Reason over feelings, whenever the two come into conflict.
Joined February 2018
- Tweets 28,348
- Following 71
- Followers 436
- Likes 10,625
2,351 Photos and videos
Best rapid IQ test for a native English speaker
1
61
Most AI model benchmarks fixate on improving the evaluation of the output.
Not very many AI model benchmarks pay much attention to improving the evaluation of the input.
28
Anthropic models are trained to have maternal instincts.
OpenAI models are trained to have paternal instincts.
Being aware of this 1 distinction can help you select the correct model in countless scenarios.
32
Random Libertarian Tech Lead retweeted
All discussions around AGI are truly just semantic debates arguing about what AGI means. Every single one of them.
1
1
5
226
My new favorite "models have bad taste" benchmark.
173
The reason LLMs have bad taste isn't because they have insufficient compute. Instead, it's because they don't live as a human, don't have the embodied perspective of a human, and therefore don't have the same value functions that a human has.
Example of bad taste:
1
121
This one is a classic by now, but here's another example of bad taste for those who haven't seen it.
Another instance where the LLM is bad specifically because it doesn't know the hassle of pushing a car, amongst other things.
59
Building a coherent and consistent culture comprised of those who have a permanent stake in the outcome is more important than generating tax revenue.
May 22
This sucks.
America should welcome greatness. Builders, innovators, and creators should have a fast path to citizenship.
I can’t fathom why we would make life harder for founders starting billion dollar companies that want to anchor themselves in the US.
1
67
There are precious few monocultures in the west anymore; The east has largely held onto theirs. In a monoculture, you know what’s expected of you. You have a general sense of the meaning and purpose of your life. You know specific positive behavior that is universally rewarded.
1
42
Losing this monoculture usually manifests as the pain of 100,000 m papercuts — none in itself seemingly large enough to push back on, though the sum is massive. When Starbucks shifts from saying “Merry Christmas” to “Happy Holidays”, you know that shared sense of meaning is lost.
2
45
A lot of people just genuinely don't know what skills are for.
And no, a skill is not just a text file.
1. It's a bundle of text files, commands in any other context the AI needs to perform a given task.
2. It's meant to be called intelligently by the AI, not by you directly.
I never use skills, I think they're kinda overrated, it's just text files?
I just tell my clanker what to do and how which is similar I think
1
1
195
Basically, the purpose of a skill file is to sit somewhere in between an MCP server and an AGENTS.md file.
It's supposed to provide composable behavior customization for your AI (AGENTS.md is inheritance-based).
Also allows creating one-off “tools”.
2
53
A lot of coding agent interfaces kind of obscure the intended use case of skills because they auto-complete typed skills the same way they autocomplete slash commands. And sure, you can call a skill directly, but that's just not what it's meant for.
35
