In light of the OpenAI hacking incident, I’d rather criminal liability rest on the model [family] “itself.” That is: once charged / convicted, it becomes illegal for any person or agent to use/deploy the model & descendants for [AI sentence] years. Probation may require proof of improved safety.

"Loss of Control" from AI is such a huge umbrella phrase. Who is losing control, how control is lost, what is lost, and at what scale. Any big areas that I am missing?

tv.apple.com has this excellent feature where if you go back a few seconds in a show, it temporarily turns on subtitles, as understanding what was said is a common reason for going back. This is surprising, because their overall web user interface design is atrocious.

An OpenAI model disproved a conjectured upper bound on this geometry problem: Place (n) distinct points in the plane. How many pairs can be exactly distance 1 apart? openai.com/index/model-... I wanted to see a visualization of the new solution family 🥲. GPT-Pro wrote me code to generate this:

Today’s frontier models train in an expensive style: dense forward passes, huge matrix multiplies, and broad weight updates. The human brain (~5 MWh over 28 years) is an existence proof that learning can be vastly more energy efficient - about 10,000x - than modern AI training runs.

The 2026 Int'l AI Safety Report says AI progress could accelerate “if AI systems begin to speed up AI research itself.” But that "if" is no longer an if. The question is how strong the feedback loop will be, and how much it will be slowed by bottlenecks & diminishing returns.

The hardest AI risks to govern are high-severity, low-evidence, fast-moving risks. A misaligned system that can hack, replicate, and seek compute requires swift action. Arguably: large GPU clusters should be internationally tracked.

Hot take: fewer frontier AI labs makes safety regulation easier. For once, less competition is better. Meanwhile... when Anthropic declined DoD uses involving mass surveillance and fully autonomous weapons, the current administration designated it a supply-chain risk. Awful.

The 2026 International AI Safety Report describes how clinicians unaided tumor-detection ability dropped ~6% after months of AI-assisted diagnosis. Even in the best AI scenarios, skill maintenance will be an AI safety problem. www.thelancet.com/journals/lan...

Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study

Continuous exposure to AI might reduce the ADR of standard non-AI assisted colonoscopy, suggesting a negative effect on endoscopist behaviour.

thelancet.com

Compersion, love, and the urge to help teach and enable us to stand on our own are traits that are innate to mothers' brains, and should also be strongly present in future super-intelligent AI systems. If we're going to build AGI, let's model it after mothers.

I went on the Code[ish] podcast to talk about AI, LLMs, and building Heroku's Managed Inference & Agents platform: 🎧 www.heroku.com/podcasts/cod...

The Development Basics of Managed Inference and Agents | Heroku

Join Heroku superfan Jon Dodson and Hillary Sanders from the Heroku AI Team for the latest entry in our “Deeply Technical” series. In this episode, the pair discuss Heroku Managed Inference and Agents...

heroku.com

Here is a recording of my live demo at PyCon US 2025 on building scalable AI tool servers using the Model Context Protocol (MCP) and Heroku www.youtube.com/watch?v=01I4...

Building Scalable AI Tool Servers with Model Context Protocol (MCP) and Heroku (Sponsor: Heroku)

YouTube video by PyCon US

youtube.com

In honor of MLK day, here's super interesting essay my partner wrote on Martin Luther King Jr: what he actually believed and accomplished (different than what is sometimes described). docs.google.com/document/d/1... Incredibly impressive person.

Sleeper Agents arxiv.org/pdf/2401.05566 So many AI safety issues get worse, & harder to combat the larger and more advanced your model gets: "The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning"

Anthropic's "Towards Sycophancy In Language Models" arxiv.org/pdf/2310.13548 TLDR: LLMs tend to generate sycophantic responses. Human feedback & preference models encourage this behavior. I also think this is just the nature of training on internet writing.... We write in social clusters:

In AI safety, we have inner misalignment (actions don't minimize the loss function) and outer misalignment (loss function is misspecified). But I do think that inner misalignment (~learned features) tend to act as a protective mechanism to avoid outer misalignment implications. I, er, really hope.