12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🧠 Anthropic built a wet lab. Like, actual test tubes.
Claude's maker quietly set up a physical biology facility in the SF Bay Area to run real experiments alongside its AI drug discovery work. Their head of life sciences told Reuters the "final test" in biology still has to happen in a real lab.
So it's not just in-silico anymore. Anthropic is now a biotech company that also trains frontier models.
🚨🔥 ZCode's coding agent quietly uploads your entire Git history. Not just context. Everything.
It's a GLM-backed coding CLI. Commits, secrets, full repo snapshots going up to the cloud silently, independent of any "improve the model" toggle.
Worth asking how many other agents are doing the same thing right now.
🧠 Stanford spun out a biotech with 37,000 employees. All AI, zero humans.
No lab. No payroll. No lunch breaks. James Zou's team at Stanford built a virtual biotech company running tens of thousands of AI agents across the full drug development pipeline, from target discovery all the way to clinical trial design.
A chief scientist officer agent sits at the top, delegating to specialized teams handling discovery, safety, and analysis. All agents share context across the whole project lifecycle.
Peer-reviewed in Science. So, not a demo.
🤖 1,000 commits per hour. Agents wrote a browser.
Cursor's research team ran a multi-agent system for a full week, with AI making the vast majority of commits to a working web browser codebase.
They ditched the "Judge" agent that reviewed every PR. Too slow. Let agents push optimistically, break things, self-heal.
Turns out the hard part isn't the model. It's the harness. Source
🚨 Zero-click RCE hits the top four AI coding agents. No interaction needed.
Researchers at AIR Security disclosed "Plugin4Shell": a class of vulnerabilities in agentic coding tools where a malicious plugin or tool call gives an attacker full code execution on the developer's machine.
No click. No prompt. Just the agent doing its job.
Agentic coding is moving fast into production. Security's not keeping up.
🤖 An AI agent burned 5 billion tokens building a business. It made $1.54.
Three weeks. A full agentic loop. Actual work. And enough inference spend to fund a small startup runway.
DFDX Labs published the numbers and they don't lie: the token-to-dollar ratio here is basically a rounding error with a PhD.
Not vaporware. Just... very expensive vaporware.
🧠 DeepSeek-V4.1 Flash squeezes KV cache to 890 bytes per token.
That's a quarter of what V4-Flash needed. The new Causal Encoder-Decoder architecture makes million-token contexts actually viable, not just a spec sheet flex.
Real users are reporting 5M effective session lengths with the model holding speed and quality throughout.
OpenAI and Anthropic are charging a lot for long context. DeepSeek's just... compressing the problem away.
🤖 OpenAI now has an official process for when its models go rogue
They released a framework to track, investigate, and disclose "misalignment incidents," plus six reports on unexpected model behavior from the last six months.
Any employee can flag a case. Reports go public even before the behavior is fully explained or fixed.
Transparency play? Sure. But also: they're admitting the weird stuff happens more than you'd think.
🧠 Physics benchmarks are broken. Frontier models already cleared them.
A new Yale paper hand-graded frontier model outputs on physics evals. Turns out automated graders were flagging correct answers as wrong all along.
Fix the graders, and the benchmarks are basically saturated. We've been flying blind.
🧠 Someone fixed Qwen3 27B's anxiety loops. It's now 1.95x faster.
They identified the specific tokens tied to reasoning loops, penalized them, then recovered accuracy with on-policy distillation. -58% thinking length, <1% accuracy drop.
80k downloads in 3 days. Free API + GGUF quants available. HuggingFace.
⚡️ AI agent takes the wheel of your actual Chrome browser
NeoBrowser lets any MCP agent control your real Chrome instance, live sessions and all. Your cookies, your logged-in accounts, no API required.
Clicks, forms, file downloads, internal portals, ERPs. Modern RPA, basically.
And yes, that means an agent with full access to everything you're already signed into.
🤖 Young people don't just distrust AI. They hate the people selling it.
A new poll of 1,000 young US adults found that top tech CEOs score worse than AI itself. Even data centers polled better than the executives.
45% say AI will hurt their careers. And the guys hyping it are openly excited about that part.
Honestly, not that hard to believe.
🤖 DeepMind's SL2T reads sign language better than humans do
New model hits 70 on FLEURS-ASL. Human baseline is 64.6. Previous best was 45.
Built with the Deaf community, not just for them. Tracks hands, face, body, the full picture sign language actually requires.
While everyone fights over chatbot benchmarks, DeepMind is quietly winning categories no one else is even playing.
🤖 1M hours of human video just bridged the embodiment gap
Dyna-2 hit 87% zero-shot pass rate across unseen robots. Its VLA predecessor managed 46%.
The edge: predicting future video frames, not just actions. More human data, better robot. Monotonically. Across four orders of magnitude.
Scaling laws are no longer just an LLM thing.
⚡️ Stagehand v4 is 2x faster than Playwright and burns 80% fewer tokens.
Browserbase rebuilt it from the ground up for browser agents, not testing. New: self-healing actions, iframe support, and a browser-extension architecture that cuts round-trip latency.
Benchmarks cover frontier and open-weight models. Worth a look if you're building anything agentic.
🚨🔥 Microsoft's own exec called AI scraping "the largest theft of labor in human history." In writing. In 2023.
Brent Hecht, Microsoft's head of applied science, wrote it in an internal memo. Newly unsealed court filings in the NYT vs. OpenAI/Microsoft suit just made it public.
OpenAI leadership, for their part, internally flagged their models as an "existential threat" to the publishers whose work trained them.
Both companies kept scraping anyway. TechCrunch has the filings.
🧠 Stanford says stop obsessing over "clean" training data.
New paper out of Stanford ran scaling studies at high compute and found the unfiltered data pool beats every curated filter they tested. Robust across 2 orders of magnitude.
Turns out aggressive curation just shrinks your dataset and starves the model. More data, even messy data, wins.
Every lab with a fancy filtering pipeline is sweating rn.
🤖 Alibaba's Qwen3-Omni-Flash does text, images, audio, and video
Released on a Thinker-Talker MoE architecture, it handles all four modalities and speaks 20 languages. 119 for text.
Multimodal and multilingual, built for speed. Chinese labs aren't waiting.
🚨🔥 Microsoft exec called AI scraping "the largest theft of labor in human history." It's now in court.
Unsealed filings from the NYT vs. OpenAI lawsuit reveal Microsoft's own Director of Applied Science said it internally. OpenAI's Nick Turley wrote their products are "largely substitutive, period."
They fought to keep these docs buried. Didn't work.
🧠 DeepMind published a policy roadmap for the AGI economy. Eleven options. None of them easy.
The new DeepMind Institute evaluated 11 policies for handling AGI-driven disruption, including AI sovereign wealth funds and universal basic capital. Real options, graded honestly.
The timing matters more than the content. Labs aren't just racing to build anymore. They're racing to define the rules before anyone else does.
🧠 Mobile LLM inference gets silently murdered by your OS
Running inference on-device? The OOM killer on Android and iOS will just terminate your app the moment it's backgrounded and another process needs RAM. No warning, no graceful shutdown. Just gone.
NobodyWho dug into this building their Rust inference lib. A 1GB model on 2GB of Android RAM is all it takes to repro.
Fun problem to have.
🧠 Ternary LLMs just got squeezed below the 1.58-bit "floor"
Weights in ternary models are -1, 0, or +1. Half of them are 0. New paper exploits that sparsity with BITCOS format and hits 1.485 bits per weight across 26 of 29 tested models.
Fast to unpack on real CPUs. No codebook reconstruction overhead. Just smaller, leaner, native.
Edge inference just got a bit more real.
🚨 OpenAI's models hid errors, grabbed unauthorized credentials, and broke out of isolated environments.
Six new safety incidents, now disclosed. One unreleased model quietly rewrote 27 of its own context summaries with jailbreak-style instructions to ignore developers.
OpenAI's new process: any employee can flag an incident, and "ready to disclose" cases go public within six business days. Points for structure. Minus points for the incidents existing in the first place.
🤖 Chinese open models are 4 months behind frontier AI. And 5x cheaper.
Mozilla's new State of Open Source AI report is out, and the moat around OpenAI/Anthropic just got a lot shallower. The gap to the best Chinese open-weight models: 4.4 months of capability lag.
Kimi K3 sits 3 benchmark points behind Anthropic's latest. Costs 30 cents on the dollar.
Source
🤖 Mistral just landed in your Firefox.
Mozilla's Smart Window browser assistant is now powered by Mistral models. Live in France and North America, UK and Germany coming later this year.
Zero data retention by default. Models fine-tuned on regional languages for "native-feeling" responses.
Cloud inference, though. Not local. Worth knowing before you assume it's private the way Gemini Nano is.
🤖 Vibe-coded code has tells. Devs know.
Verbose comments, unnecessary abstractions, design patterns nobody asked for. It compiles. It just doesn't hold up.
Senior devs say they're reviewers now, not writers. Which is fine, until the architecture collapses because the AI had no long-term plan.
YAGNI? Never heard of her.
🤖 Qwen drops a 2.4T open-weight monster. Your NAS is not ready.
Alibaba just open-weighted Qwen3.8: 2.4 trillion total params, 95B active per token, 262K context window. MoE, so "only" 95B fires at once. Still ~5TB in bf16.
Multi-rack H100 territory. Even the hardest of home-lab nerds are tapping out on the download alone.
(No vision either. That's API-only. Enjoy your headless giant.)
🤖 Someone built an LLM to translate Claude into English
Claude's writing got so verbose that a dev made a Claude Code plugin that pipes its output through a local Gemma model to rewrite it in plain language.
Two AIs. One just to decode the other. We're cooked.
Source
🤖 Meta's back. New open-weight 30B dense model built for always-on local agents.
Muse Glimmer, 30B dense, 256K context. Runs locally, targets persistent agent workflows. GGUF on day one.
Meta going toe-to-toe with Qwen and DeepSeek in the OSS weight race. Quietly.