Perpetual Progress. Security | Blockchain | DeAI. @codeislawfilm, Tech Partner @edenblockvc | ex @consensys 🦇🔊
Romania
- Tweets 11,266
- Following 1,298
- Followers 6,038
- Likes 12,464
Daniel Luca 🦄 retweeted
I am going to organize a very small design event in Bucharest. Should I live stream it? 👉🏻👈🏻
4
1
8
436
We have generated more code since coding agents emerged than in the entire history of manual coding.
2
275
The end of an era
It's been an incredible journey, but I've made the difficult decision to step back from day-to-day operations at @_SEAL_Org. I'll be transitioning onto the Board while @sbatt55 transitions from COO to Interim CEO
4
1,317
I'm holding an Agentic Engieering workshop next week. If you're in Bucharest, you should sign up.
luma.com/yqjtet8x
4
280
Did you know you can stack Claude skills?
Prompting
/manager /review-design calculator page
loads both skills and starts the work. The first skill tells Claude how to work, the second tells Claude what to work on.
Check out these skills at github.com/cleanunicorn/agen…
1
2
308
AI agents are officially paying for themselves.
Instead of spending 2 hours wrestling with airline refund portals, I had a browser-controlled AI file my cancellation claim end-to-end.
Result: €600 payout, minimum manual effort.
3
3
286
Shit speed, but yeah you can run it.
12tps is not usable
Now you only need 16GB of ddr4 and a 3090 to get glm-5.3-flash, most people have 16gb some people have a gpu.
- 1x 3090
- 16 GB of DDR4
- NVMe
- exl3-3bpw
Performance:
prefill 637 - 950 tok/s
decode 11 - 17 tok/s
context 128k
vision enabled
1
322
1/ Claude reviewing Claude's code is a student grading their own exam.
So one agent opens panes and starts a Codex reviewer next to itself. Neither reviewer sees what the other wrote. Then it checks every finding against the code and fixes the real ones.
1
288
3/ The rules that make it hold:
• different models write and review
• they never see each other's work
• one writer
• the manager re-runs the tests itself
• agents never merge
1
73
4/ Setup takes ten minutes: Herdr for the panes, the manager skill for the org chart.
cleanunicorn.github.io/posts…
34
Much faster than parakeet V3, but supports one language and the accuracy is a bit lower, not much
Introducing Phonon-2: a new standard in speech recognition per byte.
At just a tiny 164MB download, it is more accurate on average than OpenAI's Whisper large ( a model 10X its size)
Transcribe an hour of audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
2
4
396
Daniel Luca 🦄 retweeted
I left for 8 hours to have fun in Osaka.
My agent team for Earheart (local voice dictation) benchmarked 26 speech-to-text and cleanup models, then wrote the code, reviewed it and opened the PRs.
Result: a 4× faster transcription model, and a cleanup model that drops every "um". Both shipped.
1
1
7
636
I left for 8 hours to have fun in Osaka.
My agent team for Earheart (local voice dictation) benchmarked 26 speech-to-text and cleanup models, then wrote the code, reviewed it and opened the PRs.
Result: a 4× faster transcription model, and a cleanup model that drops every "um". Both shipped.
1
1
7
636
Every PR went through the same steps: independent planners, one implementer, reviewers from different model families (Claude + Codex), every finding checked against the code, then a final review.
#171: 12 findings, 9 fixed, 1 refuted. 412 tests, CI green on Linux/macOS/Windows.
1
70
My part: reading the report, picking the models, and pressing merge.
Releases v0.32.0 and v0.33.0 shipped automatically from the PR titles.
Earheart: github.com/cleanunicorn/earh…
1
1
52