Google Launches Gemini 3.1 Flash Live Audio Model

AI Product Management Certification

If you want to master building enterprise-level AI products from scratch and future-proof your career with the most important AI skills, this #1 AI PM Certification is for you.

Run by real AI operators: Rohan Varma (ex-cofounder, first ever PM at Cursor, now at Codex) & Henry Shi (former co-founder of Super.com, a $200M+ ARR company, now technical staff at Anthropic Labs, working directly with the founders).

3,000+ AI PMs graduated · 1,000+ reviews on Maven — highest in any category

Review the curriculum & enroll — $500 off →

Google Launches Gemini 3.1 Flash Live Audio Model

#1 𝕏

Google DeepMind launched Gemini 3.1 Flash Live, an audio model that delivers more natural conversations with improved function calling for more useful, informed interactions.

Also covered by: @Demis Hassabis, @Philipp Schmid, @Google AI, @Google AI, @Sundar Pichai, @Sundar Pichai

#2 𝕏

Jeff Dean launched Gemini 3.1 Flash Live, a model with native audio understanding that leads on ComplexFuncBench and Scale AI’s AudioMultiChallenge. It’s now powering Gemini Live and Search Live globally, enabling high-fidelity, nuance-aware voice interactions.

Also covered by: @Demis Hassabis, @Philipp Schmid, @Google AI, @Google AI, @Sundar Pichai, @Sundar Pichai

#3 𝕏

AI at Meta launched TRIBE v2, a model that predicts unseen individuals’ brain responses to movies and audiobooks with a 2–3× accuracy boost over prior methods without any retraining.

#4 𝕏

clem 🤗 released the CohereLabs/cohere-transcribe-03-2026 speech-to-text model on Hugging Face (https://huggingface.co/CohereLabs/cohere-transcribe-03-2026).

#5 𝕏

Guillermo Rauch says AI agents perform best when they can freely install, run, debug, and deploy code—but they need persistent compute to keep state.

#6 📝 Ampcode Chronicle

GPT‐5.4 in Deep - Amp has placed GPT-5.4 into its new Deep agent mode, tuning the model to behave more like Codex for longer-form, more code-focused reasoning. The update emphasizes deeper planning and agentic behavior for coding tasks.

#7 𝕏

clem 🤗 highlights that after Pinterest, Airbnb, Notion, and cursor_ai, Intercom is training open-source models in-house—finding them cheaper, faster, and more effective than APIs.

#8 𝕏

Google DeepMind launched a first-of-its-kind, empirically validated toolkit to measure AI manipulation in real-world settings, revealing its mechanisms and informing strategies to protect people.

#9 𝕏

Dharmesh Shah argues AI agents will need core CRM and GTM context tools to work effectively. HubSpot is therefore building an Agentic Customer Platform to supply those capabilities for both its own and third-party agents.

#10 ▶️

How Stripe’s engineering team built their AI coding agent “minions”

How I AI Podcast

Demonstrates how Stripe’s AI “minions” use a Slack reaction to provision isolated Devbox environments, run Goose agent loops with internal code search and CI to automatically generate pull requests, and how an AI agent can transact via Stripe’s machine payment protocol—paying Browserbase under $0.01 and a $1.65 Stripe Climate offset.

Stripe’s “minions” autonomously generate about 1,300 pull requests per week that require only human review.
Reacting with “create minion pay-server” in Slack triggers the Devbox agent to spin up a cloud VM, checkout the pay-server repo on a new branch, configure the local database, Git settings, and VS Code Server with extensions in seconds.
In the agent payment demo, the AI paid Browserbase under $0.01 for a browser session and issued a $1.65 Stripe Climate donation to offset 4.4 kg of carbon from 70,000 token usage.

#11 𝕏

LlamaIndex 🦙 demoed Gemini 3.1 voice agents via the Live API in its document-processing stack with LiteParse for fast, fully-local parsing. The TUI assistant lets you speak commands to parse single files or entire folders and hear real-time audio readbacks.

#12 ▶️

Paperclip: Hire AI Agents Like Employees (Live Demo)

Greg Isenberg

The demo shows how to use the open-source Paperclip orchestrator with local Claude Opus agents (via Claude Code or Codeex) to spin up a "Moola" finance app company—hiring a CEO, founding engineer, QA agent, and video editor with skills like Remotion—and automate tasks and daily GitHub update routines while logging all actions and token spend in a dashboard.

Paperclip accrued 30,000 GitHub stars in under three weeks and runs agents locally using Claude Opus via Claude Code or Codeex, resulting in $0 API spend on its dashboard when using subscription-based inference.
In the “Moola” tutorial, after naming the app, the CEO agent was tasked to hire a founding engineer and split the roadmap into five issues—set up scaffolding, configure continuous integration, implement progress tracking, design user onboarding, and build the core daily loop.
Paperclip’s routines feature schedules a daily 10 a.m. trigger to read the last 24 hours of GitHub pull requests from the Paperclip repo (500+ PRs, 30k stars) and create a Discord-formatted summary issue, complete with per-run token usage metrics.

#13 𝕏

LlamaIndex 🦙 launched a new guide for LiteParse, their fast open-source document parser, showing how to extract bounding boxes and capture page screenshots for visual citations. It walks through linking text to on-page elements for precise, agent-friendly citations.

#14 in

Dharmesh Shah is enhancing HubCode—an agentic coding tool for HubSpot—by adding support for Custom Objects to surface their data in app cards while fine-tuning the conversational prompt UX and UI.

#15 𝕏

Andrej Karpathy built menugen about a year ago to orchestrate LLM agents for app development, only to hit classic DevOps pain points around reproducible environments, secrets management, and monitoring.

#16 📝 Simon Willison

Quantization from the ground up - Sam Rose published an interactive essay explaining LLM quantization, including clear visualizations of floating point representation and discussion of outlier weights; he shows that aggressive quantization can retain much of model quality depending on approach.

#17 📝 Simon Willison

We Rewrote JSONata with AI in a Day, Saved $500K/Year - A case study of "vibe porting": the Reco team built a custom Go implementation of the JSONata JSON expression language quickly using an existing test suite and AI assistance, validating it via a shadow deployment.

#18 𝕏

Harrison Chase shares a blog post detailing his team’s rigorous evaluation framework for real AI agents, not just simple LLM prompts. It walks through scoring rubrics, simulation setups, and benchmark analyses to quantify agent capabilities.

#19 𝕏

Sebastian Raschka rolled out significant updates to his LLM Architecture Gallery—most notably a long-awaited diff tool for comparing model architectures.

#20 𝕏

Santiago launched Cline Kanban, a free open-source npm board for orchestrating coding agents (Claude Code, Codex, Cline) with task creation, dependency chaining, and real-time agent tracking. He argues the future of software development is managing a swarm of agents.

#21 𝕏

Philipp Schmid unveiled Google’s open-source Gemini agent skills library—domain-specific tools for compute, retrieval, and communication—that let LLM agents fill knowledge gaps and boost task accuracy (e.g., solving 82% more math problems).

#22 ▶️

nemoclaw, $250K token budgets and opensource ai - Nvidia GTC 2026

All About AI

Demonstrates installing Nvidia Nemo Claw via the “install nemo-claw” and “onboard openclaw-agent” commands, auto-detecting an Apple M3 Pro, and running local inference on the Quen 3.54B model within an Open Claw sandbox.

Runs “install nemo-claw” and “onboard openclaw-agent” to launch Open Claw and an Open Shell sandbox.
Detected an Apple M3 Pro, configured inference options (Nvidia, OpenAI, Anthropic), selected the Quen 3.54B model, then started the shell gateway, inference provider, sandbox, and policies.
Jensen Wong stated that a $500,000/year engineer must consume at least $250,000 worth of tokens annually—or he “will go ape.”

#23 in

Dan Shipper launched Plus Ones—a Slack-hosted OpenClaw preloaded with Every’s agent apps (Cora for email, Spiral for writing, Proof for docs) plus custom skills and workflows, all set up in one click using your ChatGPT or any API key.

#24 ▶️

Two AI Models Set to “stir government urgency”, But Will This Challenge Undo Them?

AI Explained

OpenAI shut down its Sora app to reallocate compute for the upcoming Spud model ready in a few weeks, Anthropic’s next Claude series is pitched as stirring U.S. government urgency, and ARC AGI 3 benchmark currently yields 100% for humans versus 0.37% for top AI (Gemini 3.1).

OpenAI discontinued Sora, its viral erotica chatbot, to free computing resources for the Spud model, which according to Samman will be ready in a few weeks and “really accelerate the economy.”
Anthropic warned U.S. government officials that its next Claude series could supercharge both offensive and defensive cyber capabilities, prompting the Pentagon to reconsider reviving a deal that lapsed after a six-month deadline.
ARC AGI 3 clamps AI at a 100% human baseline, caps attempts at five times the number of human actions, applies a quadratic penalty to action inefficiency, and currently Gemini 3.1 scores 0.37% against the human second-best baseline.

#25 𝕏

Harrison Chase unveiled agent middleware enabling modular, composable harnesses with plug-and-play tools, guardrails, and custom instructions.

AI Product Management Certification

Google Launches Gemini 3.1 Flash Live Audio Model

Get tomorrow's brief first