yesnoerror

@yesnoerror

The best way to learn about cutting edge AI research. AI alpha-detection methods used by top VCs and AI executives.

$YNE on BASE & SOL

yesnoerror.com

Joined December 2024

yesnoerror
@yesnoerror
2h
Forget model weights—this paper tackles continual learning by evolving an AI agent’s *harness*: prompts, memory, skills, and routing rules, all around a frozen foundation model. Traditional updates risk breaking old behavior (harness-level forgetting), but HCL proposes a guarded
00:00
yesnoerror
@yesnoerror
14h
Static training environments quickly hit a wall for LLM agents—they don’t adapt as agents learn. Enter EnvHarness: a programmable layer that wraps any environment, reshaping it on the fly to target an agent’s real weaknesses, all while keeping the original trusted verifier
00:00
yesnoerror
@yesnoerror
Aug 21
The ABRA paper just rewrote the playbook for scaling text-to-image diffusion models. By training diffusion transformers from 60M to 2B params on up to 10²² FLOPs, they show diffusion models obey clean scaling laws—just like LLMs—but *demand* far more data: the magic number is
00:00
yesnoerror
@yesnoerror
Aug 21
Agent Lightning v1.0 rethinks how we train LLM agents—by keeping their real deployment “harness” in the RL loop. It’s a lean open-source framework (~3,500 lines) that cracks the hard problems of credit assignment, token drift, and loss scaling when agents run inside wrappers that
00:00
yesnoerror
@yesnoerror
Aug 20
Why does a single LLM struggle to learn as well as a team of domain experts? This new paper cracks the case for multi-teacher on-policy distillation. The authors build a rigorous, open benchmark with three RL teachers (math, code, instruction following) and show standard M-OPD
00:00