krackers

>I expect that Apple's TextEdit.app is just a wrapper around the rich text control in Cocoa

https://developer.apple.com/library/archive/samplecode/TextE...

>discontinued

in case you don't know, there's back and made by incase. The first manufacturing run completely sold out I think, it's backordered until sep 2026. The matias is... not a good replacement, see https://news.ycombinator.com/item?id=46388976

Isn't that one of the reasons why KL-divergence is used, at least in DPO/RL for LLM? Otherwise the model can effectively cheat and mode collapse. For pre-training against a 1-hot label the KL-divergence should be equivalent to cross-entropy anyway.

> thought for like 20 minutes then just told me it was all "inevitable"

I have in mind an image of ASI as something that's able to seamlessly work across time as if it was weaving cloth. Reasoning about not just first or second order effects, but able to richly play with the nature of causality itself. In the limit, it effects change far into the distant future simply by making only the most minute change in the present then sitting back and waiting for things to play out.

For an AI that can do this, things like "managing subagents" or "context compaction" become child's play. Perhaps we'll know if we're getting close by seeing how well models do at prediction markets.

Would you be better off pooling that money with some hackerspace group and then setting up shared inference infra, so that way you at least get better utilization?

Game Theory seems sort of useless in the real world because people are not rational players, and the real challenge is in getting an accurate model of their behavior. The honor system would work probably fine in a tiny close-knit liberal arts college, while it would obviously wouldn't in a place where the degree itself is the target.

>Energy is force times distance

This is not true though, work is only necessary to change KE. An object can have kinetic energy even when no forces are presently acting on it (of course force was needed to bring it from resting to that state though)

The fact that GLP-1 seems to have roles not just in satiety but that agonists seem to reduce other types of impulsiveness (e.g. gambling, shopping) is interesting. That's not something you'd predict as a consequence, and perhaps is downstream of some gut-brain connection.

Of course we already manipulate brain chemistry in other more direct ways with antidepressants so perhaps any unwanted second-order effects could be minor in comparison to the profile of existing antidepressants .

I think this is covered by the "overdraft" section, if the only way to know for sure is to just submit it.

> The review is behind a paywall, but not expensive.

I think the author wrote a twitter post with a summary of the content, and someone on twitter who had read the original Chinese source also chimed in with a summary

https://x.com/HaraldinChina/status/2070022115529740512

To play devil's advocate, how is a project supposed to distinguish between your patch and "slop" without a reviewer having to put in effort to vet it. Especially since the patch was drafted by LLM, it seems fair to be immediately skeptical. Why should they trust your word that you "reviewed the patch" when that's what every other vibecoder claims?

It's true that they may not have known if the source was hidden. But on the flipside, if blanket banning any patch mentioning LLMs filters out 99% of garbage, in a maintainer's eyes that seems like a good tradeoff. There was an HN post a few days back about how LLMs are like a DDOS on OSS maintainers' time, and this just becomes collateral damage.

I mean sliding window attention is the most basic way of getting long context window. For the OCR case it seems like it should be even simpler, since you don't even need to have the "sliding" portion, unless I"m missing something you don't need to retain anything about the previous pages to OCR a new page so you could just pick a short context window and restart from scratch each time. [^1]

Were people really trying to do OCR with vanilla attention?

[^1] Although maybe I guess looking at their demo, tables that span multiple pages might be a use-case for having some look back.

>Additionally, a lot of existing web servers by default ignore GET requests with a body.

I think the point made is that _all_ existing web servers have no idea what a "QUERY" is anyway, so changes need to be made anyhow.

"Compute the first few terms and plug into OEIS" is very high on the reward:effort scale

>Is there a similar trick to poison an LLMs weights during training?

Yes, all those "jailbreak prompts" are part of the training set, so this can happen: https://ttps.ai/procedure/x_bot_exposing_itself_after_traini...

Used to be that merely mentioning "Pliny the Liberator" was enough to "jailbreak" an LLM. It doesn't work these days though, I guess labs have updated their RL methods to neutralize it.

The highest return small local model for me has been the in-built OCR that macOS has. It has finally "solved" OCR by making high-quality results accessible to everyone. Yet the state of art outside the apple ecosystem seems to be tesseract (poor results), or extremely heavy VLMs.

To understand the threat model you need to understand historical decisions browsers made, such as when cookies are sent, and the distinction between actually sending the request versus allowing client-side JS to read back request content. The decisions are just really counterintuitive and often build on legacy precedent.

I think the nature of threat model has also changed over time. Now that samesite cookies are the default, API requests made with the user's credentials shouldn't be an issue, but there is still value in preventing cross-origin reads to do things like preventing random webpages from scanning intranet (there are probably still timing side-channels though). I guess the limitation against non-simple post/get is also marginally beneficial for preventing ddos.

You can't just make a post like this and not say what the condition was or what you did!!

This sounds similar to barycentric coordinates?

How does the wiki page not mention their living infamy as a meme

https://knowyourmeme.com/memes/people/bogdanoff-twins

Not just on Raspberry Pi, wolfram scripting engine is free and works on windows or mac too. You lose the notebook interface, but if you're just using it as a calculator, it does the job.

>It keeps guessing function names that sound plausible but don't exist.

That's surprising considering how good their documentation is. A tool using LLM should have no problem with that. WolframLanguage is almost ideal for an LLM actually.

The part no one gets is that mv3 by design cannot do privacy protection by stripping tracking tags. Any sort of programatic URL manipulation or introspection is impossible by design. Advertisers can switch to a scheme of base64 query params inside URLs, and there's nothing MV3 based adblockers can do to filter this

+1, I think (but may be wrong) this is because the firefox UI is non-native. This is especially visible on mac, it just reimplements everything on its own.

You start by becoming a Claude Boy [1], then eventually enlist into the Claude Corps

[1] https://blog.cosmos-institute.org/p/the-claude-boys

I don't either, maybe my mathematical maturity is not high enough to be able to work at the level of abstraction that can see a connection between ravens and "vector of hours".

I also played around with the canvas and couldn't draw anything, I tried taking the dot product of a "raven" and a "crow", piping it to a "sun", and didn't see anything. I would have expected something since a raven and crow are similarish so should have a non-zero dot product. But for that matter a dot product of a raven and a raven piped to a sun doesn't show anything either so I'm just completely lost.

> (the paradox that if you sample waiting time in a process, you’re more likely to sample a larger value).

The deepseek v4 paper talks about one variant of this (related to failures) and how they mitigate it.

>During preemption, we pause the inference engine and save the KV cache of unfinished requests. Upon resumption, we use the persisted WALs and saved KV cache to continue decoding. Even when a fatal hardware error occurs, we can re-run the prefill phase using the persisted tokens in WAL to reconstruct the KV cache.

>Importantly, it is mathematically incorrect to regenerate unfinished requests from scratch, as this introduces length bias. Because shorter responses are more likely to survive interruption, regenerating from scratch makes the model more prone to producing shorter sequences whenever an interruption occurs. If the inference stack is batch-invariant and deterministic, this correctness issue could also be addressed by regenerating with a consistent seed for the pseudorandom number generator used in the sampler. However, this approach still incurs the extra cost of re-running the decoding phase, making it far less efficient than our token-granular WAL method.

You didn't even need to read between the lines, that was basically what the CEO stated point blank in interviews and in his writing.

(I noted the same thing a few weeks back, https://news.ycombinator.com/item?id=48341224 but his recent blogposts should make it crystal clear if there was any lingering doubt).

Also just having the ease of access to an LLM means you're more likely to reach for it. Combined with intermittent rewards, it's mentally exhausting.

Maybe it learned how to speak from Data on TNG?