The Hallway Track
23 tracked signals on post-training.
GLM-5.3 proves post-training RL on long-horizon tasks beats parameter scaling for reasoning
“Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.”
DeepSeek 4 Pro achieves near-frontier quality with MIT-licensed open weights, pressuring closed AI labs.
“DeepSeek has MIT licensed open weights. Anyone can run the exact same model at their own price.”
DeepSeek's updated flash model beats its larger pro version via post-training improvements alone
“Just the post-training step changed?”
Opus 4.8 and Fable regressed on business-task evals after Anthropic removed business skills from post-training
“they removed a part of the post-training recipe that was meant to do business skills”
2026 frontier post-training has shifted to Multi-Teacher On-Policy Distillation (MOPD), merging many specialist models into one.
“The shape of a post-training recipe has changed more in the last year than in the prior three.”
LangChain launches LangSmith fine-tuning in public beta with SmithTune, a CLI to post-train models from agent traces.
“today we're launching LangSmith fine-tuning in public beta with SmithTune, a CLI to allow you to post-train models from your LangSmith traces in one workflow”
LLMs are stagnant at long-form non-fiction writing, a prerequisite for autonomous science
“Models being stagnant in long-form, non-fiction writing should be alarming to those reliant on models autonomously solving grand, open science problems in the near future.”
Mercor scaled to $2B revenue run rate in 4 months as primary agentic data vendor to frontier labs
“this giant transition away from the low-skilled crowdsourcing era of behavior cloning data and moving towards the agentic era of data”
Harvey built a domain-specific research lab by leveraging frontier ecosystem rather than competing with it directly.
“it's an unfair game competing with the frontier labs if you're an application layer company.”
Modal proposes decoupling RL rollout workers from trainer clusters to use distributed GPU capacity across datacenters.
“IO wants all four of these at the same time. Enough GPU, same region, fast fabric, and available now. Any of these like is manageable, but all four of them that are pretty hard to get at the same time.”
DeepSeek V4-Flash 0731 matches GPT-5.6 performance at 60% lower cost via post-training alone
“Terminal-Bench 82.7, up +25.8 from the April preview's 56.9”
OpenAI post-training pioneer argues Claude Code and ChatGPT are the same era, not successive ones
“the team I was part of basically invented post-training as a concept”
Post-training must evolve to let agents adapt to enterprise harnesses without source code access
“there will be these kind of agentic citizens, which you can just deploy once, and they'll be able to adapt to many different types of out of distribution tasks and learn from their interactions”
Post-training is the primary moat for AI application companies over black-box APIs
“right now, with one person, a few weeks, uh without understanding how to write a single line of code, you can do that”
On-policy distillation uses RL policy shift signals to train student models without full RL cost
“the student might be stronger than the teacher in this case”
Nathan Lambert publishes definitive RLHF post-training textbook, free online with 12-hour course
“This is the book I wanted to read when I was getting started a few years ago!”
New research uses chess as a testbed to derive scaling laws for post-training compute allocation
“In pre-training we had the chinchilla scaling laws uh which is roughly like uh uh you would allocate 20x the tokens per parameters of u of your model. But in a we didn't know like how much you would allocate.”
Bespoke Labs says AI post-training has shifted from knowledge data to RL environments for agentic tasks
“we have moved on from knowing to doing right so that's the idea of agents”
AI agent capability gaps in infrastructure reasoning stem from missing training data, not model limitations.
“like with everything in NML, uh the gap in models is usually a gap in data”
The traditional base-model paradigm of web-scale pre-training is being displaced by post-training
“RL was mostly just a cherry on top, shaping the flavor of the interactions more than conferring extra knowledge or quality onto the base model itself.”
Prime Intellect is extending RL to real-world tasks lacking verifiable reward signals.
Nathan Lambert keeps Interconnects intentionally raw and independent, disclosing new advising roles at Arcee AI and Mercor.
“Interconnects is the tip of the spear of all of my missions in AI.”
Low-quality RL training environments and broken harnesses actively degrade models and ruin training runs.
“researchers don’t want your broken RL environments because they will make our models worse”