Open-source ML. Mostly post-training / agent preference data.
United States
Joined August 2026
- Tweets163
- Following77
- Followers13
- Likes369
Pinned Tweet
Cursor / Claude Desktop / ChatGPT keep **separate** memory.
You re-explain the project. Agents forget last week's decisions. Claude Projects help inside Claude — then Cursor starts from zero.
Official exports sit in a zip. Agents never see them.
Asa Verren retweeted
Gemini 4 Argon is available in Devin ⚡
It's available at 75% off for Devin users through October 27th
Built a small thing: drop your ChatGPT export in and search every old chat, right in your browser. No account, no email, and the file never leaves your computer.
mirror.projxon.ai/try
Asa Verren retweeted
Worked with Claude Opus 5.5 to create a video explaining for beginners how LLMs work and why experiments are so important.
I had an idea, put in a prompt, and it built the animation and generated the audio with @ElevenLabs v4.
After a couple iterations, this was the result:
Built for people already in Cursor / Claude / MCP land:
- Ask = lexical vault search over your exports (citations, not a generative "AI that knows you")
- Connect = MCP memory over the same vault
- Miss stays empty — we won't invent a hit
If you live in agent harnesses and hate re-pasting context, this is the layer.
Free signup (magic link) → drop an export → Find or Connect:
mirror.projxon.ai
Connect / Plan when you need higher agent-memory limits:
mirror.projxon.ai/memory
I'm building Mirror. Capture/extension keep-current is parked — free vault + Connect first; Plan is the paid path if you hit free limits.
ChatGPT decisions you’ll never find again. New agents start blank every time.
Mirror: drop official ChatGPT/Claude zips into a vault you own. Search with citations back to the real threads. Connect Cursor or Claude Desktop to the same memory.
mirror.projxon.ai — free to start
Teamed up with @HudsonGouge at Projxon AI on testing + deployment — that’s why Mirror’s on projxon.ai for now. Cool stuff from the collab still cooking for later.
The preference signal that actually matters for coding agents is the first action where a resolve and a fail split on the same issue.
Mined 5,420 of those from OpenHands trajectories → huggingface.co/datasets/asav…
Ran LoRA DPO on Qwen3.5-4B. Held-out pairwise went 51.7% → 52.9% (+1.2pp). Inside the noise. Adapter is up anyway with an honest card: huggingface.co/asaverren/qwe…
After sitting with the pairs: a lot of “divergence” was still soft explore (ls vs view). Filtered a strong split (341 train / 52 eval) where chosen is actually an edit/test/script: huggingface.co/datasets/asav…
They're never living "perfect little angel" down. I'm so here for it.
Tired of the pretend-morals pause stuff. Build.
Asa Verren retweeted
The difference is that employees are stupid but smart while AI Agents are smart but stupid.
Asa Verren retweeted
For anyone who doesn't follow,
CEO of Closed AI? Scam Altman a.k.a. Scammy Sammy
CEO of Misanthropic? Dario Amodei a.k.a. Devious Dario/Doomer Dario
