spidersouris

You don't have to be in school to learn!

Would also be interested in an invite, if anyone has any. Email on my user page.

Note that a similar idea had already been suggested by Shen et al. (2025) in Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism (https://arxiv.org/abs/2506.01979), but with lower performance.

Gen AI reached 39% adoption in two years

Source?

I'm not sure what you'd call a "pioneering scientific advancement", but there is an increasing amount of examples showing that LLMs can be used for research (with agents, particularly). A survey about this was published a few months ago: https://aclanthology.org/2025.emnlp-main.895.pdf

If you want to bind Tab to Accept suggestions:

Set-PSReadLineKeyHandler -Chord "Tab" -Function AcceptSuggestion

And plug.dj!

What we also learned after GPT-3.5 is that, to circumvent the need for new training data, we could simply resort to existing LLMs to generate new, synthetic data. I would not be surprised if the em dash is the product of synthetically generated data (perhaps forced to be present in this data) used for the training of newer models.

For type only though.

FWIW, there is an open-source collaborative editor for Typst that was posted a couple of weeks ago on HN: https://news.ycombinator.com/item?id=45481791

Maybe you can use Touying Exporter: https://github.com/touying-typ/touying-exporter

To add on what's been said already on slide decks, another great slide creation package in Typst is touying[1]. I've used it to create my own academic theme[2] for courses or conference presentations.

[1] https://touying-typ.github.io/ [2] https://typst.app/universe/package/touying-unistra-pristine/

The project was recently granted new funding so the research can get to market and benefit patients.

How is it now? Has this been extended to real use outside of research?

There is a repo: https://github.com/google-deepmind/synthid-text

There is no inherent need for humans to be "trained". Children can solve problems on their own given a comprehensible context (e.g., puzzles). Knowledge does not necessarily come from direct training by other humans, but can also be obtained through contextual cues and general world knowledge.

For the AI to say this or to produce the correct answer would be easily achievable with post-training. That's what was done for the strawberry problem. But it's just telling the model what to reply/what tools to use in that exact situation. There's nothing about "self-awareness".

What's the issue with "… disproportionately affects …"? It seems to be a correct English construction (even though the frequency in COCA is relatively low; 72).

Some examples:

Indeed, the recent cases of hyperinflation in Brazil, Argentina, and Poland illustrate that although hyperinflation is harmful to savers and disproportionately affects the poor (The Independent Review)

A hearing is set Thursday on the new version of a legislative bill to eliminate scheduled pay increases for state employees that nixes a section that disproportionately affects rural legislative information offices (USA Today)

Suicide is a key mental health issue which disproportionately affects men. (london.gov.uk)

FYI, the only English article at the time of posting was Entrevue's, which is why it was initially chosen. But indeed, Le Monde's article is much better.

I feel like this is clickbait. There is no way to check for now.

edit: I mean, there's always the good old https://haveibeenpwned.com/, but there is no guarantee the leaked data is already in there.

me getting challenged helps as well (I ask for it)

Can you elaborate? Do you have any examples of such interactions with LLMs?

For *ACL you'd have to justify your wish to change reviewers, though; and you need a good reason for that. I don't know how much reviewers changes for a resubmission are solely due to reviewers' unavailability but it seems unlikely all three of them got removed from the reviewer pool.

A quick Google search for "L1 attrition file:pdf" or "first language attrition file:pdf" returns tons of results, so it doesn't seem to be that understudied. I think it mostly depends on what you want to focus on: do you want to know how a specific language or group of languages come to be lost by native speakers (e.g., indigenous languages)? Or are there some linguistics characteristics that you're more interested in analyzing (e.g., writing attrition, phonological attrition, grammar attrition)?

Here are some the things that I found; I can't guarantee they're all scientifically sound though, you'll have to do your own checks:

[1] Schmid, M.S. 2011. Language Attrition. Cambridge: Cambridge University Press. https://www.cambridge.org/core/books/language-attrition/E01D...

[2] Gallo et al., First Language Attrition: What It Is, What It Isn’t, And What It Can Be (December 23, 2019). Higher School of Economics Research Paper No. WP BRP 113/PSY/2019. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3508640

[3] Francis, 2023. When does second language learning lead to first language attrition? https://www.researchgate.net/publication/365372235_When_does...

Thanks for the wiki -- I have always been interested in hardware hacking but I have always felt overwhelmed as I didn't know where to start. I believe this kind of resource can greatly help with that, especially the case studies.

However, I can't help but feel that a major part of the content is LLM-generated, or at least LLM-rewritten. It feels off and uninteresting to read, honestly. Is it the case? To support my case, I see that the case study page (https://www.hardbreak.wiki/introduction/case-study-led-to-a-...) has very similar paragraphs next to each other, the second one seemingly being the "genuine" one, and the first one being the LLM-rewritten version.

I'm not against using LLMs to help fix typos or reformulate things, but you should definitely keep some of your style. The LLM that you used (if you used one) made the content super bland, and as a reader, I'm not really incentivized to browse more.

While this still uses text at some level, it's no longer regurgitation of human-produced text, but something more akin to AlphaZero's training to become superhuman at games like Go or Chess.

How did you know that? I've never seen that anywhere. For all we know, it could just be a very elaborate CoT algorithm.

I'm frankly tired of this problem being a thing. I'm working on a dataset full of entries for every letter in every word of /usr/share/dict/words so that it can be added to the training sets of language models.

Interesting, but it should take a while to generate the data, no? Will zero answers be part of the dataset as well?

touying[1] seems quite great to create slides too.

[1] https://typst.app/universe/package/touying/

I wouldn't say that they can be used to do large-scale surveillance, but they can definitely facilitate it, especially with CV integration. I think one can easily imagine the following scenario: you fill a LLM with photos from people (taken from a public camera for instance), it finds the closest matches (via a web search for instance, as Gemini does). From then, you can easily gather the most essential information: first and last name, age, usernames... And then use this information to structure even more precise prompts and find even more potentially interesting data: posts on forums, relatives... And with this data, you can create an exhaustive database with a plethora of information and data about these people.

That's what any good stalker or person experienced with social engineering is able to do right now, but it takes a lot of time and energy. Resorting to LLMs would considerably decrease both. And it gets easier the more people you have information about.

I'm using Telegram's "Saved Messages" channel to bookmark things, and I'm pinning what's really important. The advantage is that I can access it anywhere and anytime (with the Android app), all while sharing all types of content, including files. I have a Python script running every 24 hours that uses the Telegram API and a SQLite database storing all my pinned messages in that channel. Each time the script is ran, it sends me two types of emails: 1) a "new" type of email that gathers all the messages that I've pinned recently, and 2) a "reminder" type of email that randomly shows X pinned messages. This forces me to reduce my backlog and unpin what is no longer relevant or things that I can deal with in a few minutes.

If it's really only the UI that's bothering you, why not use a web UI such as Open WebUI?

LaTeX requires retaining a large installation.

Or just use Overleaf.