AI Daily — September 14, 2026
2026-09-14
TODAY'S NEWS
rss.arxiv.org
Harness or Model? Isolating the Harness Effect in Agentic Coding
A contamination-controlled study of 256 private repo and post-cutoff contest tasks finds no statistically reliable advantage for vendor-native harness/model pairings (e.g., claude-agent-sdk with Claude Opus 4.8, or openai-codex SDK with GPT-5.5) over third-party harnesses like deepagents. Out of 800 planned runs across Claude Opus 4.8, GPT-5.5, Gemini 3.5 Flash, and DeepSeek V3.2, neither same-model harness contrast resolved a consistent performance edge. This challenges the common assumption that vendor-tuned scaffolding is a meaningful differentiator in agentic coding benchmarks.
rss.arxiv.org
The Cost of Compression: A Rate-Distortion Limit on Factual Hallucination
This paper establishes a theoretical lower bound on hallucination error in closed-book QA, showing that finite model capacity (bits) forces lossy compression of training facts even when facts were observed — a distinct error source from simple coverage gaps. Using a formal coverage-compression model with N queries, K answers, M training facts, and B bits of storage, the authors derive a rate-distortion bound on minimum achievable error. The result implies that hallucination has an irreducible floor set by information-theoretic limits, independent of training data quality or volume.
LAST WEEK'S TOP STORIES
GPT-6 Astra: The next generation in intelligence for work
OpenAI released GPT-6 Astra, its first GPT-6-series model with advanced reasoning, computer use, and GUI/browser automation, marking a major generational leap in capability targeting agentic enterprise workflows.
OpenAI claims AI-generated solution to Navier–Stokes Millennium Prize Problem
OpenAI published a formally Lean-verified AI-generated solution to one of mathematics' hardest open problems, triggering immediate controversy in the math community and raising fundamental questions about how AI-produced proofs should be validated and credited.
ChatGPT for Financial Services launches with GPT-6 Astra and built-in financial data
OpenAI entered domain-specific enterprise verticals with a finance-focused ChatGPT product bundling GPT-6 Astra with curated financial data feeds, directly competing with Bloomberg Terminal AI and reducing the need for operator-side RAG setups.
OpenAI and GSA offer $0 licensing and 50% usage discounts to US governments
A landmark procurement deal positions OpenAI as a primary AI vendor across US federal, state, local, and tribal governments, with zero-cost licensing and a cyber defense component that embeds OpenAI models into government security operations.
Quantifying grokking onset: scaling laws for the memorization-to-generalization transition
Researchers derived the first quantitative phase diagram for grokking across 384 configurations, finding data complexity—not model size—is the dominant factor in generalization onset, overturning common assumptions and providing actionable guidance for training dynamics research.