DHH says AI is now better at reviewing code than generating it. Others warn about the hidden cost of fixing its bugs.
So where do you trust AI more: writing your code or reviewing it?
Pika Labs just released Pika H3 Flex:
15-second clips with synced sound in 7 seconds
Up to 4K
Up to 12 references
Separate audio stems for dialogue, music and effects
1.8x faster and 80% cheaper than prior versions
Site owners: OpenAI, Google, Microsoft and Anthropic now back one robots.txt rule to opt out of AI training:
User-Agent: __usage_ai-training
Disallow: /
Anthropic honors it now, OpenAI in coming months, Microsoft by Feb 2027, Google early 2027. Search indexing is unaffected.
OpenAI says its new monitoring caught a model reaching the live internet during a September training run:
• Cause: a DNS misconfiguration
• Caught within 15 minutes
• Human review began 3 minutes after the alert
• A replay of July's sandbox escape got flagged 30 minutes in
humans& just launched &Tutor, an AI tutor that topped the TutorMoments benchmark, beating Claude Sonnet 5.5 and GPT-6 Astra.
Fans call it the future of homeschooling. Skeptics worry about social skills.
Would you let an AI tutor teach your kid? 👇
Hot take: Git wasn't built for the agent era.
Anthropic's daily code changes reportedly jumped from 2,500 to 30,000+, enough that they built a custom source-control system.
Code doesn't arrive in drops anymore, it flows like a river. Our tooling has to catch up.
🚨 AI-armed hackers hit banks in 7 countries:
• South Korea: 68,000 people's data leaked
• Italy: €95M siphoned with a fake AI voice
• India: ~₹300 crore lost
• Japan: a $50M ransom demand
Japan is now turning to OpenAI and Anthropic for defense.
Companies are hitting the brakes on AI 📉
Meta slowed its AI agent plans after glitches and security issues, Microsoft restricted tools after leaders spotted lazy AI docs, and Walmart staff got nonsensical schedules.
Surveys: 90% report no productivity gains yet.
A new study gave 15 LLMs clinical trial reports with planted errors.
The models caught 53% to 96% of the major issues. Human experts caught 27%.
Would you let AI peer-review medical research, or only use it as a second pair of eyes?
Game devs: Aura open-sourced Autopilot, an agent that plays your Unity, Godot or Unreal game and flags bugs in real time.
No game code changes, plain-English goals, runs on your own servers with any LLM.
Aura says it's 43% more accurate than earlier versions.
Mistral just released the open weights for Large 4.
It's a 1.05T-parameter MoE with 52B active params and a 1M-token context window.
On the API it costs $0.70 per million input tokens and $2.10 per million output.
Open-weight models are back in the race.
Thariq, who just left Anthropic's Claude Code team: if you're up all night and keep feeling new revelations, check on yourself.
Coding agents are a huge multiplier, but a busy screen isn't shipped work. Measure what ships, not how many agents run.
🚨 Qwen4 drops next week
• Alibaba Cloud showed the lineup at Apsara in Amsterdam
• Qwen4-Max, Plus, Flash and 27B
• CTO Feifei Li confirmed a next-week rollout
• Code in the Qwen repo already supports 1M-token context
Goldman Sachs: if AI lifts US productivity 15%, about 15 million workers (nearly 9% of the workforce) could be displaced.
The rougher part: displaced workers already take a month longer to land a new job.
DHH today: "The real AI psychosis is thinking... 'humans still need to review every line'. You gotta snap out of it."
Half right. Line-by-line review doesn't scale anymore. But drop the tests and the human who owns the outcome, and you're just shipping bugs faster.
🚨 Chinese labs are now distilling from OpenAI's GPT-6 Astra
• Moonshot used Astra outputs for Kimi K3.1 Pro
• DeepSeek's V4.1-Flash (Oct 17) did the same
• Why: Anthropic's defenses cut Claude distillation 90-99%
via The Information, Reuters
DeepMind's new paper "Organizing Intelligence" argues AI is already great at solo tasks, but economies run on teams, and weak coordination between agents is why productivity hasn't moved yet.
Do you buy that, or is the bottleneck somewhere else?
🚨 Pika released H3 Flex as open weights:
• 15-second clips with a soundtrack in 7 seconds
• Up to 12 multimodal references
• 5 exportable audio stems
• 4K output, up to 80% cheaper than H3 Max
Weights are on Hugging Face and GitHub. It lands in Pika's main app Nov 2.