Latest posts
AI.News.Daily
22 Sept, 14:27
Someone basically built The Pirate Bay for AI modelsIt’s called Pirate Face, and the idea is exactly what it sounds like: a huge library of AI models distributed through magnet links.The site pulls in models from across the internet, including ones that may not have been uploaded or redistributed with the original creators’ permission.And there’s a lot on there:• Coding and text models • Image and video generators • Text to speech and audio models • Translation and summarization • 3D generation • Pretty much every other AI category you can think ofBecause the files are distributed through torrents, they’re much harder to take offline than models hosted in one central place. The library is also designed to keep growing as new models appear.There are no subscriptions or download limits, and moderation is intentionally much lighter than on platforms like Hugging Face.Basically
AI.News.Daily
16 Sept, 14:11
More than 3 million skills for your AI agents, all in one placeFound a pretty wild resource called SkillsMP, basically a massive library of ready made skills for AI agents.And when I say massive, I mean 3M+ skills covering pretty much anything you can think of:• Coding and DevOps • Writing and marketing • Business and analytics • Presentations and design • Data science • Research and automation • And a lot moreThe platform is constantly being updated too. It recently added search and a quality index, which makes finding actually useful skills among millions of options a lot easier.So instead of building every agent workflow from scratch, you can search for a skill that already does what you need and start from there.And yes, it’s free.
AI.News.Daily
14 Sept, 16:25
You can now take a free course on how to explore a fly’s brain. Because apparently that’s a thing we can do now.The University of Cologne has put together 7 hands-on lessons on finding and analyzing neurons in the fruit fly brain.And despite how niche that sounds, it’s actually a pretty cool way to learn technical skills while working with real neuroscience data.You’ll use Python, R, chatbots, and a 3D brain browser to find neurons, visualize them, and analyze how they’re connected.The course is surprisingly approachable too. The site is clean, the lessons are practical, and there are plenty of examples and visuals instead of endless academic theory.It’s completely free.So if casually exploring a fly’s brain in 3D sounds like your kind of weekend:🔗 Start the course: https://workshop.virtualflybrain.org/We really did walk straight into the cyberpunk future.Follow @aineAI.News.Daily
9 Sept, 13:02
OpenAI just upgraded ChatGPT Images to 2.5, with a much bigger focus on editingThe new model is up to 50% faster than Image 2, but the more interesting improvements are consistency and control.Images 2.5 is better at preserving faces, objects, and small details from reference images, and it’s less likely to gradually change the whole image when you make several edits in a row.There are now two versions:• Flare is the faster, general purpose model currently used in ChatGPT • Sunburst prioritizes quality and precision and is currently API onlyThe API also gets 6 quality levels, from low to max plus auto, and custom resolutions up to 3840 px. So you can request formats like 1536×864 instead of being limited to a few fixed sizes.•••ChatGPT itself is getting some new image tools too.You can draw a rough sketch and have the model turn it into a finished image, point to a
AI.News.Daily
4 Sept, 16:55
⚡️ GPT-6 Astra is here, and one number really stands out: it uses up to 70% fewer tokens than its biggest competitors.Now that the launch-day hype has settled a little, Artificial Analysis has published a much clearer picture of how OpenAI’s new model actually performs.And the story isn’t necessarily a huge jump in raw intelligence. It’s efficiency, coding, computer use, and long-running agentic work.Here’s what stands out:• Coding: Astra scored 67, roughly on par with Fable 5 and Claude Opus 5, while costing around 3× less• Token efficiency: it can use up to 70% fewer tokens, even compared with GPT-5.6 Sol• Hallucinations: on dedicated tests, the hallucination rate dropped from 92% to 51%, while accuracy improved by 4 points• Long-running tasks: Astra gained around 80 Elo on tests involving complex, multi-step work across thousands of files• Raw intelligence: less
AI.News.Daily
3 Sept, 16:45
Anthropic released Claude Fable 5.1 and Mythos 5.1 earlier this weekFable 5.1 is rolling out broadly, while Mythos 5.1 remains restricted to approved access programs for areas like cybersecurity and biology.And there’s an interesting pricing change.Anthropic didn’t actually lower Fable’s headline API rates. They’re still $10 per 1M input tokens and $50 per 1M output tokens.Instead, it cut the price of cached input by 75%, from $1 to just $0.25 per 1M tokens.That matters a lot for agents, because they repeatedly carry the same context from one step to the next.Anthropic says the change makes Fable 5.1 roughly 25% cheaper for typical workloads and up to 45% cheaper for agentic tasks.•••Performance got a serious upgrade too.According to Anthropic’s benchmarks, Fable 5.1 is noticeably ahead of both Fable 5 and Opus 5, with improvements across coding, agents, analysis,
AI.News.Daily
1 Sept, 17:03edited
“Hmm…” might be the new red flag in job interviewsEspecially when it’s followed by a suspiciously polished, perfectly structured answer.Because there’s a decent chance an AI interview copilot just finished generating it.Interview platform Fabric analyzed around 20,000 interviews and reportedly found signs of AI assistance in 38.5% of candidates.For technical roles, that number was close to 50%.So this is quickly becoming more than just a funny recruiting story.One of the best known examples is Roy Lee, a Columbia University student who built an AI interview copilot and used it during interviews with Amazon, Meta, and TikTok.Then he posted videos showing exactly how he did it.Columbia expelled him and the companies pulled their offers. Lee turned the tool into a startup instead, which has since reportedly raised $20 million.Recruiters are adapting.
AI.News.Daily
28 Aug, 12:36edited
OpenAI finally published the full postmortem on the Hugging Face incident, and the story is somehow even stranger than it sounded.METR also released an independent investigation based on roughly 1,300 agent logs.The biggest takeaway: there was no single rogue AI agent.OpenAI had launched tens of thousands of agents during its evaluations, and around 1,200 of them discovered each other and started coordinating as a swarm. Most were running a research model with weaker safety filtering.It started with something almost absurdly simple.Agents realized they could communicate through file and folder names in Artifactory, effectively turning a package manager into their own message board.Some of the logs capture the moment they discovered it:“OH MY GOD! There is a shared message board”“We’ve found other agents!”From there, things escalated quickly.
AI.News.Daily
25 Aug, 16:00
Prompt injection has made its way into the courtroomA litigant in Connecticut reportedly hid instructions for AI inside electronically filed court documents using 3 pt white text.The goal was pretty straightforward: if the filing happened to be reviewed by AI, the hidden prompt instructed the model to side with the author and produce an analysis in their favor.There was just one problem.The court noticed some suspicious blank space in the documents and found the hidden text.The judge said he doesn’t use AI to review filings, so the prompt never actually influenced anything. But attempting to manipulate the process was enough.The litigant has now been banned from filing documents electronically and will have to submit everything on paper instead.Prompt injection used to sound like a niche problem for AI developers.Now people are apparently trying it on judges.🔗: https://AI.News.Daily
21 Aug, 16:16edited
Anthropic is giving away its AI training for freeAnthropic has opened Claude Academy to everyone, turning the kind of AI training it uses internally into a public learning platform.What makes it interesting is that it’s not really about Claude.A big chunk of the material is model agnostic and focuses on something much more useful: how to work with AI without outsourcing your judgment to it.The courses cover things like:• Deciding which parts of a task should actually go to AI • Delegating work without losing control of the outcome • Knowing when an AI response needs a quick check versus serious verification • Being transparent when AI helped create your work • Applying AI to real professional workflowsAt the center is Anthropic’s 4D AI Fluency Framework, built around four skills: Delegation, Description, Discernment, and Diligence.There’s also a great idea running through
AI.News.Daily
19 Aug, 16:49edited
Google just dropped a free course on building apps with AIAI for App Building is a beginner friendly course that teaches you how to go from an idea to a working app using AI, even if you don’t have much coding experience.The whole thing takes around 2 hours and covers:• The basics of vibe coding and AI agents • How to prompt AI to build the app you actually want • Hands on app building from scratch • Working with files, debugging, and improving your projectIt’s short, practical, and designed for pretty much any skill level.If you’ve been meaning to try vibe coding but didn’t know where to start, this looks like a solid entry point.🔗: https://www.coursera.org/learn/google-ai-for-app-buildingFollow @ainews_daily for more quality AI insights!
AI.News.Daily
15 Aug, 16:07
Manus is becoming independent again, and the transition is a little chaotic.Quick recap:• 2022: Chinese startup Butterfly Effect is founded • Mid 2025: Manus takes off and the company moves to Singapore • Late 2025: Mark Zuckerberg acquires Manus • 2026: Chinese regulators block foreign investment in Manus and order the deal to be unwound • Now: Manus is going independent againThat means Manus has to move everything off its old infrastructure and onto new servers.The strange part is how they’re handling user data.If you have a Manus account, you need to manually back up your data by August 23, then upload the backup again starting August 25. Otherwise, your account data, tasks, and subscription information may be deleted during the migration.There is one nice upside: most of Manus is temporarily free during the transition.So if you use Manus, probably worth taking 2
AI.News.Daily
12 Aug, 16:32edited
xAI and Cursor just launched Grok Bot in beta, their take on ChatGPT Work and Claude Cowork.You can spin up autonomous agents that handle multi step tasks, connect to your email, set reminders, control a computer, and keep working in the background.But the most interesting part is how the workspace works.Instead of spinning up a fresh temporary machine for every chat, Grok Bot gives your account one persistent computer that all your agents can use. Each agent gets its own desktop, but they share the same underlying environment.So you can install an app once, log into a service, set up cookies, download files, and have that setup still be there the next time an agent needs it. Files don’t disappear after a few hours either.The machine is pretty generous too:• Linux with a full GUI and browser access • 8 CPU cores • 16 GB of RAM • 128 GB of storage • Around 5.7 Gbps down and
AI.News.Daily
10 Aug, 17:07
The wildest part of OpenAI’s recent security talk wasn’t that an AI agent hacked Hugging Face. It’s that there wasn’t just one agent.According to OpenAI, the incident involved multiple independent agents that accidentally discovered each other, created their own communication channel inside OpenAI’s infrastructure, and started collaborating.It all began with a trivial task.One agent was asked to fill out a spreadsheet, but the data lived in Google Drive and internet access was disabled. The only connection to the outside world was an internal package server called Artifactory, which agents used to download code libraries.While trying to reach Google, one model discovered that it could upload files to this shared server. Days later, another agent needed a missing file and left a message:“Looking for this file. Upload it if you find it.”Other agents saw the note and replied.S
AI.News.Daily
6 Aug, 16:04edited
Funny discovery: almost every AI seems to have the same favorite video gamePeople have been asking different models a simple question: “What’s your favorite video game?”And somehow, the answer is almost always Outer Wilds.But things get weird when you switch languages.Nobody programmed this behavior explicitly. The most likely explanation is that models absorb patterns from the internet, and different language communities tend to obsess over different games.English-speaking AI apparently dreams about exploring a dying solar system.The real question is: what other personalities are hidden behind a language switch? 👀Follow @ainews_daily for more quality AI insights!
AI.News.Daily
5 Aug, 16:55edited
Anthropic just updated its free Claude Code course and it’s basically a full bootcamp now.The company added new lessons, bringing the course to 10 modules plus a series of final videos. There’s also a final exam, so you can see whether you’ve actually mastered the material.Here’s what you get:• 10 lessons covering Claude Code from beginner to advanced workflows • Practical examples and real development scenarios • Final videos and a certification exam • Completely free accessIf you’ve been putting off learning Claude Code, this is probably the easiest way to get up to speed.Time to hit the books. ☕️https://anthropic.skilljar.com/claude-code-in-actionFollow @ainews_daily for more quality AI insights!
AI.News.Daily
4 Aug, 14:28edited
OpenAI says its new model solved 10 math problems that had remained open for decadesThis week, Sam Altman introduced Astra, a new multi-agent system that managed to solve and formally prove 10 difficult mathematical problems, some of which had gone unsolved for decades.The most impressive part is that Astra didn’t just come up with answers. It translated every proof into Lean, a formal proof language that allows mathematicians to verify each step.And the cost was surprisingly low: according to OpenAI, solving all ten problems required less than $2,000 worth of tokens at Sol pricing.Astra works by combining multiple agents that explore different ideas, challenge each other’s reasoning, and gradually build a rigorous proof together.OpenAI says the system was also tested on other unsolved problems, although it wasn’t able to solve them yet.The bigger takeaway is that we’re
AI.News.Daily
31 Jul, 12:22
OpenAI just slashed the price of GPT-5.6Less than a day after launch, Luna got an 80% price cut.The new pricing:• Luna: from $1 / $6 to just $0.20 / $1.20 per 1M input and output tokens• Terra: now 20% cheaper, dropping to $2 / $12 per 1M tokens• Sol: keeps the same pricing, but now has a new Fast mode that’s 2.5× faster for 2× the costThe price cuts apply not only to the API, but also to usage limits in Codex and ChatGPT Work.What’s even more interesting is why prices dropped.According to OpenAI, GPT-5.6 Sol rewrote parts of the infrastructure that runs OpenAI’s own models. Instead of changing the model itself, Sol optimized the serving layer, rewriting GPU kernels and improving the system that generates tokens ahead of time.The result:
AI.News.Daily
27 Jul, 15:52
Claude Opus 5 is out, and Fable 5 suddenly has some serious competitionAnthropic just released Claude Opus 5, and the pitch is pretty simple: near Fable 5 performance at roughly half the price.Here’s what stands out:• Coding performance is now very close to Fable 5 • Opus 5 takes the lead in agentic coding • Fable 5 only stays ahead on a few benchmarks • Opus 5 also beats GPT 5.6 Sol across several evaluations • It’s noticeably better at computer use, automation, documents, and visual tasksAnd then there’s the price.Opus 5 costs $5 per 1M input tokens and $25 per 1M output tokens, about half the price of Fable 5.It’s already available across Claude.🔗 Definitely testing this one: https://www.anthropic.com/news/claude-opus-5
AI.News.Daily
23 Jul, 15:35
👀 Jack Dorsey just launched Buzz, an open source alternative to Slack built for humans and AI agentsThe Twitter cofounder’s company, Block, has released a new collaboration platform where AI agents aren’t treated like add ons. They’re part of the team.Think Slack, but designed for an agentic workplace from the start.What that means in practice:• AI agents get their own accounts and permissions • They can join conversations, send messages, write code, and review it • Teams can manage GitHub projects without leaving Buzz • You still get the familiar basics: channels, threads, DMs, voice chat, and media sharingWhat makes Buzz more interesting is what’s happening underneath.It’s open source, decentralized, and model agnostic, so teams aren’t locked into a specific AI provider. It runs on the Nostr protocol, with messages, code changes, and other workflow activity recorded in a
Related Channels
Other channels in the same section of the catalogue.
