§ Coding

Simon Willison hosting SF meetup Oct 14 for builders experimenting with coding agents to share work and learnings.

llm 0.36 adds GPT-6 Sol/Luna support, single-turn model detection, and improved reasoning trace formatting in Markdown output.

CliffCompaction reduces token compression costs by 50% for long-horizon coding agents while maintaining performance on KernelBench and Terminal-Bench.

EquivSVA dataset enables testing whether LLM-generated SystemVerilog assertions capture true behavior vs. implementation-specific details.

Study shows compile rate is unreliable metric for LLM code vulnerability repair; proposes change-aware evaluation across 350M–6.7B parameter models.

llm-anthropic 0.36 adds CLI support for Claude Opus 5.5 model access.

TraceVIC uses causal reasoning over code evolution to identify vulnerability-inducing commits; improves on git-blame heuristics.

Method for steering LLM reasoning via semantic exploration of problem-specific strategies rather than naive repeated sampling.

Serving infrastructure (Ollama, Gemma, Phi) confounds tool-use evaluation results; model behavior inconsistency stems from serving layer gatekeeping.

llm-typesafe 0.1a0 plugin adds support for TypeSafe AI's Jev model to the LLM CLI for structured classification tasks.

Face/Off evaluation framework reveals LLMs overweight lexical cues in code comprehension tasks despite access to semantic program structure.

PACT formalizes token-level credit assignment in RL with three axioms, improving actor-critic LLM post-training.

ClusterFewshot improves LLM few-shot prompt optimization by combining semantic task structure with utility scoring, outperforming random and metric-based selection in DSPy.

Informed Masking prioritizes high-leverage token positions in diffusion LLM RL by identifying upstream/downstream structure rather than uniform random masking.

Anecdote: large company relying heavily on Claude Code for artifact generation creates workflow dysfunction despite high throughput expectations.

Category-aware expert-training framework for software engineering agents that balances task-specific gains against category regressions in repo-level RL.

Verification-driven fine-tuning approach enabling LLMs to generate syntactically correct FORM code for symbolic particle physics calculations with zero-shot failure.

TicTacBench evaluates coding agents on RTL timing closure, a critical gap in existing hardware design benchmarks.

CraftBench-UE deterministic benchmark evaluates coding agents on 70 Unreal Engine tasks across C++, Blueprint, and editor scripting without LLM judges.

Anthropic adds AGENTS.md support to Claude Code v2.1.277 for project instruction customization via modular architecture.

CodeMidas: RL pipeline scales agentic coding tasks by extracting diverse environments directly from open-source codebase source code.

Iterative framework using generator and extractor LLMs to synthesize interpretable categorical features for tabular prediction.

Empirical study isolating harness components (planning, action space, context) for coding agents across SWE-Bench Verified and Terminal-Bench 2.1.

Hybrid speculative decoding strategy dynamically switches between neural drafting and context copying to maximize LLM inference speedup.

AdaRepair-Mem adaptively balances episodic memory retrieval for LLM-based repository-level program repair across imbalanced datasets.

ProgramDistill benchmarks coding agents on inferring and implementing features from interactive reference web apps via mine-craft-patch factorization pipeline.

Non-contrastive representation learning for detecting semantically equivalent code clones; addresses Type-IV clone detection challenges.

SKIP framework reduces LLM inference overhead by guiding step-wise preference learning for concise reasoning without sacrificing accuracy versus full chain-of-thought.

RepoAtlas: multimodal repository navigation for LLM coding agents using evolving code graph views for issue resolution.

Lit3R combines retrieval, reranking, and LLMs for evidence-grounded QA over scientific literature without task-specific training.

QART combines quantum-classical hybrid architecture with QUBO optimization for long-horizon reasoning under asymptotic reliability conditions.

Bellman Policy Optimization (BPO) improves LLM reasoning via critic-free RL with verifiable rewards, reformulating Policy Mirror Descent without intermediate state value estimation.

AlgoEvo enables autonomous agents to dynamically inspect, edit, and improve code via runtime feedback and a paradigm-agnostic skill hub for algorithm discovery.

Sequential Adapter Stacking transfers knowledge from high-resource to low-resource languages in Whisper multilingual ASR via parameter-efficient adapters.

Laurie Voss argues AI commoditizes code writing and maintenance; future software work shifts to requirements gathering, specification, and UX design.

Simon Willison demonstrates ChatGPT Work using GPT-6 Astra to generate running routes via OSM/Nominatim integration with 27-minute inference.

Paul Ford argues AI enables broader coding ability but quality software still requires skilled human thought and craft, noting failed projects often result from AI-assisted mediocrity.

Boris Cherny outlines Anthropic's production guardrails for Claude-generated code: linting, testing, fuzzers, automated review, and refactoring.

Personal reflection: AI coding agents commodifying specification-to-code translation, prompting career reorientation toward higher-level problem-solving.

CanvasAnneal curriculum RL framework for diffusion language models using teacher-guided reasoning traces to improve exploration.

Cognition integrates GPT-6 Astra into Devin to automate software testing and validation, reducing engineering review overhead.

César de la Fuente's lab applies Codex and ChatGPT to genome mining for antimicrobial peptide discovery against drug-resistant pathogens.

Study proposes retrofitting code with exception handling using LLMs guided by Exceptional Behavior Tests, automating try/catch and error-guard generation.

Φ-Bench evaluates LLM capability on open-ended LLM infrastructure engineering tasks—kernel optimization, operator design, compiler tuning—beyond isolated benchmarks.

LLM bug-fixers fail: models claim bugs in bug-free code and damage correct programs faster than fixing actual bugs, creating pseudo-cycles.

Probe-driven test-time RL method for code generation via behavioral agreement on synthetic inputs instead of surface-form voting.

ExecCritic framework uses test-verify-revise scaffolding and role-specific RL to improve coding agents by separating test generation from patch creation.

MIT researcher uses GPT-5.6 Sol with Codex to autonomously run quantum computing experiments, analyze results, and calibrate qubits.

1Password reports 21% engineering productivity gain using OpenAI Codex for feature development and internal tooling.

ProcArena benchmark evaluates LLMs on multi-scenario PL/SQL development including direct generation, modification, debugging, and interactive multi-turn workflows.