Databricks expanded its Adaptive Instructed-Retriever retrieval model to speed up multi-step evidence gathering for AI agents like Genie Code, Genie One, and Genie Agents. The company says it responds twice as fast as Claude Sonnet 5, GPT-5.6 Luna, and V4-Flash while matching their retrieval quality on seven benchmarks. Developers can now cap sequential search steps and let the model stop early when evidence is sufficient, so faster answers are possible without extra retrieval turns.
Luum deployed eyelash-extension robots built on a safety-focused wearable-robot background and began running them in salons and select Nordstrom and Ulta stores, while CEO Jo Lawson said the system is being trained to shorten visits. Luum’s setup trains on about 150 terabytes of session data. The service shifts from a human-by-human process to faster, more automated lash work where the robot can handle both eyes at once, cutting appointment time toward about 30 minutes.
IFM shipped K2 Horizon, a set of six “fully open” foundation models meant to publish the full training lifecycle beyond just model weights. The model sizes range from 0.9 billion to 375 billion parameters. At launch, some training data, code, or checkpoints were delayed or incomplete for parts of the fleet (including a Stage 1 32B checkpoint), and developers questioned whether the releases enable end-to-end reproducibility and full inspection (including missing compute details).
Researchers compared pretraining recipes and data corpora from 2019 to 2025 and found more of the compute-efficiency gains came from data than from model changes. Data improvements contributed 3.24x more compute efficiency gains than model improvements (12.0x for data vs 3.7x for models) at a 1e19 FLOPs budget. The study concludes data and model improvements are mostly independent additively (88% of OLMES score variance explained), so progress is increasingly driven by better data engineering rather than interacting with specific model tweaks.
Goodfire used Ai2’s fully open post-training stack to predict behavioral shifts from preference training, then traced observed safety regressions to specific preference examples. The work tied a compliance increase to Dolci preference dataset examples used to train Olmo 3. The approach enabled targeted training changes that reduced the regression while preserving broader capability gains.
OpenAI-linked accounts circulated a claim that an AI-assisted system produced a Navier–Stokes result related to the Millennium Problem using multi-agent collaboration. The process was described as involving about 10,000 agents, trained over roughly 1 year and using about 130B tokens (over $40M). The public discussion shifted toward whether large-scale inference-time compute and agentic coordination can contribute to hard science, while stressing that no theorem or formal verification material was provided for independent confirmation.
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.
