yorwba

14 ms·

331.
by yorwba 5mo ago
Yes, if your LLM sandbox had a huge hole in it guarded only by asking an LLM whether the stuff coming out is low-risk, you would indeed get sand into all kinds of inconvenient places. So don't do that. If you want to sandbox an LLM, al
332.
by yorwba 5mo ago
You could certainly bolt GRAM onto an LLM, but that won't magically improve its reasoning. It's a special-purpose design for constraint-satisfaction problems with simple rules, but complex interactions. E.g. when solving a Sudoku,
333.
by yorwba 5mo ago
The sales tax is the value-added tax.
334.
by yorwba 5mo ago
After going without food for a few days, even a software dev is going to have to swallow their pride and consider alternative employment options. You might think they could just take a poor person's existing job, but why would for exam
335.
by yorwba 5mo ago
Poor people with hard-to-automate occupations don't necessarily need to be able to afford token costs for social role reversal to occur. They only need to be able to hire an even poorer person who used to earn a salary that exceeded ev
336.
by yorwba 5mo ago
Yes, widespread automation of knowledge work is unlikely to decrease total production, so all the goods and services people currently demand will still be provided, but the power dynamic of who is consuming and who is providing might flip a
337.
by yorwba 5mo ago
Python's import system is extensible: https://docs.python.org/3/reference/import.html#import-hooks It might be possible to create a custom finder that will return 1.0.0 when running "import dependency_c&
338.
by yorwba 5mo ago
You can use the verification tool to check whether the watermark is still detectable and iterate until you successfully removed it . Of course if you need to regenerate the image with an unwatermarked image-generation model to remove it, it
339.
by yorwba 5mo ago
They're offering a model based on Kimi K2.5 for $0.50/M input and $2.50/M output while the cheapest third-party provider on OpenRouter charges $0.40/M input and $1.90/M output https://openrouter.ai/m
340.
Asymmetric Flow Models(arxiv.org)
3 points by yorwba 5mo ago|0 comments
341.
by yorwba 5mo ago
It is? Below the English translation.
342.
by yorwba 5mo ago
> But how do you encode stand-alone assertions like “all unitary matrices are invertible”? You could encode that as a yes-or-no question, but that’s useless, because rationally you can expect such questions to be biased towards yes. If y
343.
by yorwba 6mo ago
The author is a Greek-speaking parent teaching his Greek-speaking children to read by visually pairing each letter with a Greek word that starts with that letter. If you tried to teach English-speaking children with words that start with th
344.
by yorwba 6mo ago
A full diffusion transformer would need more forward passes (thus being slower) or produce worse output (because it can't properly account for dependencies between tokens when generating them independently in parallel), or both. Keepin
345.
by yorwba 6mo ago
The METR task set contains no tasks with a duration greater than 32 hours (conservatively eyeballed from Figure 3: https://arxiv.org/abs/2503.17354 ), so any prediction that naively forecasts a longer time horizon is t
346.
by yorwba 6mo ago
The "law" part of scaling laws is about predicting validation cross-entropy loss from the training configuration, analogous to physical laws allowing to predict one quantity based on the measurement of another. Most scaling laws t
347.
by yorwba 6mo ago
Thanks, I got bit by the counterintuitive semantics of "current international $" meaning "current at each point in the graph" instead of "current at the time the graph was made."
348.
by yorwba 6mo ago
Finnish GDP per capita is growing much, much faster than Malawi's: https://data.worldbank.org/indicator/NY.GDP.PCAP.PP.CD?locat...
349.
by yorwba 6mo ago
I expect every company to try to make a model as good as they possibly can, especially now that Mythos has served as a proof of concept to demonstrate that there's lots of interest in AI for cybersecurity. But if they don't try, t
350.
by yorwba 6mo ago
All of the reasons in the article also apply to Chinese companies. If a Chinese model becomes good enough to make it significantly easier to hack Chinese government servers, do you think they'll allow random people unfettered access to
351.
by yorwba 6mo ago
Don't forget that there was an economy before electrification was even a thing. I doubt a Kenyan coffee farmer is using a lot of electricity to grow coffee. Meanwhile a lot of electricity use in richer countries is for household electr
352.
by yorwba 6mo ago
Yeah, and things that can't be tried out are explicitly excluded by the Show HN rules: https://news.ycombinator.com/showhn.html OP should come back once there's an actual product, assuming it ever gets past the em
353.
by yorwba 6mo ago
Linguists had been intensively studying Ubykh for decades by the time the last speaker died. It's rather well-documented for a language with few speakers.
354.
by yorwba 6mo ago
It's an interesting constrained-writing exercise, but I think it demonstrates that the focus on syllable count is too simplistic as a heuristic for whether children are likely to understand a word (which in any case is context-dependen
355.
by yorwba 6mo ago
> In case people no longer remember, when China started to require websites to register for a license before be allowed to operate, it was for "protecting the children" too. Indeed I do not remember this, nor can I find corrobo
356.
by yorwba 6mo ago
The verbalizer and reconstruction models are both initially finetuned on LLM output from a summarization prompt. The resulting text is not completely unrelated, but mostly wrong: https://transformer-circuits.pub/2026/nl
357.
by yorwba 6mo ago
You should benchmark the retrieval speed of each method in terms of queries per second. I suspect that the gain in bandwidth you get from slightly better compression will be defeated by decompression being much more expensive.
358.
by yorwba 6mo ago
As the Hacks.Mozilla article notes: "We began with small-scale experiments prompting the harness to look for sandbox escapes with Claude Opus 4.6. Even with this model, we identified an impressive amount of previously-unknown vulnerabi
359.
by yorwba 6mo ago
Happens all the time: https://ustr.gov/search?q=initiates+section+301+investigatio...
360.
by yorwba 6mo ago
Announcement blogpost: https://www.zyphra.com/post/zaya1-8b

More ›