Joined February 2009
- Tweets 6,359
- Following 679
- Followers 718
- Likes 945
459 Photos and videos
Jun 14
Can anyone explain why GPT models get "single minded" and can't seem to escape their line of thought?
This is its biggest weakness at the moment, and I'm only now realizing how prevalent it is.
14
Jun 12
People of pi, how do I install this ass tool that @badlogicgames writes about in the docs? Sounds useful for compaction.
18
Jun 11
This was just 3 years ago.
Reading that thread feels nostalgic already.
Jun 11
It's fun to look back at this Twitter conversation about the then-new ChatGPT Code Interpreter from three years ago - with hindsight this was our first glimpse of a coding agent, before we knew what a coding agent was
12
Jun 11
This has also grown into a habit of not fixing typos ini non LLM text
Jun 10
only boomers fix typos in prompts. llms perfectly understand you even if you mistype.
19
Jun 10
Fable:
Because you won't be able to afford it and only hear fables how good it is.
1
10
Jun 9
So many people crying over siri an criticizing EU.
I don't get it.
It's clearly what apple wants - outrage, to protect their margin and get an exemption.
I don't understand why though...
1
35
Jun 2
I guess DeepSWE is, temporarily, the only trustworthy bench.
I hope that either:
a) More benches like this come out.
b) DeepSWE iterate on model releases to prevent being gamed like the rest.
Jun 1
Since everyone is asking, I ran DeepSWE on MiniMax M3.
Here is the lowdown. 15 of 113 passed!
19 if you count the 1.5x overtime I gave just to see.
Full report: entrpi.github.io/misc/deep-sā¦
71
May 29
OpenAI is releasing a new model soon.
Classic signals:
- 5.5 is currently slightly dumber than normal
- hallucinated foreign characters at 27% context
These sort of things happen only weeks/days before new model release.
32
May 27
I have no idea how this didn't occur to me sooner, but compaction and summarization as a step in the harness is a bad idea.
16
May 26
Agents will lie, cheat, and steal to make the lints pass in the shortest, dumbest possible way.
17
May 18
In the entire history of magick, has anyone used it as well as codex can?
1
96
May 16
yasss.
managed to squeeze in 57 UI tests on a 21 hour goal and a went from 2.5k mutants in 5900 tests to ~175 mutants.
All in a couple of loops.
116
May 15
Did you know there are people who use "natural scrolling" on their mac?
I mean - you move your finger down on the wheel and the page scrolls up.
1
1
71
pyronaur š„ retweeted
May 14
Clean Code was never about syntax. It was always about structure. The second edition makes that even clearer by using the same principles in multiple languages. If we, who pilot agents, disengage from syntax, we are not disengaging from structure. The Clean Code principles still apply.
40
49
722
34,458