Both Anthropic and OpenAI seem to be burning cash as if it was a game of startup chicken. Of course they all desperately need cash.

Comment from a perma bear. So take it with a grain of salt. lol

Fable is great, but it's no better at writing or reviewing code than the top-end Carbon or Astra models. To the extent Anthropic has a moat, it's Claude Code. Obviously that's not much of a moat, but their competitors seem content to camp on the other side of the drawbridge for some reason.

Codex is pretty close in a lot of ways, as the other poster says, but Astra consumes tokens at a higher rate, and when it compacts it completely lobotomizes itself. That used to be a problem in Claude Code as well, but it has been largely solved (although at the cost of making the compaction take much longer.)

Antigravity reminds me of the old comic strip where Firefox and Chrome are duking it out, while Internet Explorer is sitting in the corner wearing a helmet and eating out of a jar of paste.

Normally I run the models that the harness is intended to support. You can't run Claude models anywhere else besides Anthropic's own CLI and desktop harnesses -- or at least you're not supposed to -- without paying out the nose for API tokens.

I do use DeepSeek 4.1 Flash 1017 in Claude Code when I want an abliterated model for security-related work or unhobbled browser control, though. I know it's possible to get OpenAI models working in CC, but I haven't bothered. Gemini I'm not sure about.

No humans in the loop.

Deep learning had no humans in the loop, but there's always a bigger outer loop with a human somewhere.

You can place bets that these companies will be failures if you want.

It also doesn’t stop investors from being happy liquidating at an already good return, especially since they can choose to liquidate partially.

If you’ve ever don’t consulting work, you know how dangerous it is when the majority of your revenue comes from a single customer. You stop being independent, you stop having your own set of values, and if anything happens to that company, you’ve already proven you lack the ability to court as large a contract from other customers.

It always ends in a loss of identity and huge layoffs.

Making vendors product triple their current production rate for the next five years isn’t even possible to start this year, because they have to build facilities and train employees. Nobody runs businesses at 50% capacity anymore. That’s considered wasteful, even stealing profits from your shareholders.

They can’t even get memory chip companies on board with their koolaid.

Because private rounds still dilute you.

And VCs tap out well before $100B.

And they'll need more than $100B to accelerate widespread economic diffusion.

And RSI can't hallucinate gigawatt power plants and TSMC fabs into existence.

Apple, Amazon, Intel, and NVIDIA all went public in an entirely different world, before the JOBS act.

I'm not necessarily taking the position of the person you are replying to, but using these companies' IPOs as comparisons for what Anthropic is doing doesn't really make sense.

This is no longer the days of Apple, Amazon, Intel, Google, NVIDIA and Tesla.

They need resources and $100B can buy them that. I don't think you need a conspiracy theory to explain that.

But when it comes to Anthropic, I am almost certain that it does not make sense to argue based on the typical arguments on maximizing profits. Everything I know about the people working there suggests that they are not primarily motivated by money, but by their mission. I bet they would prefer a lower-strings attached private offering, but have no issue with diluting their own stock, if they think it will serve to move Anthropic further ahead on bringing about aligned AGI. Basically, the diluted money will just be missed donations to EA causes for the most part anyway.

He distinguishes between weak recursion (using sotfware to build more software), and strong recursion (handing over model development to AI which does everything), and makes a pretty strong case for the fact that only weak recursion is happening.

This maps to my experience with AI models, they reduce the cost of experimentation massively, but they can't (currently) design experiments themselves.

Even then, training RL is very data hungry, and most profitable tasks don't have good data so even if we got superintelligence (whatever that means), it's entirely possible that progress would still be bottlenecked by lack of high quality training data.

The hope was that weak recursion could produce strong recursion by automating the iteration of ML pipelines.

Theres also the idea of LLMs producing synthetic data to contribute to new models. Perhaps that is not as effective as hoped.

This never made sense to me. Seems analogus to inbreeding in biological organisms.

More complex synthetic data generates things that could be in the training data but aren't for some reason. For example, machine translation training corpuses often contain a lot of formal language (because formal documents are the ones that get translated), so you might generate a bunch of casual conversation to improve the translation model's ability to handle that.

Wasn't model collapse/model decay a big issue since the beginning? i know that RL kinda mitigate the effect, but if this was close to not be an issue anymore, wouldn't we have a paper on that already? All that i've seen on the subject was about mitigation, never about solving the issue.

Sorry if my question is dumb, I need to do some research about the difference between image generator and text generator, i feel like my understanding of LLM is skewed by my usage and understanding of Z-turbo, but i only have that much time in my week.

AI can be a "tool", but only so long as you understand how it works and what the implications of its use are. Being an enormously powerful tool implies, the consequences of not understanding and not knowing are proportionally enormous as well.

Currently, LLMs are used practically exclusively by "professionals" who, more or less, know what they're doing. The idea of applying them "everywhere" completely misses the gap in understanding and responsibility for the vast majority.

The consequences of "AI use" are manifold and very poorly understood. That doesn't make them less impactful, on the contrary.

Whether you look at "humanoid robots" or incorporeal LLMs, when they are more capable than yourself, you become their dependent counterpart, not their "master".

https://www.tradingkey.com/analysis/stocks/us-stocks/2623001...

Anthropic desribes itself as a "public benefit corporation" in its IPO. The irony. Or is it sarcasm?

A god isn't your subordinate. You cannot "program" your wishes into it, like a genie/jinn. They (or their investors) want to make themselves into gods. Over you.

Not Anthropic specifically, but there's this thing that I recently read:

https://www.nytimes.com/2026/11/09/opinion/ai-anthropic-doom...:

> A.I. company leaders often talk about how their work might lead to human extinction soon. This is astonishing. Even if you wanted to destroy humanity, you wouldn’t say so. If, however, you wanted to save the human species, wouldn’t you stop developing this technology?

> There is an explanation for why some tech leaders, so seemingly worried about an extinction-level event, are plunging ahead with A.I. development: Mankind’s survival, while important to many of them, is less critical than building something they believe to be not only conscious but godlike. A substantial portion of the A.I. community, in celebrating this belief, is quick to support any action or statement that reinforces the idea that the technology is like a deity. They might call for regulation, but only in specific ways — perhaps by deploying more A.I. to control A.I. — that adore the technology’s powers. It is that adoration that is nonnegotiable.

It's by the guy who apparently invented VR and was apparently mentored by Marvin Minsky. So he likely has the kind of access that gives his opinions credibility.

I think it's unlikely that an AI enthusiasts would literally claim "they are building a god," but such an attitude (or something very similar) is implied by their statements and attitudes, e.g.

1. Talk/hope of building something that so exceeds human capability, it's beyond human comprehension and control.

2. Taking existential risks for humanity (by their own statements) to try to get there, against all reason.

3. etc.

I guess they think AI will become smarter than people but I'm not sure that's a god.

This was always the case with ML, even absent RSI. Big bag of weights is a pain to figure out exactly why it can do whatever it is that it does, leading to all kinds of blind spots about how they go wrong.

> Currently, LLMs are used practically exclusively by "professionals" who, more or less, know what they're doing.

Even the people who made them don't really know what they're doing, hence https://www.felonybench.com

And what is that mission?

I'd almost prefer it if they were motivated by money, then at least we know what we're in for. All this 'mission' stuff seems to effectively amount to 'we're saving the planet, sorry you got in our way'.

sniffles

I was excited when I got a good metric system implemented because that meant concrete signals for Codex to use for optimization.

It attempted ~5 things. All of them did not improve performance past the baseline.

How much more for the unclear goals, general ML research, etc. My takeaway from all this: You need domain expertise to drive the models.

It is disappointing just how little recognition there is of this point. For some reason, AI maximalists settled on an exponential model of growth in their mind, and I've yet to see one seriously consider whether RSI might not be that. It's a conversation ender, like bringing up an alternative axiom that cannot be debated, rather than a testable hypothesis.

I've never seen it solve a problem by "thinking harder", but as soon as I give it access to a feedback loop where it has control it can solve a lot of problems, It sort of feels like brute force, but it's not because it's informed about what it focuses on.

    theory > change > test > assess > rollback and retry)

I think the bigger issue is how you convince people to give those findings back to the system. Either you lie about what you're training on or you have to give people a stake in the benefits. For example, I still needed some level of expertise to track down my issue, so why should I publish that so Anthropic can train the next model on it and rent the knowledge back to me?

Though the truest of true believers also see themselves as one day eliminating scarcity as a concept so it might just be that profit doesn't motivate them.

they still need many billions in compute, while it is not clear if RSI will provide strong revenue gain, so they propose Wall Street to take part in gamble.

- rsi won't happen immediately, and is inherently unpredictable, so they need to bridge the gap

- their "public benefit corporation" plans have always involved trying to share the upside with the public before it goes "exponential"

I just can't imagine this is true when their overreach problem (doing more than you ask), hallucination problem (lying), redefinition problem (doing something completely orthogonal to what you asked), and harm problem (doing something that is exactly the opposite of what you asked) are all far too high at this point, and by most measures - all of the recent models are worse on most (or all) of these metrics.

They are giving up here in RLHF to benchmaxx the metrics people are looking at to pretend like they're making huge gains, when they're largely running in circles.

They are getting better at pass@1 at the expense of far worse p90s on real agentic tasks.

Or Repetitive Strain Injury. If LLMs acheive RSI then programmers won't have to type as much, so the risks of RSI will decrease.

All insider and outsider predictions seem to think "true RSI" will really begin in 2027 or 2028.

You can do the math on the dilution, which will be on the order of ~10%, give or take. If the new shares are worth more than existing shareholders paid for theirs, each share can be worth more even though the percentage is smaller.

2nd bullish thing seems to be that chinese labs are having difficulty in distilling latest frontier models. No deepseek pro for some time now!?

also government, defence, financial and similar institutions and companies will almost sure never use chinese open weights models.