@danmanai
iAccount based inRomania
About this account
- Account based in
- Romania
- Connected via
- Romania Android App
Account-level information from X, not a live location or the device used for a specific post.
Full stack software developer. When I'm not working I enjoy traveling around the world. Sometimes I do both :)
Cluj, Romania
Joined June 2009
- Tweets613
- Following878
- Followers164
- Likes77.4K
My Jev maze solving experiments are not going well
Poor guy is more lost than the guy from Memento without his tattoos and polaroids 🙈
Can Jev play poker? Yes
Can Jev beat OpenAI? Maybe
Here is a tournament between Jev, Terra and Astra
Here's another longer tournament with all models:
Jev, Luna, Terra, Sol, Astra
Overall I think Terra and Astra have the best play styles
Is anyone else seeing a lot of Chinese characters in Opus 5 output lately?
"The Chinese characters 描述 (miáoshù) mean “description” or “to describe.”"
Here is Jev from @typesafeai classifying 300 full chat sessions from our custom @vercel eve agent by failure reason.
323 sessions, 2M tokens done in 14s and costing $0.08
nitter.cf/CompleteSkeptic/status…
Vercel Eve - Jev classification
Classifying vercel eve agent sessions by failure mode using Jev
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Bonus use-case for Jev - redacting PII information (inefficiently 🙈)
Here is Jev from @typesafeai classifying 300 full chat sessions from our custom @vercel eve agent by failure reason.
323 sessions, 2M tokens done in 14s and costing $0.08
nitter.cf/CompleteSkeptic/status…
Vercel Eve - Jev classification
Classifying vercel eve agent sessions by failure mode using Jev
If this style of Bash edits is driving you crazy in claude code, add this in .claude/settings.json
{ "env": { "CLAUDE_CODE_THRIFTY_SONIC": "0" } }
Apparently while in Auto Mode claude code adds this instruction.
It drove me crazy that I couldn't read the diffs.
And it also breaks /rewind
More info github.com/anthropics/claude…
You don't need a network admin if you have Claude code.
Traced a bad wifi connection at the office down to a faulty wire between the Gateway and PoE+ injector
It wasn't something I couldn't do on my own (probably faster but less accurately measured), but it was a nice test of Fable's logic and capabilities
First experiments with Astra compared with Fable 5.1. Initial results look good, but not quite Fable level.
Same prompt, two interactive voxel worlds of Piața Unirii in Cluj-Napoca.
Prompt adapted from @petergostev: Rome → Union Square in Cluj-Napoca.
A few differences below.
My takeaway from these two runs: Astra gets some spatial facts more right, but feels more procedural; Fable captures more of the character and life of Piața Unirii.
Both make lovely little worlds.
I enjoyed the sunset views, especially Fable - feels warmer.
Try the demos: piata-unirii.vercel.app
Finally, speed and costs (xhigh)
Fable 5.1: 1h20m · $33.60
4.3K in / 347K out
Cache: 23.9M read / 623K write
Astra: 36m · ~$7.01* (estimate)
117.8K in / 61.1K out
Cache: 2.776M read / 0? write
Videos recorded by Astra.
Code and prompt:
github.com/danmana/piata-uni…