Section
Models, research breakthroughs, and the infrastructure behind machine intelligence.
12 articles
Loading…
AWS added a charge per agent action on top of token costs, and companies that built multi-step workflows are finding their projected bills doubled overnight.
The document covers capability benchmarks, red team findings, and refusal behavior. It omits training data description and pre-deployment testing timelines, which is where critics focused.
Fine-tuning returns to OpenAI's current model generation with transparent training costs and no hosting fee, which changes the calculation against open weight alternatives.
After struggling to compete on consumer-facing capability, Cohere rebuilt around on-premises and national deployments and now counts four government customers and eleven banks.
The team behind Stable Audio leaves as an independent company with licensing deals already signed with two major labels, an arrangement its former parent could never negotiate.
A steeper batch discount plus a 12-hour completion window targets the document processing workloads that have been running on cheaper models for cost reasons alone.
The headline model is incremental. The 3 billion parameter variant designed to run on the Ray-Ban Display's coprocessor is the release that changes what wearables can do offline.
Automated scraping by agent frameworks pushed Hugging Face egress past sustainable levels. Anonymous downloads now cap at 50 gigabytes per day per address, and the community is split.
NIM microservices can now run in air-gapped environments with an offline entitlement file, removing the phone-home requirement that blocked deployments at defense and health customers.
