Thilak

Top 10 AI highlights from yesterday (Thu, Oct 29). Google had its worst stock day ever. OpenAI rewrote the rulebook for how its models should behave. And Google is shutting down one of its own coding agents. Thread.

Valid.

OpenCode 2.2.1 adds /goal: the session keeps working until a reviewer confirms the objective is met. - /goal pause, resume, clear - status under the prompt - also new: opt-in /schedule My take: a vague goal just burns tokens longer. How do you define done?

Google's export guide for Jules: open a CLI agent like Claude Code and paste our prompt. - Nov 11: sessions stop, read-only - Dec 31: shut down - export lands in ~/jules-export - env-variables.json holds real secrets, keep it out of git Moving to Antigravity, or elsewhere?

My take: yesterday was about what a lead is worth. Google plans $195B+ of capex this year and got punished for it. Mistral shipped a 1T open model. Two labs reportedly trained on GPT-6 Astra. If a frontier lead can be distilled in weeks, what is it actually worth?

10/ Shane Legg and Alex Imas of Google DeepMind argue AI hasn't changed the economy yet because agents can't coordinate. Their model: smarter agents shift wages gradually. Agents that coordinate past a tipping point make firms abruptly huge. Most benchmarks don't measure that.

9/ Canva launched Brush, its own vision-language model for editing designs. It reads layout and layers, edits by writing and running code, and keeps everything editable. Canva says it's up to 69% faster and 46% cheaper per edit than its previous design agent. Pro users first.

8/ Pika open-sourced Pika H3 Flex, a video model post-trained on MiniMax H3. It makes video with synced audio in six sampling steps and takes up to 12 references. The open weights run at 480p and 768p. 4K and the five audio stems stay on Pika's hosted API.

7/ GitLab Orbit is GA: a context graph of code, merge requests, pipelines, vulnerabilities and owners that agents query through MCP or glab. Compare the Market tested an AI code reviewer on 79 real merge requests: ~70% of expected comments with Orbit, ~58% with RAG.

6/ Google is shutting down Jules, its async coding agent. Sessions work until Nov 11, then it goes read-only, and it shuts down Dec 31. Google's export guide has you install the Jules CLI and let another coding agent pull your data out.

5/ Qwen4 may be days away. An attendee at Alibaba Cloud's Apsara conference reported CTO Feifei Li saying "Qwen4 next week," and a stage slide listed Qwen4-Max, Plus, Flash and 27B. Alibaba hasn't confirmed a date officially yet.

4/ The Information reports DeepSeek and Moonshot leaned heavily on GPT-6 Astra outputs to train Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017. Earlier this month Anthropic said its defenses cut distillation harvesting from Claude by 90% to 99%. Your API is someone's training set.

3/ Mistral released the open weights for Mistral Large 4 ("Le Chonk"): about 1T parameters, 52B active, 1M context. Three more weeks of RL took it from 28% to 41% on Terminal-Bench 4.0, the top open-weight score there. The license limits model-as-a-service offerings.

2/ OpenAI rewrote its Model Spec around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. Faithfulness is the foundation: understand what people want, not just the literal instruction. Most style rules are gone.

1/ Alphabet fell about 13% on Thursday, its worst day since Google went public in 2004, wiping out roughly $500B in market value. The trigger: Q3 free cash flow of -$9.07B (analysts expected -$2.7B), with 2026 capex held at $195B to $205B and more coming in 2027.

The fastest way to stop an agent fixing the wrong bug: make it fail first. Before any fix: 1. write a test that reproduces the bug 2. run it, show it fails for the reported reason 3. fix 4. rerun it plus the suite No red test, no fix. Which step does your agent skip?

An AI code reviewer with RAG did worse than one with no added context. Compare the Market, 79 real merge requests: - GitLab Orbit graph: ~70% of expected comments - RAG: ~58% My take: who changed the code beats what looks similar. What does your reviewer see?

GitHub Agentic Workflows is GA: agents in Actions, written in Markdown. Day one: gh extension install github/gh-aw gh aw add-wizard githubnext/agentics/repo-status gh aw compile gh aw run repo-status gh aw logs repo-status gh aw forecast Commit both the .md and the .lock.yml.

A lost race doesn't crash Linux, it only fails the handshake. So an unreleased Claude model kept retrying: 48 min to leak a kernel pointer, 5 more to root on an MPTCP server. CVE-2026-98432, CVSS 9.8, kernels 4.4+ Fixed: 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56

3 more weeks of RL took Mistral Large 4 from 28% to 41% on Terminal-Bench 4.0. Weights are out. - ~1T params, ~49B active - 1M context - 47% fewer tokens per task than GLM-5.3, per Mistral My take: top of the open pack, still ~20 points behind Mythos 5.1. Self-host it?