Please support this podcast by checking out our sponsors:
- Prezi: Create AI presentations fast - https://try.prezi.com/automated_daily
- KrispCall: Agentic Cloud Telephony - https://try.krispcall.com/tad
- Lindy is your ultimate AI assistant that proactively manages your inbox - https://try.lindy.ai/tad

Support The Automated Daily directly:
Buy me a coffee: https://buymeacoffee.com/theautomateddaily

Today's topics:

Tiny model, huge benchmarks - Sina Weibo’s VibeThinker-3B posts standout reasoning scores (AIME 2026) despite only 3B parameters, fueling debate about benchmark validity, post-training, and real-world reliability.

Million-token open-source coding model - Z.ai releases GLM-5.2 under an MIT license, targeting stable 1M-token context for long-horizon coding agents, with new training focused on messy, hours-long engineering workflows.

Agent tooling inside the browser - OpenAI adds Chrome DevTools Protocol support to Codex browser-use, letting agents read console logs, network traffic, and page state—key for debugging web apps with AI assistance.

Voice AI gets truly interactive - OpenAI is reportedly preparing a new bidirectional voice model (GPT-Bidi-1) designed for natural interruptions and real-time conversation, pushing voice toward a primary AI interface.

Anthropic pauses agent billing shift - Anthropic pauses its planned token-based billing shift for the Claude Agent SDK after developer backlash, highlighting rising sensitivity around agent usage costs and pricing models.

Windows local AI on RTX - Microsoft experiments with running Phi Silica locally on Windows using Nvidia RTX GPUs, expanding on-device AI development beyond NPUs while exposing uneven feature tiers across hardware.

NVIDIA Blackwell tops MLPerf - NVIDIA’s Blackwell platform leads MLPerf Training 6.0 with fastest time-to-train across workloads, influencing data-center buying decisions for frontier-scale AI training.

Android 17 becomes agent-friendly - Google ships Android 17 to Pixel devices and AOSP, expanding AppFunctions for agent-discoverable actions and enforcing adaptive-first UI rules for foldables, tablets, and desktop mode.

Durable streaming to stop re-billing - A proposed ‘durable buffer’ between agents and LLM providers can resume streaming after crashes, preventing duplicate token charges and improving reliability for long-running workflows.

Discipline replaces vibe coding - Charity Majors argues AI makes code cheap, so teams must invest in specs, invariants, tests, observability, and continuous evaluation—turning 2026 into a ‘return to discipline.’

AI trust gap in America - A Pew survey finds Americans are pessimistic about AI’s long-term impact and distrust regulation and corporate safety, even as daily chatbot use and AI-generated summaries rise.

Wearables as next AI platform - Qualcomm pitches AI wearables—glasses, pins, earbuds—as the post-smartphone platform, launching Snapdragon Reality Elite to bring more on-device AI to mixed-reality devices.

Language-driven robot world models - Qwen-RobotWorld introduces a language-conditioned video world model and an 8.6M video-text dataset, aiming to unify planning and prediction across robots and vehicles via language.

Text-to-CAD goes open source - CADAM is an open-source, browser-based text-to-CAD tool generating editable OpenSCAD models with real-time preview, lowering barriers to parametric design and maker workflows.

Cursor moves into Git hosting - Cursor announces Origin, a forthcoming Git hosting and code storage product, signaling competition to own the full AI-assisted developer workflow beyond the code editor.

-Z.ai Releases Open-Source GLM-5.2 With Stable 1M-Token Context for Long-Horizon Coding
-Cursor Announces Origin, a New Git Hosting and Code Storage Service
-As AI Automates Transactions, Human Connection Becomes the Real Competitive Moat
-Microsoft Experiments with Phi Silica Local AI on Nvidia RTX GPUs for Windows 11
-AI & Tech Sandbox and PMG Launch First Global Advertising-Tech Hackathon
-OpenAI Adds Chrome DevTools Protocol Access to Codex Browser Mode
-Qualcomm Unveils Snapdragon Reality Elite and START to Power Post-Smartphone AI Wearables
-Charity Majors: Cheaper AI Coding Means More Rigor, Not Less
-Pew: Americans mostly expect AI to harm society despite rising chatbot use
-Weibo’s VibeThinker-3B Sparks New Fight Over AI Benchmark Credibility
-Report: OpenAI readies GPT-Bidi-1 to overhaul ChatGPT voice mode
-Mercury launches Command, an AI assistant to run banking and finance workflows
-Qwen-RobotWorld Proposes a Language-Conditioned Video World Model for Embodied Prediction
-Anthropic Pauses Token-Based Billing Change for Claude Agent SDK
-NVIDIA Blackwell Leads MLPerf Training 6.0 Across Speed, Scale and Submissions
-Durable Buffers to Prevent Re-Billing When LLM Streams Get Interrupted
-Android 17 launches with AI AppFunctions, mandatory large-screen resizability, and tighter privacy and performance rules
-CADAM launches as open-source browser-based AI text-to-CAD tool

Episode Transcript

Tiny model, huge benchmarks
Let’s start with that small-model surprise. Researchers from Sina Weibo released VibeThinker-3B, along with open weights and a technical report, claiming it can match or outscore much larger systems on several reasoning benchmarks. The headline numbers are strong enough to reopen an ongoing argument in the AI community: are modern benchmarks still good signals, or are teams optimizing so hard that scores stop correlating with real-world usefulness? The more interesting takeaway may be strategic—if verifiable reasoning can be compressed into small models, you could pair a cheap “reasoning engine” with a larger knowledge model and cut deployment costs dramatically, provided the reliability holds up outside the test suite.

Million-token open-source coding model
Staying with open models, Z.ai has released GLM-5.2, positioning it as a flagship open-source model built for long-horizon engineering and coding—specifically targeting stable performance with a one-million-token context window. The company’s framing is worth noting: long context isn’t impressive if the model falls apart halfway through an hours-long agent run filled with detours, partial failures, and messy debugging. They say training focused on real coding-agent scenarios, and they’re also adding knobs to trade latency for stronger results when you actually need the extra effort. If this translates into day-to-day reliability, it’s a meaningful step toward making “million-token context” more than a demo—especially because it’s released under an MIT license.

Agent tooling inside the browser
On the agent tooling front, OpenAI added Chrome DevTools Protocol support to Codex’s browser-use feature. In plain terms, that means the agent can see the kinds of signals developers use when a web app misbehaves—console errors, network requests, performance traces, and the rendered page state—rather than guessing from screenshots and page text alone. It’s opt-in, can be disabled by organizations, and it’s not available everywhere yet, but the direction is clear: browser agents are moving from “click around” automation to something closer to a junior developer with real de...