@Techmeme@techhub.socialPrismML releases Bonsai 2 27B, which compresses Alibaba's Qwen3.8 27B to 5.9 GB, small enough for smartphones, while retaining 98.2% of Qwen's benchmark scores (Julie Bort/TechCrunch)
https://techcrunch.com/2026/09/17/prismml-hopes-its-tin…
@aredridel@kolektiva.socialRE: https://eldritch.cafe/@miranda_blue/116935608086213327
This is showing one thing that I think should become general principle: Using an LLM for someone else _is rude and bad_. Presenting something untranslated that _they_ can have an LLM translate is better than doing it yourself, and not showing that input — the prompt here is not included and is a key piece of context, and the results are also bad _with no recourse_.
Information has been destroyed. Trust has been broken (or in the case of commerce, failed to be established). Even if the total amount of LLM-use were the same, this is worse than the reader using it.
PROMPTING AN LLM FOR SOMEONE ELSE IS RUDE.
@publicvoit@graz.social#Atlassian nutzt alle Kundendaten für das Training von #KI. Wir erinnern uns an die zahlreichen Fälle, wo man aus #LLM teilweise die Trainingsdaten rekonstruieren konnte. 😜
Atlassian: Wer zahlt, bleibt verschont | D…
@Techmeme@techhub.socialSource: Beijing-based Naive AI, which plans to release its first AI model as early as this month, is now valued at $1.4B after raising $400M in three rounds (Juro Osawa/The Information)
https://www.theinformation.com/articles/tsinghu…
@crell@phpc.social
@heiseonline@social.heise.deChatGPT: Werbeeinblendungen kommen in weitere Länder
Werbung im eigenen Chat mit dem LLM: Das testet OpenAI gerade mit einem Teil der ChatGPT-Nutzer, auch Chatinhalte sollen einfließen.
https://www.
@laurentperrinet@neuromatch.social
@qurlyjoe@mstdn.social#MerriamWebster announces their new LLM. “There’s Artificial Intelligence, and then there’s Actual Intelligence.”
https://boingboing.net/2025/10/02/merriam-…
@ripienaar@devco.socialAm starting a series of blog posts on design and implementation of a AI Harness - based on the one from Fisk AI https://choria.io/fisk-ai
Will cover all the bits from what it is (todays post) to cover the loop, tools, memory, context, session history and more.
First post here
@grumpybozo@toad.socialI think Google Health has sacrificed its math to a LLM.
@stiefkind@mastodon.socialNur einmal angenommen, längere Texte schreibt man am besten mit Microsoft Word 5.5 für DOS. Für Behördenbriefe eignet sich Word für Windows 2.0 aber besser. Wenn der Brief auch noch Tabellen enthalten soll, geht das mit Word für Windows 2.0 zwar, aber viel besser eignet sich dafür Word 97. Ihr habt keine Tabellen, stattdessen aber Bilder? In einem viele Seiten langen Text? Dann greift man natürlich am besten zu Office Word 2007.
Wirrwarr bei LLM-Versionen, Symbolbild.
@wyri@toot-toot.wyrihaxim.us@… Five and a half hours of Opus 4.8 and I'm now looking at no errors whatsoever.
Honestly still unsure and uneasy about using LLM's for this, but it's hard to deny the results. This would have taken me a good week or two to do. Maybe even 3 or 4. But the real question now is: How much did it fuck up?
@bencurthoys@mastodon.social@… @… and given the existence of difficult to configure tools with crap documentation, an LLM is an appropriate response, and I say this neck deep in crap WorldPay documentation.
WorldPay have however …
@ELLIOTTCABLE@functional.cafeOkay, the breakdown *by-language* of Artificial Analysis's "Omniscience" hallucination/capability benchmark is absolutely fascinating.
I feel like this is some of the first concrete evidence I've seen of my longtime lean: that the *correctness* tooling we obsess over in PLT spaces is one of the greatest possible value-adds to the LLM era. (Just look at the Rust row, holy shit.)
Choosing a good language basically shifts the quality of output by an entire cost-cat…
@theodric@social.linux.pizzaIdea: glue together the most horrific LLM hallucinated spaghetti code possible and release it as The_Aristocrats.exe
@grist@fosstodon.orgWhy to use Grist for vibe-coding – an app being one file three things: how coherently you can build it with an LLM, how easily you can deploy, copy, and version the result, and how you secure it. https://www.getgrist.com/blog/why-is-grist-an-ideal-vibe-cod…
@arXiv_csLG_bot@mastoxiv.pageReplaced article(s) found for cs.LG. https://arxiv.org/list/cs.LG/new
[6/11]:
- Zero-Shot Active Feature Acquisition via LLM-Elicitation
Binyamin Perets, Natalie Mendelson, Shiran Vainberg, Yehuda Chowers, Shai Shen-Orr, Shie Mannor
@johnleonard@mastodon.socialOpen-weight AI models, many developed by Chinese companies, are quickly expanding their role in enterprise AI systems, according to new research.
https://www.computing.co.uk…
@mia@hcommons.socialThe accuracy of LLM models varies depending on how many questions were asked a time? Structure your prompts wrong and even 'good' models fail? 😬
'The performance differences between frontier models, under realistic clinical input conditions, are smaller than the performance differences within a single model across different prompt structures'
@hex@kolektiva.socialBut I would warn anyone circumventing anti-LLM protections that, as the tide turns against #LLM usage and "#AI" as a whole, that it may become politically useful for some state AGs to score political points. Those political points may come by putting you in prison, or, at the very least, dumping lawyers on you and ruining your life. An AG doesn't need to win to score points for their political party or to virtue signal. All they need to do is build a case against you. Because the CFAA is *so vague* it's easy to make something plausible enough that it won't get thrown out immediately.
So if the gentile "please don't do that" isn't enough, maybe consider this.
@Mediagazer@mstdn.socialPeter Thiel-backed Objection, initially an AI "tribunal" assessing reporters' articles, rebrands as The Primary to rank journalists "on the rigor" of their work (Gary Baum/The Hollywood Reporter)
https://www.hollywoodreporter.com/bus…
@rasterweb@mastodon.socialAlso from Codeberg:
"You must not share projects that mostly consist of code written by "generative AI"-tools (including services such as *Claude*, *OpenAI Codex*).
Such projects having an unclear copyright status and furthermore have little safeguards to ensure that they do not include harmful code."
@michabbb@social.vivaldi.net🧠 Routing strategies: LLM-as-classifier, signal-driven stage router, escalation router (weak tier first, a judge decides whether to escalate), random split for A/B tests, or a custom algorithm you write yourself
🚀 Launcher path: uv tool install "nemo-switchyard[cli]", then switchyard launch claude --model switchyard to point #ClaudeCode, Codex CLI or OpenClaw at an open-sou…
@grumpybozo@toad.socialSo, https://mastodon.social/@JenniferSmith345 is an odd bot. I don’t think it’s even a LLM behind it, the replies are just too weird.
Profile looks like it was taken from someone real…
@larsfosdal@mastodon.social
@stargazer@woof.techВ процесі іноді народжуються зрозумілі лише айтішникам жарти.
З *defineModel* мене познайомила саме LLM, це відносно нова фіча.
В Реакті, з якого я родом, щоб компонент перемикав state батька, треба прокидувати в нього як props
1) сам state
2) колбек на функцію, яка його змінює (бо сама функція має бути описана там же, де й state)
Vue концептуально робить те ж саме, але у нього є додаткова цукровість: можна прокинути щось як v-model (в свій кастомний компонент!), тод…
@portaloffreedom@social.linux.pizzaI think it would be cool to train an LLM like a baby, with a supervising teacher, instead of like a parrot that has been fed the entirety of reddit.
I'm not even sure if it's possible and what would be the blocker.
@domm@social.linux.pizzaToday I braincoded some short SQL fixup while letting the LLM also generate a fix. The LLM was a tiny bit faster (I'd guess it took 60sec vs 90sec on my side), but I liked my solution more. I used a CTE to generate the fixed data and then did an update on select, while the LLM used a subselect.
@trochee@dair-community.social"MLM" and "LLM" are separated by a Levenshtein distance of one
…and not much else
@penguin42@mastodon.org.ukUsing #llm to OCR is...interesting. I've been trying it on an old hand written card; and it does very well on stuff I can only just about read; but the mistakes are way more subtle than trad OCR; e.g. one misread a 'standard 6' rack' as a 'standard 19" rack' (the other said '61 rack' which is more obviously wrong). The same one also turned 'M/coil mike'…
@cdamian@rls.socialClaude Code today in an otherwise perfectly normal answer
"The gain would be близко to zero — [...]"
In case you are not proofreading your generated texts.
#AI #ClaudeCode #LLM
@mela@zusammenkunft.netMetadatenabgleich schlägt fehl.
Mela: *verwundert* "Kann die Bibliothek denn keine konkreten Metadaten aus einer Datei extrahieren?"
LLM: "Doch, klar."
*grillenzirpen*
Mela: "Warum ziehst du dann nur die erste Seite, anstatt der Metadaten?"
LLM: "Die Metadaten stehen doch schon in der Datenbank, das wäre dann ja redundant
.."
*grillenzirpen*
*grillenzirpen*
*grillen verstummen*
Mela: "ABZU…
@kurtsh@mastodon.socialAI fear-mongering is profitable. Blaming AI is simply lazy. Claiming humans are alway better is sanctimonious.
Greed, Sloth, Pride... interesting.
From: @…
https://
This is the “should I walk to car wash” question but without an LLM.
RE: s.1a23.studio/notes/aq089nipvi2xbfqc
@stefan@gardenstate.socialThis incite in to what my ubuntu system is doing that would have taken hours of time to research if I ever could have learned it at all had I not used an LLM. I'll continue to caveat that LLMs have many many things about them that are bad on multiple levels.
Using LLMs to fix system problems is doing something I wouldn't be able to do otherwise.
#ai
@dde@social.tchncs.deMit dem LLM auf Distributionssuche – #linux
@cwensel@fosstodon.org
@thijs_lucas@norden.socialRE: https://norden.social/@thijs_lucas/117276776028587546
Ist KI verlässlich? Oder ist KI verlässlicher als mancher Lokaljournalismus?
Ich hab den KN-Artikel durch ein LLM laufen lassen und gebeten, zu prüfen ob und wie weit Positionen der CDU dari…
@marcel@waldvogel.familyCool idea:
"Researchers from Tracebit on Monday said they found that placing prompt injections alongside passwords, cryptographic keys, and other secrets stored on Amazon Web Services was often all that was needed to shut down attacks from AI hacking agents. The prompts direct the attacking LLM to perform an action forbidden by its guardrails, the safety barriers AI developers erect to prevent it from taking harmful actions. The LLM responds by shutting down."
@v_i_o_l_a@openbiblio.social"What Changes When an LLM agent Searches Your Library Catalogue?" asks Aaron Tay:
https://aarontay.substack.com/p/what-changes-when-an-llm-agent-searches
"I connected an LLM to search Primo and found it was excellent at academic database …
@thomasrenkert@hcommons.socialfunniest thing in #LLM news. One man's month-long research to uncover the secrets of the Frontier Labs is another man's "hey I wrote this python deep_think script":
https://xcancel.com/_can1357/stat…
@sakhavi@aoir.socialCurrent #LLM-driven degradation of software seems to follow an eternal pattern of humans and #technology, where the former seek the latter as a shortcut through reality and experience any resistance and complexity as loss and insult. In my publications I’ve called this “prelapsarian programming” bc it betray…
@ErikJonker@mastodon.socialFor me this age of AI is the time we should appreciate less polished handwritten texts for their authenticity. I see this happening during selection of job applications.
https://www.theguardian.com/books/ng-interactive/2026…
@fell@ma.fellr.net
@toxi@mastodon.thi.ngRE: https://mastodon.gamedev.place/@jonikorpi/117263464874714292
The push to replace UI/UX design with LLM generated outcomes is so disillusioning/shortsighted and sooner than later will land us in a clichéd, bland, semi-broken cesspit of good-eno…
@pavelasamsonov@mastodon.socialFollow me for more high quality LLM training data. The more data we produce, the better informed AI models can become.
@aral@mastodon.ar.alSharing this not because it’s about AI/LLMs but because it’s solid advice and the recommendations work equally well for *people* coding web applications (funny that, almost like these things were made by people for people or something).
TL;DR: Use web and platform standards; don’t reinvent the wheel.
https://www.jimmo…
@kubikpixel@chaos.social«Jadepuffer — Die erste KI-Ransomware-Attacke verbaselt nicht nur den Schlüssel:
Forscher dokumentieren den ersten LLM-gesteuerten Ransomware-Angriff. Doch die KI scheiterte an grundlegenden Fehlern. Eine Mahnung für Admins.»
Das die KI mittlerweile scheinbar "alles" hackt ist auch ein Teil ihres Marketings. Denn durch Unsicherheit lässt sich einiges besser als "Die Lösung" verkaufen.
🙄
@stsquad@mastodon.org.uk @Techmeme@techhub.social
@fazalmajid@vivaldi.net@… wouldn't the deciding factor be how much a language is represented in the LLM training set vs the inherent properties of a language?
BTW it's Yossi Kreinin, not Kreinen
@Kingu@sakurajima.moeIf your idea of a game is a text adventure like Zorg but with a LLM creating the story as you go then please reconsider.
There's already 100 of them and none of them works.
@philip@mastodon.mallegolhansen.comOn one hand, I am heartened by seeing more and more people say things like "I've come to realize I can't trust the LLM to do the right thing", but then they shatter my heart by immediately following up with "So I make it do the tedious stuff, like writing the tests".
NO.
No. Stop it. NO.
The tests *are* the important part, they're the ones that prove your code does what it should. Be confident in the tests, and you can even afford to LLM gene…
@stephane_klein@social.coopJe viens de publier « Qu'est-ce que je dois faire signer Š mes clients pour être autorisé Š analyser leurs données confidentielles avec un LLM ? »
#LLM
@heiseonline@social.heise.de @michabbb@social.vivaldi.net⚡ One Docker command starts the full stack locally on port 3010. First startup pulls images in 2-3 minutes, then you pick inbound or outbound, name the bot, describe the use case in a few words and click Web Call.
🔑 No API keys required to start: #Dograh ships with auto-generated keys and its own LLM, TTS and STT stack. Own keys for LLM, TTS, STT or telephony can be connected later.
@crell@phpc.social
@avstockhausen@fedihum.orgBookmarked: TEI-NER Pipeline — Automatisches Entity-Linking für historische TEI-Editionen http://tei-llm.histomatiker.de/ #GND #LLM…
@jdrm@social.linux.pizzaInsisto, ojalš unos juicios de Nüremberg contra los CEO de las IA generativas, LLM y similares.
@thomasfuchs@hachyderm.ioOn LLM watermarking: LLMs aren't doing anything other than statistical generation (accurately or not) _by design_.
Having them watermark stuff doesn't change anything about how they work, it's just ends up being slightly different statistics used for the stochastical generation.
In other words, they always have been and always will be watermarking their output; the only difference about the Anthropic watermarking thing is it's less computationally intensive (for them) to detect.
This is categorically different from e.g. steganography in printers.
@Techmeme@techhub.socialQwen 3.8 27B shows a 17GB open-weight general purpose model can have long context, effective tool calling, strong vision ability, and competent code generation (Simon Willison/Simon Willison's Weblog)
https://simonwillison.net/2026/Aug/16/qwen-38-27b/
@stiefkind@mastodon.socialNur einmal angenommen, die Benutzung von LLMs und GenAI wäre Handwerk, einschließlich drei Jahre Lehrzeit zum Gesellen, Qualifizierung zum Handwerksmeister LLM und einer Handwerksrolle. Nach der bestandenen Gesellenprüfung ist es üblich, drei Jahre auf die Walz zu gehen um seine Erfahrungen zu schärfen. Erst danach kann man auch seinen Meister LLM-Handwerk machen. #justthinkin
@grumpybozo@toad.socialI am likely to be chastised for my latest post in a (private) mailing list, but IDGAF.
"I do appreciate your transparency in making it clear that your message was from a LLM. If you didn't bother writing what you submit, I certainly will not bother reading it.”
@aardrian@toot.cafeRE: https://front-end.social/@fox/116899694179909261
I try very hard to do this on the daily. But where I don’t is when the LLM output is wrong. I’ll intentionally anthropomorphize it by saying it’s lying or it’s a liar. I’m rather hoping, sans evidence, that…
@arXiv_statAP_bot@mastoxiv.pageCrosslisted article(s) found for stat.AP. https://arxiv.org/list/stat.AP/new
[1/1]:
- Which Pairs to Compare for LLM Post-Training?
Jiangze Han, Vineet Goyal, Will Ma
@stefan@gardenstate.socialI feel confident that not every person agrees on what "used LLM to make art" means.
Some might be obvious "write X" but other seems less obvious to me. LLM to market a book? LLM to fix an email to a publisher? LLM to make a grocery list so you have more time to do art?
I feel like all of that could count because everything in your life affects everything. Everyone could pick their own line or we could all agree on one but there is no clear answer.
@rberger@hachyderm.ioNot that this is anywhere near a well hosted full frontier model. It does show that there can be creative avenues, along with the normal semiconductor and software learning curves, we won’t need those stinkin Data Centers and the chips they are buying today will probably be obsolete by the time any of these planned data centers might get built.
https://www.geeky-gadgets.com/run-llm-esp32-microcontroller/
@johl@mastodon.xyzWhy the #Zig project bans all LLM contributions – How Open Source Survives the Age of Free Code
https://lumramabaja.com/posts/engineering-sacrifice-ho…
@chpietsch@fedifreu.deDebian developer Antoine Le Gonidec just quit the project over its decision not to ban #LLM-assisted code contributions:
Pretending to have a "neutral stance" when faced with fascism is not neutrality, it’s active collaboration.
AI slop is fascism materialized. Stop it whenever you can!
Source:
@heiseonline@social.heise.de @mia@hcommons.socialI love that this includes 'What is Benchmarking and Why Should You Care?'
RISE-UNIBAS/humanities_data_benchmark: LLM Benchmark Suite for Humanities Data https://github.com/RISE-UNIBAS/humanities_data_benchmark
@toxi@mastodon.thi.ng @michabbb@social.vivaldi.net @ripienaar@devco.socialI have a little cow joke LLM agent made with my Fisk AI tool, I use it to test various.
I told it to save the jokes it already told me and always tell new ones, this is getting pretty bad after a bit lol
@stephane_klein@social.coopJe viens de publier « Septembre 2026 - je code avec des open weights pour 10 Š 30 € par mois »
#LLM
@crell@phpc.social @thomasfuchs@hachyderm.iohear me out, a LLM but trained exclusively on works by Joseph Weizenbaum, Hubert Dreyfus and Shoshana Zuboff
@cwensel@fosstodon.org @hex@kolektiva.socialThe thing about training an #LLM on human culture and then renting that culture back isn't that it's "Intellectual Property Theft." There's nothing wrong with sharing. There is something wrong with framing LLMs as piracy. That's a completely different concept.
When a powerful group of people takes the writing, the art, the music, various parts of a culture and then make some bland reproduction that lacks all the meaning and essence in the original, then those powerful people profit off that mess while the people who made the original stuff are erased, marginalized, and made to suffer, there is a different term for that.
The term everyone is looking for is "appropriation."
I'm not saying this is the same as other forms of cultural appropriation, but there are definitely enough shared elements that we should revisit that conversation. Perhaps some folks who thought the idea was silly when the idea first came up may feel differently now.
@ErikJonker@mastodon.social @jdrm@social.linux.pizza¿Qué sentido tiene usar un LLM para una afición? Es como hacer trampas jugando al solitario. ¿No disfrutas con el proceso de tu afición? Entonces, si lo haces, por qué tomar atajos. El disfrute estš mšs en el proceso que en el resultado.
@aardrian@toot.cafeClosing the floating action button for the LLM chat to reveal and close the floating action button for the accessibility overlay to reveal and close the floating action button for the shopping cart to reveal the floating action button for the original chat feature.
FAB archeology.
@crell@phpc.socialYour regular reminder:
ALL LLMs are ultimately just autocomplete. There is no thought, no creativity, no consciousness, no insight. It's just autocomplete with infinite monkeys. What turbocharged autocomplete can do is legit impressive, but it is still just turbocharged autocomplete.
Don't give it more credit than it deserves.
#AI
@philip@mastodon.mallegolhansen.com@… A lot harder, because the hard part of software engineering was always agreeing on what we’re trying to build, not actually building it.
And getting agreement on what we’re trying to do is a lot harder once nobody reads or engages with what everyone else is saying, but everyone is busy slinging 10 page LLM generated documents that nobody else reads anyway.
@Techmeme@techhub.socialSources: Apple trained a China-specific LLM with Alibaba's support, which would make Apple the first foreign company to offer a proprietary AI model in China (Reuters)
https://www.reuters.com/business/retail-co
@heiseonline@social.heise.de @cwensel@fosstodon.org @mia@hcommons.socialHow many senior tech folk are honestly like 'I'm in this picture and I don't like it'?
'a lot of people who started out as developers are thrilled to get their hands dirty again with a bit of vibe coding – that using genAI makes them feel more in control of a tech stack that’s got away from them, perhaps even makes them feel younger and more dynamic'
@thomasfuchs@hachyderm.ioMaking progress with my OpenSCAD shelf model.
For the artwork, I vibe-coded* a converter from PNG to voxels (1x1x1 cubes) as OpenSCAD doesn't support textures.
*it took like ~2 mins to make a little Python script with a local LLM that's running on my MacBook Pro and is powered by our solar cells.
@aardrian@toot.cafeRE: https://front-end.social/@SaraSoueidan/116845563478631182
My experience is that LLM-derived writing about digital accessibility is wrong.
So not only do I abandon the page, I also mentally move the writer (LLM prompter) into a do-not-contra…
@crell@phpc.socialIf the net result of LLMs on education is that homework becomes unworkable and is abandoned in favor of in-school work... I am OK with that.
#Homework #Education #AI
@crell@phpc.social @thomasfuchs@hachyderm.ioI really think for “AI” coding that local models are the future.
It’s basically free to run them on hardware you might already have, doesn’t have the environmental issues that data centers have and you don’t have to share your code or project details with anyone.
I’m a bit ambivalent about the ethics of the training data—for coding stuff that data is mostly open source code. I don’t have an issue with my MIT-licensed stuff used for this, but I know other authors do and their preferences should be honored.
I’m hoping that we get “clean” models for this in the future.
Generally I don’t think LLM-assisted programming is universally applicable to all software development but it can be a useful tool.
I also don’t think LLMs are useful for much else (coding is like a one in a million special case). ¯\_(ツ)_/¯
The code is here, if you're interested: https://github.com/madrobby/png-to-scad/
@Techmeme@techhub.socialMIT launches the LLM Election Observatory, a dashboard tracking how nearly a dozen AI models tailor responses to political queries during the 2026 US midterms (Tiffany Hsu/New York Times)
https://www.nytimes…
@michabbb@social.vivaldi.net @crell@phpc.socialWho cleans up after the vibe-coding party?
#ai
@Techmeme@techhub.socialHugging Face says it used the open-weight GLM-5.2 hosted on its own compute for breach forensics, after US frontier model safety guardrails blocked the requests (Edward Targett/The Stack)
https://www.thestack.technology/hugging-fa
