GPT-OSS-20B was fine-tuned for multilingual reasoning on Amazon SageMaker HyperPod after the base model mostly ignored non-English reasoning-language instructions. On the 50-case eval, reasoning-language correctness rose from 22% (11/50) to 100% (50/50) after SFT, and GRPO further increased answer-language correctness from 62% to 70% (35/50). This shifts the model from defaulting to English mid-thought toward consistently reasoning and answering in the requested language for French, German, Spanish, and Italian.
Meta has been ordered to identify executives who authorized torrent downloading tied to copyright disputes in AI training cases. The discovery order was issued by Magistrate Judge Thomas Hixon, and it forces Meta to clearly name key decision-makers tied to access to shadow libraries like Anna’s Archive and LibGen. As a result, Meta’s discovery process may expand beyond staff to include a broader chain of supervisors and leadership involved in early AI training dataset work.
Eka Robotics launched a physical AI manipulation system and completed a $150 million Series A that includes a force-sensitive gripper and a vision-force-action deep-learning model for dexterous grasping.
Vectorial AI raised $3 million in seed funding for an AI platform that simulates human behavior to produce feedback for consumer research. The platform is trained on a database of more than 2.5 million real consumers and claims its SAPIENS model is 30% more accurate than general-purpose models. This funding expands its behavioral dataset and roadmap toward SAPIENS 2.0 while pushing simulated audiences into companies’ product development workflows instead of relying on real-person surveys and interviews.
DeepMind’s DiscoRL uses a neural network to represent a reinforcement-learning update rule and searches for rules that make new agents learn better from experience. Disco57, a learned rule found on Atari57, was applied to train fresh agents on ProcGen and DMLab-30 without additional meta-training. The work shifts learning from a hand-designed update procedure to a discovered one, enabling zero-shot transfer of the rule while highlighting the need to show continued gains under real compute and evaluation costs.
Metis Reasoning, a London AI company co-founded by former Google DeepMind lead Thore Graepel, is launching with a focus on compute-efficient AI reasoning rather than frontier large language models. Metis is initially raising tens of millions of dollars, and it claims results on frontier reasoning problems at a fraction of the cost of frontier models. It aims to improve AI reliability and verification via a more verifiable reasoning approach, with no set date given for when its first public model will launch.
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.
