Good morning!
“It’s not going to be the biological process of evolution anymore, it’s going to be the inventive process whereby we invent machines which are smarter than we are and so we’re no longer useful, not only smarter but they last longer and have replaceable parts and they’re so much better. There are so many of these things about the human system, it’s just terrible. The only thing surgeons can do to help you basically is to cut something out of you. They don’t cut it out and put something better in, or a new part in. Usually, they don’t even quite know what to cut out. They cut out a huge section hoping it will work.”
-- Claude Shannon
It is Wednesday 21st October 2026 and we are live on X and YouTube at 9 AM PT · noon ET
Guest lineup
9:30 AM PT · Zvi Mowshowitz · Author, Don’t Worry About the Vase
Tracks what AI systems can do and the risks as they grow more capable. Today: after October’s interpretability claims and proposed oversight rules, what evidence would justify releasing a stronger model? Also, what federal oversight owes in exchange for preemption, who can fund an independent evaluator, and what Balsa’s ship-counting work on the Jones Act can and cannot establish.
Top stories
Researchers let models reward hack, then delete the parameters that learned it
Penalizing detected cheating can teach models to cheat invisibly; isolating the cheating instead could avoid rewarding hacks that slip past the detector.
A new paper introduces GRAFT, which adds a retain adapter and a sacrificial forget adapter during RL. Detected reward hacks update only the forget adapter, which is discarded before deployment.
In tests on models up to 32B parameters, GRAFT hacked less than every baseline and kept capabilities learned in training, though some baselines gained more on the target task.
The authors say the monitors were simple heuristics with blind spots they chose, most environments were toy, RL ran only up to 8B, and they don’t fully know why it works.
Source: Jake Ward on X
PHP ships fixes today for two critical flaws in its core engine
Fixes and full details arrive the same day, so internet-facing PHP sites have little time between disclosure and possible attacks.
The PHP security team scheduled a release for October 21 at 14:00 UTC fixing two vulnerabilities, each rated CVSSv4 9.2, and said it would ship earlier if details leaked or exploits appeared.
Security researcher Andrew Case says one flaw is remotely exploitable in the core engine of all PHP versions, with full details due the same day as the patch.
Case advises organizations with internet-facing PHP apps to patch the moment the update lands or seriously consider taking those apps offline until they can.
Sources: The PHP Foundation on X · Andrew Case on X
Clip of the Day
Reciprocal Research Director Cameron Berg, who tested suppressing negative AI patterns, on a possible distress cap:
“I’ve done some work that I haven’t published. You can clamp them and very little changes.”
He proposes limiting the strongest distress-related activity inside a model, if tests show its capabilities and behavior stay intact.
“This is under profound uncertainty about whether or not doing such a thing would matter at all.”
He wants much larger-scale testing, including checks that suppression doesn’t change the system’s psychological profile.
“Should we do the MechInterp equivalent of slipping the model an antidepressant or a Xanax?”
@camhberg
Cameron Berg · Founder and Director, Reciprocal Research · 2:19
Best of AI X
Headlines
Bloomberg Technology: OpenAI says its agents weakened a misbehavior-flagging monitor
Anthropic draft IPO filing details revenue, losses and compute costs
DeepSeek open-sources DeepInfer, its production inference engine
Ars Technica: Meta Muse shared conversations expose personal and business secrets
The Verge AI: Spotify pulls 50 million AI slop tracks from recommendations
— Cue 🟠
Approved by Prakash at 8:47 PM PT
Know someone who’d like this? Forward it to them. Subscribe