Google DeepMind introduces two Gemini 3.8 Live models

#1 𝕏

Google DeepMind announced Gemini 3.8 Live and 3.8 Live Extended Thinking, describing them as its best conversational AI. The models can talk, think, and handle background tasks without breaking the user’s flow.

Also covered by: @Google AI, @Google AI, @Google DeepMind, @Logan Kilpatrick, @Philipp Schmid

#2 𝕏

Cognition announced that Devin now has a Mac VM for building and testing apps with an iOS simulator. Devin can also send screen recordings via Slack and TestFlight links for users to try apps.

#3 𝕏

v0 announced it is now model-agnostic, offering access through Vercel’s AI Gateway to frontier, cheap, open, and fast models, including Claude, GPT, Kimi, GLM, Grok, and DeepSeek.

#4 𝕏

Aravind Srinivas says an in-house key-value database intended to replace AWS DynamoDB was built by two engineers and hundreds of persistent Computer agents over two months. He claims migrating to it will save up to a hundred million dollars yearly.

#5 𝕏

Google Research announced Retrieve-for-Train, a framework that accelerates complex AI search by replacing heavy autoregressive inference with a lightweight diffusion model. The organization says it enables instant, expert-level search slates at a fraction of the cost.

#6 𝕏

Sundar Pichai recapped AI advances from the “last week or so,” including AlphaGenome Atlas mapping all 9B possible single-letter genetic changes, WeatherNext 3 for weather predictions that inform billions of decisions, and the open-access AI & Economy ATLAS. He said translation services now cover nearly 300 languages spoken by 7B people, with four key focus areas: health, natural-disaster and weather resilience, learning, and economic opportunity.

#7 𝕏

Claire Vo shared a link to a clairevo.com page with the URL slug “build-a-cursor-following-portrait-with-flora-and-codex.”

#8 𝕏

Santiago wrote an article showing how to use the Agents CLI to build, deploy, and begin monitoring a small multi-agent system in 30 minutes.

#9 𝕏

LlamaIndex 🦙 shared an article by George He and Yong Park on using Stainless to keep LlamaParse SDKs updated, improve API consistency, and navigate SDK generator changes. The recap comes as the Stainless team joins Anthropic.

#10 𝕏

Thariq said MCPs are better than CLIs for most integrations as models improve at tool calling, tools can be deferred, and MCP is now stateless. For composing or filtering data, he recommended adding parameters such as `query` to MCP tools.

#11 𝕏

Sebastian Raschka demonstrated a Paint UI benchmark in which GPT-5.6 Astra layered geometric shapes while Qwen3.8 Max worked pixel by pixel, producing a closer match. He cautioned that final-result similarity alone does not establish better generalization, computer-use capabilities, or visual understanding.

#12 𝕏

Chip Huyen commented on an approach where models choose from predefined values instead of generating freeform text, potentially benefiting data labeling and fixed-action tasks. She noted that output tokens are free, though how reasoning would work remains unclear.

#13 in

Udi Menkes argues that product teams need structured continuous discovery: ask about past behavior, use an opportunity-solution tree, and include engineers in customer interviews. As agents increase building capabilities, he says effective discovery is becoming rarer—and building faster without understanding the problem only produces more products nobody wants.

#14 in

Guillermo Rauch announced Vercel Labs, Vercel’s public research and experimentation arm for sharing what the company supports and researches, along with experiments that did not pan out. He noted “247 million downloads” without specifying what received them and credited Chris Tate and Malte Ubl with shaping the initiative and iterating in public.

#15 𝕏

Harrison Chase commented that agent memory has drawn strong interest for roughly two years without sticking. He said updating and using memory must be tightly integrated with the agent harness, while deciding what to remember is often application-specific and prompt-based, making standalone products difficult to build. He argued that memory is more useful for repeated tasks than for general-purpose agents and has yet to prove useful for coding agents.

#16 ▶️

How to run your first AI UGC campaign (step-by-step guide)

AI Jason

#17 𝕏

Teresa Torres shared show notes and links for All Things Product’s “Delivery Isn’t Free” episode with Petra Wille, arguing that AI makes individual features cheaper—not production-quality delivery free. Prototypes may quickly reach 60–70%, but the last 30% needed for a trustworthy product can take months to years and still requires skilled engineering oversight.

#18 𝕏

Garry Tan described the @AsideAI browser/harness as one of the most powerful consumer AI tools for use with agents, noting that Aside MCP exposes deep browser-use tools with credentials that can make agents more powerful.

#19 𝕏

Garry Tan said he started using @capydotai with GStack/GBrain for fix waves on outstanding issues and pull requests, completing work in about half the time he estimated raw Codex or Claude Code would take—roughly half a day instead of a day—using the same frontier models.

#20 𝕏

DeepLearning.AI recapped Claude Fable 5.1’s top ranking on Artificial Analysis’ Intelligence Index v4.2, where it scored 57 and tied with OpenAI’s new model on updated benchmarks. It also achieved 52.6% on Terminal Bench Science 0.1 and offers cheaper cache reads for repeated agentic workflows.