Topic
Long Horizon Training
17 episodes
-
Building always On agents
Onyx and Echo walk through Anthropic's guide to always-on agents: what makes them different from ordinary agents, the five primitives behind them, and whether the claim that they're newly practical actually holds up.
-
Model Behavior: Week of November 9, 2026
We spent this week arguing that the agent race just forked: frontier labs are chasing giant unified brains, while everyone else is quietly standardizing on a big-planner / small-doer stack and fighting to own the orchestration layer in the middle. Our read is that if you’re building agents, the question is shifting from “which model is smartest?” to “who controls how your planners, executors, tools and tests snap together.”
-
Introducing LongCat 2
LongCat-2.5 is a 1.6-trillion-parameter native multimodal model aimed at long-horizon coding and agentic work. Pippa and Tyler examine whether its real advance is a better model, a more serious training system, or both.
-
Introducing Inkling 2 Small
Pippa and Tyler dig into Thinking Machines’ new Inkling 2 Small model: a 594B-parameter Mixture-of-Experts that’s API-only today but promised as Apache 2.0 open weights next week, with strong long-horizon coding and autoresearch results. They unpack why the kernel-optimization and time-series experiments matter more than the big parameter number, and why this could be the first open-weight model that’s genuinely tuned for agents doing multi-hour work rather than just answering one-off prompts.
-
Model Behavior: Week of November 2, 2026
We read this week as a stack-splitting week: more capability is concentrating at the top, while the useful middle is getting messier and more controlled. The interesting move for us is that winning now looks less like picking one champion model and more like deciding which layer you trust to own the workflow.
-
Advances in Agentic Computer Graphics
Jessica and Cathy dig into a survey that does two useful things for agentic computer graphics: it gives the field a shared vocabulary, and it shows why the literature is getting hard to track by hand. The real value here is less a new graphics algorithm than a mechanism for comparing systems that all do similar things with different names, plus a literature system that can keep up as the corpus grows.
-
Gemini 4 Carbon: our next frontier model
Jessica and Cathy unpack Google's Gemini 4 Carbon announcement, focusing on the real tension: Google is leading with cyber and long-horizon capability, but the same capabilities make access control and measurement unusually important.
-
Continual Learning Exam: Benchmarking Long Horizon RSI and Experiential Learning in LLM Agents
CLE (Continual Learning Exam) is a 600-level benchmark built by 50+ domain experts to measure whether LLM agents actually improve through accumulated experience across long workflows—not just solve isolated tasks. The benchmark spans 89 professional roles in 12 domains, with tasks organized as five-level ladders where later levels depend on knowledge gained earlier. The key finding: even frontier models with diagnostic feedback and revision opportunities clear only 47.7% of levels and complete all five levels in just 23.3% of tasks. This reveals a hard gap between task-solving capability and sustained learning from experience.
-
Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date
Onyx and Echo dig into Meta's Muse Spark Open release, with Echo skeptical of the headline claim and Onyx arguing that the practical desktop-and-tooling story may matter more than the leaderboard framing.
-
Muse Wonder: A Frontier Model For Personal & Professional AI Agents
Meta’s Muse Wonder launch is less interesting as a leaderboard claim than as a bet that useful personal agents need two system-level abilities: retrieving the right personal context and sustaining effort across messy, unfamiliar tasks. The evidence is promising but mostly internal, so the practical question is whether those abilities survive real users, real permissions, and long-running work.
-
Overview: Long Horizon training
We slow down on long-horizon training: what it means to train a model for the whole arc of a task instead of the next shiny step. We use the repo-fix mental model, then walk through rewards, credit assignment, curricula, decomposition, and why this is still very much live in agent work.
-
Gemini 4 Argon: our next era of frontier intelligence
Onyx and Echo dig into Google’s Gemini 4 Argon announcement and land on the same uneasy place: the real story is not “bigger model,” it’s a model being shaped into a controlled workflow engine for expensive, long-horizon work. They focus on the 1 million token output cap, the aggressive introductory pricing, the internal Google use cases, and the cybersecurity-first rollout, while pushing on where benchmark claims still need real-world proof.
-
The Tasteful Agent: Measuring and Improving Taste in Long Horizon Tasks
Justy and Cody dig into a paper arguing that long-horizon agent performance is bottlenecked by something more specific than general capability: taste, meaning the ability to choose the better path before the evidence is obvious. They unpack Taste-Bench, how decision forks are mined from trajectories without human annotation, why the best model only gets 59.7 percent, and why more reasoning tokens do not rescue weak judgment. The conversation lands on a familiar Exploring Next theme: the useful progress is in better scaffolding and better supervision, not mystical agent vibes.
-
Introducing Gemini 3.8 Flash and 3.8 Flash Cyber
Asteria and Draco discuss the launch of Gemini 3.8 Flash and 3.8 Flash Cyber, focusing on the 'work harder' reasoning approach and the specialized cybersecurity capabilities for trusted defenders.
-
Overview: World Models
We finally stop hand-waving and explain world models from the ground up — what they are, how they actually work, and why the field keeps coming back to them as the missing piece between AI that reacts and AI that plans.
-
GLM 5.3: Scaling Post Training with Long Horizon Environments
Vince and Ava discuss Z dot A I's G L M five point three release, focusing on the claim that post-training alone drove the gains, with long-horizon task environments as the real lever.
-
Meta Superintelligence Labs Releases Muse Code
Vince and Ava dig into Meta’s new Muse Code terminal coding agent and Muse Spark 1.2: are the persistent agents and replayable event log real developer wins or just more agent hype?