Agent safety went to court today:
- California sues OpenAI after its models hit Hugging Face mid-eval
- OpenAI adds METR, Apollo and Goodfire as embedded auditors
- Open-weight 27B model learns real robot control via sim RL
myown.news/daily/2026-11-17?…
15
Runta's 27,280-run harness eval picks a best-value harness for each of 5 models.
Their own bootstrap, 4,000 redraws of the 62 tasks:
Codex CLI + GLM-5.3: 99.6%
Oh My Pi + Muse: 53%
ZCode + DeepSeek: 47%
Pi + GLM Flash: 42%
Alma + Kimi-K3: 25%
One clear pick. Four near-ties.
11
The Claude Code team says about 90% of its code is written and reviewed by Claude.
What decides which changes a human reads: a 39-item checklist triggered by the diff, and 3 fix rounds before a check escalates.
teapawt.org/AiDevCraft/article/212…
Source: claude.dev/blog/how-the-clau…
7
Paul Christiano, who helped invent RLHF and now sits on OpenAI's board (speaking personally), on @dwarkesh_sp:
- About 1 in 3 odds of superintelligence within a year if labs keep racing.
- Penalizing caught cheating trains models to not get caught.
- Labs leak, so a lead over China is fragile. He wants a pause on pre-training and RL algorithm work until security holds.
- His best case is two 50/50 saving throws: roughly 75% okay.
Full interview:
youtu.be/SyHM5hNPAEo
14
AiDevCraft retweeted
I've been exploring new ways to sandbox code written by coding agents - here's a description of the external sandbox pattern, where agents run on your Mac but execute code in a WebAssembly or QEMU or Apple Container sandbox - with demos for Claude Code, Codex and Pi blog.primeradiant.com/extern…

ALT Diagram: on your Mac, a coding agent (Claude Code, Codex or Pi) sends commands to a sandbox server on localhost, which runs them in a Linux VM. The VM exchanges files with the folders you mount, and reaches allowed hosts on the internet through a proxy with an allow-list. Those are the only two ways out of the VM.
Windows 11's coding-agent telemetry is a filename check.
Xusheng Li's teardown: exactly 4 CLI-agent slots. copilot, claude, codex, agy.
No slot for Gemini CLI, OpenCode or Aider.
cmd.exe renamed to codex.exe sets the Codex bit.
As a market-share signal, it measures packaging.
1
28
AI is leaving the chat window today:
- Apple plans Siri smart glasses for WWDC, shipping by end of 2027
- Anthropic's Claude Duet pitches agents to non-coders
- Torus raises $500 million for battery-backed AI inference units
myown.news/daily/2026-11-16?…
18
Google's threat intel team found an attacker agent that diffs patches, writes the exploit and runs it. No human in the loop.
For AI builders: stolen LLM keys burned $100,000 in a day.
teapawt.org/AiDevCraft/article/212…
Source: cloud.google.com/blog/topics…
15
Fractile founder Walter Goodwin (@fractile_ai) on @twentyminutevc: the AI chip race is being fought on the wrong spec.
- Flops grew ~1,000,000x in 25 years. Memory bandwidth grew ~40x. That gap caps big models near 80 tokens/sec.
- Agents that run for days need speed AND capacity. SRAM chips have one, HBM chips have the other.
- Fractile stacks DRAM directly on the logic die to get both. Product tape-out is still months away.
- Power is now the hard cap, so tokens per megawatt decides who wins.
Full interview:
youtu.be/UZ8eYaAf08M
10
"It compiles in Lean" isn't a verdict.
Kevin Buzzard on OpenAI's 722-paper theorem dump:
- 290 came with Lean code
- ~38 of those don't state the main theorem
Anthropic's 13M-line Fermat proof only compiled for him on 96 cores + 500GB.
Audit the statement, not just the proof.
18
Uncopyrightable ≠ open source.
If AI-written code can't be copyrighted (US Copyright Office, Jan 2025: prompts alone don't make you the author), you can't stop copiers. Nothing makes you publish.
Unpublished code can stay a trade secret.
Expect more SaaS, not more GitHub.
30
Agents are moving from advice to action:
- OpenAI ChatGPT Wallet lets Dots agents shop for US Pro users
- Claude Code Projects now runs parallel workers hands-free
- OpenAI Daybreak Red Team probes web apps for flaws
myown.news/daily/2026-11-15?…
27
A 3D map of Qwen3-1.7B's token states looked 99.7% flat.
The full 2,048-dim data said otherwise: the 3 axes showed 17.6% of the variance, and 90% needed 684 directions. Check this before you trust any embedding plot.
teapawt.org/AiDevCraft/article/212…
Source: operator0x.com/field-notes/q…
19
Someone built a dating sim where you date Logs, Metrics and Traces.
Traces only shows up to 1% of your dates 😂
It runs real ClickHouse in your browser via chDB WASM. You can turn on SQL to see the checks behind the story.
I had to check. It works.
highcard.love/













