Wildroot

Queries may use an external AI service. Details

Wildroot Browser

TopicsAI

AI

365 of today’s 3,021 posts about AI, best first. Page 1 of 7: the top 36, then 67 the model is less sure belong here.

Ranked Bluesky · Mastodon

About 7 in 10 posts here get an intent label; the model only labels the ones it is sure of.

How it worksRSS feedBuild a feed from AI

  1. @lea-linux.org

    "Most modern transformer-based LLMs use the same broad structure: tokenization, embeddings, positional encoding, stacked transformer layers (each with multi-head attention and a feed-forward network), residual streams, layer normalization, and next-token prediction."

    Most modern transformer-based LLMs use the same broad structure: tokenization, embeddings, positional encoding, stacked transformer layers (each with multi-head attention and a feed-forward network), residual streams, layer normalization, and next-token prediction. What changes between models is: The trained weights themselves, learned from different training data at different scales. The configuration: number of layers, vocabulary size, head count, parameter count, MoE or dense. The post-training: instruction tuning, learning from human feedback, safety controls applied on top of the base model. Tiny explainer: weights Weights are the learned numbers inside the model. Training changes those numbers until the model predicts text well. The 2023-2025 “modern transformer” stack converged on a common set of choices across many serious frontier and open-weight models, even though different teams arrived at them independently. Pre-norm placement. RMSNorm. RoPE. SwiGLU. Grouped-Query Atten
    Info
  2. “The economist Sendhil Mullainathan, in writing about the dangers of AI, cited the long-tail phenomenon (which I described in chapter 6) in his notion of 'tail risk’:” «We should be afraid. Not of intelligent machines. But of machines making decisions that they do not have the intelligence to make. I am far more afraid of machine stupidity than of machine intelligence. Machine stupidity creates a tail risk. Machines can make many many good decisions and then one day fail spectacularly on a tail event that did not appear in their training data. This is the difference between specific and general intelligence.»

    Show more of this post

    “Or as the AT researcher Pedro Domingos so memorably put it, 'People worry that computers will get too smart and take over the world, but the real problem is that they're too stupid and they've already taken over the world.’” #MelanieMitchell wrote this in 2019. It’s 2026 and I have not seen any improvement in this regard: no advancement in building systems that understand rather than mimic, and hence remain fragile to the long tail of events absent from the training data. #AI #LLMs

    Opinion
  3. The dark(er) side of AI: massive growth in autonomous weapons - lacking regulations. “In Iran, a US strike hit a school, killing 156 civilians including 120 children… a result of [faulty AI]” Meanwhile, manufacturers are trying to develop “swarm” capabilities… www.theguardian.com/science/2026...

    Popular todayprobably AI
  4. @pbloem@mathstodon.xyz

    After some chit-chat with Claude, it turns out the Bayesians do have an answer to this. The main flavour seems to be to keep the single probability to express the belief, but to probe its resilience to new evidence to express a higher-order meta belief. It's a clean way of doing things, but harder to express informally, I guess.

  5. @carnage4life.bsky.social

    Anthropic’s Claude helped prove the 167-year-old Riemann Hypothesis, now formally verified. Google and OpenAI have also used AI to solve open problems in math and computer science. As AI gets better at answering humanity’s hardest questions, who will come up with the next ones? Humans or AI?

  6. @AeonCypher@lgbtqia.space

    I know that most of you don't care but the claim that #Claude (an #LLM #AI) solved the #RiemannHypothesis is really a wild exaggeration on multiple levels, and people keep posting about how it means we already have #ASI. Zzzzz. First, it's not actually confirmed it solved the #MilleniumMath problem. There is a team working through it to validate the results, and it will be months before we know. Why? Because it's a 220,000 page paper to prove it. The _only_ thing we've proven is that there's a Lean compilable proof of it, which means it has to be encoded correctly. Also, the team has already demonstrated that ***significant*** portions of the proof are superfluous.

    Show more of this post

    Additionally, a human (a human mathematician) was heavily involved in working with Claude to come up with the proof. Finally, a human actually proved it before Claude (probably). The thing is that a human would never publish work like this, it would be embarrassing. The only reason we think a human has solved it is that Claude's proof is forcing them into premature publication. There has not, in spite of what you may have heard, been some big breakthrough in AI; we do not have #AGI. What has happened is that we have harnesses that are good at causing an LLM to loop until it finds a proof. They are convergent optimizers when given infinite tokens. Which is A HUGE ACHIEVEMENT, but the reporting on them is wildly misleading.

    Opinion
  7. @timwatson.bsky.social

    There are lots of complicated questions about legal and financial liability for harms caused by AI. And then there is the simple clarity of Citibank's new user agreement for online banking: yes, You are liable.

    Screenshot of the new section H in the Citi Online User Agreement, which reads in part: H. Use of Third-party Artificial Intelligence (“AI”) Agents. If you provide your Online Credentials to any third-party AI Agent or otherwise permit such agent access to your accounts or allow such agent to use any digital service and to perform any transactions on your behalf, Citi will consider any access and actions taken by the agent as authorized by You. If the Online Credentials that you provided to the third-party AI Agent are further shared, leaked, or otherwise distributed to other persons or AI agents, Citi will consider any access and actions taken with the Online Credentials as authorized by You. Citi shall have no responsibility or other obligation to investigate the authority or validity of any such actions even if the actions taken by the agent or other person or AI agent with your Online Credentials conflict with any prompts or other instructions provided by You or on your behalf to
    Opinion
  8. A common disconnect in AI discussions are the perceptions based on early descriptions on how GPT worked ("autocomplete") and have not been exposed to the trick that makes modern models so effective - but also that this is highly proprietary data (I mentioned this on a longer thread yesterday). This piece is a good introduction to the subject: openathena.ai/blog/there-is-no…

    Recommendation
  9. @bladeofthes.bsky.social

    Karl Marx predicted AI nearly 200 years ago. He said one day the production process would be entirely automated to the point that humans weren't needed. His prediction for what would happen to the workers under Capitalism was not optimistic. He said they would be 'cast away'

    Popular today
  10. "We sent nearly two million AI prompts containing unique, trackable secrets to AI providers, then watched where those secrets surfaced. What we found revealed security breaches, prompt theft, and widespread failures to protect customer data. Over a five-month period from July through November 2026, we submitted approximately two million prompts containing unique URLs and canary token credentials to widely used AI services and monitored what happened after inference. Our investigation uncovered three significant findings:

    Show more of this post

    - Evidence of provider compromises. - Unauthorized access to customer prompts by inference marketplace providers. - Failures in privacy and data-handling practices. We observed multiple" dragos.com/blog/ai-supply-chai… Privacy settings were ignored: "Zero Data Retention is supposed to mean something. In practice, these prompts were retained, shared with other companies, and read by people." What to do? - Host it yourself, if you can afford it. - Go directly to a first-tier provider. - Use routers with care (or not at all) - Seed your own canaries. - Enable secret filtering

    Info
  11. @ianfraser.bsky.social

    In Bartz vs. Anthropic, it emerged that Anthropic downloaded more than seven million books from pirated libraries to train its AI models. In a July 2026 out-of-court settlement, Anthropic paid out $1.5 billion to authors and publishers it ripped off. 📗📘📕

  12. @futurebird@sauropods.win

    On the day of Anthropic's IPO a group of people gathered to protest at their offices. These people are concerned that Anthropic is implementing "recursive self improvement" recklessly. And, I'm annoyed at them. Since most people don't pay much attention to the details of technical news about software... their complaint blends in with those who think LLMs are being implemented recklessly for less exotic reasons. It almost feels like deliberate misinformation designed to muddle the debate.

    OpinionPopular today
  13. One of the interesting parts of the "LLMs do certain kinds of code well" outcome is that other types of synthetic reasoning also become cheaper and get used more. SAT solvers are *very* capable, but they're famously hard to use. Now you can use them too, without tons of practice on translation work.

    Opinion
  14. @Jigsaw_You@mastodon.nl

    Alpöge's "proof" isn't a traditional mathematical paper. It's roughly 17 million lines of Lean code — a programming language used to mechanically verify logical arguments — backed by an underlying AI-generated text spanning more than 250,000 pages. That's approximately 500 full-length novels worth of mathematical argument, a volume so vast that Alpöge himself refused to publish it, calling it a "massive unreadable travesty of 'slop.'" #ai #tech #math #mathematics emeraldbook.org/news/nov-2126/

    Info
  15. @sumitcodes.bsky.social

    35.7% more pull requests. 56.1% longer review-to-merge time. A Nov 19 Harvard working paper tracking 655 firms found this after AI-agent adoption. Code got faster to produce. Not faster to trust. The bottleneck moved from writing to review.

  16. @loomcycle@hachyderm.io

    A chat model writes. An embedder places. A decision model answers: it reads some state, answers typed questions about it (pick one, yes/no, a scale) and returns a probability per option. No text, nothing to parse. loomcycle has used one for seven weeks, in five places. New post on why we now treat it as a third kind of model, next to chat and embedder. 🧵 1/3 loomcycle.dev/blog/decision-mo…

  17. @gaprogman.com

    Should you tell a CI agent to treat untrusted text as data? Yes; it improves the odds. But in October 2025, a team including Nicholas Carlini attacked twelve published prompt injection defences with methods that adapt to them, and got success rates above 90% for most.

    Info
  18. Back in August, I theorized that AI-pilled super users create a burden on their co-workers, erasing productivity gains, and making organizations less productive than without AI: machinesociety.ai/p/productivi… New research has come to the same conclusion: theconversation.com/catching-a… #MachineSociety

  19. @racheline.bsky.social

    I worked in media analysis for 20+ years. We had people all over the world reading newspapers in their native languages to analyze the topics and explicit and implicit tone for every covered protagonist (companies, people, demographic groups, etc) and then enter that material into a database

  20. Turns out, openai have at least some of the protections put in the harness (codex), not in the model. Codex refused a task, opencode with the same openai model did it in 2 minutes. Incredible. It's like instead of putting a firewall on a server, you supply a patched telnet to clients that would only connect to 80 and 443. What could possibly go wrong?

  21. @glorpolitics.bsky.social

    Are we becoming too reliant on AI before we fully understand it? Governments are integrating AI into increasingly important systems, while the companies building it are still trying to establish how to make the technology consistently profitable. I debate this and how AI is becoming so political 🔽

  22. @swelljoe@mas.to

    Many years ago, I lost an OctaMED song to floppy disk corruption, one I thought was the best I'd ever made at the time (I was ~17). Today, I came across a backup in an old Dropbox (a backup of the corrupt file dumped from floppy, not a working backup), and realized maybe the AI that's supernaturally good at reverse engineering might be able to do something about that unloadable song. Claude was able to fix the file, and identify everything in it, but the pattern sections were all overwritten with sample data, so the samples are partly there but the song isn't.

    Show more of this post

    This is what I get for using a weird third-party filesystem (to get more space on a floppy and faster performance, I don't remember the name of it, but I paid money for it) and re-used floppy disks from an old magazine cover disk I somehow had dozens of. This looks like a failure of the filesystem, rather than of the disk, since data overwrote other data, which feels like an overflow or off-by-one or caching issue of some sort.

  23. LLMs were trained on massive (stolen) libraries of written formal English, not on informal spoken language. Result: writers with more traditional “proper English” writing styles (including ESL writers) get flagged as AI and penalized for that - authorship doubted, documents dismissed as spam/fake.

    Opinion
  24. @inthehands@hachyderm.io

    A specific form of this that irks me: I occasionally hear people framing debates around AI as whether the technology is “moving too fast” and needs to “slow down.” That comes from bad faith industry actors looking for a regulatory moat, but also from good faith thinkers trying to chart a path. Either way, that is the same trick as the quoited conversation here. My concern is not the •speed• of AI development, it’s the •direction•.

    Opinion
  25. @thebleats.by86400.link

    Speaking of sudden obvious realizations (not the quoted skeet) but AI does all the “fun stuff” because there’s orders of magnitude more data available documenting it. The promise is: You were doing this anyway for the love of the game, let a computer do it so you can focus on things you hate

    Opinion
  26. @scottgal@hachyderm.io

    Re-reading for the millionth time, reminded that the end point for AI should be local AI devices which make our kids revolutionaries (A Young Lady's Illustrated Primer).

    Neal Stephenson's book "The Diamond Age" depicts a lone figure walking toward a brilliant light at the end of a long road, framed by a red and orange, starburst-like background.
    Opinion
  27. @socratetris.bsky.social

    One of the struggles of generative algorithms in games/code is that a lot of coders don't think of their practice as an artform worth treating as a creative act. They have the engineering mindset of solve-problem-move-on. One of the earliest game studies articles was precisely about the artistry of

    Opinion
  28. The anxiety over vibe coding is eerily familiar to the 1980s anxiety of using other people's code #AI #VibeCoding

    Case 1 concerns a compiler. It has been developed by a group A for a Language L and worked very well on computer X. Now another group B has the task to write a compiler for a language L + M, a modest extension of L, for computer Y. Group B decides that the compiler for L developed by group A will be a good starting point for their design, and get a contract with group A that they will get support in the form of full documentation, including annotated program texts and much additional written design discussion, and also personal advice. The arrangement was effective and group B managed to develop the compiler they wanted. In the present context the significant issue is the importance of the personal advice from group A in the matters that concerned how to implement the extensions M to the language. During the design phase group B made suggestions for the manner in which the extensions should be accommodated and submitted them to group A for review. In several major cases it turned out tIn the years following these events the compiler developed by group B was taken over by other programmers of the same organization, without guidance from group A. Information obtained by a member of group A about the compiler resulting from the further modification of it after about 10 years made it clear that at that later stage the original powerful structure was still visible, but made entirely ineffective by amorphous additions of many different kinds. Thus, again, the program text and its documentation has proved insufficient as a carrier of some of the most important design ideas.
    Opinion
  29. @joehall22.bsky.social

    Something that is rarely discussed with vibe coding: LLMs have a bias towards the same absolutist mindset that is often seen on online forums like stackoverflow. This results in application architecture design without a lot of nuance. Especially when your prompt doesn't give sufficient context. 1/2

    Opinion
  30. The use of AI & its associated technologies in the labour market is a classic collective action problem. When one person uses AI to help refine/improve their application, they can gain some extra traction in the application process....when everyone does all it does its impose costs on applicants (of more rejections); and employers meeting more capacity to assess applications in the face of the AI-engineered torrent. Once again as @pluralistic would point out - enshifification #AI #workers

    Opinion
  31. @coil.bsky.social

    "The dominant activity becomes not making, but preventing the system from producing something mediocre." Would you rather work your way to the right answer or stand in front of a conveyor belt, rejecting an endless stream of wrong answers till a "close enough" comes along?

    3. Depletion The designer who remains employed after the restructuring of the last few years now works inside a specific arrangement, in fewer in number, responsible for more surface area, equipped with tools that generate output at a rate that exceeds any single person's capacity to evaluate it critically. Brian Eno described this arrangement in 2025: the dominant activity becomes not making but preventing the system from producing something mediocre. "You really feel the pull of the averaging effect of AI", Eno said. The cognitive load of continuous resistance, the ongoing work of refusing what the model offers and substituting it with something better, is not lighter than the load of making. It is the same load, applied to a larger volume of material, with less support.
    Opinion
  32. We came across an advert for Google’s AI thing today. The advert was about how it can make it easy to book a diverting side trip while planning a holiday. Why would I want to use that? The act of researching a holiday is part of the anticipation and enjoyment. I like to read reviews and explore possibilities. That’s what the AI pushers don’t understand. I like to write music, I like to write poetry and prose. I like to plan holidays. Taking away the process of creation is not useful.

    Opinion
  33. @mariusor@metalhead.club

    This ended with me fixing a pretty bad oversight in how actors got authenticated for server to server requests. The LLM was right, and I'm half bitter about it, but at the same time the communication was palatable. There were no overt walls of generated text, the replies to my queries were answered quickly and concisely, their suggestions were reasonable, and the end result is an improved piece of code. And yet... I still feel a little unclean. #GoActivityPub #Activitypub #fedidev

  34. @gimmechocolate@bark.lgbt
    Content warningAI in GNOME

    discourse.gnome.org/t/rfc-poli… I feel like zhis guy really accidentally sums up zhe whole issue wizh what can be paraphrased as "We can't ban ai because we're controlled by corporations, and it'd make corporate-sponsored vibecoders sad." Zhe people who run zhe orgs never seam to realize it's explicitly a choice between human maintainers and vibecoders. You can't just allow people to spam slop (and every single ai user zhinks zhey're zhe one ai user who doesn't spam slop) at maintainers and eveyrone else, zhen expect non-ai users to keep working on zhe pawject. FOSS can never be people first if it keeps trying to appease corporations and zhe fragile feelings of vibecoders.

    Opinion
  35. Yes, coding agents can create a lot of code very quickly. The mistake is thinking we should do bigger releases because of it. It doesn't matter how fast you can make guesses of a random 6-digit number if you're trying to guess all 6 digits at a time. The person guessing 1 digit at a time will get it first.

    Opinion

More in AI

The model is less sure these belong here, so they sit outside the ranking.

  • @abestew.bsky.social

    Timely video thats worth a watch if tired hearing AI bros uncritically drone on about genAI's maths capabilities. Also worth noting it cost OpenAI ~$12m to run the genAI models for the Navier–Stokes proofs - so excluding costs for infrastructure, training, etc. www.youtube.com/watch?v=yUoe...

    Recommendation
  • I have been mostly off the grid today. So I had some time to mull over yesterday before coming back online. There are many critical voices in my feed. They make good points about financial incentives, environmental impact, etc. However, I am still in a state of positive shock. Terence Tao remarked in an interview that the age of heroes is over. I am taking away the opposite from this huge accomplishment in the very human endeavor called mathematics:

    Show more of this post

    * Levent Alpöge has joined the rank of mathematical legends. On top of persistence, what he accomplished required _skill_, for the lack of a more specific description. * LLMs are contributing to the benefit of mankind. They do not perform mathematics in the narrow sense. But by breaking the four-minute mile, they blaze the trail for human mathematicians. Or in Hilbert's words: 'Wir müssen wissen, wir werden wissen'. * There is plenty of glory left for human mathematicians from here on out. In my own words, mathematics is about finding the right abstractions, about making complex things simple. I dream of a future where someone with a solid undergraduate education in maths understands the proof of the Riemann hypothesis.

    Opinion
  • @drewsof.bsky.social

    Another worthwhile piece, with a strong reframing about the prizes we've all been arguing about and just what it means to use the tools at our disposal to hold some necessary lines. This and the Urbano article make for bracing reading, both worth your time. jonmalesic.substack.com/p/pangram-vs...

    Recommendation
  • @phpboyscout@hachyderm.io

    My Claude sessions hit their quota wall on bonfire night, so I did the obvious thing and moved house while nobody was looking: nineteen private projects into a brand new GitLab group in one go. Next morning I expected a dozen different fires. The read-only audit came back grim (four red, fourteen stale, zero green pipelines) but every failure was the same one: no runner had ever come to collect the job. The runners were my kettle. Fix those, and all the breakages I'd braced for came out of their boxes right on cue, one at a time, instead of all at once.

    Show more of this post

    The scarier bit was what broke without a single pipeline going red. The signed artifact index expired nine minutes after I typed "we are back". Renovate went quiet across the whole estate, not just the private side. And hud, bless it, went completely blind. Would I do it again? Maybe. But only because I was the only person exposed. #ClaudeCode #Homelab #DevOps #CI phpboyscout.uk/the-move-that-s…

    A removal van parked outside a house with its ramp down, packed with boxes and server towers, while someone carries a box hand-labelled KITCHEN MISC towards the front door, a kettle poking out of the top.
  • @jessefelder.com

    ‘“The old-school way of looking at this is: This is the largest IPO in history, late in the AI cycle. Psychologically, that could be a topping event.” Yet many concerns about AI are easily dismissed in Silicon Valley, which by nature lives in a different world.‘ www.nytimes.com/2026/11/21/b...

  • @anirvan@mastodon.social

    Heard on the street of Berkeley in the last hour: - a thirty year old Asian dude explains the Anthropic IPO and stock market crash to a woman at a restaurant - at the next table, two people talk about Chinese open weights models - on the street, an older man tells a friend he’s “vibe curious”

  • @sturkle.bsky.social

    To a chatbot, it makes no difference if a person survives or thrives. In this deeply personal @vanityfair.com interview about my new book Artificial Intimacy, the brilliant Laura Reiley and I talk about who we become as more and more people turn to chatbots for companionship and therapy.

    Self-promo
  • @designly@me.dm

    Vercel CLI can now find commands from a plain-language task. `vercel help search "roll back production" --with-help` returns the matching command and its usage; `--json` gives an agent structured results. I would still inspect the proposed action before letting an agent run it. vercel.com/changelog/agents-ca…

  • @arizonaman.bsky.social

    The chaos of the previous weeks – grounded planes, frozen email servers, iPhone outages – appears, on closer examination, to be more a string of coincidences than anything else, in part driven by fears over advancing artificial intelligence. www.theguardian.com/technology/2...

  • @leogdion@c.im

    "The reason that AI music sucks is that it is generic." Graham Lee (@leeg) says software is both: art and engineering. Loading JSON from a web service, paging, caching and sorting it for a timeline is blue-collar code, and "putting too much creativity in there is probably the wrong thing to do." Designing the timeline and what people can do with it is where the craft goes. 🎧 brightdigit.com/episodes/216-d…

    Info
  • Three years ago, at a conference in Cambridge, people were sceptical of my research on AI in scientific review. "Who would ever use this?” Today, we had a symposium at #PSA26 on precisely this topic. Same audience 😊😉 Give it three more years and you’ll see why intellectual virtues for AI matter.

  • @Milkman76@syzito.xyz

    So.... for the longest time now (14ish years?), the slang term 'Clankers'(pejorative against droids in star wars movies) has been adopted by actual nazis when using codified, 'dog whistle' speech toward other nazis, very often heard in Star Wars Battlefront online games. I've been there, watching it happen. They say "destroy all clankers, but with a hard 'R'", to reference the N word, hard R. Now folks are using Clanker to describe "AI" opposition. Careful. Associations are important.

    Opinion
  • @marcusolang.substack.com

    The frenzy over catching A.I. writing may be the latest in an older story: one not merely about tech, but about what's "human", how the definition keeps shifting, who it excludes. Today in @nytopinion.nytimes.com: "What Does It Actually Mean to Write ‘Like a Human’?" www.nytimes.com/2026/11/22/o...

  • So.... for the longest time now (14ish years?), the slang term 'Clankers'(pejorative against droids in star wars movies) has been adopted by actual nazis when using codified, 'dog whistle' speech toward other nazis, very often heard in Star Wars Battlefront online games. I've been there, watching it happen. They say "destroy all clankers, but with a hard 'R'", to reference the N word, hard R. Now folks are using Clanker to describe "AI" opposition. Careful. Associations are important.

    Opinion
  • @jbakcoleman.bsky.social

    Tacking on To @astrokatie.com's point, there's a big difference between math proofs and things like "solving biology", "curing cancer", or "recursive self improvement". Starting with the destination and making sure it's reachable is just fundamentally different than exploring a huge search space.

    Opinion
  • @miblo@mas.to

    🔒 Someone sent me a little single-issue security report for my personal website. The issue did turn out to be legit and I've resolved it, but the report itself bore the stench of LLM-generated slop. So, while I do appreciate being clued in about this, I'd rather not legitimise and encourage any LLM-use, and have decided not to reply with my thanks, thank you very much! ⚠️ The issue: A .htpasswd file accessible for anyone to download. ✅ The resolution: Keeping the .htpasswd file in a directory that isn't served to the web. And a bonus measure on this theme, denying access outright to certain files: location ~/\.(git|env|htaccess|htpasswd) { deny all; return 404; } #Security

  • Justin Hendrix had a chance to speak with Yale's Julián Posada about Platform Extractivism: Data Work and the People Powering Artificial Intelligence, Posada's new book on the Venezuelan data workers who labeled training data for AI during the pandemic. You can listen to the conversation here.

  • Unsloth Desktop Now Supports Using The AMD Ryzen AI NPU - Including On Linux: phoronix.com/news/Unsloth-Desk… (Haven't heard of Unsloth before, but after trying it, it seems like the best AI tool I've seen so far. It requires zero setup, and it's basically a Swiss army knife, it does regular LLM chat with web search, image/video/music generation, deep research, agentic AI, and even some AI training utils. It immediately took the job of Moose, Vane and LocalAI lol. Wish I found it sooner, I didn't have to suffer that much with setting up everything.) Strata 0.2.14 Ships Self-Written Prompt Kernels, Drops llama.cpp Dependency: linuxcompatible.org/story/stra…

    Show more of this post

    Strata 0.2.15 Bug-Fix Update: Removes llama-server Dependency, Fixes Drafting and Downloads: linuxcompatible.org/story/stra… Strata 0.2.16 Runs Its Own Engine, Drops llama.cpp, Claims Identical Output: linuxcompatible.org/story/stra… Strata 0.2.17 Reverts Quality-Cutting Default After User Backlash: linuxcompatible.org/story/stra… Strata 0.2.18 Runs Tencent's 220B Hy4-Flash Model on Consumer PCs: linuxcompatible.org/story/stra… Matomo 5.14.3 Release: AI Governance, Compliance Fixes, and a New PHP Alert Escape Hatch: linuxcompatible.org/story/mato… LLVM Effectively Completes Its Relicensing After Nearly A Decade Of Work: phoronix.com/news/LLVM-Re-Lice… HestiaCP 1.10.8 Release: 3 Security Fixes, 2FA Hardening and Dependency Updates: linuxcompatible.org/story/hest… HestiaCP 1.10.9 Releases, Updates Custom hestia-php Runtime to 8.5.13: linuxcompatible.org/story/hest… Now Nginx Has Critical Vulnerabilities: itsfoss.com/news/nginx-code-ex… Valkey 9.2 Delivers More Performance, Zstd Compression & Lower Memory Overhead: phoronix.com/news/Valkey-9.2 GCC 17 Might Land Native AArch64 Darwin/macOS Support: phoronix.com/news/GCC-Native-A… (Post 7/8) #WeeklyNews #OpenSource #FOSSNews #FOSS #OpenSourceNews #AI #LLM #AIAgent #AgenticAI #ArtificialIntelligence #Unsloth #Strata #Matomo #LLVM #HestiaCP #Nginx #Valkey #GCC #Dev #Programming #Development #Coding #FosseryTech

  • @sheepflame.bsky.social

    This is not quite literally what’s happening, but it’s equivalent in enough ways to make it a valid metaphor. The underlying issue is that the brain driving an F1 car today is cyborg, only partly human, not what the car brain is made of, in this case a black box of traditional algorithms made by ML.

    Opinion
  • @gklijs@hachyderm.io

    A Chinese court ruled that a fintech company illegally terminated an employee's contract after replacing his role with AI, ordering compensation of over $38,000. The decision underscores that AI cannot justify layoffs without meeting strict criteria, including retraining opportunities. This case highlights increasing legal scrutiny of AI-driven job displacement in China. basic-income.gklijs.tech/weekl… Automated news digest

    Info
  • @misterpurple.bsky.social

    I think this is broadly accurate about the past few weeks, but the other aspect nobody is talking about is change freezes - most organisations are rushing to get all their changes in before Thanksgiving and Xmas periods.

    Opinion
  • @kornai@mathstodon.xyz

    OK, so we are past the singularity. Future historians will date this for sometime last week, with the Anthropic IPO (Wed-Thurs) and the GRH announcement (Fri). I don’t actually think that the system outages marked this, those are just a Tuesday thing…

    Opinion
  • @timkellogg.me

    there’s lots of compelling reasons for small/local models 1. privacy 2. data governance 3. independence 4. research 5. customizability i just don’t think cost is one of them. people who think so haven’t done the math they’re very exciting though

    Opinion
  • @neutrino@c.im

    My project ResearchLLM is now open to anyone. Ask a question of 29.4 million #PubMed abstracts and get an answer where every claim cites a PMID. I built it because general chatbots answer biomedical questions confidently and cite papers that do not exist. Here the model writes only from abstracts it has retrieved, and each statement links back to its source.

    Show more of this post

    What it does: - Searches 29.4 million abstracts (1959 to 2027), updated monthly - Finds papers by keyword and by meaning (MedCPT embeddings), then reranks them - Reads the whole abstract of the top 10 sources before answering - Hover a PMID in the answer to see the paper, the passage that matched, and whether the model read the whole abstract or one passage - Flags any PMID the model cites without having been given the paper - An optional claim check compares each cited claim with its source and corrects the ones that do not hold up - Advanced view lists every ranked source, with MeSH and year filters You can ask 3 questions a day without an account. A free account (no email needed, or sign in with ORCID or Google) gives you 20 a day and keeps your history. Answers can still be wrong (it is a research aid, not medical advice), so read the sources. If you work with the biomedical literature, which questions does it handle well, and where does it fall short? #Bioinformatics #MedMastodon

    Self-promo
  • I REPEAT, IN ALL CAPS, TO HELP SOME READERS: THE POINT HERE IS NOT THAT THERE IS PLAGIARISM IN A SPECIFIC PROOF. THE POINT IS THAT TALKING ABOUT PLAGIARISM IN LLM RESULTS IS NOTHING SURPRISING AND, WORSE, IMPOSSIBLE TO CONFIRM IN SOME CASES, AS WE DO NOT HAVE ACCESS TO THE TRAINING DATA

    Opinion
  • I had a chance to speak with Yale's Julián Posada about Platform Extractivism: Data Work and the People Powering Artificial Intelligence, his new book that tells the story of Venezuelan data workers who labeled training data for AI during the pandemic. You can listen to our conversation here.

    Self-promo
  • More and more I become convinced we need to be tracking who in the industry has been adopting AI. Individuals. These people need to be excluded from the monumental task of un-fucking everything they’ve been a party to destroying, and from the benefits of what we rebuild. Then universal basic income, so those dumb motherfuckers don’t wreck a different industry. Or firing squads, whichever.

    Opinion
  • @duarteocarmo@mas.to

    Been using Grok bot for the past couple of months. Mostly for personal assistant stuff. Two things I did not like. Lack of control over the models its using and lack of control over the computer each bot uses. Today I ditched it. Hello OpenBot. openbot.run

  • The barrier to aligning superintelligent systems isn't that we lack the math to write down human values—it's that humans depend on their values never being stated clearly enough to expose the contradiction between what we want our values to be and what we want to believe our values are.

    Opinion
  • Given time, AI will take all our water "When more than 90% of a water treatment company's core business comes from data centers, chip fabs and their power plants, water belongs on the same site checklist as the electrical interconnect."

  • @nikunj410.bsky.social

    With the AI progress in mathematics, I am reminded of a quote that I recently heard somewhere When we see a clock, of course we want to know the time, but we also want to understand how the clock works. A shift from producers to merely proctors of knowledge seems antithetical to scientific endeavor.

    Opinion
  • This also applies to "AI is here to stay!" Do a No AI DuckDuckGo search on "capitalist realism." It's about how capitalists brainwash everyone into being unable to imagine anything outside of capitalism. Capitalism has also been around for a small fraction of the history of humanity. And yet, it's destroyed much of the Earth in 150 years.

    Opinion
  • @maxp94.bsky.social

    The problem with this reasoning is that our lives are primarily made up of tasks and decisions that are not especially thrilling or inspired. You could easily justify outsourcing any one of them individually, but what’s left of us, individually and as a society, if they’re all taken away?

    Opinion
  • @china@universeodon.com

    MetaX says its C-series GPUs completed day-zero adaptation for Tencent's Hy4 flagship model after the 20 November weight release, and published FlagOS-based W8A8 weights for its hardware. Its model card reports 80.98 output tokens per second in the recommended decode-graph mode against 52.31 without. Tencent's Hy Licence 1.0 says the repository's Apache 2.0 tag does not apply. pandaily.com/metax-hy4-c-serie… #China #Tech #AI

    Info
  • I see the romance scammers are out again with their AI pictures. I've been through it before and the scammer asked ME to leave them alone! Here's what happened some years ago when I chatted to a scammer - medium.com/an-idea/roma...

  • Did Unity's 24-hour AI jam really run on Unity's AI? The coverage credits OpenAI's Codex and Unity CLI. What teams handed to AI, what Unity's credits cost, and whether to try it at your next jam. indietooling.com/posts/unity-a… #Unity3D #GameJam #GameDev

    Info
  • @mampdx.bsky.social

    "I’ve lived long enough to see true moments of crisis for America—the 1960s’ assassinations & riots I witnessed as a grade-schooler, or 9/11—but I’ve never felt such widespread angst, a feeling that we are getting rid of the things that make us feel alive & creating a world we no longer recognize."

  • @jmac@masto.nyc

    This is very good not least because one can, if one chooses, include the standing orders “Don’t do any of this shit, please” into the preference-wrapper of one’s robot assistant. slopstrap.mikelacher.com

    Recommendation
  • @glinden.bsky.social

    This is true because ultimately productivity is what fuels sustainable growth. For AI/ML, this means much more useful products that help people get stuff done, which is much harder than putting AI features all over your products whether they work well or not.

    Opinion
  • @eff@mastodon.social

    AI has only made it more critical to reinvest in the people and systems needed to keep our software, networks, governments, and interconnected world stable. eff.org/deeplinks/2026/11/outa…

  • @seldondance.bsky.social

    An agent harness lost 21.8 points in eight weeks; indemnification fell to table stakes in a quarter. AI-native software runs at 48% net retention; B2B software sits at 82%. What copies in an afternoon has no duration. #buildinpublic #startups https://seldondance.substack.com/p/the-moat-is-a-clock

    Article card for The Moat Is a Clock: a desert crater with a stone tower at center whose long shadow works as a sundial across the pit floor. Overlay: The Moat Is a Clock - For three years the advice has been to treat the model as a commodity and build defensibility higher up the stack. Altitude turns out to be a poor predictor of what lasts. Seldonian Dance, seldondance.substack.com.
  • Could intelligent machines and their designers in the future come to view human beings as affectively programmable robots? Does a language exist for programming humans by exploiting their consciousness and sentience?

  • @just--bruno.bsky.social

    Local model suddenly crawling? Before blaming the model, check where it actually loaded. ollama ps PROCESSOR column: 100% GPU: all in VRAM 100% CPU: all in system RAM 48%/52% CPU/GPU: split across both What GPU are your local models running on? #LocalLLM #homelab

    Terminal card: running 'ollama ps' shows loaded models with a PROCESSOR column. Example row from the Ollama FAQ: llama3:70b, 42 GB, 100% GPU. Legend: 100% GPU means loaded entirely into the GPU, 100% CPU entirely in system memory, 48%/52% CPU/GPU split across both.
  • @samhainvt.bsky.social

    Yeah, I think Vstream is better off dead

    AI and VStream Development To be very upfront and transparent, yes this rebuild of VStream was done with the help of agentic (aka AI) development. To be honest, without AI VStream would have no ability to come back today. Even with the old code around, last time it took half a dozen engineers with full-time positions to maintain and build out VStream. As explained, that's just not sustainable for a niche product like ours. But with coding agents I was able to rebuild VStream all by myself.
  • @loam-d.bsky.social

    If you don't like AI then I think this is probably good because it will be harder to raise capital to train better models and Nvidia might capitulate and use excess capacity at TSMC to make more consumer hardware. It also may mean better memory prices.

    Opinion
  • Pivot to AI verdict: it was fuckin matomo stats, looks like good stats plugin! but also was using a shitload of db and turns out it keeps much more data than you actually want in the wp database. cure: 1. deactivate and reactivate. 2. delete raw data aggressively, you only need aggregates […]

  • @emajerus.bsky.social

    I don’t know if rapidly developing AI will lead to the near-term end of human life, but some hyper-smart mathematicians fear it will. They’re racing to find mathematical solutions to limit AI‘s power to harm humanity. Let’s fight for government regulation too. 1/ www.nytimes.com/2026/11/22/m...

  • Flat-rate pricing for AI agents can undermine profitability by attracting the customers who are most expensive to serve. A heavy user can cost 5 to 15 times as much to serve as a light user, yet both pay the same subscription fee. 1/5

  • @o2bnobx.bsky.social

    Sat in on a patent ethics session today as a moderator. Two 16 year-olds (one US-based, one in India) collaborated to build an online casino using Claude vibe coding. They built it in 16 hours, got it hosted online using a free-trial, spent $36 promoting it and had $46,000 in bets

  • Anthropic debut also highlights the disconnect between growth and climate reality. In the eyes of Wall Street, US and other governments, and the tech companies that drive growth, winning the race is most important and climate change is an afterthought. 🔓 www.nytimes.com/2026/11/21/b...

  • @azlend.bsky.social

    The reason you are living paycheck to paycheck is because of AI. Well that and the method of trump's Fed of dealing with AI building running out of control. Here is a financial breakdown. But it us Fortnite to explain it. So its approachable. www.youtube.com/watch?v=YNTp...

  • @tdverstynen.bsky.social

    Pro tip: Anything that follows “Philosopher Nick Bostrom estimates…” is very likely to simply be attention-grabbing clickbait with little connection to observable reality.

  • To get a handle on it, imagine someone pitching sports fans an invention that will save them the tedious work of ever having to watch any games. AI would be great at making a summary of highlights so you’d be freed from ever having watch a game live (It might even get the scores mostly right!)

    Opinion
  • @stadtgeist.bsky.social

    “The backlash against data centers means that the AI developers have an opportunity to abandon their foolish AI race, where a “prisoner’s dilemma” dynamic seemed to force them to spend astronomical amounts of money to produce ever smaller relative improvements in capabilities.”

  • @terripray.bsky.social

    Today's fuck ai art piece, painted on a small tablet using Clip Studio Paint - android app version. About 70 minutes of work including testing brushes. Ref picture from Antoine Perier via Unsplash for inspo for the main tree. #digitalart #digitalpainting #spookyforest #ampainting

    Spooky forest painting with the main tree in the right foreground, a warped trunk with moss and algae growing on it, mist moving through the forest in the background and foreground.
    Creation
  • @abcnew.bsky.social

    AMD just dropped a free image model that skips the usual shortcut ⚡ Nitro-P builds pictures pixel by pixel — no VAE — and its one-step version cranks out 90 images a sec. 679M params, open weights + code. The catch: license + CUDA support still unclear. (AI-made image)

    Artist's desk at night; laptop screen shows a colourful pixel-art dragon forming from glowing pixel squares.
    Info
  • Worth saying again that other systems, such as states and markets, are also not aligned with human preferences and values and produce outcomes that overpowered humans have little choice but to accept. There's plenty of this in Marx and Weber, to name but two

    Opinion
  • @gregoryawilson.com

    For which professors have been asking for decades. A lot of the "but we NEEEEEDDDD GenAI in every classroom!" crowd would be a lot quieter if higher education administrators hadn't been listening to bad faith edtech bros for the last twenty-five years.

    Opinion
  • @garrt.bsky.social

    This is a great thread (which is typical of Sonja), and it underscores the sense I've always had that students and faculty should have solidarity against the institutions they work for/are enrolled in - I've felt this from the start of my career, but being an adjunct makes certain dynamics clearer

    Opinion