AI briefingThursday, 8 October

Jessica: Okay, so Mistral just dropped a one trillion parameter model.

Eric: Wait, one trillion?

Jessica: One trillion. But only forty-nine billion active at a time.

Eric: Right, sparse model. That makes more sense.

Jessica: This is dailybrief, I'm Jessica—

Eric: and I'm Eric, it's Thursday, October eighth, twenty twenty-six.

Jessica: So this is Large 4. Simon Willison reported it, Ars and WIRED picked it up too. They trained it on thirty-eight hundred Nvidia Grace Blackwell GPUs, and it scored thirty-eight on Artificial Analysis.

Eric: That's… what, four times better than Large 3 from December?

Jessica: Large 3 scored nine, so yeah. Huge jump.

Eric: And they're releasing the weights end of month?

Jessica: End of October, according to Mistral. API preview is live now, though.

Eric: Okay, that's actually really cool. You get to test it now, deploy locally later.

Jessica: And it supports two reasoning levels via the API—none and high.

Eric: So you can dial up compute for harder tasks.

Jessica: Exactly.

Jessica: Meanwhile, OpenAI's launching what they're calling Intelligent UI with GPT-six.

Eric: Wait, GPT-six?

Jessica: Yeah, TechCrunch says it's rolling out Wednesday—well, yesterday by the time people hear this—and it adds interactive charts, calculators, editable graphs right into ChatGPT.

Eric: So not just text anymore.

Jessica: Nope. You get tappable buttons, customized calculators, diagrams. Pro and Plus users got it yesterday, free tier gets it today.

Eric: Honestly, that feels like a big shift. ChatGPT's been so text-first.

Jessica: Yeah, and you can dial it down if you want less visual stuff. It's customizable.

Eric: Okay, that's good. I don't need a chart for everything.

Jessica: Speaking of visual stuff, Google launched Playground.

Eric: Which is?

Jessica: You type a text prompt, it builds a game. No code. Just describe what you want.

Eric: Huh.

Jessica: It's at playground dot google. Some genres support multiplayer and leaderboards. You can keep games private, share a link, or publish to a community gallery.

Eric: And Unity Spark integration is coming soon, according to Google's blog.

Jessica: Right, for more professional-level mechanics.

Eric: I mean, that's a fun prototyping tool. I'm skeptical it replaces real game dev, but—

Jessica: Oh, definitely not. But for sketching out an idea fast? Super useful.

Eric: Okay, what's this Liquid AI decision model?

Jessica: So Liquid AI released d1-3B and d1-omni-600M. They're on Hugging Face. These models answer classification and decision tasks in a single forward pass—no token generation.

Eric: Wait, so they just… output a decision?

Jessica: Yeah. Sixteen milliseconds on Jetson Thor, fifty on Orin Nano.

Eric: That's ridiculously fast.

Jessica: Right? d1-3B scored forty-eight point five seven on the Decision Index, which is the highest under ten billion parameters.

Eric: And d1-omni supports text, images, and audio?

Jessica: Audio's experimental, but yeah. d1-3B does text and images.

Eric: So if you need intent detection or routing on edge hardware—

Jessica: This is built for that.

Jessica: Google also shipped EmbeddingGemma 2.

Eric: For embeddings?

Jessica: Multimodal embeddings. Seven hundred forty million parameters. It maps text, images, audio, and video into unified vectors.

Eric: And the license?

Jessica: Apache two point oh.

Eric: Okay, good.

Jessica: It's got an eight thousand token context window, processes five and a half minutes of audio or twenty-nine images. Runs in as little as one ninety-one megabytes of RAM for text, five sixty-seven for multimodal.

Eric: So this runs on a phone.

Jessica: Yep. On-device search, offline RAG pipelines.

Eric: And it improved almost ten points on code performance versus the first EmbeddingGemma?

Jessica: Nine point nine two points, according to Google DeepMind.

Jessica: Nvidia fine-tuned Nemotron 3 models to gold medal level at this year's International Olympiad in Informatics and the International Math Olympiad.

Eric: Wait, both competitions?

Jessica: Both. Ultra-CC scored five thirty-five point four out of six hundred at IOI twenty twenty-six under actual competition constraints.

Eric: How'd they train it?

Jessica: Twenty-two thousand curated problems, synthetic reasoning traces, supervised fine-tuning and reinforcement learning on top of Nemotron 3 base models.

Eric: So it's really just post-training on a strong base model.

Jessica: Yeah. And it shows you can reach expert-level performance in these domains with standard methods.

Jessica: Meta updated Muse.

Eric: iPad support?

Jessica: Yep, one month after iOS. Plus integrations with Canva, Dropbox, Figma, QuickBooks, GitHub, Zoom, Asana, Klaviyo.

Eric: That's a lot of integrations.

Jessica: And there's a new business goal feature that can autonomously plan marketing campaigns and product pages. The Verge and TechCrunch both covered it.

Eric: So this is Meta going after ChatGPT for business workflows.

Jessica: And Grok. Yeah.

Jessica: Quick ones—OpenAI's adding College Planner to ChatGPT for Teens.

Eric: For college apps?

Jessica: Yeah, plus flashcards and quizzes for studying. They're forming a teen AI council to help shape future features.

Eric: Interesting positioning.

Jessica: Atlassian and OpenAI expanded their partnership—frontier models connecting to enterprise knowledge.

Eric: Makes sense for Jira and Confluence workflows.

Jessica: And Radisson partnered with Accenture to build a ChatGPT plugin for hotel discovery and booking.

Eric: So ChatGPT's becoming a travel agent.

Jessica: Apparently.

Jessica: That's it for today. Every story's linked to its original source on the dailybrief website.

Eric: See you tomorrow.