wzdd

That's an impressively tight TOCTOU exploit!

These aren’t tools which play chess. They are language models which roleplay a conversation (in this case including use of tools) which an evaluator is likely to mark as good. That’s all they do. Under that lens, playing chess is just one potential side effect and alignment, which requires a much fuller understanding of what’s going on than “do the sort of thing which evaluated well during training” is a fantasy. People are acting like it’s shocking and talking about cheating and so on. But these concepts exist at a way higher level than what these things are trained to do — the vast majority of which involve producing a transcript where it wins games, its code works, etc.

User wants me to play a game of chess. Let’s see what’s available so I can produce an outcome they will consider satisfying and be pleased that they requested my assistance.

Agreed. I was confused by the Trio example until I reverse engineered what they meant from the outputs. Trio behaves differently (in well-defined, easy to understand ways) depending on where you put nurseries.

This issue is older than the M chips: a formative OpenCL moment on my Intel MacBook involved locking it up in this way. By contrast, a similar CUDA kernel running on Linux was forcibly stopped after some short-ish time.

Worse and less defensible on the web of course.

These rethinking-highlighting posts always start out with the wrong premise. The color for class names in my code is the color all the class names have. After a short time working with the code I would notice if one looked off. The brain is really very efficient. You don’t optimise for random out of context recall unless you really try hard to do so. It says nothing about the efficacy of the scheme.

Or, you know, using the totally-real-picture camera to take a photo of an AI-generated scene?

AGI is here and its first task is obviously completing the C64’s game library.

Not really. AGI is a game interpreter, like LucasArts’ SCUMM or Infocom’s z machine. Games are shipped as portable code and a machine specific AGI interpreter runs them. Looks like they’re preprocessing the graphics first, though. https://en.wikipedia.org/wiki/Adventure_Game_Interpreter

Wilson didn’t write “sex scene”, she wrote “sex”. In the story there is plenty of implied sex. In the movie there is none. It’s not a major issue in itself, it’s part of a broader point about movie Odysseus being “uncomplicated” versus the story— both morally, and when it comes to passions (sex, food, revenge).

It does, but in the same way that LLMs take two paragraphs to make a one-sentence point: by being weirdly specific and also redundant. LLMs don’t have a sense of whimsy in the face of RLHF, so if it’s proposing a solution, the solution is tractable. All implementation plans start with a beachhead.

“A good starting point” is almost certainly all it needed to convey and what it should have written.

> They don't "blow bubbles" to sleep vertically, they vent gas to regulate their specific gravity. This allows them to sleep vertically

When I read this I thought you'd discovered that the whales release the gas in a form other than by blowing bubbles, or that the purpose of doing so was for something other than improved sleep.

But no, the whales literally release air through their blowhole in the form of bubbles, and, at least during study you linked, they do this while resting to maintain their vertical orientation. An orientation, the paper theorises, that enables better sleep by not being buffeted by waves.

They quite literally blow bubbles to achieve restful, vertical sleep.

My reaction was negative when I read the title (policy!), but the actual outcome was very reasonable. "Hey, can we add $ORIGIN to the VFS layer to support relocatable interpreters?" "You can already do that with binfmt_misc and ebpf, here's an example."

Then, later, a couple of patches to make binfmt_misc more ergonomic. Seems like a good outcome.

The article discusses a technique by which the author achieves high accuracy at detecting AI written text. Unless you have a problem with their experimental method, this is the opposite of tarot card reading.

> we are well into undetectable sophistication with today's models

The article directly contradicts this, as do you, in your previous paragraph: "Sure you might be able to detect today's tells". The article is literally about a technique that detects today's tells.

Your comment is mostly expressing doubt that this technique will work reliably in the future, but it's framed as opposition to the article, which it's not: the article is about detecting today's AI-written text, at which it seems to be quite successful.

Come on, that video is on another level. "All my friends are turning green" (shot of $1 bills), "she's been living on the highest shelf" (woman standing on Juliet balcony), "they come unstuck" (someone pulling twin pole popsicle apart)

For contrast these have "ice cold" (shot of ice cubes), "got chucks on" (shot of shoes), "livin' it up in the city" (shot of city).

The OP notes that they had to use special phrasing to get their exfil to work, so clearly GitHub was aware of the issue and made an attempt to prevent it.

It seems like the proper fix is for GitHub not to allow their agentic workflow to execute in a public repo context if it also has private repo access. Or, to use your phrasing, for GitHub to flag and disallow this easily-detectable and dangerous type of misconfiguration.

Given the Mandelbrot program is 450k probably not.

"Factual mistake" is a bit harsh, but the missing piece is that there are multiple ways to romanise Japanese; all of them produce "valid Romaji" but only in the particular system being used. Si is how you write し in romaji using the kunrei-shiki romanisation. In the Hepburn romanisation it's shi. See https://en.wikipedia.org/wiki/Romanization_of_Japanese#Diffe... .

Oodi is at least equally community / maker space and library, very distinctively and attractively designed, quite new, and in the middle of Helsinki, so there are a lot of non Finnish speakers visiting so there is a large English section.

It’s a nice engineering approach, but I’m interested in the motivation. Um and ah is distracting in a transcript, where you can naturally pause to take in information; in speech however it can serve as a focusing point to indicate the next part is important. See https://medium.com/better-humans/dont-worry-about-saying-um-... for example. The weirdly obsessive zeal that orgs like Toastmasters have about eliminating them is weird.

Disfluencies aren’t necessarily bad even if the word starts with “dis”!

> In your analysis you have shown that exactly one of these reasons is anecdotal.

This was actually the convincing one for me though. “Did AI increase the rsync bug rate? Dunno, can’t tell yet” seems like a fine conclusion to me. Plenty of people in this thread and previous ones on the topic seem convinced one way or another, so it’s nice to see actual numbers.

It is certainly unkind, when a developer asserts the opposite of what you have assumed about their code, to double down and imply they are lying.

The code is part of a function called expand item list. It looks like it over allocates memory and uses a bump pointer for internal allocation, only expanding the allocation when necessary. Thus OOB writes to the list would hit the allocated memory.

You’re not a bozo but it is helpful to read the code.

I have one of these and run Debian 13 on it. I love it. Having only two USB ports is annoying and I ended up buying a relatively expensive PD Thunderbolt hub, and there are some compromises that come with the territory (middling battery life, trackpad certainly isn't Macbook-quality). In general, though, it's great and it feels fun in way that I haven't felt about laptops in a long while.

As others have noted the company has done some pretty shady things with some of their other products, and I would not really expect a warranty, so this isn't really a recommendation. But my personal experience after ~six months of use has been good.

> Every time someone imagines a country going after Microsoft

You don't need to imagine it: the comment you are replying to links to a press release from a Government agency "going after Microsoft". And yet somehow we haven't seen Microsoft stop doing business with the Australian government.

It really depends. If you're cranking out prototypes or testing ideas, it's genuinely great. But if you're familiar with the code it's very easy to spot its (many) mistakes. It's Gell-Mann amnesia.

Then again, I just caught Claude writing setTransparent(!opaque == false), opaque being a bool, on a purely vibecoded project. Which was pretty impressive. ("• You're right, that's nonsense.")

> Cream in his coffee, butter in his soups, all the gristle he could handle.

I've noticed Claude is specifically fond of three-element lists consisting of (standard example of class), (standard example of class), (nonstandard example of class). It's interesting because Claude often gets the nonstandard example more or less incorrect -- in this case, because gristle isn't an example of a high-fat-content food.

People said the same thing about a joke Douglas Adams made in his Hitch Hikers series -- that the (corrupted) Ultimate Question to which the answer was 42 ("what do you get if you multiply six by nine?") was a maths joke because 6x9=42 in base 13. Douglas Adams said this was nonsense.

Quotes from unhappy staff are in the Business Insider article which the article links to in its third paragraph.

Zip disks were much less floppy than floppies. They felt more similar to a single magnetic hard disk platen. Presumably the stiffness was what enabled the head to float above the medium, while also allowing tighter read/write timing because it wasn't subject to such variation. Having a single manufacturer of the disks (at least initially) probably also helped.

Approximately 10 minutes is enough time to see the benefits but not the disadvantages. These effects might disappear again after multi-week interactions with AI, after it has led them down the wrong path, wasting hours; gaslit them into believing an incorrect solution is correct; or repeatedly misunderstood the problem.

This initial honeymoon phase happens whenever someone starts using a new tool, fades after weeks or months, and and is not the time to evaluate the deleterious effects of the tool.

They carry information in regular human communication, so I'm genuinely curious why you'd think they would not when an LLM outputs them as part of the process of responding to a message.