jsnell

I don't know that I got anything out of this other than that it took a long time. I think there is an interesting article to be written on this. Like, what were the concrete requirements? Why were those the requirements? Maybe show how one letter evolved over time as the project went on, and explain why.

But instead the article is effectively the Gravitational Pull of Pepsi[0].

[0] https://www.goldennumber.net/wp-content/uploads/pepsi-arnell...

That is not actually a typical playbook. Normal businesses do not want to be regulated, and will fight it tooth and nail. So regulation happens when bad things happen, the public wants change, and the businesses can't fight it or set favorable terms. (At most they can minimize the badness of the terms). Regulatory capture happens much later, and slowly. It's not where you start at.

FT clearly didn't do their due diligence, since the auxiliary information provided with the $70B leak (about e.g. growth rates and enterprise sales) made it clear that the number could not possibly be correct.

It just wasn't clear exactly what the error was (e.g. was a projection of a $70B ARR by end of year being misinterpreted as $70B ARR now -- that would have been stupid, but less stupid than the "investors added a fudge factor to the numbers" story that they're now going with).

If that were the case, the hyperscalers would be ramping down capex, not increasing it.

The not-in-service numbers on the financials are just the normal lag, and have been stable in proportion to capex for years, to well before the start of the boom.

We do in fact know:

> The chips in the proposed deal were bought or leased by Amazon. They are installed in more than a dozen US data centres across five states, including Nevada and Virginia, the report said.

The "warehouses full of GPUs" thing might be the most absurd of the AI economy conspiracy theories.

Your analogy make zero sense. Atom bomb manufacturers do not market their products.

Big businesses really do not like to talk about their products being harmful or dangerous, and there is ample evidence to that. Danger only sells in artisanal quantities to niche audiences.

Yes, famously tobacco companies wanted to advertise their products as addictive and deadly. That's why they funded and heavily publicized research to that effect.

Climate change is just a scam by the fossil fuel industry to hype up their market cap. Look at the power of co2, and the damage it can do without even trying.

Medicines being pulled during trials or after public availability for side effects are just Big Pharma making their products more desirable via scarcity.

Boeing did really well out of the 737 max. Airlines were just lining up to buy an airplane that could give the passengers an experience they would remember for the rest of their life. That's why they absolutely made clear that it was their product that was dangerous, rather than blaming operator error.

No. They tasked them with answering questions by retrieving data from public sources. This obviously requires internet access. So the incredulity about "how could they fail to block internet access" is just inane. That was the task.

They did not task the agents with hacking into systems. Not monitoring for that was a mistake, but maybe a forgivable one the first time around. The subsequent coverups are inexcusable.

These were evaluations or training runs dealing with the ability to do web searches. The entire reason the bots were targeting this site is that it had some of the data they'd been tasked with finding. Access to the internet is not really optional for that, and providing internet access doesn't demonstrate neglicence.

> This is why it smells like marketing

It doesn't. "Our product commits felonies" is not marketing. If something is marketing, you don't engage in repeated coverups of the true extent. If something is good news, you don't release it on a Friday evening (in this case) or wait for 3rd parties to find and publicize the evidence (the past cases).

The impact from a model not being included in the training set is increased false negatives for text generated by that model, not increased false positives for text written by humans.

The latter are not doomers at all, that's just a misuse of the term.

A year ago I would have described that group as AI skeptics or AI deniers depending on how charitable I felt on any given sany. At this point AI truthers seems more appropriate.

You should probably look at the cost/score graph by effort level instead:

https://artificialanalysis.ai/models/claude-opus-5-5#intelli...

It is most of the pareto frontier.

> Half of the 832K LoC are mostly tests she answered:

No? That quote is clearly saying the opposite of your summary.

Flagged.

This is like a web page asking for your email password to check password strength, promising that it won't leave your browser. Maybe it doesn't right now, but who knows about tomorrow.

Discussed two days ago: https://news.ycombinator.com/item?id=49672281

A keyboard/video/mouse switch.

But the proposal the article is reacting to is for literally none of that! It is quite literally the opposite, with its proposed measures applying only to frontier labs rather than grandfathering them. It doesn't say anything about open source training, self-hosting, open weights, or startups. It does not suggest a ban on Chinese models (just better enforcement of chip export controls).

> Any AI model a company offers to the public has to be released as open weights.

The obvious outcome of this plan would be frontier models not being released at all to public It'd be an amazingly bad outcome. Models accessible to the public would stagnate (no capital, no access to frontier model tokens to distill from), while internal models would keep improving at their previous pace, use them in-house, and eventually eat the whole economy.

I honestly don't even understand what straw man Doctorow is arguing against here.

But he is wrong on the facts: these incidents were not merely the models already being in a infosec context and escalating beyond the intended parameters. They happened also with no kind of security elicitation. So the task was something like searching the internet for economic statistics, not to hack into a system.

I can't believe we're finding out about this from 3p researchers again (but nice job on the investigation!). OpenAI had two great opportunities to disclose this. The HF incident report, and in response to the German Wiki issue.

It seems impossible to believe they didn't know. This must be the same training run the HF incident was about, and this should have lit up like a Christmas tree in the investigation. How many more incidents do they know about and didn't disclose?

A customer willing to make a multi-decade deal for half a reactor at a stable price is perfect for incentivizing new nuclear buildouts. The basic financial problem of nuclear is that it' capex heavy and opex-light, so the costs of the entire enterprise are set during the buildout stage, but the revenue is volatile.

> Claude wrote it in my voice

That is not your voice. That is the most Opus 5 writing imaginable.

But thanks for at least stating right up front that it was AI-generated slop rather than at the end. (Still flagging it.)

The ability to appoint a board member doesn't mean you get to control their every decision. Nor is their fiduciary duty just one shareholder.

Of course usually hand-picked boards don't go against their benefactor, and they're not keeping their seats.

That they did in this case says something about just how bad the situation must be.

Native americans had like a 95% population mortality rate from infectious diseases in the first generation of European settlement, no? Measles, smallpox, flu, etc.

The mortality rate from those diseases was higher than today because the population had no immunity at all, leading to both higher transmission rates and higher mortality rates. This would be true of an engineered virus as well.

And it wasn't just the one virus, it was an entire continent's infectious disease arsenal landing in overlapping waves. Again, something you'd do if you were trying to maximize damage.

So I don't find the bulk of the article about it being impossible for a virus to have this effect all that compelling.

4) There are non-monetary limits on how many qualified people OpenAI can hire for these roles.

What do you mean by "most writing", how are you scoping it? Most HN comments aren't LLM prose. Nor are most HN frontpage submissions. But by reputation, most substack articles or or linkedin posts are.

This is a case where we normalisation of deviance has not yet started biting. And as long as the community manages to make it clear what the norms are and enforce them, we can keep it that way.

Now, if 25% of the frontpage was LLM prose at all times, the site is probably unrecoverably dead. Which is why at least I personally flag anything that I think is ai-written and Pangram concurs. (And write a comment to the effect, or upvote an existing one.)

And it doesn't matter if you say that the ideas were your own, and just the prose was LLM. We can't tell what the idea mix was. But we can tell whether you weren't willing to do your own writing. If you want people to put in the time to read your ideas, human writing is the signalling you need pay for.

The apparent advantage is exaggerated by them running Astra at six different effort levels, and almost everything else at just the maximum available effort.

I don't really understand why they keep doing this. Either run and report everyone at multiple effort levels, or run everyone at only one.

But alsi, token efficiency seems pretty artificial? For example tokenizers are different from model to model. The cost/perf Pareto frontier seems a lot more meaningful (and Astra does very well at that too, just to be clear. It seems to be a great model.)

The "it's all just marketing" conspiracy theory is always totally detached from reality, but particularly so in this case. Your quote shows OpenAI is denying it being a hacking attempt, the opposite of what you say.

> In the same way that YouTube videos have a 0% chance of phishing, cross-site scripting, malware, and viruses.

So, the same as in not actually 0%? Even though the video streams themselves were never used for attacks[0], the contents of the videos frequently have been. The closest analogue to phishing would be the fake "SpaceX launch" streams that were actually a carrier for a crypto scam payload. Or for malware, it doesn't matter that the video itself wasn't malware as long as a video showing a fake Fortnite hack could get the viewer to install malware thinking they got the hack.

AI agents need to be substantially better than humans at this, but they don't need to be perfect.

[0] As far as I know! It seems very hard to smuggle a vuln past the transcoding, but I guess technically possible.

There are no ten year old H100s. The first production shipments happened exactly four years ago.

I'm pretty sure your claim about TPUs is similarly exaggerated, only a v1 (barely) qualifies and would have no utility today.