Sure, ya can't rely on it, but sometimes inner misalignment (following standard methods of paperclip production even you could make more by taking over humanity) will help deal with with outer misalignment (oops, we trained it to maximize paperclips, which is obvs terribad)

Formatting email text well is an undervalued skillset, ya'll.

When I've read a fair amount of research / technical blog posts from someone, and then I watch a video of them speaking about research, and they swear like a sailer. A very well-informed technically accurate neural network specialist sailor. I like it. 😁

Tried to google "Why do I look so young on Zoom"... That didn't work 😂; I am not searching in the same direction as most people. But I guess that means there's not an automatic face filter??

I thought GPT usually picked 7 as a random number! Well Dall•E 3 really likes 1, I guess.

OK, completely unrelated to AI, but why do most baby socks come in variety packs? Like, surely most parents care more about being able to find matching socks easily VS. making sure their baby's socks match their daily outfit? I am bemused. bit.ly/3RlK5NA

Just found out about Milton Friedman’s Thermostat analogy. Very good explanation of an important concept. I think scott cunningham has a similar analogy with a sailboat staying on course in the face of wind in the Mixtape.

Open source software is GREAT. But "open source" AI is NOT like software - it's VERY different. Rob Miles cuts through the bullshit: ROB: Oh, hey, Meta. I heard Llama's weights leaked. That's rough, man. Information security's hard. How you holding up? META: Oh, we're

"LLMs can't do logical reasoning at all, they've just memorised some basic rules of logic, and use pattern matching to imperfectly apply those rules to new situations"

It's been months and Google Firebase's definition of a p value is still wrong. 😳😔 cc Firebase

AGI is hard to define. my preferred definition of AGI is a computer system that can can accomplish a task impossible for 100 human geniuses working together, such as publishing a blog post with a single canonical spelling of GPT-4o / gpt-4o / gpt4o

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

I assume there's a good chance this will be algo-suppressed, but: bsky.app/profile/meatle… There I am :)

Trying to imagine aspirin company CEOs signing an open letter saying "we're worried that aspirin might cause an infection that kills everyone on earth - not sure of the solution" and journalists being like "they're just trying to sell more aspirin."

A 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/1Ol… Incredibly impressive guy.

Recently gave a live demo at PyCon US 2025 on building scalable AI tool servers using Model Context Protocol (MCP) & Heroku Recording is here: youtube.com/watch?v=01I4yw…

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

Lots of emojis in PRs screams that an LLM wrote it all 😂. Not that that's always bad... But sometimes yeah, it's bad.