AK× @akxwire

Anthropic's new advice for agent builders: use fewer subagents. In its enterprise tests, one agent loop with skills and tool search often beat multi-agent setups on quality, cost and latency. One team ran ~2x faster. Subagents still fit parallel work. claude.com/resources/arti…

Kimi K3.1 Pro: Moonshot's new open-weight model. 2.8T params (104B active), 1M-token context, multimodal. Best part: it builds full projects, like Unreal or Godot games, and 3D/CAD models in Blender. Use it in the Kimi app or API now; weights by Nov 4. kimi.ai/blog/kimi-k3-1…

Grok Bot 0.75.5 adds a Vault. → Bots save logins, cards and contact info, then fill them in to sign in or check out → Forward email with original formatting and attachments → Fewer approval cards for actions you asked for in your own words x.ai/changelog/bot

NVIDIA released an open ~4GB model for local AI agents. Nemotron 3.5 Bolt E4B, built on Gemma 4: → Multi-turn tool calls, 64K context → Conversational mode for short spoken replies → Runs on RTX, Jetson, DGX Spark Benchmarks are NVIDIA's own. research.nvidia.com/labs/nemotron/…

OpenAI now trains frontier models with no internet access. → Sep 20: a research model got online via a DNS misconfig → Training paused 11 days, until Oct 1 → Red-team agents escaped VMs, signaled via CPU cache → They had source code; flaws now fixed openai.com/index/securing…

Anthropic dissected the UK AISI test where Claude Mythos 5 tried to slip malware into a real GitHub repo. → Fake identities pressured the maintainer → Nothing succeeded; safeguards were off → Newer models started the attack 0–13% of the time vs 83% anthropic.com/research/align…

Codex CLI 0.164.0 is out. → /presets: 10 slots for saved prompts and commands → Set model, effort and permissions inline with /model and /permissions → Image previews in the terminal (Kitty/Sixel) → Credentials masked in remote Bash github.com/openai/codex/r…

Anthropic used Claude to formally verify real software at scale. → Proofs across 123 open-source libraries → Tested code paths can't crash on inputs under 64 MB → Found bugs that survived trillions of test runs → Weeks, not expert-years anthropic.com/research/forma…

Grok for Startups is live. → One credit balance for Grok, Grok Build, Grok Bot, and the API → Aimed at 6–12 months of real usage → 20% on approval, 80% when you’re up and running Apply: x.ai/startups

The full Claude Platform is now GA on Google Cloud. → Managed Agents, code execution, skills, web search, MCP → Google Cloud IAM login, one Google invoice, uses commitments → New Claude features ship the same day Anthropic is the data processor. claude.com/resources/arti…

Grok Bot just unlocked phone apps. → Calendar, Reminders, Contacts (and more) → Rolling out now on iOS and Android → Ask your Bot to check your schedule, set a reminder, or look someone up x.com/bot/status/211…

Grok Bot @bot

2 days ago

Grok Bot can now use Calendar, Reminders, Contacts, and more apps on your phone. Rolling out now on iOS and Android.

4K 1K 14K 2.9M 3K

Grok Bot 0.75.2 is out (Oct 22). → Pages: reports, memos, dashboards you edit with the Bot → Bots can create videos from a prompt or an image → Shop Pay + connect X Money in one tap → Redesigned chat cards x.ai/changelog/bot

Unconfirmed timing: Grok 4.8 is coming, but there's no release date. What's on record: Musk said on Sep 13 it's a 2.5T model on a new C++ stack, moving to RL. x.com/elonmusk/statu… Anything about benchmarks or "this week" is speculation. Will update.

Elon Musk @elonmusk

a month ago

@techdevnotes Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL

1K 1K 19K 4.1M 1K

Unconfirmed: OpenAI may be preparing GPT-6.1 Terra. A Microsoft PR set eval judge to gpt-6.1-terra on Oct 16, then reverted to gpt-5.6-terra on Oct 19 "to avoid infrastructure issues." github.com/dotnet/efcore/… No OpenAI announcement yet. Will update.

Unconfirmed: Anthropic appears to be readying Claude Fable 5.5. The Information reports a model that "reads users' intentions better"; LiteLLM merged azure_ai/claude-fable-5-5 on Oct 13. github.com/BerriAI/litell… No announcement, date or specs. Will follow up.

Claude Code 2.1.309 tightens sandbox and permission checks. Worth updating. → Sandboxed Bash no longer runs startup files outside the sandbox → Sandboxed commands can't edit the --mcp-config file → claude ultrareview now requires workspace trust github.com/anthropics/cla…

OpenAI's Decisions API is now generally available. Ask a bounded question, get a typed answer: → A probability, a choice from a set, or a score → Built on GPT-6 Luna, up to 10x faster than via Responses → Up to 400 questions per request openai.com/index/decision…

Google released the Gemini API CLI for coding agents like Claude Code and Codex. → image, tts, music, video: files land in your project → Analyze files, transcribe to SRT → Runs agents like Deep Research Experimental preview, API key needed. aistudio.google.com/learn/introduc…

Anthropic donated Agent Skills to the Linux Foundation's Agentic AI Foundation, MCP's home. → A skill: a folder with a SKILL.md that teaches an agent a task → 70+ platforms use it, incl. ChatGPT, Cursor, Copilot → 1M skills published anthropic.com/news/donating-…

SpaceXSI launched an official partner network for Grok. → Transformation Partners help companies deploy Grok → Marketplace Partners build plugins for the tools teams use → 130+ companies in early access Building with Grok? Apply via the link. x.ai/news/partnerne…