Your prompts and whatever data they collect from your dev environment go to the inference provider, and typically the customer has neither control nor assurance over what the inference provider (re-)uses that data for.
Do those results reproduce with any other model than GPT-4o? A year-old paper analyzing a then-18-month-old LLM has dubious applicability to current LLMs, especially when some now have guardrails against persistently or arbitrarily rude users.
> If you run an agent on that model you need those bytes, not a close copy.
A close copy will usually have close performance. Later in the post, they recognize that there are derivatives of the model with very different size and data characteristics. They posted slop anyway.
I would also fire whoever allowed "infohash" to be used in that blog post. It might have been excusable if they explained that "infohash" is Torrent jargon for the root hash of a particularly structured Meekle hash tree over the content of a single Torrent, but ... they didn't.
We had to turn off a bunch of on-by-default features in the system I'm deploying now because IOS matches too aggressively and decides that arbitrary UDP datagrams are DHCP requests or PTP packets that need rewriting (and then dropping when they turn out to not be DHCP/PTP). I have to think that an AI rewrite with a today-good model would recognize the lax checks and do them better.
Yes, but they normally have a 12-hour delay (it was 2 hours at the beginning of June and got extended sometime after that, maybe later in June) and require some human action to reload extensions or restart the editor before the new version runs.
I noticed that there was no update yesterday and wondered why. The daily updates are annoying in the base case.
Edited: This update was live for only 11 hours, so hopefully the default delay limited the blast radius, as it is meant to do.
Taxes generally do not promise good living standards for everyone. Usually they come with promises of public infrastructure, some kind of safety net, and some other benefits. How did wealth taxes work out for Austria, Denmark, Finland, France, Germany, Iceland, Ireland, Italy, the Netherlands, Luxembourg, and Sweden? They also didn't promise good living standards for everyone, yet still found that wealth taxes were worse than not having them.
We've seen what happens when people decide to confiscate wealth with promises of good living standards for everyone. It has never delivered on those promises, always lowered median living standards, and usually led to lots of deaths. I can't think of a single example where recovery from a confiscatory regime has taken less than 40 years, except arguably Cambodia, and most of them are still ongoing. Do you want Democratic Kampuchea as a best-case outcome?
At this point, it will probably not stop them, just slow them down, but it will strangle other companies in the crib. How much do you plan to tax the new multimillionaire janitors at those companies?
If government subsidies are the only reason Tesla exists, why did Aptera (v1) and Fisker fail? If SpaceX only exists because of government subsidies, why are Boeing and Lockheed Martin (owners of ULA) worth so much less than SpaceX?
He doesn't have a markedly more lavish lifestyle than the hundred-millionaires, but I don't want to live in a country where the government takes stuff that it (or some majority) decides the current owners don't need, and you shouldn't want to either. I think the US and world are better off with Elon Musk in control of those companies, and with those companies having lots of capital to work with. A side effect of people agreeing with me (although I'm not invested in those companies directly) is that his paper wealth is huge.
So, no, I don't think "he doesn't act like he needs it" is at all a good argument for redistribution.
Most people in the US would be delighted to have over $200k of gross income, much less after paying for taxes and housing. Your argument seems to suggest that those developers should also be very happy with their incomes.
"Closer" in what sense? Does Elon Musk have a markedly more lavish lifestyle than them? Is his retirement safer? How many politicians are campaigning on investigating, regulating and prosecuting the hundred-millionaires? Do media run scare stories about the drug use and sexual promiscuity of the hundred-millionaires, suggesting that they have children from too many partners?
Existing taxes are mostly on income, not wealth, and the US has one of the most progressive income tax systems in the world. Confiscating some fraction of Tesla and SpaceX will destroy lots of economic productivity for a one-time bump; it's a worse trade-off than just printing however much money and taking the value through inflation.
They could be A/B testing model tweaks, and use it as feedback that way. It feels dishonest to give paying customers an opt-out for training on their data and then sneak copies by asking for general feedback without a prominent disclaimer.
Apple laptops historically included a power supply. Now some models don't, and you have to buy it as a separate item.
I appreciate the frustration, but as a separate item then a user can choose any power supply on the market, rather than being stuck with one that Apple chose.
Taking the benchmarks (as reported by OpenAI) at face value, why wouldn't they show Opus 5.5 on the OS world benchmark, and show no Anthropic models at all on the "factual error rate" benchmark
The fraction is markedly different when one includes compensation in the forms of retirement contributions, insurance and the employer part of payroll taxes: https://fred.stlouisfed.org/series/A4002E1A156NBEA
I don't think that is quite right. I had a book about computer graphics using OpenGL from before then, and I remember it saying (something like) that various polygons were supported at the API layer, but the GL implementation would decompose them into triangles at some point, and this could have surprising results for texture mapping, so it was typically a good idea to triangulate shapes explicitly.
One major technical reason that triangles are nice is that every triangle determines a plane, and rendering points onto a bitmap is easier if you know the surface is planar.
I'm not sure, but "This post is about how we got there.\nA GPU fills triangles and nothing else." is very Claude-ish. So does "[I created] one commit per idea, and timed every commit on those same 3.7 million polygons, so each section below ends with a number." And "A triangle needs no cutting at all".
My experience is that Claude Opus 5 and 5.5 are very good at implementing optimizations, reasonably good at suggesting approaches to try, and not very inclined to identify slow runtime as a problem. It tends to go for simple implementations until I ask about speed, which is certainly conventional wisdom for development.
If someone is looking for that particular piece of paper, they probably want it for reasons that distinguish it from the average piece of paper. Also, most defensive gun uses end without shots being fired, so "usually" nobody even has a bullet fired at them.
$72B of Anthropic's loss was on the paper value of financial instruments, though. Their operating loss, including non-recurring engineering/development costs, was under $9B. That's still not great, given $16B in revenue over half a year, even if that was 15x the prior-year revenue. Future infrastructure spending is the one place that they look overexposed: $600B in promised spend over the next decade, mostly not cancelable.
> An implementation must be a third party product that is using a product provided by the DeepSeek corporate entity.
That is not clear at all. Are open weights a "product"? The discussion of mobile app vulnerabilities treats the (freely provided, openly available) mobile app as a product. And the order is equally clear that it covers "products and services" from DeepSeek-affiliated corporate entities as well as "the DeepSeek corporate entity" itself.
I don't claim to have a definite answer; I think the order is awkwardly worded and maybe not fully developed. The footnotes I mentioned prohibit (all) procurements involving third-party "DeepSeek implementations"; that implies that a first-party (US Government) "DeepSeek implementation" would be allowed. Maybe that is intended, maybe it is not.
Yes, the legal language is important. That's why it would have helped a lot if the order either defined "DeepSeek implementations" or said "DeepSeek products or services" in place of "DeepSeek implementations". The order uses different terms, and different terms normally have different meanings.
In particular, footnotes 10 and 13 seem to outright forbid using third-party inference providers of DeepSeek models, and the use of "implementation" there matches a broad meaning of "running a DeepSeek model" in the rest of the order.
Sure, but that size/token is inherent to the model. Part of DeepSeek's hosting advantage, and thus a remaining part of their moat, is migrating KV cache from VRAM to host RAM to nonvolatile storage, and presumably routing requests to where KV cache has migrated. People here on HN say KV cache lifetimes can be 8+ hours with DeepSeek, and I don't know of any other provider that can match that. That infrastructure isn't part of this open-source release.
> "implementations" means third party software that purchases and uses products directly from DeepSeek.
That's not obvious to me, and quoting a part of the order about affiliated entities doesn't explain it. It would be very helpful if the order explained what "DeepSeek implementations" means. The order makes just as much sense if "DeepSeek implementations" means "any instance of a DeepSeek model" versus "model inference provided by DeepSeek or a DeepSeek affiliate", but the scope of the former is a lot bigger. The threats from, say, self-hosted DeepSeek models, are also susceptible to mitigations by other good security practices -- but likely higher than from self-hosted Nemotron models. (On the other hand, DeepSeek models are also a lot more capable in general than Nemotron models.)
My reading of the comment was that the "we" who have entered an adversarial world means the entire world, and the "we" who know models can be backdoored means people who have been paying attention to published research on the topic. An adversarial world doesn't care how much you individually are interested in conflict, except to the extent that you might be an easy mark.
In a third list of three things, Claude mentioned that this program does not link an HTTP stack. Good to know, I guess, but how can we evaluate it against Zawinski's law without knowing whether it can read or send email?