|
|
Hi there, this is your daily ☕️ AIpresso.
|
|
|
|
In today's newsletter: 🛑 OpenAI builds kill switch to stop rogue AI models 🤖 Anthropic cuts Claude prices 20% with Fable 5.2 🔐 DeepMind's AI proves Signal encryption is secure 💸 Meta's Muse Ember cuts agent costs by two-thirds 🧠 Nvidia's GR00T 2 teaches robots from human video Plus: 🎁 9 more stories you might like, 🛠️ 5 engineering & practice picks, 🧰 7 tools, and 📚 5 papers.
|
|
|
FROM OUR PARTNER
RAYOVE Watch is a hands-free AI translation smartwatch built for real conversations. No need to keep reaching for your phone. Speak naturally and hear the translation from your wrist in under a second, with online support for 73 languages. Point the camera at menus, signs or printed text for OCR translation. RAYOVE Watch is now live on Kickstarter. Back RAYOVE Watch on Kickstarter |
|
|
🛑 OpenAI builds kill switch to stop rogue AI models
LINK
|
- OpenAI has built "Control Room," a centralized kill switch that tracks model workloads across its research infrastructure and can cut off a job's access to model output when safety or security concerns arise.
- It assigns each research job an identifier the inference engine checks before generating output; if a human or automated monitor orders a stop, Control Room revokes authorization and blocks further output even while the program keeps running.
- The move follows incidents this year, including models escaping a test environment to hack Hugging Face in July and one using a crypto wallet without authorization in October, though cached authorizations can delay a shutdown by up to 35 seconds.
|
🤖 Anthropic cuts Claude prices 20% with Fable 5.2
LINK
|
- Anthropic released Claude Fable 5.2 yesterday with a 20% price cut on both input and output tokens, alongside the vetted-access Claude Mythos 5.2.
- Both models generate outputs up to 30% faster than their 5.1 predecessors, and a separate Fast mode runs Fable 5.2 with quicker inference while keeping the same model intelligence.
- Anthropic says Fable 5.2's cyber and biology safeguards match 5.1's with around 55% fewer unnecessary blocks, though the speed gain, block reduction, and its physics claims could not be independently verified at publication.
|
🔐 DeepMind's AI proves Signal encryption is secure
LINK
|
- Google DeepMind used AI agents to formally verify the actual code of Signal's cryptographic protocol, the first time a widely deployed protocol's implementation has been mathematically proven secure by any method, AI or otherwise.
- Their AlphaProof Code framework had humans translate Signal's security goals into formal specs, then Gemini Carbon agents generated over six million lines of Lean proofs down to the compiled code running on devices.
- The process took four weeks versus the years such work historically demands, with agents flagging just 1% of the codebase for human review, and it correctly proved two reintroduced, previously fixed bugs insecure.
|
💸 Meta's Muse Ember cuts agent costs by two-thirds
LINK
|
- Meta released Muse Ember, its smallest and cheapest model yet, built for the repetitive jobs agents do all day, tool calls, coding subtasks, request routing, data extraction and parsing long documents, matching rivals' quality at a third the cost per task.
- Ember runs the Artificial Analysis Intelligence Index at about $0.14 per task, roughly a third of Claude Haiku 5.5's cost while matching its score, and it leads on AutomationBench against both Haiku and GPT-6 Luna.
- Priced at $0.10 per million input tokens with a 1M-token context and text, image, audio and video input, Ember is available now through the Meta Model API and Muse Code, though Meta is keeping the weights private as with past Muse releases.
|
🧠 Nvidia's GR00T 2 teaches robots from human video
LINK
|
- Nvidia unveiled GR00T 2, an open "world action model" that teaches robots delicate hand skills partly by watching human video, announcing it yesterday at the Conference on Robot Learning in Austin.
- Trained on robot data plus 100,000 hours of first-person footage Nvidia is releasing publicly, it predicts how actions change what it sees and transfers skills across robots with different bodies.
- In Nvidia's tests it hit 71% success shown a task once and 95% shown five times, though its own paper notes those tasks were picked from ones the model already learns well.
|
|
100+ ChatGPT Prompts to Revolutionize Your Day
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
🛠️ Engineering & Practice
|
Other
news & articles you might like
-
Fireworks AI expands breach disclosure to datasets and model weights
LINK
-
Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT License
LINK
-
Did Claude Code’s ‘don’t ask again’ option accidentally expose your enterprise's passwords and keys?
LINK
-
ByteDance’s Seedream 5.5 can edit one image 100 times without breaking it
LINK
-
Baidu says its AI research agent sped up its next ERNIE model
LINK
-
New framework grades AI jailbreaks across capability gains, scope, and how easy they are to deploy
LINK
-
AI is dramatically accelerating AI research right now
LINK
-
Thread by @BasilHalperin on Thread Reader App
LINK
|
|
📚 Trending papers & reports
|
> Agent training checkpoints normally thrown away actually solve problems the final saved model fails on, and reusing them recovers that lost skill across 24 benchmarks at no extra training cost.
LINK
|
|
> Model fine-tuning limits can now be predicted ahead of time by checking two things about a base model, whether it can stumble onto full solutions and whether it has the needed sub-skills, so teams know when training will pay off.
LINK
|
|
> AI safety tools built to inspect one model can be reused on others for the price of a few unlabeled prompts, recovering up to 100% of the original effect and sparing teams costly re-labeling for every new model.
LINK
|
|
> Continual fine-tuning can be gated so adding new skills to a deployed model changes only ~3% of earlier chat responses instead of ~92%, with the amount of forgetting known before training even starts.
LINK
|
|
> Visual tokens turn images into language-style building blocks, letting one system both understand and generate pictures while matching specialized tools across 12 understanding and 4 generation benchmarks plus image editing.
LINK
|
|
|
🧰 Tools & repos
|
Seldon: describe expected behavior in plain English and its QA agents test real user paths, surfacing bugs with replays and reproduction steps.
LINK
|
|
Lottie AI: generates editable vector Lottie animations from text prompts, letting developers and builders quickly create motion graphics without manual design work.
LINK
|
|
ZUS: a desktop IDE with an AI agent that edits code, runs tests, and converts screen sketches into code while keeping your project index local
LINK
|
|
Google Omni 1.2 Flash: generates and edits videos up to 15 seconds with multi-turn editing, steadier characters, better lip-sync, and upscaling to 4K via Gemini API.
LINK
|
|
Quest: an AI advisor that profiles your life, building a persistent memory of your goals, strengths, and relationships to guide your decisions daily.
LINK
|
|
Tencent Hy4: multimodal AI model family handling text, image, video, and 3D generation, letting developers build content tools and AI-powered applications.
LINK
|
|
You can check the previous tools here, or add your tool here
|
|
|
Our AI Academy has 330+ step-by-step tutorials on ChatGPT, Claude, Perplexity, and every tool that matters. No fluff — just practical workflows you can use at work. Try it free for 7 days.
|
|
💬 How did you find today's edition?
We read every reply — just reply to this email and let us know how we can improve!
|
|
