dannyw

You can tell Claude few-shots by eye; and asking for verification, such as a few screenshots isn’t unreasonable.

Your inference server is just an inference server that spits out tokens. You can use the same harness, etc. you’re basically just changing the API base URL.

I'd say it's both intentional friction, as well as just some degree of load distribution throughout the day, instead of peak hours (US work hours).

At OpenAI/Anthropic scale, they're can't just scaling up-and-down on demand on rented autoscale GPUs (to be honest; that's not really practical anymore for even medium size volume); the peak hours are basically the same for everyone.

Rough hand-wavey guess, assuming:

* 85% inference margins = $375

* 50% abandonment (since it does expire; and general prepaid credit abandonment is pretty high; especially given how much credits they dished out; so probably on the low end) = $188

Plus, it's unclear how many users opt out of training, so more usage and RLHF data on their best models from long-term power users, might work out to be a net benefit to them; since these kind of "non-abusable" (i.e. not announced beforehand) announcements don't suppress monetization as much.

We don't even have laws, against a storefront taking away your purchased shows/games/music due to a licensing issue or whatnot. Or if we do, they're clearly not enforced enough.

As for the 'mystery nature' of subscriptions; you can always pay API rates if you would like. If their API rates are too expensive, very cheap models like DeepSeek V4.1 Flash, etc, are available.

They actually have built and tested a feature called “Luna Reserve”, where you have an extra pool of Luna to tap into only when you’re out of usage.

I had it for a couple weeks before it disappeared. It probably decreased upgrades and monetisation too much.

The friction is intentional. Esp with the broader ‘AI trade’ showing jitters; they need good financial numbers.

Recently, HuggingFace has already taken down multiple models abliterated for cyber and with such descriptions, at their own initiative.

Example: https://huggingface.co/audnai/penclaw-GLM-5.3-abliterated-fo...

Previously, I recall HuggingFace only took models down if someone complained / sent a DMCA or cease and desist.

He's a PhD student, so a researcher. Researching how to build better AI is still research; not commercial.

Considering academic research to be competitive is an incredibly dangerous slippery slope. It didn't use to be this way; e.g. just look at the GPT-2 paper: https://cdn.openai.com/better-language-models/language_model...

The linked article in Spanish contains substantially more detail. Natzir Turrado tested hundreds of prompts.

For this example; he tested 30 prompts at 5 pass-pairs each; so 300 times.

https://natzir.com/posicionamiento-buscadores/que-es-como-fu...

They are doing similar things, yes; but they are not the same especially when it comes to `cyber` classifiers, or the extent to which these models will assist you with ML and AI research.

I find that logic similar to thinking both political parties are the same.

The FAA is federal, and may pre-empt state/local laws (see: CFTC, Kalshi/Polymarket/etc).

It's also pretty well accepted that the FAA is very cozy to SpaceX, and partially because NASA is basically dependent and beholden on SpaceX at the moment.

I’d love model drops early, but I’d also love to see them drop something the day before Anthropic’s IPO.

We need a counterbalance.

Fortunately it is not a one horse race. GPT 6.1 Sol is an excellent model, sure it has some weaknesses (e.g. frontend design), but it also has strengths too (e.g. computer use / control; which is great for automated E2E QA).

On software engineering I find it's generally close enough to Opus 5.5 where it's more situational and even personal preference / skills / prompting style on which is better.

Most importantly, while neither company are saints, I don't curse at OpenAI models for classifiers anywhere as much as Claude models.

The "take over any instagram/meta account through their AI password reset chatbot" was going around and sold on shady forums for ~1 month before being patched.

That’s not backed by data at all. We’re discovering more existing vulnerabilities.

Anthropic needs to issue a cease and desist to Cursor for claiming official “by Anthropic” extensions they don’t make or control.

They would have done this in a heartbeat to anyone that they’re not reliant on for compute.

Honest question: does that even remotely matter, or is it an outdated prior? All software has bugs.

The goal and objective is to deliver value for your user/customers; to the speed, reliability, and assurance requirements your market segment expects.

Anthropic is desperately dependent on SpaceXSI's compute (which either party can cancel without cause, on short notice), they can't really tell SpaceXSI to do anything without ruining their IPO; or more Claude outages leading to users flocking to OpenAI, etc.

Of course not! But my question/point is why they allow a little bit to be scraped at all?

Obviously, the YouTube team is aware of yt-dlp, and with Google's engineering density, a small team can completely block it & win in the cat-n-mouse game against community volunteers that make all their code public.

You’re right, but the risk here could be reduced significantly by not loading dozens of third party marketing, analytics, and tracking scripts; one of which was supply chain compromised in this case.

Of course, that goes against commercial interests.

Another similar observation: if you ask Claude to download pirated content, of course it'd refuse.

But during the middle of a long training project, when it became clear I was data bound, Claude offered recommendations for BitTorrent clients for me to "acquire data".

yt-dlp works fine for low-volume requests from high-reputation IPs.

But if you try to download in significant volume, their systems know it's automated and probably have signatures specifically for yt-dlp.

I think it's a bit of an escape hatch: they know they can't close the hole (after all, there is always the "analogue hole", or now AI doing browser control); and probably prefer leaving a small monitored and rate-limited door than trying to patch everything.

See: Google changing their sponsored link DOM/CSS/classes matchers every (few weeks - few months) to break ad blocker lists, but they're certainly capable/smart enough to do it daily or hourly, etc.

People here copy and paste random "curl $x | bash" commands all the time.

Reading between the lines, it seems like Anthropic isn't involved in the "official by Anthropic" extension at all.

Cursor seems to be scraping extensions from Microsoft's marketplace, and repackaging it and keeping the metadata.

I suspect the 1/4 (?) ownership by Microsoft plays a factor into OpenAI.

Otherwise I reckon they’d have cloned GitHub already; possibly with a “Mirror to CodexHub? Free usage reset”.

This is incredibly cool! Fascinating to see looped transformers be necessary for mini-Doom, but not for simpler games.

GCBD operates the iCloud China service with Apple, and I imagine Apple, Inc and its associated LEA portal/systems does not have access (otherwise there's no chance China would let iCloud China be a thing).

This is also where the words of law and pragmatic reality intercede.

What digital services do you pay for that don't require sign-up, or an account?

"The requirements set forth in Section 2 do not apply to: (a) internal use of the Model or its derivative works, defined as any use that does not make the Model or its derivative works available to third parties, [...]"

Works for me. I can host it, I can run it, I can use it; but if I want to sell access to it as an inference provider, I'd have to get a license.

The economics of open-weight models has always been shaky. It's not an OSI license yes, but metaphorically it's kinda like AGPL.

AA (like most other VC-funded benchmark sites) is pay-to-play.

Want a private benchmark and see where you'd rank? $$$.

Want to show up day 0? $$$$.

Want to show up day -0? $$$$$.