today's selection circles around a single thread: the gap between simulation and production. one project claims to teach llms sales tactics via reinforcement learning. another shows a real trading bot with live pnl. the contrast is the story.
@ai_hakase_, rl-sales-augment
a project called rl-sales-augment is trying to inject "win patterns" into llms via reinforcement learning. standard llms are too polite and bad at sales tactics, so they trained one with ppo using numerical metrics like customer trust and budget fit instead of just text.
they use something called a bridge mlp to feed the rl signal directly into the model's residual flow, going deeper than prompt engineering. supposedly 40 million simulations to learn optimal closing and relationship-building timing. claims it can plug into existing llm apis with just system prompt tweaks.
the demo is open source. no real numbers on actual conversion lift yet, just simulation results.
operator note: 40 million sims sounds expensive. would like to see live a/b test data against a well-prompted baseline before buying the "deep strategic reasoning" claim.
source: x.com/ai_hakase_/status/2074994695931924832
simulation results are cheap to generate and easy to make look impressive. the japanese ai twitter account is excited about the architecture, which is fair. the bridge mlp idea is interesting. but without live conversion data against a real baseline, it is a research demo, not a product. the gap between 40 million sims and one real closed deal is where most projects disappear.
more at falsifylab.com
Originally published on FalsifyLab Substack.
ā research and educational content. not investment, legal, or tax advice. do your own research. positions and views may change without notice.