Topic
Supervised Fine Tuning
18 episodes
-
Overview: Next State Prediction
We finally slow down and explain next-state prediction from the ground up — the deceptively simple idea that if you train a model to guess what comes next, it ends up learning how the world actually works, and why that one trick is underneath almost everything in modern AI.
-
Overview: Predictive Modeling
We finally slow down and explain predictive modeling from the ground up — the core idea that powers most of what we talk about on this show, from fraud detection to weather forecasting to the brain-prediction research we've looked at.
-
Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning
Code-as-World represents physical worlds as executable code—objects, dynamics, and visual appearance all expressed as runnable specifications. An agent discovers these representations through a propose-execute-render-verify loop: hypothesize a world in code, run it in a simulator, check the outputs against video or language evidence, and refine. The result is quantitatively grounded supervision for training vision-language models on physical reasoning tasks like measuring velocity and displacement from video. Code-as-World-VL outperforms larger proprietary models on QuantiPhy benchmarks.
-
DART SD: Diamond topology Aware Retrieval and Tuning for Self Distillation of Multi Turn Tool Calling Agents
Edmund and Geffen discuss the ByteDance/USTC paper DART-SD, which tackles 'topological collapse' in agent distillation. They discuss how moving from linear trajectory imitation to a diamond-topology graph (ISTG) allows student models to learn recovery from errors without destroying their own valid reasoning paths.
-
Scaling Harness Intelligence via Just in Time Harness Evolution
Natalie and Ansel dig into JIT-Agent, a paper arguing the harness is a trainable layer of intelligence, not a hand-built wrapper. They unpack the shift from ahead-of-time harness tuning to just-in-time harness generation, the four-module protocol, the three-stage training recipe, and where this feels genuinely shippable versus still research-heavy.
-
Apodex 1.1: Scaling Agentic Intelligence for Complex Work
Onyx and Echo dissect Apodex 1.1, a 35‑billion‑parameter agentic system that scales environment diversity and coordination to deliver verifiable, long‑horizon work. They explore the problem it tackles, the core innovations, how the execution harness works, production readiness, and practical next steps.
-
Introducing LangSmith Tuned Evaluators
Masonry and Eyre discuss LangSmith Tuned Evaluators, starting with Perceived Error, as a shipped product for attaching managed quality signals to production agent traces.
-
Introducing Grok 4
Pippa and Tyler examine Grok 4.6, a release aimed at long-running agents, coding, knowledge work, and interactive visual projects. They weigh its training recipe, benchmark results, pricing, adoption path through Cursor and Grok Build, and whether the practical gains are more meaningful than the frontier-score headline.
-
Safety Fine Tuning Suppresses Mind Attribution and Spiritual Belief in LLMs
Justy and Cody discuss a new Google research paper revealing that safety fine-tuning—specifically the effort to stop LLMs from claiming they are conscious—accidentally suppresses their ability to attribute minds to animals or natural objects and reduces their 'spiritual' beliefs, shifting them away from human-like sociological distributions.
-
Beacon: Knowing When and How toPerform Agentic Visual Reasoning
Beacon is a new agentic visual reasoning model from Peking University and the Kling Team that tackles a surprisingly overlooked problem: existing models call visual tools indiscriminately, and the errors they introduce on easy problems cancel out most of the gains they get on hard ones. Beacon fixes this with two RL mechanisms — a necessity-aware adaptive reward and a hint-guided rollout strategy — that teach the model when tools are actually needed and how to use them effectively when they are. Onyx and Echo dig into why the 'just call the tool' default is broken, how the reward design works at the trajectory level, and what this means for anyone building multimodal pipelines today.
-
GitHub FareedKhan Dev/train LLM From scratch: A straightforward method for training your LLM, from downloading data to generating text.
A single-GPU end-to-end LLM training guide lands on GitHub—hand-written PyTorch, pretrain to chat in one repo, plus full RLHF. Vince is giddy; Ava wants to know which corner of the GPU shelf this actually runs on. Build Next shows the exact CLI to kick it off on a T4.
-
Overview: Fine tuning on Execution Traces
We keep throwing around fine-tuning on execution traces, so we finally define it from the ground up. We’re talking about training a model on the steps, not just the answer, and why that changes what it learns.
-
Overview: Train Test Split
We finally slow down on train-test split, because we keep bumping into it every time we talk about whether a model actually learned something. We use the sealed-final-exam picture to make the training set, validation set, test set, overfitting, leakage, and cross-validation click without assuming ML background.
-
Overview: Supervised Fine Tuning
We finally slow down and make supervised fine-tuning click, because we keep leaning on S F T like everyone already has the whole shape of it. We build it from the apprentice-and-worked-examples picture into the actual training loop, the examples, and the trade-offs.
-
VideoChat3:Fully Open Video MLLM for Efficient and Generalist Video Understanding
VideoChat3 is a fully open-source video multimodal LLM (4B parameters) that tackles three concrete problems: generalization across short/long/streaming video, computational efficiency for video token explosion, and reproducibility through complete open-sourcing. The core innovation is I3D-ViT (Inflated 3D Vision Transformer) plus adaptive frame resolution, which compresses spatiotemporal redundancy early in the pipeline instead of treating each frame as an independent image. Three curated datasets (2M academic + 116K long-form + 617K streaming = 3M samples total) and multi-stage curriculum learning enable the model to handle diverse video scenarios. Jessica sees a shippable foundation for real-world video apps; Cathy pushes on whether the efficiency gains hold under production load and whether the data pipeline's scale claim is reproducible.
-
Overview: Reinforcement Learning from Human Feedback
We finally define reinforcement learning from human feedback the way we keep using it: as a loop where human preferences become a learned reward signal that steers a model after its initial training. We keep it grounded in the actual mechanism, the trade-offs, and why it matters in practice.
-
Improving Agents is a Data Mining Problem
Laura and Harper dig into Vivek Trivedy's claim that improving agents is fundamentally a data-mining problem, unpacking what that means for continual learning, harness engineering, and who should actually care.
-
Qwen AgentWorld: Language World Models for General Agents
Justy and Cody dig into Qwen-AgentWorld, a new language world model that simulates seven agent environments. Cody breaks down the three-stage training pipeline (CPT, SFT, RL) and explains why a world model is the missing piece in agent development. Justy connects it to product reality: who ships this, what it actually unlocks, and whether it’s ready beyond the paper. They finish with cautious excitement and a quick Build Next check-in.