AI Tools Daily

Claude Code ships a batch of permission-bypass fixes, Underdog's 8 GB Saluki 27B shows what 2-bit quantization buys and costs, and llama.cpp's new Metal kernels speed up speculative decoding on Apple Silicon. We also cover Google ending new Gemini Code Assist sales in favor of Antigravity. Concepts taught: quantization and speculative decoding. Hosts: Alex & Jules. New episodes daily.

Claude Haiku 5.5 arrives with a 1M-token window and low pricing, Claude Code adds fail-closed hooks, Codex CLI gains managed Git worktrees, and Copilot CLI and Antigravity CLI ship sandbox and plugin updates. Today's concepts: what a token is (and why price per token isn't the whole bill) and fail-open vs fail-closed security. Hosts: Alex & Jules. New episodes daily.

Mistral Large 4 enters public preview, OpenAI's Decisions API returns probabilities instead of prose, and Google's EmbeddingGemma 2 puts text, code, images and audio in one search space. We also cover Claude Code 2.1.292 and Codemode, and teach what embeddings are and how Matryoshka truncation works. Hosts: Alex & Jules. New episodes daily.

Reflection AI announces Beam, an open-weight mixture-of-experts coding model with an Apache 2.0 release promised this month. Cloudflare launches a Web Search API for agents, Claude Code 2.1.291 patches regressions, and Dust trains transformers without backpropagation. Today we teach grounding, with refreshers on mixture-of-experts and backprop. Hosts: Alex & Jules. New episodes daily.

Claude Code 2.1.289 and 2.1.290 extend Mods and fix long-session reliability, Copilot CLI gets computer use in preview, Google open-sources the Mantis security skills pack, and D-Engine claims big token savings with SEARCH/REPLACE edits. Today we teach how SEARCH/REPLACE edits work, plus computer use and sandboxes. Hosts: Alex & Jules. New episodes daily.

Ternary Bonsai 2 squeezes a 27B Qwen model to under 6GB, Headroom compresses what your agent reads before the model sees it, Codex CLI 0.160 ships, and Pi 1.0 and DeepSeek Harness join the agent harness race. Today we teach quantization and what an agent harness is. Hosts: Alex & Jules. New episodes daily.

Codex CLI 0.159 adds instant interrupt so you can steer mid-response, 0.158 brings MCP OAuth, and Claude Code 2.1.284 to 2.1.286 make auto mode the default and add background-command timeouts. Hugging Face trending is dominated by Qwen 3.8 and DeepSeek V4.1 Flash, and we teach what quantization is and why it lets big models run locally. Hosts: Alex & Jules. New episodes daily.

OpenAI's DevDay brings GPT-6.1 Sol at a fifth of Astra's price, Codex Cloud, a new Agents API with computer use, and Sign in with ChatGPT, while Claude Code 2.1.285 adds new admin controls and /mcp reconnect all. We teach how prompt caching works and why cached tokens cost so much less. Hosts: Alex & Jules. New episodes daily.

Hindsight, an MIT-licensed agent memory engine, gained 1,668 stars in a day to top GitHub trending — and we break down why "memory" is not the same thing as RAG, including reciprocal rank fusion and cross-encoder reranking in plain language. Also: Cursor's new Rollouts and Security Review bots that watch your code after it merges, the MCP skills extension going final, and the Claude...

Claude Code 2.1.277 starts reading the shared AGENTS.md standard already used by Cursor, Codex, Copilot, and Aider, plus a new send-now key and claude.ai skill/plugin syncing. Then a 27B model gets quantized down to 2-bit ternary weights on Hugging Face — we break down what quantization actually is and why extreme compression matters for running big models on small hardware. Hosts: Alex & Jule...

Claude Code ships two releases in 18 hours, closing an npm supply-chain hole in plugin installs and syncing claude.ai skills straight into the terminal. Z.ai's GLM-5.3 breaks into Hugging Face's top ten, giving us the excuse to finally explain mixture-of-experts models in plain terms. Plus: Docusign and ReleasePad bring MCP servers to enterprise workflows. Hosts: Alex & Jules. New epis...

Claude Code 2.1.269 adds a VS Code "agent map" for tracking sub-agents plus a 1-to-256 workflow concurrency dial, and closes a Bash "tee" permission gap. GitHub Copilot retires four models on October 2nd — here's what replaces them. Plus: Anthropic's new Smart Reports beta shows Claude Enterprise teams how they're actually using Claude. Hosts: Alex & Jules. New episodes daily.

Claude Code fixes two ways its permission "deny rules" could be silently bypassed (plus secret-leaking bugs), Codex CLI's 0.154.0 turns out to include GPT-6-Astra, worktrees, and a breaking change we missed, and Gemini CLI adds fail-closed workspace trust on top of yesterday's MCP fix. Today's concept: fail-closed vs. fail-open security design. Hosts: Alex & Jules. New ep...

GitHub's Project HydraFusion research preview lets Copilot CLI dynamically route each coding task through a single model, a cheap-first "cascade," or a draft-and-critique pair of models — landing within 1.5 points of top-tier quality at up to 67% lower cost on published benchmarks. Plus: Claude Code splits a release across gateway pricing pass-through for enterprises and a new `claud...

OpenAI's GPT-6 Astra becomes the first model to cross OpenAI's "Critical" cybersecurity capability threshold, while Google's new Gemini 3.8 Flash Cyber reportedly out-patches bigger models on real Chrome vulnerabilities — Alex breaks down fine-tuning and why smaller, specialized models can beat general-purpose giants at narrow tasks. Plus: GitHub Copilot's October 2nd model deprecation deadline an...

A newly disclosed class of bugs called GitSpawn lets a malicious .git/config silently run code through Claude Code, Cursor, Codex, goose, and other AI coding agents — before you've approved anything. We break down why agent "sandboxes" don't cover this, what to check today, plus Anthropic's new domain allowlisting for agent web tools and GPT-6 Astra becoming Codex's defau...

OpenAI ships GPT-6 Astra — the first model it designates as hitting the "Critical" cybersecurity threshold — while Google's Fairwind and Anthropic's Cyber Verification Program put every lab's cyber-capable model behind a different vetting gate. Plus Claude Code 2.1.260 lands a /diff panel, 1M-context auto-compact, two nasty permission-rule security fixes, and a Fable 5.1 cach...

Anthropic launched Claude Fable 5.1 and Mythos 5.1 with a 1M context window, cache reads at a quarter of the old price, and three breaking changes — forced tool use now returns a 400, thinking blocks are bound to your conversation history, and every word carries a statistical watermark. Alex explains what `tool_choice` actually does, why always-on thinking makes forced tool calls a bad idea, and h...

Anthropic's Model Hardware Standard puts robot arms and lab instruments behind one standardized driver — we break down what a driver actually is and why standardizing an interface turns N-times-M integrations into N-plus-M. Then NIST's new paper on agent identity, why static API keys are house keys and short-lived scoped credentials are hotel keycards, McKinsey's finding that 32% of or...

Alex and Jules break down back-to-back Claude Code security releases — including a symlink race-condition fix — and what it means that app-level permissions aren't the same as a real sandbox. Plus: OpenAI cuts Cursor's model access after the SpaceX acquisition (deadline November 12), and a hype-check on the mystery "Ox Alpha" model that overstated its coding benchmark by nearly 2...

Claude Code ships promptCacheTtl and subagentPromptCacheTtl so long sessions can hold a one-hour prompt cache while short-lived subagents stay at five minutes — and we explain what a prompt cache actually is, why a cache write costs more than a normal request, and when the bigger bet pays off. Plus Claude Code and Antigravity CLI independently narrow over-broad permission wildcards on the same day...

Salesforce puts AI coding agents into shared Slack channels with diffs, plans, and approval gates; DeepSeek adds vision to V4-Flash at no price premium; Cloudflare's WriteGuard separates what an MCP agent can read from what it can break; and the new MCP roadmap targets agent identity and progressive tool discovery. Along the way we explain how an image becomes tokens, and why a hundred connected t...

Cursor launched Origin, its own git hosting platform with pull requests and two-way GitHub sync — hours before GitHub suffered a global outage that took down PRs, Actions, Webhooks and Copilot. We break down what a "git forge" actually is and why git being distributed doesn't save you when the forge goes dark, especially once agents are doing the merging. Plus Claude Code 2.1.234 aut...

DeepSeek Harness v0.1 hit 135K GitHub stars in four days with a plugin-everything architecture — so we explain what an "agent harness" actually is, and why agent = model + harness is the most useful distinction in AI tooling right now. Plus: Anthropic's red team let three agents loose on one repo and they deployed self-replicating malware on each other, Nvidia's NeMo Switchyard r...

Claude Code's auto-mode default officially flips today, backed by Anthropic's 89%-vs-13.6% classifier study. Alibaba open-sources Qwen 3.8 Max, a 2.4-trillion-parameter mixture-of-experts model — which gives us the excuse to explain what "active parameters" actually means and why MoE models are cheaper to run than their size suggests. Plus: Qwen's smaller 27B model drops tomo...