gelisam
@haskell_cat
AI Safety ∩ Programming Language Theory. Part-time technical alignment researcher, full-time Haskell software engineer at http://well.co, opinions are my own.
You might like
Mathematically-define "safe behavior", not "humanity's utility function". After training, but before giving it enough power to be useful, move the trained program in the latent space to the closest safe program. youtu.be/-2nFTfXAsmU
youtube.com
YouTube
2022 09 26, Samuel Gélineau, Can we Prove Facts About Machine...
Somebody on my timeline recommends a book and it is _not_ If Anyone Builds It Everyone Dies??
This is a really good book. I like it because it covers both ends of the spectrum: 1. How LLMs work 2. How to build using LLMs It's a really nice one-two punch: start with the theory and use that right away to implement something useful. The second half of the book is what I…
> my prediction is that auto-regressive LLMs are doomed Yan LeCun, notorious AI Doomer 🤣
1. "Nobody in their right mind will use autoregressive LLMs a few years from now." The technology powering ChatGPT and GPT-4? Dead within years. The problem isn't fixable with more data or compute. It's architectural. Here's where it gets interesting...
One concept I wish more people were aware of is the Tocqueville Effect. Named for Alexis de Tocqueville, this concept describes the curious phenomenon by which people become more frustrated as problems are resolved: As life gets better, people think it's getting worse!🧵
Technology is generally really good. Why should AI be any different? A new video: (youtube link in the reply)
When I was a teen, my government held a public consultation about switching from 1st-past-the-post to a proportional system. I did my research, went, presented approval voting, and was laughed off the stage. I never did politics again. Today, we're still using 1st-past-the-post.
If you liked this, follow @ElectionScience for more information on a better way! It's called "approval voting", and it's so simple: you can vote for multiple candidates, and the candidate with the most votes wins. While not perfect, I think it's better than ranked-choice
I like to learn about neural networks by working on tiny problems for which there exists a 100% correct solution. Here is an interactive experiment showing how in practice, backprop doesn't find this solution: gelisam.com/local-minima/
My first MCP server, which allows the agent to pick from a selection of shell commands. GitHub Copilot can natively run shell commands, but VS Code asks you to confirm each command. With mcp-cli, you only have to authorize the use of the tool once! github.com/gelisam/mcp-cli
Some people say they liked this one better than previous podcasts. youtube.com/watch?v=0QmDcQ…
youtube.com
YouTube
Eliezer Yudkowsky: Artificial Intelligence and the End of Humanity
New video, about how to work in technical AI Safety research! (link in reply)
This is the Frog Fractions of PuzzleScript! 🐸⅜ 🤯
I created a programming language prototype that harnesses bidirectional type inference to infer JSON schemas for LLM prompt chains that use structured outputs haskellforall.com/2025/05/prompt…
Let's all write a short story with the same prompt! Here is my attempt. gelisam.blogspot.com/2025/03/metafi…
I said the AI writing was shit; somebody challenged me to do better based on the same prompt; and so you know what, fine. CW: grief, suicide. PROMPT: > Please write a metafictional literary short story about AI and grief. It is like a dream, always, everything half-real. I…
I recommend today's haskle.net puzzle. It looks impossible, but in retrospect, it's blindingly obvious!
United States Trends
- 1. D’Angelo 199K posts
- 2. D’Angelo 199K posts
- 3. Brown Sugar 16.5K posts
- 4. Black Messiah 7,755 posts
- 5. Voodoo 16.1K posts
- 6. #PortfolioDay 10.4K posts
- 7. Young Republicans 4,712 posts
- 8. Happy Birthday Charlie 123K posts
- 9. Powell 38.3K posts
- 10. How Does It Feel 7,392 posts
- 11. Pentagon 99.4K posts
- 12. Osimhen 132K posts
- 13. Alex Jones 28.6K posts
- 14. #BornOfStarlightHeeseung 82.5K posts
- 15. Neo-Soul 18.3K posts
- 16. CJGJ N/A
- 17. VPNs 1,091 posts
- 18. Sandy Hook 11.4K posts
- 19. Untitled 6,491 posts
- 20. Baldwin 17.2K posts
You might like
-
João Forte Carvalho @Conste11ation
@bgamari -
Tweag
@tweagio -
Well-Typed
@welltyped -
Neil Mitchell (neildmitchell @🪡)
@ndm_haskell -
Andres Löh
@kosmikus -
Matt Parsons
@mattoflambda -
Joachim Breitner
@nomeata -
Nikita Volkov
@NikitaYVolkov -
Michael Snoyman
@snoyberg -
Vladislav Zavialov
@int_index -
Tikhon Jelvis
@tikhonjelvis -
Veronika Romashkina
@vrom911 -
Dad×2_jack
@Iceland_jack -
Oskar Wickström
@owickstrom
Something went wrong.
Something went wrong.