ai research + people || prev. academic, future ??
Joined July 2025
- Tweets16
- Following43
- Followers0
- Likes9
woolgatherer retweeted
I get emotional every time I go to the library like the fact that it’s free and they have so many services
got interested in code + design bc of winamp skins and nothing will hit like that again, i fear
I think it's sad none of these applications will ever be built again.
I no longer want cleverly packaged utility. I want silliness and whimsy delivered through convoluted high skill ceiling bloatware if that's not too much to ask for
12
new hobby: themed book clubs
requirements:
* low-commitment character costumes (e.g. "I will be X. Looking tired and a little suspicious of murder.")
* food relevant to the book
* yapping
making room for play/fun with good people is the spice of life!
10
cannot underscore enough how deeply unserious it is to develop production-grade systems without evaluations. if you want to get somewhere, you need to define where you're going
imo, evaluations are a key pillar of alignment 🤷
10
the pace of ai is plateauing (see: GPT-5) because downstream usage patterns are in flux and therefore the correct objective f(n) is ??
but it's def not these random benchmarks. ppl care about how models perform on actual real tasks. that's what makes something useful🤷
5
woolgatherer retweeted
chatgpt agrees with you, your therapist agrees with you
no one is left to tell you the truth besides toddlers and drunk people
Big +1. Relatedly, I love @AlisonGopnik's take on LLMs as a cultural technology: youtube.com/watch?v=k7rPtFLH…
More than just STEM folks ought to be involved in conversations around what innovation with LLMs looks like, what new norms should exist, etc.! Therein lies the fun.
Don't leave AI to the STEM folks.
They are often far worse at getting AI to do stuff than those with a liberal arts or social science bent. LLMs are built from the vast corpus human expression, and knowing the history & obscure corners of human works lets you do far more with AI
1
7
I see this all the time. There's a tension between wanting to move fast and figuring out which direction to move in.
There might be resistance to manual trace review because it feels inefficient, but it's usually where you discover what actually needs fixing!
I once gave a talk on evals to a big tech company. Engineering audience. > 50% of their questions could have been answered by "well, did you actually look at a few LLM traces and flag the turn at which the LLM failed?" Most of the time they looked at me like I was suggesting something that wasn't in their job description. One asked if there's a tool that can automate error analysis.
5