Slade

Lost in the right direction.

[Slade @R_S1ade]

I am spending lots of time on my @getsolari project in hopes of landing an interview.

Even though its super unlikely, I am loving every bit of it. Seeing how AI has progressed, being able to test differences run locally vs renting out GPU's for massive compute, and just exploring the unknown has all been amazing.

No matter what happens, I'm glad I decided to take this project on.

[Slade @R_S1ade]

Although this isn't a task result yet, it's certainly real progess. I will do a live 30-seed run next using base and v3 on the same input.

Everything so far (what has worked, what is broken) is all in the repo.

Week 1 thread with the setup and the reset benchmark: https://t.co/HbryknogSA

[Slade @R_S1ade]

Building in public for the @getsolari challenge.

Forkloop: a training loop for computer-use agents. Real OpenEMR + a payer portal inside one Solari desktop snapshot. Every reset is one revert(). Every reward comes from the DB rather than a LLM

Goal: a held out learning curve for a 4B vision only policy.

Week 1 numbers ↓

[Slade @R_S1ade]

Solari/Forkloop Project Update: I starting fine tuning for a 4B vision agent and in an attempt to provide results, it started to lie...

First LoRA checkpoint results shown in the image.

It starting lying because it was getting rewarded with a DB query which looks good on paper.

Repo: https://t.co/8j6DeEalpR

Read more below:

[Slade @R_S1ade]

The actual bug: The number is visible in the previous screenshot. Verified with OCR vision on all 25 training examples.

v3: On this version, we give the model the previous frame too. The image shows results from identical inputs to both models:

exact auth number typed: 2/20 → 14/20
navigation matches teacher: 8/20 → 13/20

[Slade @R_S1ade]

Cause: We had no reasoning(it was getting stripped) so the model learned "type an auth-shaped string here."

As a fix, in v2 we put the teacher's reasoning back. OpenEMR path returned (12/30 reach the chart). Most of the numbers entered were memorized from previous work so we still have 0 verified.

[Slade @R_S1ade]

@getsolari Repo: Includes everything that broke and what currently works well: https://t.co/nSbvIPrIh7

[Slade @R_S1ade]

Building in public for the @getsolari challenge.

Forkloop: a training loop for computer-use agents. Real OpenEMR + a payer portal inside one Solari desktop snapshot. Every reset is one revert(). Every reward comes from the DB rather than a LLM

Goal: a held out learning curve for a 4B vision only policy.

Week 1 numbers ↓

[Slade @R_S1ade]

@getsolari 3. Next: LoRA on the verified trajectories with checkpoints at 25/50/94. Once we get a held-out eval with the same oracle. The curve is the deliverable.

Everything measured is in the repo (including what broke)

@harrychow_

[Slade @R_S1ade]

2. Base Fara 1.5 4B, same tasks: 0/30.

It logs in 30/30 once the instruction says which word is the password. Then it navigates OpenEMR by inventing URLs and drops its own session. The 9B gets 4/30 to the patient chart and no further.

The web agent struggles to click through an EHR it's never seen. That's what fine tuning is for

[Slade @R_S1ade]

1. Teacher model on an annoying task (find the auth number in the patient's documents, file the appeal): 94/100 verified at $0.061 per trajectory.

Verified = the oracle checked the appeal row, the attachment hash, the insurance ID, and that no unrelated record changed. 0 LLM-graded rewards.

[Slade @R_S1ade]

Reported a 409 on revert() for desktop snapshots to @getsolari on Tuesday. Fixed within a day.

Before: 409 "Not revertable" on every machine and a failed revert destroyed the VM. After: revert to a snapshot of a running desktop in 21.5 s p50 (state restored 3/3), and 10/10 reverts to our 8.5 GB golden image on one machine id.

That golden is an Ubuntu desktop + OpenEMR + a payer portal. Now every RL episode resets with one API call.

If anyone is wondering, I am using GPT luna with this. I didn't want it to be checked by another LLM so it gets checked with a SQL oracle.

As far as the chart goes, restore time is bimodal. ~22s or 70–160s and the same applies for revert() and forks, so it's host-side, not our client. Thanks @harrychow_ and the rest of the team.

[Slade @R_S1ade]

@KrisTalksAI @Da7_Tech Was likely just a power hungry task. The same exact thing happens to me on claude code and codex. Doesn't even matter what agent. The openAI SSD issues have been fixed now but the damage that was previously done doesn't go away.

[Slade @R_S1ade]

@TomJBelfort @1Umairshaikh This. The community loves to talk big about ideas until its actually on the market. Real people using your MVP is how progress is made.

IMO this is why so many vibe coded products fail. Founders take signicantly too long "perfectly" their MVP.

[Slade @R_S1ade]

@TomJBelfort Cool tool but assuming it runs off of API cost, isn't it signicantly more expensive than users having multiple max subscriptions?

Your certainly not dumb so Im assuming you have an answer I don't. Do you mind sharing what I'm missing?

[Slade @R_S1ade]

@slakpsd Not offering refunds when you can't complete a commission for 3 weeks is kinda wild 💀

[Slade @R_S1ade]

@DiverseSnipes You still alive?

[Slade @R_S1ade]

@lyriqsele Ngl this are all sick

[Slade @R_S1ade]

@rapunksnft @Boobanubba @Shshekfeh @jamesmentes

[Slade @R_S1ade]

@nickcreates R