Sources

Company@X — 2026-03-26#

Signal of the Day#

Google executed a massive, ecosystem-wide rollout of Gemini 3.1 Flash Live, their highest-quality audio and voice model optimized for real-time dialogue at scale. The deployment pushes voice-first AI agents into global production, launching simultaneously across the Gemini App, Google Cloud enterprise tools, AI Studio, and enabling multimodal Search Live in over 200 countries.

Key Announcements#

Google · Source Google released Gemini 3.1 Flash Live, delivering significantly lower latency, faster response times, and improved instruction-following in noisy environments. The model boasts 2x longer conversation memory and is rolling out immediately to both end-users via the Gemini App and developers through the Gemini Live API.

Cursor · Source Cursor published a new technical report on Composer 2, detailing their pipeline for training new checkpoints. By leveraging real-time reinforcement learning, the team claims they are now shipping improved versions of their model every five hours, highlighting an exceptionally tight iteration and deployment cycle.

Meta · Source Meta introduced TRIBE v2 (Trimodal Brain Encoder), an open-source foundation model trained on over 500 hours of fMRI data to predict how the human brain responds to sight and sound. Achieving a 2-3x improvement over previous methods without any retraining, the model enables zero-shot predictions for new subjects to accelerate computational neuroscience and neurological disease research.

Waymo · Source Waymo officially doubled its paid autonomous rides per week in less than a year. The company pointed to this milestone as a key indicator of responsible, effective scaling in real-world autonomous technology.

OpenAI · Source OpenAI is rolling out plugins in Codex. The update enables Codex to work seamlessly out of the box with vital developer and productivity tools, including direct integrations with Slack, Figma, Notion, and Gmail.

Google (Gemini App) · Source Google introduced a utility that allows users to import their memories, preferences, and chat history from competitor AI applications directly into Gemini. Rolling out first on desktop, this is a strategic product move aimed at lowering switching costs and capturing market share from rival platforms.

Also Noted#

  • Intercom / Hugging Face (Source): Intercom leadership stated it has become cheaper and faster to train open-source models internally rather than using generalized APIs for many tasks, signaling a broader enterprise shift toward in-house, full-stack AI built on open weights.
  • Google DeepMind (Source): DeepMind released a toolkit and a 10,000-person study measuring real-world AI manipulation, finding models had high influence in finance but failed to manipulate users in health domains due to existing guardrails.
  • ARC Prize (Source): ARC Prize 2026 (v3) officially launched with industry figures like Sam Altman and François Chollet, throwing down a new gauntlet as state-of-the-art results for the benchmark currently sit at just 0.37%.
  • Supabase (Source): Developers can now add and launch Supabase projects directly from the Stripe CLI.
  • Cohere (Source): Cohere released a new open-source, Apache 2.0 licensed Transcribe model that is already trending heavily on Hugging Face.