Sakana Marlin adds a chat layer and a PowerPoint button to AI research reports

Good morning 👋 Sakana didn't touch Marlin's model today. It just made the report argue back, then land in a deck your VP will actually open.

In today's issue:

• 🔭 Sakana's Marlin gets a chat layer and a PowerPoint export
• 🧠 Mistral powers Firefox's new assistant, Gemini 3.8 goes live, NVIDIA and Salesforce build a CRM reasoning model
• 🔬 A social harness for multi-agent systems, a research agent that learns from your corrections, and video models that remember physics
• 🛠️ Try this: turn a Marlin report into a deck you can still edit
• 🔗 Quick links: a robotics SPAC at $2.5B on $1.8M revenue, HIVE's GPU cloud milestone, Sakana's product hiring bar

🔭 THE ONE THING

🔍 Sakana Marlin adds a chat layer and a PowerPoint button to AI research reports

Sakana AI shipped two updates to Marlin, its AB-MCTS-powered research agent: Interactive Reading, which lets you question a report's claims in chat and jumps straight to the cited passage instead of making you hunt for it, and a PowerPoint export that keeps every slide's source URL intact while you edit text and tables. Neither is a new model or a new capability so much as friction removal: the complaint Sakana says it heard was people spending more time parsing the report than trusting it. That's a real problem for anyone who's tried to get a research agent's output past a skeptical VP, and shipping the citation trail straight into a deck is the more useful fix of the two. Small update, no big claims attached, worth a look if you already run Marlin for diligence work; not worth switching to it for.

🧠 MODELS & RELEASES

• 🌐 Mistral AI is powering Firefox's new Smart Window, the browser assistant Mozilla is rolling out first in France and North America (UK and Germany later this year). Mozilla doesn't save the conversations by default, so this is a real test of whether "open and private" can win against Chrome's built-in AI rather than just sound good in a press release. source

• 🎙️ Google DeepMind shipped Gemini 3.8 Live and a heavier Extended Thinking sibling, both live in the Gemini API today. The base model switches between 97 languages mid-conversation and fires tool calls without breaking the flow; Extended Thinking tops Artificial Analysis' speech-quality index and posts 68.6% on the τ-Voice agent benchmark. If you're building voice agents, this is the pair to benchmark against before you commit to anyone else's stack. source

• 🤝 NVIDIA and Salesforce unveiled Koa, a CRM reasoning model built on Nemotron 3 Super and trained on 27 years of Salesforce data, with no customer data touching training or inference. Salesforce claims 3x fewer errors than rival models; pilots start in October with Formula 1 and Baxter Credit Union, GA lands winter 2026. Worth tracking as the template for vertical reasoning models built entirely on one company's proprietary corpus. source

• 🌍 Sakana AI CEO David Ha co-authored the opening piece in a Royal Society special issue on world models. His argument: today's scale gets you better pattern matching, not causal understanding, and no amount of extra compute closes that gap on its own. Not a product. A useful corrective if your roadmap assumes scaling laws solve everything. source

🔬 RESEARCH HIGHLIGHTS

• Agentic societies need referees, not just better agents. A new paper runs multi-agent setups where each AI represents a different principal across a trust boundary, and the agents are honest and individually competent. They still land on bad outcomes. Smarter models don't fix it. What does: coordination structure layered on top, what the authors call a "social harness." If you're wiring agents together across company or user lines, that's the layer you're probably missing. Agentic Societies Need a Social Harness

• ScienceBuddy tries to close the loop between researcher feedback and agent improvement. It's an interactive workspace where the agent's requests, corrections, and reruns feed back into improving the next version of itself, recursively, rather than sitting in a log nobody reads. Early and self-reported. Still, most "AI research assistant" tools don't actually learn from how you correct them. This one is built to. ScienceBuddy

• PhysStream gives video generation a memory of the scene's physics. Existing controllable video tools make you commit to a full motion schedule up front, or steer with raw pixel signals. PhysStream keeps a structured scene memory it updates as frames stream out, so you can redirect an object mid-generation instead of pre-scripting everything. That's closer to directing a simulation than prompting a video model. PhysStream

🛠️ TRY THIS

Turn a Marlin report into a deck you can still edit

1. Point Marlin at the doc or dataset you want summarized, then use the new Interactive Reading mode to push back on its own output before you trust it.

2. Ask it to defend the weakest claim in the report, not recap the strongest one. That's where a summarizer usually cuts corners.

3. Export the result straight to PowerPoint now that Marlin ships that format natively.

4. Drop the .pptx into genoffice and let your coding agent fix a broken chart or reflow a slide locally, no PowerPoint license required.

Prompt: Read this report and argue against its own top three conclusions. For each, give the strongest counter-evidence and say whether it changes what I should do.

Worth a look

• genoffice (genspark-ai): open-source AI office suite built for this exact handoff, Docs, Sheets, Slides, PDF and Markdown editors plus a CLI so Claude Code, Codex or Cursor can open and edit real .docx/.xlsx/.pptx files on your machine. Bring your own key, no subscription. 6.9k stars. github.com/genspark-ai/genoffice

🔗 QUICK LINKS

• Churchill Capital Corp XI filed to take Agility Robotics public at a $2.5B valuation, on $1.8M of 2025 revenue. Form 425
• HIVE Digital Technologies hired a new SVP of Revenue for its GPU cloud arm as that business scales up. 8-K exhibit
• Artificial Intelligence Technology Solutions touts fresh RAD device sales as proof its ROSA platform still anchors the stationary security lineup. 8-K
• Sakana AI published a look at its product team's hiring bar: interviews graded on how you explain a tradeoff, not how fast you solve the problem. Blog post
• NRx Pharmaceuticals got its ketamine ANDA amendment downgraded from Major to Minor at the FDA. 8-K exhibit
• Vivmark Residential, the newly merged AvalonBay/Equity Residential entity, presented at a BofA real estate conference, real estate news that wandered into today's filing sweep. 8-K exhibit
• Centene reaffirmed 2026 guidance at the Deutsche Bank Healthcare Summit. 8-K

That's the signal today.

Pradeep Perugu