Wednesday / AI 101 series: On-policy distillation zeitgeist

Friday / Open Source AI series: a surprise announcement

Type I civilizations harness planetary-scale energy.

Type II harness the energy of their star.

Type III operate on galactic scales.

First, energy becomes the limiting factor for intelligence. Access to cheap, continuous power at scale – that’s what matters most. This is why AI shows up in utility forecasts, transformer shortages, and regional politics. Once intelligence hits the grid, the grid pushes back.

Second, the geometry of infrastructure starts to matter. On Earth, energy is seasonal, regulated, land-constrained, and socially contested. In orbit, solar power is near-constant and space is abundant. “It’s always sunny in space!” changes where the bottleneck lives.

Third, logistics replaces invention as the hard problem. Starship matters less because it can reach Mars and more because it is meant to move mass repeatedly, cheaply, and on schedule. That changes what is possible. A civilization does not move toward Type II by inventing one breakthrough device (or coding platform). It moves there by building systems that can move material and energy at scale, over and over again, without stopping.

Thinking out being a workforce for AI. Join me →here

The Second Pre-training Paradigm by Jim Fan

The Anthropic Hive Mind by Steve Yagge

End Game Play by Will Manidis

Claude Opus 4.6, Incrementally Better
Anthropic launched Claude Opus, an update focused on more consistent reasoning, improved tool use, and better performance on long-context tasks. The release avoids bold claims and flashy benchmarks, instead emphasizing reliability and steady progress. It fits Anthropic’s broader pattern: iterate carefully, prioritize trust, and let adoption do the talking. The most interesting case so far: Building a C compiler with a team of parallel Claudes

GPT-5.3-Codex Expands the Scope of Codex
OpenAI introduced GPT-5.3-Codex, an updated model that combines improved coding performance with broader agentic and professional task support. The release focuses on longer-running tasks, better tool use, and more reliable computer interaction, positioning Codex as something closer to a general work agent than a coding assistant. OpenAI also emphasized internal use, noting material changes in how its own teams operate.

ChatGPT Tests Ads, Promises a Firewall
OpenAI began testing ads in ChatGPT for logged-in adult users in the U.S. on the Free and Go tiers. Paid tiers (Plus/Pro/Business/Enterprise/Education) stay ad-free. OpenAI says ads are labeled, kept separate from answers, and do not affect responses; advertisers get only aggregate performance data. Users can manage personalization and delete ad data.

OpenAI Goes Agent-First, on Purpose
In a widely circulated post, OpenAI president Greg Brockman outlined an internal shift toward agentic software development. The goal: agents as the default interface for technical work, replacing editors and terminals where possible. The guidance is notably operational – roles, documentation, infra, and accountability – suggesting this is less a vision statement than an execution plan.

Agentic Coding Grows Up
A new 2026 Agentic Coding Trends Report argues that software development is shifting from writing code to orchestrating agents. The report highlights coordinated multi-agent systems, long-running agents, and scaled human oversight as the real levers of change. The message is restrained: productivity gains are real, but durable advantage comes from structure, supervision, and security – not full automation.

Anthropic Triggers a Market Repricing
Anthropic’s release of Claude Opus 4.6 and its broader push toward long-running, agentic coding systems prompted a sharp selloff across publicly traded AI tooling and dev-infrastructure companies. Investors reacted less to raw benchmarks than to pricing pressure and the implication that large labs are moving directly into territory once reserved for startups. The move forced a fast reassessment of defensibility across the AI software stack.

Model Tech Report: Kimi K2.5: Visual Agentic Intelligence
Integrates joint text–vision pretraining and reinforcement learning with parallel agent orchestration to enable scalable multimodal agentic intelligence →read the paper

ERNIE 5.0 Technical Report
Trains a unified autoregressive multimodal foundation model with elastic ultra-sparse MoE routing to support flexible deployment across scale and resource constraints →read the paper

Agents are becoming population-based and modular

RL is becoming data-scalable and behavior-aware

Memory, attention, and retrieval are being treated as policies

SWE and GUI are the real stress tests

Systems work is setting the ceiling for everything else

⭐️ Golden Goose: Synthesize Unlimited RLVR Tasks from Unverifiable Internet Text
One of the most strategically important RL papers right now. It breaks the data bottleneck for RLVR by exploiting unverifiable text at scale → read the paper

⭐️ Reinforced Attention Learning
Shifts optimization from tokens to attention distributions. This is a real conceptual step forward for multimodal post-training → read the paper

⭐️ Rethinking the Trust Region in LLM Reinforcement Learning
Argues PPO-style clipping is structurally wrong for LLMs and replaces it with divergence-based constraints. This will age well → read the paper

⭐️ GRP-Obliteration: Unaligning LLMs With a Single Unlabeled Prompt
Shows that post-training safety alignment can be reliably undone using GRPO with minimal supervision, while largely preserving model utility. Important because it treats alignment as reversible behavior, not a stable property, and uses the same RL machinery the field relies on for capability gains → read the paper

F-GRPO: Don’t Let Your Policy Learn the Obvious and Forget the Rare
Fixes rare-solution collapse in group-based RL. A clean, incremental improvement with real gains → read the paper

SLIME: Stabilized Likelihood Implicit Margin Enforcement
Addresses unlearning and formatting collapse in preference optimization. Solid alignment hygiene work → read the paper

Self-Hinting Language Models Enhance Reinforcement Learning
Uses privileged hints during training to prevent GRPO collapse, then removes them at test time. Clever and practical → read the paper

Good SFT Optimizes for SFT, Better SFT Prepares for RL
Important reminder that SFT quality should be judged by downstream RL performance, not standalone metrics → read the paper

On the Entropy Dynamics in Reinforcement Fine-Tuning of LLMs
Theory-heavy but useful for understanding why entropy control methods behave the way they do → read the paper

⭐️ Group-Evolving Agents: Open-Ended Self-Improvement via Experience Sharing
Group-level evolution beats tree-style self-evolution by actually reusing exploratory diversity. One of the clearest signals that agent learning is shifting from “single mind” to “population dynamics” → read the paper

⭐️ AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration
Formalizes agents as composable tuples and treats sub-agents as dynamically instantiated tools. This is quietly one of the most practical orchestration abstractions this year → read the paper

⭐️ MARS: Modular Agent with Reflective Search for Automated AI Research
Budget-aware planning + reflective memory for research agents. Important because it treats research as a cost-constrained search problem, not a prompt-engineering task → read the paper

WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent RL
Argues that width, not depth, is the right scaling axis for broad search. Strong empirical signal that parallelism beats ever-longer chains → read the paper

daVinci-Agency: Unlocking Long-Horizon Agency Data-Efficiently
Uses real-world PR sequences as supervision for long-horizon agency. Interesting mainly as a data lens, less as a general framework → read the paper

MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents
Treats memory operations as learnable skills that themselves evolve. Fits the broader shift toward memory-as-policy → read the paper

RE-TRAC: Recursive Trajectory Compression for Deep Search Agents
Cross-trajectory reflection instead of linear ReAct loops. A clean fix for local-optimum collapse in deep research agents → read the paper

⭐️ SWE-Universe: Scale Real-World Verifiable Environments to Millions
One of the most important infrastructure papers of the week. Million-scale verifiable SWE environments changes what mid-training and RL can even mean for coding agents → read the paper

⭐️ SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training
A transparent, end-to-end recipe for building strong SWE agents. Valuable because it’s reproducible and explicit about the full pipeline → read the paper

MEnvAgent: Scalable Polyglot Environment Construction for Verifiable SWE
Solves the multi-language environment bottleneck. Less flashy, but very necessary if SWE agents are to generalize beyond Python → read the paper

SWE-World: Building Software Engineering Agents in Docker-Free Environments
Replaces real execution with learned surrogates. Important mainly for cost and scalability tradeoffs → read the paper

Closing the Loop: Universal Repository Representation with RPG-Encoder
Treats repo comprehension and generation as inverse processes. Strong representation idea that complements SWE agents nicely → read the paper

⭐️ Reinforcement World Model Learning for LLM-based Agents
Aligns simulated and real next states instead of predicting tokens. This is a strong move away from brittle next-token world models → read the paper

Self-Improving World Modelling with Latent Actions (SWIRL)
Learns world models without action labels by treating actions as latent. Conceptually elegant and broadly applicable → read the paper

InftyThink+: Infinite-Horizon Reasoning via RL
Optimizes when and how to summarize reasoning, not just how long to think. Good evidence that CoT scaling needs structure → read the paper

No Global Plan in Chain-of-Thought
Shows LLMs plan locally, not globally. Useful as a diagnostic lens rather than a training recipe → read the paper

Research on World Models Is Not Merely Injecting World Knowledge
A meta-paper, but an important one. Argues for world models as unified systems, not task-specific hacks → read the paper

⭐️ POINTS-GUI-G: GUI-Grounding Journey
One of the clearest demonstrations that RL works extremely well for perception-heavy tasks when rewards are verifiable → read the paper

Generative Visual Code Mobile World Models
Predicts GUI states as executable code instead of pixels. Very strong idea for mobile and UI agents → read the paper

Training Data Efficiency in Multimodal Process Reward Models
Shows most MPRM data is redundant and how to select informative subsets cheaply → read the paper

⭐️ Horizon-LM: A RAM-Centric Architecture for LLM Training
Redefines the CPU–GPU boundary and makes 100B+ training feasible on a single node. This is a serious systems contribution → read the paper

OmniMoE: Atomic Experts at Scale
Pushes MoE granularity to the extreme while fixing the systems bottlenecks. Strong system–algorithm co-design → read the paper

HySparse: Hybrid Sparse Attention with KV Cache Sharing
Uses full attention as an oracle and reuses KV cache. Very clean design, very practical → read the paper

OmniSIFT: Modality-Asymmetric Token Compression
One of the better token-compression papers for omni-modal models, with real latency wins → read the paper

FASA: Frequency-aware Sparse Attention
Discovers functional sparsity in RoPE frequencies. Elegant and surprisingly effective → read the paper