Good morning, {{ first_name | AI enthusiasts }}. Both Mira Murati's Thinking Machines and Ilya Sutskever's SSI have spent the post-OpenAI era mostly out of view, making every public reveal feel that much bigger.
Murati's lab just broke the silence with 'interaction models,' a new type of AI built for real-time collaboration across voice, video, and text — in a direct counter to the agentic-first direction the rest of the field is racing toward.
In today’s AI rundown:
TML’s new interaction models for real-time AI
Google traces software attack back to AI
Build a YouTube research bot in 15 minutes
Anthropic fixes Claude's blackmail problems
4 new AI tools, community workflows, and more
LATEST DEVELOPMENTS
THINKING MACHINES LAB
Image source: Thinking Machines Lab
The Rundown: Thinking Machines Lab (TML) just introduced a research preview of interaction models, a new kind of AI system built to collaborate live across voice, video, and text — letting users talk, show, interrupt, and steer while the system keeps working.
The details:
The model takes in voice, video, and text in 200ms chunks, perceiving and responding in a streaming loop without the turn-taking pauses of other rivals.
A second background model handles slower reasoning, searches, and tool work, allowing the live model to keep talking and interacting with the user.
The system can also react to visual changes, count reps, translate live speech, and speak up at timed moments instead of waiting.
CEO Mira Murati said TML is focused on advancing human-AI collaboration, and that “the way we work with AI matters as much as how smart it is.”
Why it matters: Murati's TML has been fairly quiet since its inception, but interaction models are one of the lab’s first big differentiators: models designed around how people naturally work together, not how long an agent can run solo. Whether it carves out its own market or gets absorbed by a frontier lab's next update is the question now.
TOGETHER WITH YOU.COM
The Rundown: It happens—LLMs hallucinate. Grounding your LLM, however, can help dramatically improve accuracy. In this guide, You.com explains what AI grounding is and how organizations can implement it to achieve more reliable outputs.
The playbook covers:
A three-part approach that outperforms RAG alone
Why grounding isn't set-and-forget, and how to build audit trails
The open vs. closed platform trade-off (and what it means for your next model switch)
Image source: Google
The Rundown: Google's Threat Intelligence Group confirmed the first known case of hackers using AI to discover and write a zero-day software security flaw, catching them before they could break past login protections on a widely-used web management tool.
The details:
The hack was intended to allow the user to get around two-factor authorization on the affected app, with Google working with the company to stop the attack.
Google pointed to unusually polished attack code, long explainer notes, and a made-up severity score as clues that the exploit was written with an AI.
GTIG's John Hultquist called the find "the tip of the iceberg," with Anthropic's Rob Bair warning cybersecurity defenders' lead is "months, not years.”
GTIG detailed other hacks, including software that lets AI remotely control a device, and AI-assisted malicious prompts and code from N. Korea and Russia.
Why it matters: We’ve already started to see what Anthropic’s Mythos can do on the cybersecurity front, but attackers aren’t too far off from having similar power. Even with careful rollouts, the next step up the release ladder is about to open the door to some serious security issues that will cause chaos for the many systems not ready for it.
AI TRAINING
The Rundown: In this guide, you will learn how to build a Gumloop agent that tracks YouTube channels or search topics, reads transcripts, and turns the useful videos into a ranked research brief.
Step-by-step:
Go to Gumloop agent builder, create an agent named YouTube Scout, and enable YouTube and Google Sheets in the right-hand section under "Apps"
Prompt: Build me a YouTube scout for (niche). Check (channels/queries), find videos from the last (hours/days), read the transcript, and return a brief with title, link, 3-5 takeaways, why it matters, follow-up ideas, usefulness score, and a "what changed" summary. Track topics and videos in a Google Sheet
Start small: one niche, a few trusted channels, one or two searches, and a 24-48 hour lookback window. The tighter the scout's beat, the better the brief
Run the agent, then review the Sheet it creates. Make sure each result has a source link, concrete takeaways, and a usefulness score
Pro tip: Dial in the signal score early. If the agent calls a mediocre video an 8, tell them why that should be a 5. You can also add a User Signal Score column for future runs.
PRESENTED BY TELY AI
The Rundown: Your buyers are asking AI questions — and AI is answering with your competitors, not you. Tely makes AI like ChatGPT, Google, and Claude recommend your business instead.
With Tely AI, you can:
Get recommended in ChatGPT, Google, Perplexity, and Claude in as little as 1 week
Fully hands-off: no writers, no agencies, no managing content
Costs less than hiring freelancers or maintaining a marketing team
Ideal for niche industries where expertise matters
AI RESEARCH
Image source: Anthropic
The Rundown: Anthropic published a study detailing how it fixed Claude's previously seen blackmail behavior, highlighting the need to teach the model “why” and tracing the problem to internet fiction that depicts AI as power-seeking and self-preserving.
The details:
Earlier tests put Claude models in fictional workplace situations, with older systems resorting to blackmail and threats to avoid shutdown.
Having Claude reason through ethical choices, not just copy the safe action, cut blackmail rates from 96% in Opus 4 to nearly 0% for every model after.
Fictional stories of well-behaved AI and constitution-based documents also helped reduce bad behavior by more than 3x.
Just 3M tokens of ethical reasoning data matched 85M tokens of behavioral examples, a 28x efficiency gain that held up in deeper training.
Why it matters: AI is still far from an exact science, and eliminating blackmail via essentially positive AI stories and constitution docs is another one of the many strange training quirks. A small dataset of ethical fiction outperforming 28x the behavioral data shows how much of alignment is still guesswork, even when the guesses work.
QUICK HITS
🤖 Slackbot - Your AI assistant that searches, summarizes, and automates work right inside Slack*
❤️ Lovable Aesthetics - Vibe coding with more control over layout, typography
⚙️ Parallel Agents - Run up to 10 parallel computer-use agents in Replit
☀️ Daybreak - OpenAI’s new Codex-driven cybersecurity product
*Sponsored Listing
OpenAI launched “The Deployment Company”, a $14B business to embed engineers inside enterprises to deploy its AI, also acquiring AI consulting firm Tomoro.
SoftBank’s Masayoshi Son is reportedly in talks for a $100B AI investment into France, with plans to build out new data centers in the country.
Anthropic reportedly signed a 7-year, $1.8B cloud infrastructure deal with Akamai, adding another compute avenue to power its Claude models.
China’s Kuaishou Technology is reportedly planning to turn its Kling AI video branch into its own company, with a projected valuation of $20B and plans to IPO in 2027.
Former OpenAI Chief Scientist Ilya Sutskever testified in the Elon Musk vs. OpenAI lawsuit, revealing his current shares of the company total nearly $7B.
COMMUNITY
Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.
Today’s workflow comes from reader Sasha M. in Cape Coral, FL:
"I have a family of 5, and planning what we have for dinner was a nightmare. I have a Trello board full of hundreds of recipes that I use to plan our meals, and then I would place a grocery delivery order online. The whole process would take up to an hour.
I built a Claude plugin that includes multiple skills to help plan meals and order groceries. I have it on a schedule to run once a week. First, it asks me for details about the week: our schedule, any days that I'll have fewer than 5 people eating, etc. Using an MCP to Trello, the first Claude Skill picks out 7 recipes and presents them to me.
Once I've approved the meal plan, Claude then creates an ingredients list that I check off anything I already have in my fridge/pantry. The plugin then runs a skill that goes to my grocery store website and adds all the ingredients to my cart. All I have to do is check the cart and click ‘Order.’"
How do you use AI? Tell us here.
Read our last AI newsletter: Google’s powerful new AI co-mathematician
Read our last Tech newsletter: 'RAMageddon' is coming for your laptop
Read our last Robotics newsletter: Figure’s robots make a bed together
Today’s AI tool guide: Build a YouTube research bot in 15 minutes
That's it for today!
See you soon,
Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown
