quicklywilliam

The article is very dismissive of recent mathematical proofs. It mixes together ethical problems (plagiarism, "strip mining" proofs) with capability problems (proofs not being novel, overstating capabilities). Again, it's fair to be skeptical of what the companies themselves are saying (and why). But from what I've read, many mathematicians are taking AI quite seriously. I've recently read several thoughtful essays from credible folks who are calling for dramatic changes to the field in response. For example: https://www.daniellitt.com/blog/2026/9/13/a-beginning-for-ma...

The article is also dismissive of the Hugging Face incident and other hacking incidents. This one is trickier, because we are more reliant on what the companies have released to piece together what happened. The METR report is probably the closest thing we have right now. Here again, I think the issue is that a lot of coverage blends the ethical issues (holding AI companies responsible for what their bots do) with capability issues (whether AI is or will be capable of developing novel and dangerous attacks).

Lots of anti-AI takes seem to fall prey to the moralistic fallacy (ie because something shouldn’t be, thinking it isn’t). I’m very skeptical of the claims AI companies make, their motives, and whether their impact will be anything positive. But I am pretty tired of takes like this that are so reflexively skeptical of anything AI that they carelessly dismiss claims about the dangers and real impacts of AI that are happening.

(And to be clear: I’m not talking about being skeptical of fantastic claims about things that have not happened- ie an AI takeover, runaway super-intelligence, etc.)

Best tip in here: when tweaking details like colors and fonts, just ask Claude to make you throwaway picker tools right in the site.

This is totally obvious in retrospect but I never thought of it (probably because of 25 years of thinking about tools in terms of reusable components).

Thanks for writing!

Welp, here’s a way to make sure your content stays out of training sets

Better comparison IMHO is nuclear energy. Lots of potential for good, lots of problems. Managed by an insular, engineer-dominated industry who completely lost touch with and trust from the public. Public got the last word, industry suffered a 40 year ice age.

Ok, but do you really think this would have happened if OpenAI had a goal of never allowing an agent to do something like this? If a company making any extremely volatile explosives failed to take precautions and blew up a factory, shouldn’t we focus on their negligence?

Perhaps we might ALSO decide that this new explosive is so dangerous it needs specialized laws to protect the public, but that would seem premature if all available evidence indicated that the explosion would not have occurred if the company had taken the most basic precautions.

I see it as analogous to companies building fiber in the public ROW during the last big infrastructure bubble. Under the Telecoms Act, these companies had to allow competitors to use their fiber at a fair price.

Similarly, AI companies should be required to allow distillation at a fair price. Fair Use doesn’t make sense as a social contract if it only cuts one way!

It’s not different from any other tool. If you use a dangerous tool recklessly, you should be liable for the damages. That means holding OpenAI liable for HuggingFace hack because they ran the tests, and the same goes if someone did something similar with GLM.

Of course in cases of negligence a tool maker could also be held partially liable. That’s a matter courts can decide. The main point is we shouldn’t jump to making special laws around the development of LLMs. The starting place should be enforcement of existing liability laws. New laws take time and will be heavily influenced by AI companies seeking a regulatory moat for their business. Moreover, it is a distraction from the illicit behavior that is already going unchecked.

My takeaway: We should be not be concerned about AI bots' "goals", we should be concerned about the goals of the companies making them. Powerful but not sentient technology in the hands of reckless accelerationists is a plenty dangerous enough thing.

Just in time for midterms.

Want to be clear that this is not me endorsing the feature. It's clear Apple did think about the privacy/social implications of the design, but it's also clear to me that there are real and perceived problems with having automatic conversation notes. Real differences from written notes: - It was recorded and transcribed by a machine, not a human - There are time stamps and locations - You can't tell someone is using this feature (it is obvious when someone is writing notes during a conversation)

As far as perceptions goes, it feels very tone deaf of Apple to launch this feature in this way at this time. Read the fucking room guys.

Details the design and architecture of Apple Watch's new "Audio Intelligence" features. In particular, regarding Siri Recap which has been discussed elsewhere:

*How Siri Recap respects those around you* By design, Siri Recap does not create a recording, does not produce a verbatim transcript, and does not identify and attribute speakers. The output is a brief, high-level summary, comparable to notes a person might write after a conversation. There is no audible signal because no raw audio is retained, and there is no way to reconstruct the original audio from a Siri Recap or share raw audio with anyone.

Maybe this will be true legally (we wont really know until it is tested in court), but I think there's a huge difference in practice.

Siri Recap is an electronic record of a conversation that includes timestamps. Imagine for example that during a scheduled meeting with my boss, he threatens to fire me for not doing something illegal. I now have strong evidence that this was said (even if it cant be officially attributed to him) during the time I was in a meeting with him. Not saying this will hold up in court, but it certainly is more damning that having written notes (not to mention the fact that my boss probably wouldn't say this when I am taking written notes).

Agree with this framing, but we may find out that the answer is somewhere in the middle. Personally I think it is extremely unlikely that this is straight plagiarism in the sense of the model simply regurgitating training data from prior work by Tristan, but it is very plausible that his work (and others’) was foundational to the breakthrough. What is unfortunate is that the stakes are so high (and no I don’t mean $1M) and the timeline is so compressed.

I expect a lot more of this kind of drama in the near future.

The fact that this link is a Reddit thread should probably be taken as evidence that we don't have a clear picture of what happened yet, and speculation is rampant.

IMHO a blacklist would have been better than a whitelist

I think it’s important not to let AI companies control the framing here. This is anti-competitive behavior, plain and simple. The fact that xAI/Musk routinely break the law simply provides the fig leaf.

Distillation is a threat to frontier model companies, but that doesn’t make it dangerous or wrong. OpenAI can (and likely will) use the same logic to ban you from using open source harnesses.

We do

Tired.

This is what I did, tho I’m 41

The used EV market is a perfect storm of two factors. First, used EVs are finally reaching a price point where price-sensitive shoppers will consider them. Second, gas prices are making EVs look ever more attractive to those shoppers.

Even though this dynamic may not last, I think this is a really positive thing. Since their inception, EVs have been seen as a luxury item for tech elites. This is finally changing. More and more people start to see EVs as a good way to save money. This change will do far more for adoption than the tax credits we lost.

Long live the EV shit box!

I love this question. Most eink UIs feel like they just took interaction patterns from smartphones and crammed them in. I would love to see more innovation on eink UI patterns that works from first principles given the nature and limitations of eink!

For example, I consider dragging operations an anti-pattern on eink devices. Both latency and touch precision are just not at the levels we expect from years of using phones, leading to an experience that is prone to errors and frustration. A particularly common problem is messing up your state just as you lift your finger. I can't tell you the number of times I have carefully highlighted a passage only to have it jump to the wrong words as I release. Dragging is also just an inherently physically awkward maneuver, especially when performed prone.

I think a better pattern would be to use discrete mode changes with modal controls for refining state. As a simple example, highlighting could work like this: 1. Touch and briefly hold to enter selection mode. 2. Selection mode brings up large buttons around the current selection to refine it. Refinement happens via taps, not drags. 3. Selection mode is exited via another button tap. Confirmation is explicit, not implicit. This feels like it would be a lot more relaxing and less error prone.

I could also see a hybridized approach for interactions (like highlighting) where refining exclusively via button taps might be tedious. In this case you could use drag gestures while in selection mode to quickly add/remove large sections to/from the current selection state. This would give you a shortcut for highlighting large sections quickly, which could then be detail-refined via button taps. Having it be modal would also allow you to undo any accidents.

A further refinement for this interaction pattern would be to offer intelligent controls that give the user context-inferred options for their refinement. For example, if you have part of a paragraph highlighted, a button could appear to highlight the rest of it, and subsequently buttons could appear to highlight the adjacent paragraphs. This could even be combined with the hybrid drag gestures idea - after quickly dragging over a rough area, the system could offer multiple proposed selections if there is ambiguity what the user may have intended.

I think the big idea here is that you can get a lot more performance if you take an integrated approach. This specific model made to work with this specific inference engine made to work with this specific harness/agent. When everything is done separately, developers of a given pieces have no idea what they are targeting for all the other pieces.

This is currently a huge advantage that Anthropic has over open weights models – they control the whole stack. Indeed, they train new models against Claude Code!

It's early days on this project, but just imagine it gets enough traction that future models start training against ds4. Indeed, in the post Antirez even seems to be hinting at some sort of collaboration with DeepSeek?

Picture the early days of the PC - local models are like that. While there are scattered groups of hobbyists who tinker endlessly to get a few lights blinking, almost everyone doing “real work” is using a large, expensive, centralized product. Despite the exponential performance curve of smaller models (not to mention the rapidly increasing cost of using frontier models), we are all still stuck using mainframes and minicomputers. Why? Because few of us have the time and resources to build a computer from raw components.

The problem with local models isn't that they are local, it's that they are not integrated. Hobbyist testing the latest local models are great (thank you!), but we also need someone building the Apple I.

Same question. Given the people involved I am inclined to take it seriously, given the money and hype involved I am somewhat less inclined.

Their website doesn’t even have a hint of what the approach is.

Agreed, there is probably a theoretical world where we got enough money/compute together and had this explosion happen earlier.

Or perhaps a world where it happened later. I think a big part of what enabled the AI boom was the concentration of money and compute around the crypto boom.

This is a really cool list and repository of ideas. Seems like the focus of the work is on making knowledge legible to AI. I wonder if you (or others) have done a similar level thinking about the inverse – making AI more legible to humans?

Curious if you looked at using SwiftDown (https://github.com/qeude/SwiftDown), MarkupEditor (https://github.com/stevengharris/MarkupEditor) or any other libraries for live/WYSIWYG Markdown editing.

There’s been a lot of discussion lately about Anthropic and others turning the screws on their subscription plans, skyrocketing costs for enterprise customers. This is driving more and more folks to consider cheaper and non-proprietary models (which are getting more capable) for some of their tasks. This article goes into both of these trends.