Sabitlenmiş Tweet
Hyper-success = Alienation -> Defiance -> Obsession -> Mastery -> Impact
English
René Sultan
219 posts
René Sultan
@rene_sultan
building @ramplabs, prev @ycombinator @columbia
San Francisco, CA Katılım Temmuz 2021
785 Takip Edilen663 Takipçiler
I attended an event hosted at a Michelin star restaurant in SF.
Only one person drank alcohol.
English
what problem do you most hope AI will solve in the future?
maybe we can help!
English
@googleaidevs @RampLabs Pleasure working with the team, congrats on the launch 🚀
English
René Sultan retweetledi
Using the new Managed Agents in the Gemini API, @RampLabs built their advanced finance agents without touching the backend infrastructure.
Learn more ↓
English
San Mateo, CA 🇺🇸 English
@googleaidevs @RamP Pleasure working with the team, congrats on the launch 🚀
English
It was cool to work with the @GoogleDeepMind team on their Gemini Managed Agents service! They did a great job at integrating the agent runtime into the platform to quickly build production agents. Congrats on the launch!
English
@QiXuanLu19010 @AnthropicAI @claudeai @alexalbert__ I think you can do that already on Claude Code by manually editing your MEMORY.md file in your .claude folder. It’s not available in Claude chat because the memory file is handled server-side and only accessible via Claude chat tools
English
@rene_sultan @AnthropicAI @claudeai @alexalbert__ Same. I don't want memory to be something I have to negotiate with the model. I want a page I can inspect, fix, or delete when it's wrong.
This matters even more for coding agents, where one stale fact can keep getting reused.
English
.@AnthropicAI @claudeai @alexalbert__ feature ask: a raw text editor for Claude chat memory. Right now settings lets me view my edits but every change has to be dictated to Claude in chat. Claude Code allows it by manually editing MEMORY.md, why not with Claude chat too?
English
René Sultan retweetledi
Things every AI app startup says today to justify their defensibility:
1. We support multiple models. Our customers do not want to lock in to one vendor.
2. We have a data moat. We post-train open-source models to be much better and cheaper than closed-source.
3. We do deep integrations to help our harness use our "context graph" and build custom workflows.
In the best case, this is actually true. In many cases, it is hilariously false.
English
.@PrimeIntellect makes it incredibly easy to do good RL post-training and the team behind it is phenomenal. Every forward looking company should look into this newer paradigm
Prime Intellect@PrimeIntellect
We worked with @RampLabs to train Fast Ask using Lab A small RL-trained subagent that helps the Ramp Sheets agent find answers in spreadsheets. The resulting FastAsk model outperformed Opus 4.6, while obtaining Haiku-level speeds at even lower costs.
English
René Sultan retweetledi
We partnered with @PrimeIntellect to build Fast Ask, a small RL-trained subagent that helps our Sheets agent find answers in spreadsheets. It scores +4% over Opus on exact match accuracy at Haiku latency.
English
A great case study for when and how to do good RL-Post Training. A must read.
@RampLabs 🤝 @PrimeIntellect
Ramp Labs@RampLabs
English