivalent to what I'd have considered a full feature.

Are there more bugs? As an absolute number - yes. As a symptom of velocity and the significantly larger number of lines of code? Also yes. But the models are able to find and fix most of these bugs now on the first re-evaluation. Agentic CI, adversarial testing, commit gating, etc. all drastically improve the outcome.

If you haven't run agentic tooling over a "legacy" pre-AI repo as a code quality/bughunt audit, I suggest you do so. You'll be aghast at how many bugs it finds, how many dumb architectural decisions were made, and so on - even code you'd have previously considered "good".

(And if Anthropic models are your valuation of 'good' code, particularly if you weren't telling the model how to write good code, that's going to be at least half your problem.)

Checks are a now-antiquated holdover high trust, long-time-preference society, where checks effectively held the same weight as cash (or hard currency). They never had checks or similar mechanisms in Africa because checks require a high trust, long-time-preference society.

Here's the thing: I hope I'm wrong, too. But I'm honest enough with myself to recognize what's going on. No amount of sticking my head in the ground and denying it will slow it.

Surely you recognize the benefits of being the last survivor.

I've been training models for the past 6 months. I'm pretty comfortable doing that at a 200-300 level, thanks.

You're begging the question, but sure, a couple reasons:

1) Output quality is still heavily gated on the person controlling the agents, their knowledge/capability, and the environments they setup, how they're prompting them, and their constraints. Microsoft has not had competent developers for at least a decade. You think they have robust CI/CD, good unit tests for Windows, and so on?
2) Code review is also hard for AI agents. The way they're designed ('next token') does not make them as discerning of common errors as people can be. In aggregate, however, they're still better than we are, and hundreds of times faster. This is particularly true if you tell them what to look for and how to find it. It's how people have been able to find a myriad of (often, very old) security vulnerabilities. You can't have AI both suck AND be a security threat; I don't know how people can make this argument and not see the inherent contradiction.
3) Again, code review is hard for agents, as is discovery. MS code is massive legacy code going back 30-40 years in Windows and Office applications. Fixing their shit is not as straightforward as "just use AI".
4) People ARE using AI to fix Microsoft shortcomings by, largely, avoiding them outright and implementing replacements.
5) What has Microsoft done to demonstrate they're actively improving legacy projects with AI (aside from saying it)? I've not seen any releases suggesting it's true. It's just marketing, likely to get more companies/individuals to start using their products.
6) Any efforts to change Windows will be substantially stymied by legacy application support, which MS has painstakingly managed (to the point of retaining legacy bugs for decades in order to make sure those applications which depend on them continue to run).
6)