The challengers to the leading GPU stack are competing on software, not silicon
AMD shipped ROCm 10.2.0 on a six-week cadence, Microsoft and Intel deepened their Azure work, and a new invocation layer cut latency on AMD hardware.
29 stories, 120 Manifests cited.
AMD shipped ROCm 10.2.0 on a six-week cadence, Microsoft and Intel deepened their Azure work, and a new invocation layer cut latency on AMD hardware.
The best model on a new abstention benchmark scores only 59.5%. Other preprints show agents can skip most helper messages, and that personal-agent updates can be repaired after the fact.
One system lifts end-to-end success from 16.0% to 76.0% by scoring memory on utility. Another improves accuracy by 10.9 points while cutting cost. The catch: confidence-based retrieval only pays off when budgets are tiny.
One preprint finds the inference backend alone can move a score by 16.6 points. Three more show the same fragility in reliability, monitoring and agent evaluation.
Arda says compute and experimental throughput are the bottleneck, Nvidia says synthetic data is 45 times cheaper than road collection, and Horizon and Momenta show the models and the machines arriving together.
Four preprints posted this week attack the same bottleneck from different sides: the memory that holds the model's working state.
A vision model that runs faster than a datacenter GPU, a simulator 135 times faster than real hardware, and Gemini on low-memory phones all point the same way.
New serving work attacks the cost of running models from four directions at once: electricity, fairness, batching and automatic model swaps.
Quantization designed by agents, expert models stitched into one without training, and speculative decoding that finally uses its full budget.
One preprint's defenders reach near-zero attack success on static attacks. Another finds a model executing image-borne commands 77.96% of the time. The gap between the two is the story.
A scouting agent matches the best single model at about a fifth of the cost, and a timeout cascade cuts GPU costs by a third. The harness itself buys far less than its price suggests.
In one new audit, 72.99% of trading points show at least one gap. In another, every model leans left-progressive on US content. A third finds expert-level science is not the same as winning tasks.
Snowflake cut AI costs 16% with per-user quotas after one engineer's bots spent $1,100 in a day. UiPath now offers fixed-spend billing, and Meta says its agents saved $45 million in compute.
Oracle launched AI agents for clinical trials and construction planning. Salesforce admins are trusting agents with inbox triage but not bank accounts. The Premier League generated 13 million personalised summaries a week.
Four preprints posted this week agree on one thing: the model is the least unreliable part of an agent. The runtime around it is where trust breaks.
An enzyme that can be switched on and reshaped, a theory that sorts kinetic regimes by one rate, and odor models that don't need 3D shape — three preprints, one lesson.
Spacer fit decides a perovskite's phase, light intensity sets a polymer brush's thickness, and one amino acid steers gold into chiral fibers. Structure is the dial.
MAS issued tokenised bills natively on its test network, published a collateral framework, and opened 2027 trials. Denmark is building the same plumbing for euros and kroner.
Almost all FOMC participants said more hikes would likely be appropriate. PIMCO expects two or three more. Even the ECB is talking about tightening.
82% of AI-investing leaders worry about token costs. The answer emerging this week: budgets, audit trails, and agents that cannot improvise.
Reported cyber incidents jumped 52% above the previous peak in October. The regulators' answer is guidance on AI and quantum risk, not new rules.
980 million shares, trading under ANT from 19 November. The pricing comes as investors scrutinise AI spending and the Fed turns hawkish.
Betaworks backed nine agent startups, Jump Capital led a $41 million round for AI permitting, and agent orders jumped 229% in a month.
Flu, cancer, liver disease and H. pylori all split the same way this week: the headline number hides the group moving the other way.
Plasma assays wobble when tracking one person over time, while other brain-health numbers only make sense in context.
Nineteen unranked prospects earned 40-man spots before the Rule 5 draft, while the same roster squeeze pushed veterans off. Performance beat ranking.
Oklahoma City lost at home for the first time, the Lakers' nine-game run ended, and Milwaukee fell to 1-14. The early season is being decided by streaks.
The FCA proposed new cyber guidance and launched an AI observatory the same week the OCC pointed to lighter bank rules and the White House called recent outages routine patching.
Pennsylvania doubled a rail line, New York approved homes on MTA land and a station rebuild, and Virginia put transit-oriented development in its housing agenda.