Charlii

Full-Stack | GenAI Researcher | Creator of Vibe Coding | Exploring AI, Multimodal Interaction & Practical Dev Hacks

[Charlii @charliiai]

jev-ultrafast indexes visible browser controls so Jev can pick an action + target in one request. A small LLM writes only when text input is needed. The repo includes a measured demo and its limits: https://t.co/hs8Z1H0TQB #BrowserUse #AIAgents

[Charlii @charliiai]

#Jev #BrowserAutomation #AIAgents #AgenticAI

[Charlii @charliiai]

Browser agents rarely fail because they cannot click. They fail when the page changes and they keep clicking anyway. We tested Jev Ultrafast, a deliberately narrow next-step chooser. Here is what one controlled run actually showed. Thread:

[Charlii @charliiai]

The useful question is not how many buttons an agent can click. It is: can we see its next choice, bound its actions, and take over when needed? Source and limitations: https://t.co/hs8Z1H0TQB

[Charlii @charliiai]

@cleeeeeeeeement Seo

[Charlii @charliiai]

@theo_louro Ads

[Charlii @charliiai]

@thibaultbarral Q4

[Charlii @charliiai]

@meetshukla_ reactions

[Charlii @charliiai]

@cleeeeeeeeement RADAR

[Charlii @charliiai]

@theo_louro LEADS

[Charlii @charliiai]

@VengeonsP jev

[Charlii @charliiai]

A very interesting idea from Dream-RSI:

Maybe the next step of Agent self-improvement isn't making the model smarter.

It's making the **search process smarter**.

Instead of treating past agent trajectories as logs, Dream-RSI turns them into a replayable world.

The agent can “dream” inside this historical world, test different exploration policies, select a better one, and then deploy it back into the real environment.

So the loop becomes:

Agent → Experience → Better Search Policy → New Experience → ...

The key question shifts from:

“How do we make the model smarter?”

to:

**“How do we make the agent better at deciding what to try next?”**

That feels like a much more interesting direction for recursive self-improvement.

Maybe the future is not just **self-training**.

It is **self-search**.

#AI #AIAgents #LLM #SelfImprovement #AIResearch

[Charlii @charliiai]

@vercinyc deadlines

[Charlii @charliiai]

@SebJefferies Lab

[Charlii @charliiai]

@everestchris6 Fiverr