These are the 12 stories I found this week that did not make it into the video. Each one has what happened, why it matters, and the original source so you can read more.
Story 01 of 12
OpenAI says an investigation found they broke its rules on sensitive information. It announced this on Oct 9.
The three say they were fired for speaking up about safety. OpenAI says it was not about raising safety concerns.
Why it mattersA public fight over whether people who raise safety concerns are protected inside a top AI lab.
Story 02 of 12
It listens through your mic and system audio, so it works on any call or in person.
If someone asks a question, it answers from your own files like PDFs and Google Docs. Everything stays on your device.
Experimental, Mac only, optimized for Apple Silicon.
Why it mattersMeeting notes with no subscription and no audio sent to the cloud.
Story 03 of 12
722 manuscripts in 372 result families, all from an internal model.
Many proofs are written in Lean, which lets a computer check every step. Not every result is verified yet. Each took about 3 hours of ChatGPT Pro compute.
Why it mattersHard problem solving is starting to scale with compute, not only with human experts.
Story 04 of 12
Only about 52 billion parameters are active per answer, which keeps it fast.
It is in public preview on Mistral Studio now. Free weights land at the end of this month.
Mistral says it beats every open model built outside China.
Why it mattersMost of the strongest open models came from China. This is a serious third option you can soon run yourself.
Story 05 of 12
Ultrafast mode is up to 8x faster than Standard in the API, Codex and ChatGPT Work.
It costs $12 and $60 per million input and output tokens. Standard is $2 and $10.
OpenAI also opened a public beta of the Decisions API for quick structured choices, up to 10x faster than GPT-6 Luna.
Why it mattersSpeed is becoming a paid tier. Coding tools and agents feel snappier if you pay.
Story 06 of 12
A 740 million parameter model that puts text, images, audio, video and code into one search space. It runs fully on your device.
Apache 2.0 license. Weights are on Hugging Face and Kaggle, and it works with Ollama, LM Studio, vLLM and llama.cpp.
Why it mattersPrivate, offline search on phones and laptops with no data sent to the cloud.
Story 07 of 12
Vetted security professionals get stronger cyber abilities with fewer blocks. There are three tiers: Defense, Red Team and Specialized.
Partners in the earlier program found over 129,000 verified vulnerabilities between April and July 2026.
Why it mattersPowerful AI is being handed out by trust level. Verified defenders get the sharper tools, the public does not.
Story 08 of 12
Biohub, the US Department of Energy, the NIH and new partners are committing $1.8 billion. Google DeepMind, Isomorphic Labs and Meta put in $300 million of it.
The data will train AI that predicts how cells respond to changes. The first dataset is expected in about a year.
Why it mattersAI drug discovery needs cell data that does not exist yet at this scale. This pays to create it.
Story 09 of 12
Beam has 501 billion parameters, 23 billion active. Weights and the technical report are due later this month.
Reflection says it scores comparable to GLM-5.2 on advanced reasoning with 3 to 4x less compute. By its own numbers, bigger models like Kimi K3 are still ahead.
These numbers come from Reflection and are not independently verified.
Why it mattersA US open-weight model aimed at China's best, pitched on cheap running costs.
Story 10 of 12
Etched is reviewing offers at $40 to $50 billion. It raised at $21 billion only two months ago.
It says it has $1 billion in customer orders and claims its inference chips beat Nvidia on speed and cost.
Based on unnamed sources. Etched declined to comment.
Why it mattersInference is where AI money gets spent every day. A real Nvidia challenger would reprice the whole stack.
Story 11 of 12
Instinct, valued at $10 billion, works even if your friends have not joined Instinct. It is pitched for trips, tickets, fantasy leagues and carpools.
Your personal agent asks permission before sharing anything with the group. Early access first, then everyone.
Why it mattersAgents are moving out of solo chats and into the group chat, where real plans get made.
Story 12 of 12
He posted on X that Grok Bot will pick the best model per task: Claude Opus 5.5 for text, Midjourney for images and Suno for music.
This is Musk's claim, not a confirmed product. There are no details on how routing works, and Suno has no public API.
Why it mattersEven xAI is admitting that one model cannot win every task.