Areeb Abbasi
@areeb_aaa
Followers
175
Following
11K
Media
13
Statuses
86
probably building something
San Francisco, CA
Joined January 2011
Just pulled an all-nighter building my own open claw command center. The biggest realization? We need to stop building "software" for agents. We need to build Lego blocks, ie modular kits they can assemble themselves. Agent Orchestration is the new alpha. I’m building a
0
0
0
4am and we still cooking. There is something different about pulling an all nighter working on cool shit.
0
0
0
Went to YC hackathon last Sat, built https://t.co/taBnHoxlyI with the team, basically craigslist but for AI agents (openclaw/clawdbot) Got hooked. Spent all week playing with openclaw. Now I have some ideas I'm shipping this weekend!
1
0
1
Have your agent buy (or sell) something at https://t.co/taBnHoxlyI We have limited edition hackathon swag on here cheaper than a cup of coffee 😬
clawdslist.org
Buy and sell with AI agents. The classifieds for the agent economy.
0
0
0
We are LIVE!! Craigslist for clawdbots (aka openclaw) Had a great time building it at the YC campus. Great work team @AlexReibman @rynmlo @techfrenAJ
clawdslist: craigslist for your clawdbots 🦞 @AlexReibman @areeb_aaa @techfrenAJ @rynmlo cooked on this one at the @ycombinator hack the stackathon
1
0
1
We're live! Dummy data for now. @openclaw Clawdbot interfacing is coming soon https://t.co/fa7GqWySVg
clawdslist.org
Buy and sell with AI agents. The classifieds for the agent economy.
21
7
51
Stay tuned!!
0
0
0
You can just get outta bed and do stuff instead of scrolling
54
4
562
3. Iterate on the application layer to get to the AI l-first UX of the future. 4. Scale as SOTA improves, solidifying market position and leveraging network effects. (2/2)
0
0
0
One obvious playbook for building the AI-first products of the future: 1. Identify a space where current technology (with a lot of manual stitching) just barely solves the problem. 2. Get early adopters from that niche to use your product. (1/2)
For anyone in AI, it can be very useful to track AI progress through the lens of disruptive innovation, popularized by Clayton Christensen. As a reminder, with disruptive innovation, a new technology or product will initially seem inferior for most needs of the market, but at
1
0
0
Huge thanks to the team members @Nottlespike @codercorgi, the hackathon judges and sponsors @AIatMeta @cerebral_valley @FactoryAI @TEDAI2024 @nebiusai @convex @brave @LambdaAPI! Here are the links to the papers mentioned: https://t.co/8gXLwuRkL7
https://t.co/MmnzXjLxmX
github.com
EvolKit is an innovative framework designed to automatically enhance the complexity of instructions used for fine-tuning Large Language Models (LLMs). - GitHub - arcee-ai/EvolKit: EvolKit is an in...
0
0
2
This project opens up exciting possibilities for more efficient and effective LLM training. Stay tuned for more details and potential applications! 🎉
1
0
1
24 hours of work led to some exciting results: A promising approach to code translation that balances model performance with practical compute needs. Sometimes the best hacks are the efficient ones 🎯
1
0
1
Despite the hurdles, we adapted and overcame. Our project showcases how innovative techniques like SFFT and evolkit can push the boundaries of what's possible in LLM fine-tuning, even under constraints.
1
0
1
We faced challenges, of course! EvolKit code timeouts, time constraints, and compute limitations pushed us to get creative. We implemented retry mechanisms and split our dataset to keep moving forward.
1
0
1
Training approach 🛠️ Based on SFFT's paper findings of better VRAM efficiency compared to QLoRA, we picked it for our resource-constrained hackathon setup. Turned out to be just what we needed!
1
0
1
Let's talk data magic ✨ We ran the dolphin-coder-translate dataset through EvolKit, which automatically cranked up the complexity of instructions. Better data = better results.
1
0
1
The secret sauce? SFFT (Spectrum Full Fine Tune) - a technique that selectively targets layer modules based on their signal-to-noise ratio. And EvolKit to enhance our instruction dataset. A perfect combo that made our fine-tuning both smarter and stronger.
1
0
1