A startup founder says fine-tuning open-weight models on private data rivals frontier performance at a fraction of the cost, while a preprint warns safety guardrails can be stripped away — and a lab offers a poisoned-answer defence.
Catholic Health's voice agent handles two-thirds of MyChart calls and saves $375,000 a year. Presbyterian closed care gaps by text. Both started with the workflow, not the model.
Tool Retrieval, Gemini 4 Neon pricing, and Genie 4 all point the same way: Google wants the cost and plumbing of agent building to stop being the hard part.
Two in three enterprises would break if AI stopped, but only a third have a formal operating model. The gap is organisational, not technical — and business units are moving anyway.
Genie One accuracy jumps when curated agents sit on governed definitions, and Genie ZeroOps fixes pipelines only with a human sign-off. The pattern: control first, autonomy second.
Frontier models recreate at most 22.9% of software requirements from pixels alone, while static analysis guidance during generation can match the security of iterative repair. The benchmarks are getting honest — and the models are not yet good enough.
A preprint finds agent sessions hold environments for 15 minutes after a 33-second conversation, and three more papers attack the KV cache that grows underneath. The fix is sharing, not more hardware.
A self-improving system now holds the top rank on GPU kernel benchmarks, a framework turns papers into agents that beat coding assistants, and a training method that learns from failed attempts lifts professional-task scores.
In one study, 64.9% of agent rollouts that hit an error did something unsafe. In another, persistence alone pushed rule violations up. And a new incident log is collecting the cases.
The largest IPO in history closed its first day with a valuation near $2 trillion and demand seven times the shares on offer. The numbers underneath — $2.6 billion of revenue against $900 billion of future spending commitments — are the real story.
A study of five million Hugging Face records finds the EU AI Act barely mentioned. In California, a new law makes lawyers personally responsible for AI-written citations. And hiring models fail the law under pressure.
Small open models extract figures from SEC filings with accuracy ranging from 17% to 83%. A financial research agent gets 29% of calculations right on its own — and 100% when it uses Python.
Four preprints posted this week cover ligand design, autonomous electrosynthesis, structure identification from NMR, and lab resource tracking. Together they describe a lab that runs itself.
Three preprints posted this week trace capacity loss to reversible clumping, a hidden reaction route, and a degrading interface — and each points to a different fix.
Three preprints this week use genomic surveillance to show that resistant infections spread through long transmission chains, and that the mutations themselves arrive on a biased schedule.
Measles is climbing where exemptions cluster, typhoid fell years before its vaccine trial, and childhood TB looks rarer than it is. Three preprints show why the count alone misleads.
McDavid's season-opening point streak ended in Tampa, Ovechkin passed Backstrom for the Capitals' assists record, Stone set Vegas's points mark, and Catton scored his first NHL hat trick.
Indianapolis fell 26-3 to the Texans on Thursday night, managing 167 total yards and one third-down conversion. Daniel Jones was benched in the fourth quarter.
Eight entities were charged over fake crypto trading platforms that allegedly took more than $37 million, while two men reached final judgments in a $3.85 million historical-bond scheme.
Ohio put $28.6 million into 268 homes, New York broke ground on a 48-unit project in Rome, and Virginia's governor laid out a package to free up land and speed building.