David Heimann (@davidmeh) header
David Heimann (@davidmeh) profile picture

friend of computers, always optimistic

Brooklyn, NY
Joined March 2011
Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)Photo by David Heimann (@davidmeh)
After experimenting with orbs all weekend I might just be sold on @AmpCode. Some things that surprised me: (Caveat: I’ve been using pi almost exclusively since ~January so ymmv)
1
3
260
3. I could do _everything_ from my iPad. In the past, I’ve used Moshi to ssh into my dev machine to do work, but it’s always been clunky. Amp on my iPad never felt clunky.
1
4
87
In summary, things just works, and in the end I did more things because the amount of friction is reduced to almost zero. Very bullish.
1
1
3
2,444
There’s something unpalatable about too much test time compute. Imagine Athena standing there “but wait!”-ing herself
57
The new deepseek v4 flash model is great whenever it finally finishes thinking. Kind of sucks you have to compact so often though
74
It would be kind of lame to spend the first act of our first meeting with a new intelligence begging it not to kill us
35
I would be so happy to spend the rest of my life fighting for the Engelbartian ideals of man computer symbiosis.
30
Centroids are fascinating. Like how the average person can’t be average across more than a few traits (see: Daniels 1950)
1
1
30
So, in conclusion, nothing is in distribution! 😂
1
7
The difference between great work and slop isn’t how you made it (a prompt or fingers on keyboard) but whether or not you own the details
23
AI-2040 is a recommended read but i find it revealing that their case for continuing AI progress is basically "so we can do alignment and defend against bad actors." not "cure all disease." or "end poverty." The optimistic case for AI is somehow the one nobody wants to make?
20
You don't get trusting children by raising them in paranoia. I think about this a lot with language models.
21
Is there any counterfactual that would lead yudkowsky to admit he's wrong? I can't think of one, and that's kind of the whole problem.
2
61
Every era has a version of "you don't need to do that by hand anymore." Compilers said it about assembly. Copilot said it about boilerplate. Claude code is saying it about implementation. The details will always matter, but the difference is which details…
38
Weekly token resets and usage limits are one-armed bandit mechanics. Anthropic and openai are running the most effective slot machine in history and somehow still claiming the moral high ground on responsible AI 🤔
49