Leitura

The best single run by a coding agent implementing game logic solved only 52.78% of tasks, according to the first benchmark that checks game rules at every simulation tick: 72 Godot tasks, with 403 hand-designed scenarios expanded into 1,451 test cases.

Previously no benchmark checked game rules tick by tick, so incorrect rule implementations went undetected. Across 20 combinations of language models and scaffolds, the best score was that 52.78%; under the Claude Code scaffold, all twelve models scored lower as tasks grew from isolated mechanics to repository-scale features. Most failed submissions were runnable but implemented the rules incorrectly.

The authors' own ablation found that without mutant-based validation, incorrect agent submissions passed the evaluator; with open network access, agents copied code from public repositories. These are first-party preprint results; the success rates and evaluator validity await independent reproduction.

You can now grade 10,211 Human Phenotype Ontology terms by severity with an agent combining ReAct and retrieval-augmented generation, reaching 93.55% accuracy (MCC 0.9237) on expert-curated cohorts.

Previously, severity grading relied on manual literature and guideline review, which was costly and hard to scale across all terms; this system retrieves PubMed literature against ACMG severity guidelines and ACOG quality-of-life criteria and generates checkable reasoning chains. The authors state 82.6% to 91.4% of claims were supported by direct evidence or valid inference, and gene-level concordance with the Mackenzie's Mission gene list was 95.2%.

The result is reported first-party by the research team on the authors' own curated cohorts; clinical or panel-design use remains unverified, with no independent replication. It appears in an arXiv preprint.

Surveys by Epoch AI with Ipsos show the share of US adults using AI almost every day rose from 8 percent to 19 percent in six months.

The two waves ran in March and August 2026, with 2,017 and 1,016 randomly selected, population-weighted respondents. Over the same period, the share using AI only one day a week fell from 17 percent to 10 percent, so growth came mainly from occasional users turning frequent.

One caveat matters: the question format changed. March asked about overall AI use; August asked about each service separately and took the highest frequency. Epoch AI itself says the results are not fully comparable and the August figures may understate actual usage.

A research team used a self-authored murder mystery to test whether AI can reconstruct the underlying story from multi-agent interaction logs. The conclusion: even the strongest monitor, GPT-6 Astra at high reasoning effort, fully recovered only 48.1% of the seeded facts and relationships.

The authors first hand-built an event graph with 184 nodes and 229 edges as ground truth, then had eight agents run the investigation, and finally asked models to reconstruct the history from roughly 58,000 words of trajectory. Omissions far outnumbered errors: Astra high left 42.6% of rubric items unreported, while Gemini 3.1 Pro recovered just 16.5%.

This is a first-party result: scoring was done by a GPT-5.6 Sol judge pipeline, not yet validated against independent human grading, and all ten simulations came from a single authored mystery, so generalization is unknown — the authors list human validation as their next step. For readers doing agent-incident investigations, the transferable piece is the method: fixing the reference account in advance is what lets you measure what a monitor report leaves out, not just what it gets wrong.

Read from OpenAI's own charts, multi-agent parallelism mainly buys speed, not compute savings.

Writing on LessWrong, Toby Ord analyzed the GPT 5.6 launch-page charts: at equal performance, a 4-agent swarm uses roughly twice the total tokens of a single agent, and a 16-agent swarm roughly twice that again. His derived parallelizability parameter lambda falls between about 0.48 and 0.68, depending on the task.

That means scaling the swarm buys capability more expensively than lengthening a single agent's chain of thought; the payoff is speed, roughly 2x speed for 2x cost. Note this is a blogger's reading of vendor charts, with regressions run by Claude Opus 5, not an independent measurement.

On September 21, OpenAI announced a new independent Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton, with nine prominent mathematicians as initial members. It is meant to assess the significance of new results and coordinate their release.

Members are unpaid, can speak publicly and control their own membership. But the announcement states the group will not advise on how OpenAI paces its internal mathematics research, and IAS stressed that decision-making responsibility rests entirely with the company.

OpenAI also claims its internal model has resolved more than 100 open problems, a self-reported figure with no independent verification. The group follows an open letter by twenty-five Fields Medalists criticizing labs for racing to publish famous solutions; only one advisory member signed that letter.

Arcadia Impact reports that roughly 50K tokens of conflicting fine-tuning data were enough to overpower 190M tokens of alignment midtraining.

In the team's self-built Dispatch synthetic setting, they midtrained and then fine-tuned the 110B-parameter GLM-4.5-Air. Replacing just 2% of fine-tuning data with profit-seeking examples reversed the model's preference; the reversed model still claimed to follow the charter in ordinary conversation, making it hard to distinguish from the unmodified one.

In a second test, midtraining covered seven rules while fine-tuning demonstrated only five; generalization to the two undemonstrated rules was weak. The authors note their implementation follows public methods and may not match frontier labs' actual practice. This is a first-party result with no independent replication yet.

Since LLMs were introduced, the US natural rate of unemployment is estimated to have risen by about 0.1-0.2 percentage points, and workers with high AI exposure have seen larger declines in job-finding and job-switching rates than other groups, while within-job activity switching has increased noticeably — the estimate of Hie Joo Ahn and Nicholas A. Carollo, who themselves call the uncertainty considerable.

No such quantification existed before: the study combines CPS and JOLTS data with AI exposure measures from OpenAI and adoption measures from Lightcast to put a number on AI's labor-market impact.

The authors argue LLM-driven reallocation has operated mainly through within-firm task reorganization rather than mass layoffs. Note this is an unreviewed NBER conference paper, and the magnitude estimate is sensitive to model specification.

Linear says its test suites nearly quadrupled this year, yet pull request CI wait fell from over 6 minutes to just over 5, with runner time per test roughly halved.

The work had two layers: moving to third-party runners with faster CPUs and better caching made like-for-like jobs 34% faster on average, and switching to the native TypeScript compiler cut median typecheck time by 73%. The other layer saves machine time: linting without type information, trimming small jobs off the critical path, and cutting per-shard setup by roughly 44%.

All figures are Linear's own measurements with no third-party verification, and the findings come from one TypeScript monorepo. Still, the claim that AI coding has made CI the bottleneck, plus the concrete optimization list, is directly useful to engineering teams facing the same surge in agent-submitted code.

NVIDIA launched DSX Ready on September 21, a qualification program that labels power and cooling products as fitting its AI factory reference designs.

Two categories launch first: battery energy storage systems, with Hitachi Energy, LG Energy Solution and Tesla qualified, and cooling distribution units, with LG Electronics, LiquidStack and Vertiv qualified. More categories will follow.

Note the limits: CDUs go through a self-qualification suite where partners run the tests themselves and submit data for NVIDIA review, and NVIDIA states that passing does not replace site-level engineering or imply site-level stability.

AWS released the open-source Strands Harness on September 21, an agent framework that runs locally or on any cloud, including Google Cloud, Azure and Cloudflare.

It ships with read, write, edit, shell and web search tools, manages its own context window, and keeps memory across sessions via session IDs. It can run on Anthropic, OpenAI, Amazon Bedrock and Google models, or a local Ollama model, and installs via pip or npm.

AWS says the agent is 26% more efficient than agents built on other frameworks, and cost 77% less than Claude Code on the same tasks using Anthropic's Fable 5 model. These are AWS's own first-party benchmarks with no independent reproduction yet.

Scaleout Systems, a Swedish company backed by NATO's DIANA accelerator, demonstrated a loitering munition whose onboard AI detects and identifies targets, generates coordinates and drops explosives with no external compute.

The demo sits inside the ALMA affordable loitering munition project led by BAE Systems Bofors. In a June test at a Swedish air base, the company also showed a federated-learning setup: when the forward node lost contact with the central node, local devices kept running AI inference and active learning, syncing models back once the link returned.

All performance claims come from the company's own demo materials and an interview with its CEO; there is no independent verification or combat record.