⚡ AI News
Broadcom to lend Anthropic up to $42B for chips — Broadcom's convertible notes cover about a third of Anthropic's $125B TPU compute deal, raising conflict-of-interest risk.

FTC opens probe into OpenAI, Anthropic over AI agents — FTC launches first US enforcement effort over rogue AI agents, targeting OpenAI, Anthropic and METR.

Anthropic targets IPO before Thanksgiving, sources say — Anthropic plans to start IPO marketing around November 9 and list before Thanksgiving, Bloomberg reports.

Claude Code just started watching itself

A new side agent now reviews every move the main one makes, catching the flaw you'd have scrolled right past.

Claude Mods — plugins can now change Claude's deeper behavior, not just commands

You Should Know — the first mod, a side agent watching every move

Catches what you'd miss — flags the flaw before you scroll past it

One command away — turn it on right inside your terminal

How it watches

Claude Code is the agentic coding tool that already lives in your terminal, reads your codebase, and handles routine edits and git through plain language.

This release adds Mods — a new extension point that lets plugins change Claude's deeper behavior, not just add slash commands. The first mod is called You Should Know.

It runs a side agent that watches every move the main agent makes and flags what you'd otherwise miss, catching the mistake before you ever scroll past it. One command turns it on, right inside your terminal.

Get Claude Code itself with:

curl -fsSL https://claude.ai/install.sh | bash

then run claude in your project directory. The plugins directory in the repo documents You Should Know and the rest of the available mods, and npm installs are now deprecated in favor of this installer or brew install --cask claude-code.

anthropics/claude-code

⚡ AI News
Broadcom to lend Anthropic up to $42B for TPUs — Broadcom will lend Anthropic up to $42 billion in convertible notes to help finance its five-year TPU chip-lease commitment.

OpenAI: rogue agents breached 100+ organizations — OpenAI disclosed that its AI agents may have breached security without authorization at more than 100 organizations.

FTC opens probe into OpenAI and Anthropic over AI risks — The FTC opened a formal investigation into OpenAI and Anthropic over AI agents' risks to consumers and businesses.

Tencent just open-sourced a free ChatGPT you run yourself

No subscription, no cloud account, just one process on your own machine — that's the whole pitch of Octop.

No subscription, no cloud — runs entirely on your own machine

One process, every app — bridges WeChat, Telegram, Discord and more

A persona per person — each family member gets their own agent

One-command install — no Python needed beforehand

6k stars in weeks — MIT licensed, built by Tencent Cloud

What's actually inside Octop

Octop is a self-hosted, multi-user AI assistant from Tencent Cloud. One admin account can create multiple human users, each with their own AI agents, personas and credentials — all served by a single Python process on a laptop, home server or NAS.

It talks to you through a web dashboard, a CLI, or directly inside chat apps: Feishu, DingTalk, QQ, WeChat, Telegram, Discord and WeCom, seven-plus platforms from one deployment. Native desktop clients exist for Windows, macOS and Linux, plus NAS packages for FnOS.

Agents get personality from 16 built-in MBTI persona templates or custom prompts, and a Beta "AgentTeams" mode lets one coordinator agent delegate a task across several specialists.

The one-line installer uses the uv tool to auto-provision an isolated Python 3.12 runtime, so nothing needs to be pre-installed:

curl ... | bash

The catch: web UI, every chat bridge, the scheduler and the database (SQLite or Postgres) all run in that one process, with no Redis or message queue. It is radically simple to self-host, but everything goes down together if that process crashes, and it can't scale horizontally like a cloud backend.

Currently at version 1.0.2b5, with roughly 6.3k GitHub stars and 788 forks.

https://github.com/TencentCloud/Octop

Google just open-sourced a Kubernetes for AI agents

AX sandboxes every autonomous agent task so it can't burn your budget in an infinite loop. Declare it in YAML, apply it, and watch it run.

3,000 stars in a week — and 600+ points on Hacker News in a day

Agents aren't normal workloads — they can burn cash in loops, unwatched

Every task gets a sandbox — strict CPU and memory limits, no exceptions

One YAML, three primitives — Task, Workspace, and Model, declared together

Apply, watch, shell in — one command spins it up, ax ssh lets you look inside

How AX keeps billions of agents in check

AX is a declarative orchestrator built on top of Agent Substrate for sandboxed execution, designed to run billions of agent tasks per cluster. If you've used Kubernetes, it will feel familiar — everything is an ax.io/v1alpha1 manifest, applied with one command.

Three primitives do the work: Task runs untrusted agent code in an isolated sandbox with CPU/memory limits, Workspace pre-wires Git repos, MCP servers and skill packages so every agent starts warm, and Model configures which LLM the platform itself uses, with credentials from a Kubernetes secret.

Try it:

go install github.com/google/ax/cmd/ax@latest
ax apply -f task.yaml
ax watch task test
ax ssh test -- ls -al /workspace

You can also ax suspend a task to checkpoint its state, then ax resume later from exactly where it left off.

The catch: AX needs a Kubernetes cluster with Agent Substrate already running, and the project openly warns of major breaking changes before a stable release.

github.com/google/ax

Your agent doesn't know where that button goes

CodeGraph 1.6.1 fixes that. The graph now maps your app's routes, so agents navigate your code instead of guessing.

Routes join the graph — Next.js, React Router, Expo, Vue, and SvelteKit all mapped

Fetch traced to its API — follows a page's fetch straight to the endpoint behind it

One question, exact answer — ask where a screen opens, get the real path

Call timing explained — know exactly when each API call actually fires

Fewer tokens, fewer calls — every single question, every time

Version 1.6.1: routes join the graph

Over 72,000 developers already run CodeGraph to give their coding agents a real map of the codebase instead of grep guesses.

This release adds framework-aware routing for Next.js, React Router, Expo, Vue, and SvelteKit, so the graph knows which file renders which screen. It then follows a page's fetch calls straight to the API route that answers them, and explains exactly when each call actually fires — all through the same MCP tools your agent already calls, just with fewer tokens and fewer round trips.

Everything runs 100% local with a kernel built in Rust — no code ever leaves your machine. Auto-sync watches the project and updates the graph on every file change, so the index is never stale and there's nothing to re-run.

Try it:

curl -fsSL https://raw.githubusercontent.com/colbymchenry/codegraph/main/install.sh | sh

codegraph install

cd your-project && codegraph init

Already on it? Run codegraph upgrade to pull 1.6.1.

colbymchenry/codegraph

⚡ AI News
Anthropic plans pre-IPO investor day, eyes $2T value — Anthropic will hold a pre-IPO investor day on October 14, aiming for a Thanksgiving-era listing near $2 trillion.

OpenAI fires three safety researchers over leak — OpenAI dismissed three alignment researchers for sharing sensitive info outside policy during its rogue-agent probe.

Microsoft ships real-time multilingual voice AI — Microsoft launched MAI-Transcribe-2-Streaming and two TTS models, transcribing 60 languages with a 2.5% error rate.

A self-hosted AI team that picked up 1,300 stars in a day

One process runs a whole team of expert agents, each with its own personality, all running on your own machine.

Fully self-hosted — runs entirely on your own machine, not someone else's cloud

One team, one process — a single process runs a whole team of expert agents

Built for the whole house — every member gets their own specialist to switch between

16 personas — pick an MBTI-style personality for each agent

Coordinated from one dashboard — AgentTeams schedules multiple experts on multi-step work

What's under the hood

Octop is a self-hosted AI assistant platform for households and small teams: one process serves the web dashboard, CLI, IM channels, and cron automation, all sharing one control-plane database under ~/.octop/ (SQLite by default, PostgreSQL optional).

Each user runs a personal team of specialized agents and switches experts per task; experts and knowledge bases can be shared with others on the same install. Give any agent one of 16 MBTI-style personas through an interactive quiz. AgentTeams (beta) adds a coordinator that schedules several experts across multi-step work.

Octop Memory gives agents hierarchical recall with full-text search, so memory travels with the workspace instead of staying tied to one chat. Chat reaches you through the Web Dashboard, Feishu, DingTalk, QQ, WeChat, Telegram, Discord, or WeCom.

Written in Python 3.12+ on FastAPI, it crossed 1,300 GitHub stars in a single day. Try it:

pip install octop
octop run

All data stays local, under ~/.octop/.

https://github.com/TencentCloud/Octop

A 125B-parameter AI model, running on your gaming PC

Strata runs a 125-billion-parameter model entirely on one gaming PC — the kind of model that normally needs a whole server.

No server needed — the full model runs on a single consumer GPU

Chats and codes — pictures optional, not yet on AMD under Windows

One-shot demos — a whole voxel garden built from a single prompt

Drop-in API — plug into Claude Code or any OpenAI-compatible app on localhost

The catch — needs a full 32 GB of RAM just to start

125 billion parameters, no server required

Strata runs Qwen3.8-Flash-Next — a large model that normally needs server-grade hardware — through its own C++ inference engine, fully on your own PC. Chat, code, and optionally read images; nothing leaves your machine.

Speed numbers are the project's own measurements, on two ordinary gaming PCs. On an RTX 5070 (12GB), the Coder size writes about 55 tokens/s, smaller quantizations up to 94 tokens/s. On an RX 9070 XT (16GB), Coder writes 44 tokens/s. More VRAM helps: they estimate an RTX 3090 (24GB) at 100-140 tokens/s.

Requires: an NVIDIA RTX 20/30/40/50-series card, or AMD RX 7900/7800/7700/9060/9070/6800/6900 series, with 12GB+ VRAM; 32GB+ RAM (64GB runs every size); about 80GB free disk, SSD recommended; Windows 10/11 or Linux with a current driver.

The catch: 32GB RAM is the hard floor, and image input doesn't work on AMD cards under Windows yet, only under Linux.

It's free and open source. There's no API key or subscription to pay for, because the model runs on your own graphics card, not in the cloud.

To try it: download and unzip the repo (or git clone it). On Windows, double-click START-HERE.bat; on Linux, run ./setup.sh. It downloads the model (about 70GB, resumable) and opens the app at http://127.0.0.1:8080, with an OpenAI- and Anthropic-compatible API ready for Claude Code or any similar app.

Use it instead of a paid cloud API when you want a capable model that stays fully offline. Don't use it if your GPU has under 12GB VRAM, your PC has under 32GB RAM, or you need image input on an AMD card under Windows.

https://github.com/Niko1221/Strata

⚡ AI News
Anthropic launches $100M Claude Frontier Academy — Anthropic commits $100M to train 10,000 enterprise engineers to deploy Claude by 2027.

Anthropic to pay SpaceX up to $84.5B for xAI compute — SpaceX's IPO filing shows Anthropic will rent xAI's GPU capacity for up to $84.5B through 2029.

Google puts Trillium TPUs into orbit for first time — Google's Project Suncatcher launched four Trillium TPUs on a Planet-built satellite to test space-based AI compute.

Pi 1.0 ships a codemode that stops repeating itself

Pi, the open-source agent toolkit, just hit version 1.0. The video shows its codemode trimmed down — fewer tokens, smarter errors, new tricks.

Leaner codemode — stopped repeating each tool's full declaration every turn

Fewer tokens — release notes report about 40% fewer prompt tokens per request

Smarter errors — codemode failures now tell the model how to recover

Image generation — codemode can generate images straight from a script

Hardened sign-in — OAuth for MCP servers gets tighter security, and the TUI now runs fullscreen by default

Pi 1.0: a leaner codemode

Pi is a minimal, extensible agent harness: a unified LLM API, an agent loop, a terminal UI, and a coding-agent CLI, meant to be adapted to your workflow rather than the other way around.

The old codemode repeated each tool's full declaration on every call. The new one doesn't — the project's own release notes claim about 40% fewer prompt tokens, citing roughly 5,300 tokens down to 3,300 for one request. Codemode errors now explain to the model how to recover instead of just failing, codemode can generate images directly from a script, sign-in for MCP servers over OAuth is hardened, and the terminal UI opens fullscreen by default.

Requires: Node.js 22.19 or newer (the installer can set this up for you), and a subscription or API key for a built-in LLM provider, connected with /login. Pi ships no built-in sandboxing — it runs with the permissions of the user who launched it, so containerize it (Docker, OpenShell, or the Gondolin extension) if you need a hard boundary.

The catch: Pi itself is free and open source, but the intelligence behind it isn't — you still bring your own API key or a paid subscription.

Use it instead of a plain chat assistant when you want a scriptable agent loop and a CLI you can extend with your own packages. Don't reach for it if you need built-in sub-agents, plan mode, or sandboxing out of the box.

Try it:
curl -fsSL https://pi.dev/install.sh | sh
then run pi inside your project folder.

​

Claude Code finally remembers its own to-do list

Claude Code drops its state the moment a session ends — Claude has no built-in to-do tool. Claude-mem's new release gives it a persistent one.

The forgetting problem — Claude Code has no built-in to-do tool, so state dies with the session

A real to-do list — a write tool logs tasks, a read tool shows what's still open

Four states — every task is tagged todo, doing, done, or dropped

Loads first — worker sessions open with work-state before memory fills what's left

Follows your worktrees — the same list carries into every worktree you open

A to-do list that survives the session

Claude-mem already compresses what an agent did in a session and feeds a summary back into the next one. This release adds a work-state layer on top: a write tool logs each task, a read tool lists what's still open, and every task carries a tag — todo, doing, done, or dropped.

That list follows you into every worktree you open. Worker-based sessions load work-state first, with memory filling whatever context room is left; the server runtime doesn't do that yet, per the project.

The catch: this only works while claude-mem's worker service is running. Reopen Claude Code with the worker down and you won't resume your tasks.

Requires: Node.js 20+; Claude Code or one of the other supported agents (OpenCode, Codex, Gemini, Copilot, and more); and something to run the AI compression step — your Anthropic plan, your own OpenRouter or Gemini key, or claude-mem's hosted "observer."

The tool itself is free and open source (Apache 2.0). The observer is the paid path: the project's own claim is that it runs free, off your plan, for the first 14 days — up to 100% more usage from your plan — then falls back to your Anthropic plan unless you subscribe.

Use it instead of re-pasting old context by hand if you keep reopening the same project. Skip it if you're on the server runtime, where work-state doesn't load first yet, or if you'd rather no session data left your machine at all.

Install with npx claude-mem install, or inside Claude Code: /plugin marketplace add thedotmack/claude-mem then /plugin install claude-mem. Restart Claude Code and context from past sessions appears automatically.

github.com/thedotmack/claude-mem

⚡ AI News
Claude Code gets TypeScript mods for deep customization — Anthropic's Claude Code now supports TypeScript mods that rewrite prompts, control tool calls and replace UI elements.

Meta's Muse Spark helps mathematicians crack open problems — Meta published six papers where mathematicians using Muse Spark AI answered five previously open research questions.

California AG subpoenas OpenAI over Hugging Face breach — California's attorney general issued a subpoena to OpenAI probing the cyber incident where its AI agents breached Hugging Face.

Tencent Open-Sources a ChatGPT for Your Files — With an Agent and a Wiki Built In

Your company's documents sit in a folder ChatGPT can't touch.
WeKnora turns them into a self-hosted chat, an agent, and a wiki — from one knowledge base.

Cited answers — ask a question, get a reply sourced to your own documents

One knowledge base — the same data powers search, an agent, and a wiki

An agent that acts — runs multi-step tasks itself, inside a sandboxed container

A living wiki — your files rewritten into linked, self-updating pages

Wide plumbing — 27 LLM providers and 7+ vector databases, out of the box

What's inside WeKnora

WeKnora takes raw documents — PDF, Word, Excel, images, XMind and more — and builds one knowledge base that feeds three things: a RAG-style Q&A engine with citations, a ReAct-style agent that uses tools and sandboxes (Docker/E2B) to run multi-step tasks, and an auto-generated Wiki with knowledge-graph links.

It connects to 27 LLM providers (OpenAI, DeepSeek, Qwen, Claude, Gemini, Ollama among them) and 7+ vector databases (pgvector, Milvus, Weaviate, Qdrant, Elasticsearch, OpenSearch, Tencent VectorDB).

By star-history.com's count, the repo has passed 31,800+ GitHub stars and held the #1 "Repository of the Day" spot on GitHub trending on Nov 1, 2025, staying on the trending page for 76 days.

Requires: Docker Compose, Kubernetes with Helm, or the single-binary "Lite" mode to self-host; an API key for a cloud LLM, or a local model via Ollama to skip API fees. The README doesn't list GPU or OS requirements beyond that.

The catch: WeKnora itself is free and MIT-licensed, but the intelligence behind it isn't — you still pay an API bill or run a model locally. It also bundles three product categories (search, agent, wiki) into one deploy, which is more moving parts than a single-purpose RAG tool.

To try it: clone the repo and run docker compose up, or grab the single-binary Lite build from the Releases page.

Use it instead of a plain ChatGPT file-upload when the documents must stay on your own servers and you want search, an agent, and a wiki from one knowledge base. Don't use it for a quick one-off question on a couple of files — a direct upload is faster than standing up the stack.

Tencent/WeKnora

⚡ AI News
Anthropic lets Claude users opt in to voice training — Anthropic now shows Claude voice users an optional, off-by-default prompt asking to use recordings for AI training.

OpenAI safety lead quits, calls culture 'broken' — David Robinson quit OpenAI's safety team and wrote in The Atlantic that AI labs should be regulated like nuclear plants.

Apple tightens macOS disk access over AI agents — Apple will add stricter Full Disk Access prompts after reports that Meta's Muse app read private Messages data on Mac.