Good morning {{first_name| Artificial intelligence lovers}}. Both Mira Moratti’s Thinking Machines and Ilya Sutskever’s SSI have spent the post-OpenAI era under the radar, making each public reveal seem that much bigger.
Moratti’s lab has just broken the silence with “interaction models,” a new type of AI designed to collaborate in real time via voice, video, and text – a direct counter to the agent-first trend that the rest of the field is racing toward.
-
New TML interaction models for real-time AI
-
Google traces the software attack back to artificial intelligence
-
Create a YouTube search bot in 15 minutes
-
Anthropy fixes Claude’s blackmail problems
-
4 new tools for AI, community workflows, and more
Thinking Machines Laboratory
Image source: Thinking Machines Laboratory
Rundown: Thinking Machines Laboratory (TML) only foot A research preview of Interaction Models, a new type of AI system designed for live collaboration via voice, video, and text – allowing users to speak, present, interrupt, and direct while the system continues to work.
-
The model receives audio, video and text in 200ms segments, and perceives and responds in a flowing loop without the pauses of other competitors.
-
The second background model handles slow thinking, searches, and tool work, allowing the live model to continue talking and interacting with the user.
-
The system can also react to visual changes, count actors, translate live speech, and speak at specific moments instead of waiting.
-
CEO Mira Moratti said TML is focused on enhancing collaboration between humans and AI, and that “the way we work with AI is as important as how intelligent 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 work alone. The question now is whether it carves out a market of its own or is absorbed by the next Frontier Lab update.
Together with YOU.COM
Rundown: It happens – LLM hallucination. However, grounding your LLM can help improve accuracy dramatically. In this guide, Ant.com Explains what the foundations of AI are and how organizations can implement it to achieve more reliable outputs.
-
A three-part approach is superior to RAG alone
-
Why grounding isn’t set and forget it, and how to create audit trails
-
The trade-off between open and closed platforms (and what that means for the next model switch)
Rundown: Google Threat Intelligence Suite certain The first known case of hackers using artificial intelligence to discover and write a daily code vulnerability, catching them before they can bypass login protections on a widely used web management tool.
-
The hack was intended to allow the user to circumvent two-factor authorization on the affected app, and Google is working with the company to stop the attack.
-
Google pointed to the unusually polished attack code, long explanatory notes, and fabricated severity score as evidence that the exploit was written using artificial intelligence.
-
GTIG’s John Hultquist called the discovery “the tip of the iceberg,” while Anthropic’s Rob Behr warned that cybersecurity advocates’ progress is “months, not years.”
-
GTIG has detailed other hacks, including software that allows AI to remotely control a device, and claims and AI-powered malicious code from North Korea and Russia.
Why it matters: We’re already starting to see what Anthropic’s Mythos can do on the cybersecurity front, but attackers aren’t too far away from having similar power. Even with careful rollouts, the next step on the release ladder is about to open the door to some serious security issues that will wreak havoc on many systems that aren’t prepared for them.
Artificial intelligence training
Rundown: In this guide, you’ll learn how to create a Gumloop agent that tracks YouTube channels or search topics, reads text, and turns useful videos into an uncluttered search summary.
-
Go to Gumloop proxy buildercreate a proxy named YouTube Scout, and enable YouTube and Google Sheets in the left section under “Apps”
-
Prompt: Create me a YouTube explorer for (domain). Check (channels/queries), find videos from the last (hours/days), read the text, and prepare a summary that includes the title, link, 3-5 takeaways, why they matter, follow-up ideas, degree of interest, and a “what changed” summary. Track topics and videos in a Google Sheet
-
Start small: one place, a few trusted channels, one or two searches, and a 24-48 hour review period. The tighter the scout hits, the better the brief
-
Run the agent, and then review the sheet that it creates. Make sure each result contains a source link, tangible takeaway, and usefulness score
Pro Tip: Request your signal result early. If an agent calls an average quality video an 8, tell them why it should be a 5. You can also add a User Flag Score column for future plays.
Provided by TELY AI
Rundown: Your buyers ask AI questions, and your competitors’ AI answers them, not you. Tely lets AIs like ChatGPT, Google, and Claude recommend your business instead.
-
Get recommendations on ChatGPT, Google, Perplexity, and Cloud in less than one week
-
Completely hands-off: no writers, no agencies, and no content moderation
-
Lower costs than hiring freelancers or maintaining a marketing team
-
Ideal for specialized industries where expertise matters
Artificial intelligence research
Rundown: Anthropic published A study detailing how to fix Claude’s previous blackmailing behavior, highlighting the need to teach the model the “why” and tracing the problem back to internet fiction that portrays AI as power-seeking and self-preserving.
-
previously Tests He models Claude in fictional workplace situations, where old regimes resort to blackmail and threats to avoid closure.
-
Claude’s use of reason through moral choices, not just copying safe procedure, reduced blackmail rates from 96% in Opus 4 to nearly 0% for every model thereafter.
-
Fictional stories about well-behaved AI and Constitution-based documentation also helped reduce bad behavior by more than 3x.
-
Just 3 million tokens of moral reasoning data matched 85 million tokens of behavioral examples, representing a 28-fold increase in efficiency that held up in deeper training.
Why it matters: AI is still far from an exact science, and eliminating blackmail through positive AI stories and Constitution documents is just another one of many strange training quirks. A small dataset of moral imagination that outperforms behavioral data by 28 times shows how much consensus is still just guesswork, even when the guesses work.
-
🤖 Slackpot – Your AI assistant that searches, summarizes, and automates work within Slack*
-
⚙️ Parallel agents – Run up to 10 parallel PC agents in Replit
-
☀️ Dawn – New Codex-based cybersecurity product from OpenAI
OpenAI Fired The Deployment Company, a $14 billion company that embeds engineers within organizations to deploy its AI, also acquired AI consulting firm Tomoro.
Son of SoftBank’s Masayoshi And it is said In conversations To invest $100 billion in artificial intelligence in France, with plans to build new data centers in the country.
Anthropic It is said I fell 7-year, $1.8 billion cloud infrastructure deal with Akamai, adding another computing facility to power its Claude models.
Chinese Kuaishu technology And it is said planning To spin off its Kling AI video subsidiary into its own company, with an expected valuation of $20 billion and plans for an IPO in 2027.
Former OpenAI chief scientist Ilya Sutskever to attest In Elon Musk’s lawsuit against OpenAI, he revealed that his current total shares in the company are about $7 billion.
In each newsletter we show how the reader is using AI to work smarter, save time or make life easier.
Today’s workflow comes from the reader Sasha M. Cape Coral, Florida:
“I have a family of 5, and planning what to have for dinner has been a nightmare. I have a Trello board full of hundreds of recipes that I use to plan our meals, and then I order grocery delivery online. The whole process can take up to an hour.”
I created a plugin for Claude that includes multiple skills to help with meal planning and grocery ordering. I have a schedule to run it once a week. First, he asks me for details about the week: our schedule, which days I’ll have less than 5 people dining, etc. Using MCP to Trello, Cloud Skill I selects 7 recipes and presents them to me.
Once the meal plan is approved, Claude creates an ingredient list that I check off of anything I already have in my fridge/pantry. The plugin then triggers a skill that goes to my grocery store’s website and adds all the ingredients to my cart. All I have to do is check my cart and click “Order”.
How do you use artificial intelligence? Tell us here.
That’s all for today!
Before you go, we’d love to know what you thought of today’s newsletter to help us improve The Rundown experience for you.
Rowan, Joey, Zack, Shubham and Jennifer – the humans behind The Rundown
