JoshPurtell

6 ms·

1.
by JoshPurtell 5mo ago
There could be a leak from post-training but not pre-training
2.
by JoshPurtell 6mo ago
gpt oss 20b is not dense
3.
by JoshPurtell 6mo ago
Codex is not far behind Claude Code
4.
by JoshPurtell 6mo ago
You most certainly should not have stuck with Conda
5.
by JoshPurtell 7mo ago
Have been testing this in dev and really like the performance so far
6.
by JoshPurtell 7mo ago
I concede that it is not precisely OSS. But if I tell someone that it is source-available, they will expect some kind of license restriction for any use. If I tell someone OSS, they will expect mostly what the Sentry license entails, unless
7.
Specification Driven Development(joshuapurtell.com)
3 points by JoshPurtell 7mo ago|0 comments
8.
by JoshPurtell 7mo ago
What compile times do you work with? I use bazel and it still hurts
9.
by JoshPurtell 7mo ago
Billionaires and corporations can hire teams of people to work for them full-time. You, likely, can hire one or two (or zero!). Not to make it personal. These inequalities already exist
10.
by JoshPurtell 7mo ago
I do both and compile times are very unfriendly to AI!
11.
by JoshPurtell 7mo ago
Why, because of the Sentry license?
12.
by JoshPurtell 7mo ago
Every AI advancement liberates real humans from drudgery and allows them to create what they want more easily. The invention of the digital calculator turned human calculators into accountants, and that's great! We're contributing
13.
by JoshPurtell 7mo ago
rlm-rs: https://crates.io/crates/rlm-rs src: https://github.com/synth-laboratories/Horizons
14.
by JoshPurtell 7mo ago
Monty is the missing link that's made me ship my rust-based RLM implementation - and I'm certain it'll come in handy in plenty of other contexts. Just beware of panics!
15.
by JoshPurtell 7mo ago
Just like codex or opencode provide strong oss implementations of the core agent loop, our ambition (not achieved! hoping this is a solid start) is to provide a solid oss implementation of the context updating loop, memory, basic database +
16.
Show HN: Horizons – OSS agent execution engine(github.com)
39 points by JoshPurtell 7mo ago|8 comments
17.
by JoshPurtell 7mo ago
It's 2026, just use Planetscale Postgres
18.
by JoshPurtell 7mo ago
Completely absent of substance
19.
by JoshPurtell 8mo ago
From Sama "More Texans use ChatGPT for free than total people use Claude in the US, so we have a differently-shaped problem than they do" Facts don't care about your feelings
20.
by JoshPurtell 8mo ago
Important to note Anthropic has next to no consumer usage
21.
by JoshPurtell 8mo ago
This is super cool! Have you tried GEPA?
22.
by JoshPurtell 8mo ago
For devtools cos providing skills - we've found that using GEPA where you optimize the skill content instead of the prompt works really well to make sure the skill actually gets claude code/ codex /opencode to successfully us
23.
Engine-Bench: Evaluating Coding Agents on Writing Game Engine Code(github.com)
2 points by JoshPurtell 8mo ago|0 comments
24.
by JoshPurtell 8mo ago
As a demonstration, I've built Crafter in lean - and have hosted it on the web using Lithe https://lean-crafter-production.up.railway.app/ https://github.com/JoshuaPurtell/lean-crafter
25.
Lithe – A Web Framework for Lean4(github.com)
4 points by JoshPurtell 8mo ago|1 comments
26.
Engine-Bench: Benchmarking Coding Agents on TCG Game Engine Tasks(github.com)
2 points by JoshPurtell 8mo ago|0 comments
27.
by JoshPurtell 8mo ago
At some point in the future (if not already), claude will install malware less often on average. Just like waymos crash less frequently. Once you accept that installation will be automated, standardized formats make a lot of sense. Big q is
28.
by JoshPurtell 8mo ago
lmao
29.
Verify long-horizon tasks with GEPA on the judge(usesynth.ai)
4 points by JoshPurtell 9mo ago|0 comments
30.
by JoshPurtell 1y ago
I had a high opinion of PS before this comment. Now I have a higher opinion of PS

More ›