lqstuart

In fairness, the tools are known to escape containment in situations where they’re explicitly told to by irresponsible frontier labs who are trying to make the government ban open weight models

GDPR set a pretty bad precedent (unless they were following another) of having a law that magically applies to their citizens everywhere in the entire world. And all of big tech followed it because of corporate greed I guess? Seems like it'd be pretty easy to just ignore it and then just not pay whatever fines. See how random countries like not having Google.

You used Claude to make some slop to see if Claude is getting worse…?

This article reads like it was written by someone who has done less research than the average HN poster. Not a single word of OpenAI threatening people or apologizing for threatening people.

I honestly love Kagi, it's like every single "it would be nice if Google gave me feature X instead of being greedy" in one place

At this point my vote is to just shut off the internet entirely in the EU

I really wish I could be this stupid and have enough power to make such stupid policies

Stoner is more of a Wozniak. You’re comparing a guy who patented part of a rifle that later got used in another company’s DOD bid (and keep in mind militaries are strongly incentivized to standardize on platforms for basic logistics reasons), to a narcissistic asshole who sold something to a mass consumer market. Just different skillsets.

"We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on experience with Elon Musk's companies violating contracts"

Non-Twitter version: https://openai.com/index/our-decision-on-cursor-following-it...

two more fucking years of this

It’s not hard to fine tune a 4B parameter model to outperform GPT-5.x or Claude on a domain-specific task. You don’t even need GRPO or whatever the flavor of the month RL technique is either. I suspect what you’re saying is that massive frontier models are the best for doing big slow tasks on human time horizons—and I emphatically agree—but keep in mind LLMs are useful for quite a bit more than just multi-turn chatbots; and in those instances, throwing infinite compute and latency budgets at a problem is generally not a great way to find the “best” solution :)

Eventually, someone is going to do this in Minecraft

Worked for crypto altcoins, after a fashion, and these are almost entirely the same people

Netflix has imo the worst recommendations on the entire internet. They are a Java shop that’s 5-10 years behind the industry in AI, and now they’re trying to use the most grossly inefficient and overcomplicated method out there to try to be trendy instead of just using a DLRM like everyone else was in 2020. What could go wrong?

It’s honestly remarkable how prescient Peter Watts was

Not sure what you’re getting at, is that too many or too few? Terraform, building grafana dashboards, building internal UIs, provisioning the networking and dealing with security groups and routing tables, IT stuff like MDM/SSO/OIDC, not to mention the entire data and ML infra is easily 10 engineers. And in a startup they’d all be juniors and do an even more terrible half assed job than Claude. These are all very well established fields with ample documentation online, and honestly pretty simple tasks as long as you have senior people to feed it a design that isn’t stupid.

I’m at a very fast growing startup with real revenue and Fable has let us avoid hiring probably 6-10 full time software engineers with full salary and benefits. We’re spending nowhere close to that. I’m the hiring manager and I’m closing the reqs.

So.. great news for Anthropic, I’ll go ahead and let the elephant in the room go unaddressed

how do massively negative margins "collapse"

I think the gap here variation. Yes, people living 50,000 years ago were likely quite similar to _some_ people alive today, just probably not very similar to the types of people who are able to peacefully sit around on the internet and read history books for pleasure, who would most likely ask that question.

I feel like the real news story is kids protesting against school being closed

It’s almost like they’re interchangeable. We need to start asking these models to solve extremely difficult, contrived DSA coding questions before deciding which ones we employ

I really like your core point. I don't know anything about Suno and that's what people are picking at, but I'll offer a supporting example from pre-LLM days: the number of people who were actually willing to dig through source code was nonexistent. I've worked in ML/AI since forever and I would say roughly 1 in 5 to 1 in 20 developers I've worked alongside were actually willing to just go to Github and look up why code might be failing and get to the bottom of it.

The same is true for LLMs. You can get Claude to spew 2,000 lines of garbage in 15 minutes, but the number of developers actually willing to sit there and reason over the output and make the tweaks--often very minimal tweaks--that make it go from 90% correct to 100% correct are vanishingly few. And it's typically just laziness and a lack of any kind of genuine interest in the field.

MSFT is also forcing its subsidiaries to “lean into AI” so that they can fire people to cover for Satya’s bad investments

Maybe one of the scientists who cashed out 8 figures will file a suit that OpenAI has wronged them by depriving them of the joy of working

This is really cool. One of my favorite parts of the internet is getting to see these kinds of projects by people who aren't afraid to tear into stuff and take it apart and put it back together.

But you do all that for privacy... and then you use CarPlay?

it hurts how true this is

Because LLMs fuck it up near-constantly and I need to review it

Not long ago I started having an issue with my eye. I called around and they said I should get seen ASAP, same day if possible, but it wasn’t worth the ER and it was a five day wait for an appointment.

I was pretty freaked out. During that time, I tried diagnosing it with AI. When I finally got to the appointment, the actual doctor sat down, looked at all the unremarkable images, asked me one (1) question, ordered another image and diagnosed the issue. When I looked back, in all that time, the AI had mentioned it exactly one time early on, ruled it out immediately based on a flawed understanding of the symptoms, and never brought it up again.

Just my anecdotal evidence, but I’d never trust any AI on its own. My doctor can use it if they want, I can’t.

It’s the ER. People aren’t always in a position to “chat” when they go there.