November 2025
Theo reviews Anthropic’s Claude Opus 4.5, praising its coding reliability, tool use, token efficiency, and UI improvements while critiquing pricing, benchmarks, and competitors’ consistency.
A French-language analysis argues AI is accelerating across coding agents, scientific reasoning (GPT‑5 Pro), semantic video understanding (Meta SAM3), high‑fidelity UI generation (Google Nano Banana Pro), 3D world creation (Fei‑Fei Li’s Marble), and geopolitical infrastructure shifts (OpenAI–Foxconn, Saudi GB300), urging professionals to move from operators to system architects.
A walkthrough of seven creative, visual-first ways to use Nano Banana Pro—from explaining code with infographics and solving handwritten physics problems to auto-generating menus, summarizing PDFs on a whiteboard, visualizing code errors, turning blueprints into 3D views, and batch-creating slide decks.
A practical walkthrough showing how ref.tools and exa.ai MCP servers drastically cut context usage during AI coding workflows, including step‑by‑step setup in Claude Code, Cursor, Codex, and DroidFactory and a Tailwind v4 refactor demo with prompting tips.
How to use VS Code Insiders to add and configure OpenAI-compatible custom models—including local and remote LLMs via settings—for chatting, editing, and agent workflows, with demos from LM Studio and a large remote Kimi K2 setup.
Hands-on tests of Google’s new Nano Banana Pro (Gemini 3 Pro Image) show strong image generation and editing across posters, mockups, spatial reasoning, and realistic photo edits, with notable improvements over the previous version.
Live walkthrough where Debbie uses Goose with Gemini 3 to redesign her Nuxt site, iterating on heroes, grids, accessibility, and UX while debugging and automating tasks like image handling.
Hands-on first impressions of Gemini 3 Pro across seven tests—terminal sandboxing, drone control, UI cloning, Linux-like web terminal, Path of Exile 2 build design, image "spot the ball" reasoning, and a Mario-like game and video-powered quizzes—arguing it’s the strongest LLM so far.
Strategic analysis arguing that Gemini 3’s benchmark gains plus Google’s end‑to‑end distribution (and a likely Apple partnership) shift AI power from model quality to workflow and default access, with concrete career implications.
A detailed review and hands-on testing of Google’s Gemini 3 Pro highlighting its benchmark wins, multimodal strengths, blazing speed, strong UI/code generation, and quirks like higher cost, token usage, and occasional hallucinations.
Live test of Google’s new Gemini 3 across Cursor, Factory AI’s Droid, and Google’s Antigravity IDE by building real apps (volcano tracker, Canoe Club, AnimeLeak tweaks), highlighting design strengths, agent workflows, and rate-limit hiccups.
A step-by-step walkthrough showing how to deploy a TanStack Start app to Railway, add an MCP endpoint, and integrate it with ChatGPT to render interactive widgets, culminating in a React-powered widget that links to product pages.
Explains the difference between Playwright MCP and Playwright Test MCP, showing how the first enables general browser automation while the second powers testing workflows with Planner, Generator, and Healer agents and TS/JS-focused tools, plus how to install and use each.
The video explains how GPT-5.1 shifts from merely recognizing instructions to reliably obeying them, enabling precise automation, adaptive reasoning, and better uncertainty handling while trading off some safety categories for a warmer, more human experience, and outlines the implications for jobs and concrete actions to adapt.
The video analyzes a University of Chicago study on AI coding agents, concluding they boost output (e.g., more merges) without degrading short-term quality, while stressing that experienced developers see higher acceptance by giving precise, well-planned prompts and warning about overreliance and long‑term maintainability risks.
Theo critiques MCP’s tool-heavy approach and highlights Anthropic’s new push for code-executed agents as a far more efficient, secure, and scalable way to use external tools with LLMs.
A roundtable podcast debates Codex’s new 256-line tool-call truncation and its impact on coding agents, compares Claude, Kimi K2 Thinking, Minimax M2, GLM 4.6 and more, and explores pricing, safety, and workflow strategies shaping today’s AI tooling.
Overview of GPT-5.1’s upgrades—including faster, more accurate, more personable chat, adaptive reasoning that calibrates thinking time, better instruction following, enterprise latency and extraction gains, and front‑end coding improvements—plus rollout details and API changes like prompt caching.
Theo analyzes leaks and performance clues suggesting OpenAI’s rumored GPT-5.1 may be the stealth “Polaris Alpha” model, comparing benchmarks, UI generations, TPS/latency, and deployment patterns to argue a release is imminent.
Overview of Moonshot’s open-weights Kimi K2 Thinking model, its benchmark results, tool-using agent capabilities, and live demos showing long-horizon reasoning, browsing, and complex project execution.
The video examines whether today’s AI boom is a bubble, outlining massive spending, circular financing, demand and infrastructure constraints, and how this differs from the dot‑com era.
Anthropic’s post argues that generating and executing code to call tools on demand beats loading full MCP tool definitions, cutting token use, latency, and exposure while enabling progressive disclosure, privacy guards, and persistent skills—though it adds reliability and sandbox overhead trade-offs.
Addy Osmani discusses the "70% problem" in AI coding—how AI accelerates scaffolding but struggles with the last mile of quality, trust, testing, reviews, and real-world constraints—and shares pragmatic workflows, testing-as-feedback, and guidance for teams and juniors.
First look at Moonshot AI's Kimi K2 Thinking model with a technical overview (int4 quantization, MoE, 1T params) and hands-on tests spanning a browser OS, roleplay safety, Python low‑poly FPS, a web 3D racer, high‑pressure vs. standard PC repair websites, creative writing, and a terminal-based "Quantum Space" experience.
A step-by-step walkthrough of building a no‑code chatbot agent in n8n powered by Gemini, connecting tools like Strava, YouTube, weather, GitHub issues, RSS news, Google Calendar, and Gmail, with memory and automated email summaries.
A step-by-step guide to building a custom n8n trigger node that connects to Twitch—covering setup from the starter kit, node metadata, trigger logic with TMI.js, local linking for development, and testing within a workflow.
A practical, French-language walkthrough of setting up and using n8n to build a chat-triggered workflow that classifies emails with an AI text classifier and generates tailored responses, plus tips on data tables, webhooks, and integrations.
Theo reviews community claims that GPT‑5/Codex quality regressed, walking through OpenAI’s internal investigation (hardware variance, compaction, timeouts, constrained sampling bug, responses API checks) and fixes, arguing perceptions stem from tougher tasks and setup complexity.
A technical livestream with Luke Alvoeiro (Factory AI) explaining how Droid’s agent scaffolding, anchored summaries, and context compression enable multi‑million‑token coding sessions without losing context, plus discussion of spec mode, parallelization, and practical SDLC impacts.
In a 3h live build, Ray re-architects AnimeLeak into a multi-theme system using Factory AI’s Droid with Sonnet 4.5, showcasing spec-mode planning, stacked diffs, Convex migrations, and why Droid’s long-context management and adherence to agents.md outperformed Cursor 2.0 for large, continuous coding sessions.