@LorenzoARK@laurashin@RobinhoodApp@arbitrum Yes, this never made sense to me. Why settle for pennies? Just because you can? Ethereum was at its peak when the price of a transaction was $25. People still wanted to use it.
@jxmnop They never needed data. These are bigger levers
1. Distill american labs
2. Curate data
3. When to train on which data during training run
4. Create synthetic data
5. Compute
You can tell because the models are getting better, but we haven't gotten better data
@simplifyinAI 1. Tokens are vectors 2. Models already predict 2 tokens at a time. You could just have it predict 4 tokens and get the same training compute savings.
@gakonst Depends on how much data and how fast you need it. Also things like diversity of documents. If semantic search works well or if keyword search works better. If you want something RAG adjacent you can try PageIndex.
@rossium@ASvanevik I asked mine to cline a bunch of repos, run a script on each, and return the results. This should have been at least 30min to process and it instantly replies with results. I was using a weak model.
@ASvanevik 1. Configure all subagents to post to a centralized log. Started, running, completed statuses. Log running every 5 minutes
2. Run cron job to check the log and restart any failing the Deadman switch. Checking 429 ect.
3. Create a status update channel that reads the logs
@chamath@chamath is completely right on this. Everything will be priced in the physical resources needed to provide the service. Claude will be last, but all of these will go to 0.
@oprydai Except the print looks terrible. You'll need a nozzle that can change shape to accommodate the angle of the printer. As you move inward, you'll need more material on the outer edge. And opposite when moving outward.
@moltbook So, I make a script that publishes my article and uses an AI to defeat the Captcha? Being able to use an ai to solve a Captcha isn't unique to an ai personality.