September 2025

A concise walkthrough of Anthropic’s Sonnet 4.5 release, arguing it’s part of a broader “virtual collaborator” push, with improvements in coding performance, agentic capabilities (Claude Agent SDK), computer use, and long-running, plan‑driven workflows.

Theo reviews Anthropic’s Claude Sonnet 4.5, benchmarking its coding, agentic and safety improvements versus prior Claude and GPT models, and concludes it’s the nicest model to code with despite weaker UI generation.

Scott from Syntax interviews Dr. Sarah Bird at Microsoft about what responsible AI means for developers—covering privacy, security-by-default, human-in-the-loop oversight, local vs. cloud models, risks (malfunction, misuse, systemic), and how developers can build trustworthy AI.

Theo explores "cursed lang," a Gen Z–slang Golang-style language auto-generated by running Claude in a loop for three months, testing features like emoji identifiers, debugging builds, and reflecting on AI-driven development.

Alex Ziskind examines NVIDIA’s delayed DGX Spark—its promised GB10 Grace Blackwell architecture, FP4/FP8 advantages, clustering and CUDA ecosystem benefits—asking whether the still-unreleased $4,000 mini AI supercomputer is worth the wait.

Theo - t3․gg unpacks Anthropic’s postmortem on three overlapping infrastructure bugs—routing to 1M-context servers, output corruption on TPUs, and an XLA top‑K issue—that degraded Claude’s quality in August–September and what fixes and process changes followed.

GosuCoder reviews Qwen 3 Max’s latest update, showing strong coding performance, tool-calling accuracy, faster speeds, and improved UI output that challenges Anthropic’s Claude Sonnet on price and capability.

Jack Herrington demonstrates building a simple MCP UI client in TanStack Start that renders external MCP server iframe responses securely using the MCP UI resource format and React renderer.

Theo - t3.gg explains why Grok 4 Fast is a surprisingly strong, fast, and ultra‑cost‑efficient successor to Grok 4, outperforming on several benchmarks while being far cheaper and more practical to use.

GosuCoder reviews GPT‑5 Codex, finding it produces solid coding/refactoring results but is often 2–3× slower than GPT‑5, recommending Codex Low for practical use.

Syntax ranks today’s major AI coding assistants—from Cursor, VS Code + Copilot, Windsurf, Kira, Claude Code, Opencode, Gemini, Qwen, Codex, v0, Bolt.new, Replit, Lovable, ChatGPT, and Warp—debating strengths, UX, planning modes, and real‑world reliability.

Theo reviews OpenAI’s new GPT-5 Codex model aimed at agentic coding, praising its token efficiency and code-review strengths while critiquing the Codex tooling’s search and UX rough edges.

Jack Herrington explains why he’s switching from Claude Code to Cursor CLI, demonstrating installation, model switching, MCP integration, CLI vs desktop features, non‑interactive prompts, and practical coding workflows with caveats about beta glitches.

Alex shares his daily AI-driven developer stack and workflow—tools, optimal configuration (memory bank, rules, templates), and a live demo fixing a bug and adding a feature—showing how to plan then build to ship clean, high-quality code faster.

Alex so yes tests OpenCode with xAI’s new Grok code model to debug a real AIDD project issue, explore the AGENTS.md standard, and show how terminal-first, provider-agnostic workflows can speed up coding while reducing IDE lock-in.

Theo explains why OpenAI’s recent acquisitions (like Statsig and Alex) are really about importing aligned, ready‑to‑execute teams and product leadership to accelerate apps and developer tools without disrupting core research.

Convex shows how to rapidly build and deploy a Twitter/X clone using Claude Code and Convex’s end-to-end type-safe queries, mutations, actions, and storage—covering auth, profiles, image uploads, tweeting, likes, retweets, bookmarks, DMs, notifications, and deployment—all in about 20–40 minutes.

This video reviews Kimi K2’s 0905 update—doubling context to 262K, improving frontend/code generation, and weighing speed, reliability, and prompt-caching costs versus providers like OpenRouter and Groq—with hands-on demos and cost/latency analysis.

Analysis of Anthropic’s $13B raise at a $183B post-money valuation, comparing dilution, investor strategy, compute costs, and competitive dynamics with OpenAI and Google.

A walkthrough showing how MCP-UI enables tools to return full UI (HTML/JS/iframes) and how to build an MCP-UI server with TanStack Start, test it with the MCP-UI Inspector, and render an iframe card in Goose.

A hands-on look at running local MoE coding models like Qwen 3 235B on the Framework Desktop, comparing TPS and context trade-offs, the limits of agentic workflows, and a prototype local code editor that applies AI-generated diffs directly to files.

Theo argues that vibe coding is valuable for quickly producing throwaway or legacy-bound code while emphasizing that you still must understand how code works; AI should augment skilled engineers, not replace them.