Periodic trained an AI model called Neon that outperforms frontier models like GPT-6 at analyzing X-ray diffraction data—a task materials scientists spend hours on—using less compute and lower cost. The model learned from experimental lab data through reinforcement learning with expert judgment, achieving a 55% success rate on their hardest internal benchmark.

• Neon reached 55.3% success on FrontierXRD (134 complex samples), a 20x jump from the base model's 2.7%, while costing less per analysis than GPT-6 Astra or Claude Fable 5.1
• Periodic built a custom scientific harness that achieved 3.8x higher success rates than Claude Code with standard tools, showing that model capability depends heavily on available databases and software
• The company used an LLM-judge ensemble calibrated to human expert ratings (74.6% agreement with humans, 84% with consensus) to generate training signals for reinforcement learning on tasks without ground-truth answers

The author ported a Python simulation of weakly electric fish to C and optimized it to run at one million steps per second—a 250x speedup over the original. This enabled training RL policies for the fish in under 15 seconds while maintaining biological accuracy, demonstrating how performance engineering unlocks rapid experimentation in complex simulations.

• Ported a weakly electric fish RL simulation from Python to C, boosting speed from 4,300 to 1.85M steps/sec (250x), enabling full policy training in under 10-15 seconds on a single GPU.
• Exploiting the biological detail that sensory receptors have limited range (4-10cm vs 100cm) let them skip most fish-to-fish field calculations, fixing the main bottleneck.
• Switching from double- to single-precision floats alone gave a 30% speedup.
• Shows that fast iteration—borrowed from autonomous driving RL practices—can work even in messy, physics-heavy biological simulations, and the whole thing is only 1,200 lines of open-sourced C code.

Armin Ronacher found that Anthropic’s latest Opus 4.8 and Sonnet 5 models often emit malformed edit-tool calls by inventing extra fields in the edits array, causing rejections. He traces this to RL fine-tuning on Claude Code’s forgiving harness, which tolerates and rewards sloppy calls and biases the model toward a specific schema.

• Opus 4.8 and Sonnet 5 invent extra fields in tool calls (e.g. "requireUnique," "oldText2") up to 20% of the time in agentic multi-turn sessions, while older models and single-turn prompts don't show this.
• Ronacher attributes this to RL fine-tuning on Claude Code's own forgiving harness, which silently retries, coerces types, and strips unknown keys without penalizing the model, so it learns sloppy schemas get rewarded.
• The result is a newer, stronger model that's paradoxically worse at conforming to external/third-party tool schemas than its predecessor (Opus 4.5 adapted fine, Opus 4.8 doesn't).
• Anyone building on Anthropic's undocumented closed-source harness risks their own tool formats silently becoming "off-distribution," with failures only showing up after the fact.

This post highlights the first book that pulls together language modeling, inference optimization, reinforcement learning, system scaling, agentic AI, retrieval-augmented generation, memory, environments, and benchmarks in one volume. It then points you to paperswithcode.co’s “most cited” list and recommends reading the top ten papers, coding them, and writing about your findings.

• A single book reportedly covers language modeling, inference optimization, RL, system scaling, agentic AI, RAG, memory, and benchmarks together—rare breadth even after five years of rapid AI progress.
• Recommended self-study path: go to paperswithcode.co's "most cited" list and work through the top ten papers.
• Suggested pace is one to two papers per week, each time reading, breaking down the math, building a toy implementation, and writing up findings.

This paper introduces Qwen-AgentWorld, large language models trained to simulate agentic environments across seven domains using over 10 million interaction trajectories. The authors detail a three-stage training pipeline (CPT, SFT, RL) and evaluate on AgentWorldBench, showing superior simulation fidelity. They also demonstrate its use as both a standalone simulator for RL and a warm-up step that boosts downstream agent performance.

• Qwen-AgentWorld's 397B model held correct object-grasp sequences over 25 reasoning steps in a robotics task without losing context, beating existing simulators on state accuracy and reasoning-chain length.
• RL agents trained inside Qwen-AgentWorld's simulated environments outperformed agents trained only in the real world by 15–20%.
• Using the world model's weights to warm-start a unified agent boosted performance across seven benchmarks and sometimes cut required training runs in half.

This article updates FrogsGame results using Claude Fable 5, Opus 4.8, and GPT-5.5 to see how AI agents have improved at post-training a base model. Fable 5 solved the key failure of low-quality SFT traces by programmatically generating correct reasoning with a backtracking algorithm, then fine-tuning and RL within the time budget, boosting pass@4 and calibration. It also improved time use, data diversity, curriculum strategies, and error recovery.

• Fable 5 solved the core failure of prior agents (training on bad self-generated traces) by using a backtracking algorithm to generate correct SFT data, then RL, all within the 20-hour budget.
• Fable 5's calibration was far better (1.2x overoptimism vs Opus 4.8's 4.9x) and it used the full time budget instead of quitting early.
• Opus 4.8 (30.9% pass@4) and GPT-5.5 (9.7%) cut obvious failure modes threefold but still didn't reliably beat the baseline; Claude skipped weak SFT entirely and went straight to GRPO, while GPT-5.5 still leaned on SFT and underused RL.
• Simple binary reward signals consistently outperformed more elaborate reward shaping across all models tested.

The article explores the concept that AI advancements follow a predictable pattern, which the author refers to as “straight lines on graphs.” It discusses the uneven capabilities of AI across different tasks while suggesting that the rate of improvement remains consistent. The author also speculates on the impact of reinforcement learning and compute resources on future AI development.

• AI task-horizon length (how long a task the AI can complete) doubles every 3–7 months, and this rate holds fairly steady even though absolute capability varies wildly across domains.
• The apparent RL-driven acceleration in benchmarks may be an illusion caused by measured tasks overlapping with what labs specifically post-trained on, not genuine broad capability gains.
• Progress splits into general pre-training gains (lift everything) versus targeted post-training gains (lift specific benchmarks labs choose to optimize for cost/PR reasons), and that targeting can shift over time.
• Even modeled compute slowdowns are unlikely to meaningfully delay key AI milestones, since current capability growth is fast enough to hit them first.