Software enjoyer
Joined April 2026
- Tweets 351
- Following 116
- Followers 15
- Likes 1,439
50 Photos and videos
Jun 8
Putting it together was very fun, reminded me of my prusa (in a good way)
Now setting up the raspberry for my agent runtime and the spark for the models… A lot more to come 👀
17
Jun 5
Man i wish that @OpenAI gpt-5.5 pro and @AnthropicAI opus-4.8 would not gaslight me into thinking that i am a genius with revolutionary ideas (which i don't think i am) every time i try to explore stuff 🥲
20
May 29
Basic integration of Reachy Mini by @huggingface and @pollenrobotics in my agent runtime.
For now just pose and pre-built actions are available, will look into dynamic behavior graph generation and execution later.
Next step is a e2e voice pipeline, using the work from @andimarafioti as a baseline.
1
1
76
Diego Carlino retweeted
May 24
Anthropic onboarding day: Michael Scott introducing Karpathy like he just signed Wemby in free agency.
396
1,490
17,625
2,350,063
May 22
.@mattpocockuk grill-wth-docs is amazing but i've been grilled sooo hard in my last session 🥲
40
May 20
WIP: ds4 by @antirez running over Unix socket in my agent runtime
While i can't integrate it directly like ds4_agent does, i can still bypass most of the overhead added by OpenAI/Anthropic adapters
NOTE: playback speed is doubled the original video was 03:07m
1
99
May 20
Some initial results of this approach in my harness 👀
So I told you about this idea of tagging lines for agent EDIT. But what about instead the *harness* remembering the old view of the file? And just asking the LLM for lines/ranges, and failing the edit if the range no longer matches the old content. Likely best solution.
1
241
May 20
Benchmarking the line tag for edits idea from @antirez in my agent runtime.
Results show, at least for my implementation, that it works best for big changes and probably not worth it for smaller ones.
Need to explore how it scales on long running agentic sessions, these tests are currently one off simple tasks.
His blog post about this idea: antirez.com/news/166
1
2
15
5,540
May 19
BenchLocal results for DeepSeek v4 flash q2-imatrix served by ds4
DGX spark specific CUDA kernel, 140k ctx size
BugFind-15: 86
CLI-40: 53
DataExtract-15: 88
HermesAgent-20: 84
InstructFollow-15: 93
ReasonMath-15: 63
StructOutput-15: 80
ToolCall-15: 93
1
2
265
May 19
Gave each failed test another run
CLI-40: 53 -> 62
DataExtraction-15: 88 -> 90
HermesAgent-20: 84 -> 88
ReasonMath-15: 63 -> 78
StructOutput-15: 80 -> 90
NOTE: some tests are timing out, might re-try them with mtp later
67
May 19
Doing some local inference tests on my custom agent runtime using Deepseek v4 flash served by ds4 from @antirez on my DGX spark.
Would love to explore native integration to bypass the network layer and tool translations introduced by the openai/anthropic adapters
2
1
21
3,966
Apr 30
mind: blown 🤯
@karpathy very insightful takes, i need to sketch some POCs out ASAP
youtube.com/watch?v=96jN2OCO…
67
Apr 24
01:34:45- I 100% agree on this
> Once you break out of language, the language models starts to break down aggressively
I am actively researching a working on a solution for this, representing application state and actions as "language"
github.com/devteapot/slop
Apr 23
GPT-5.5 is here. Ben and I have had access for a bit. We have a lot of thoughts.
00:00:00 - Intro
00:01:34 - Vercel hack
00:05:38 - Kimi K2.6
00:14:29 - Cursor acquired?
00:37:24 - GPT Image 2
00:49:30 - GPT-5.5
01:22:02 - GPT-5.5 Pro
35
