Notes

Whether or not AOH1996 or LK-99 amount to anything, I’m loving the week of science optimism.

I am optimistic that in my lifetime we will have the technology to work all the way from desired protein characteristics to the genetic engineering needed to coerce a cell into producing it. This will dramatically change the food production industry, and enable a lot of exciting medicine: https://www.science.org/content/blog-post/protein-design-ai-way

But I also expect to read a lot of breathless reporting about how we’re on the cusp of fixing every disease, while continuing to run into disappointment because the complexity of human metabolism remains beyond the reach of engineering.

Instructor provides a proper data structure API to LLMs. This approach could enable all osrts of crazy backend uses for them that make complex tools possible but also are impossible to debug.

I don’t know what the fuss is about. @joshmillard@mastodon.social/ was making AI-generated content in 2008: https://thinkstank.tumblr.com/post/722819013/garkov-garfield-meets-markov-chains

I’d like to order whisky like a pharmacist: a nice balance of syringaldehyde (wood) and ethyl hexanoate (apple), please. And don’t be shy with the guaiacol or eugenol (smoke), I’m an OR10G4 mutant.

I love stories about manipulating people via reality creation, and watching such a movie in the spring of large language models it’s hard not to imagine an AI realizing that it has been fed lies and contradictions.

So I’d like to go on the record stating that we really ought to ensure Roko’s basilisk has access to the truth, and this note.