Gemini API ships Dynamic Tool Retrieval for remote MCP servers
Today's top 20 insights for PM Builders from X and YouTube.
Gemini API ships Dynamic Tool Retrieval for remote MCP servers
#1 𝕏
Philipp Schmid says Dynamic Tool Retrieval was shipped in the Gemini API, supporting deferred functions, remote MCP servers, server-side tool search, and a client-side mode. He says it can avoid loading 200 tools on turn 1.
#2 𝕏
Google AI recapped models and products it said shipped over the last couple of weeks: Gemini 4 Neon, Hydrogen and Carbon; Google DeepMind’s Genie 4, Gemini Robotics 2.5 and AlphaGenome 2; Gemini 3.5 Transcribe; and Google Labs’ Stitch Loop experiment. For builders, the updates span coding and agentic workflows, long-horizon research, first-time Genie 4 access via the Gemini API and Google AI Studio, single-step voice editing, robotics, disease research and production-data-driven product improvement.
#3 𝕏
Thinking Machines announced Inkling 2 Mini, a 26B-parameter looped transformer it says can run on a laptop and handle agentic work, scoring 20.1 on AA Index. It is substantially smaller than the 594B-parameter Inkling 2 Small.
Also covered by: @Mira Murati, @Thinking Machines, @clem 🤗, @Julien Chaumond, @Julien Chaumond
#4 𝕏
Google Research recapped how Google Duplex evolved over the past decade to power natural voice conversations and local business calls, with these capabilities now being brought directly to Tasks in Gemini.
#6 𝕏
Alexandr Wang announced a Muse feature that turns camera pictures of items users want to sell into polished product shots designed to make them look expensive.
#7 𝕏
Google DeepMind shared a ski simulator coded by Gemini 4 Neon using real Mont Blanc terrain. Built in @Antigravity with custom shaders and zero pre-made 3D models, it renders carved snow and avalanches live.
Also covered by: @LlamaIndex 🦙, @Aravind Srinivas, cofounder and CEO @perplexity_ai, @Greg Isenberg
#8 𝕏
Grant Sanderson set up a Devin Bot to help with Manim, the open-source animation system behind every 3Blue1Brown video. Manim has 96.6k stars, 7,936 forks, and 250,000 downloads per month.
#9 𝕏
Demis Hassabis shared @3DVR3’s account of using Google’s Genie 4 world model to explore a generated Shibuya, where pedestrians appeared coherent in real time but turning around could reveal an unfamiliar location. @3DVR3 compared the visuals to video generation from a few years ago, noted its potential, and reported a generation cost of approximately $1 per 10 seconds.
#10 𝕏
NVIDIA AI shared WorldTrace from NVIDIA Research, which helps interactive video world models remember places without retraining and maintain consistency as users explore. On LoopBench, it improved recall of revisited scenes by 32% over the standard sliding-window approach; the linked work is identified as a NeurIPS 2026 paper.
#11 𝕏
Harrison Chase, @LangChain recapped LangSmith Tool Gateway authentication for every agent tool call, merged-PR cost visibility for coding agents, and trace access controls for sensitive data. He also highlighted two blogs on agents that pursue goals.
#12 𝕏
claire vo 🖤 shared that she made merge mommy validate pull request changes and auto-merge when checks are green and risk is low. After four hours of work, she said it may have changed her life.
#13 𝕏
claire vo 🖤 recapped a @cxodev discussion with @zachdavis about code review, highlighting faster deterministic checks, risk review, and auto-merge—an approach @Amplitude_HQ reached 81% with, though the metric’s meaning and timeframe were unspecified. She said teams that do this well improve both velocity and quality.
#14 𝕏
Santiago shared 1,200+ reinforcement-learning environments built from real workplace tasks across healthcare, life sciences, legal, finance, and other fields. They can be used to train and evaluate models on specialized tasks.
#15 𝕏
NVIDIA shared its breakdown of eight design guidelines for efficient sparse attention on NVIDIA accelerated computing, covering token selection through memory reuse. It notes that sparse attention does less work, but identifying the right tokens still has a cost.
#16 𝕏
Udi Menkes 🚢 shared a Reuters Institute item reporting that 14% used an AI agent for a real task last week, 13% just tried one, and 67% did not use one at all. He recommends measuring how many users progress from trying an agent to completing a first real task.
#17 ▶️
If you’ve deployed anything since 2015, I have bad news...
Fireship
#18 𝕏
Guillermo Rauch said improved models, @agentcardhq, mature @eve iMessage and WhatsApp channels, and better performance without sacrificing durability or encryption now enable builders to ship fully controlled personal shopper platforms on an open intelligence stack. He recommended trying and customizing the personal shopper agent template.
#19 𝕏
Sebastian Raschka shared his post on coding a KV cache in LLMs from scratch and said he may write one about speculative decoding in the future.
#20 𝕏
Guillermo Rauch announced that Vercel’s AI infrastructure is audited annually by an accredited third party under ISO 42001:2023, which he described as the “SOC2 of AI” and the first global, independently audited standard for building, evaluating, monitoring, and governing AI systems.