Cognition AI used always-on AI agent infrastructure to connect software development work with continuous inference, feedback, and training so Devin can improve from production outcomes. The effort targets 99.99% reliability for distributed training across thousands of GPUs, and Nvidia’s early access to the Vera Rubin platform supports better price-performance research. CoreWeave Forge adds tracing, distillation, and RL rollout services that can hot-load updated checkpoints into live deployments, changing the workflow from occasional retraining to an ongoing learning loop.
Olmo released Olmo-core 3, an open training framework with a redesigned mixture-of-experts system for scaling large language-model training. In one benchmark, it increased the expert pool from 8 to 128 while keeping about 3.2B active parameters per token and limiting throughput drop to under 5%. It replaces an earlier FSDP-based MoE approach with a DDP-style expert-resident training stack (and other routing/precision optimizations) to improve throughput and enable scaling up to trillion-parameter MoE configurations.
Clef announced two open-source decision models, Clef and Clef-flash, hosted on Workers AI for fast classification and agentic workflows. The reinforcement learning platform lets developers fine-tune these decision models using their own data. This adds both prebuilt decision-model options and an RL fine-tuning path for custom workflows.
Olmo-core 3 was released as an upgraded open training framework for large mixture-of-experts (MoE) language models. In benchmarks, increasing the expert pool from 8 to 128 while keeping 4 experts selected per token held active parameters per token to about 3.2B and reduced training throughput by less than 5%. The update changes the training stack by switching from FSDP to a DDP-based approach that keeps experts on GPUs and adds routing, throughput, and MXFP8 optimizations to scale efficiently toward trillion-parameter MoEs.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
