Harvey AI Corp. raised $550 million at a $15.5 billion valuation to expand its cloud platform of AI tools for legal teams. The round followed 6 months after its last nine-figure funding and came from Diffusion and Lightspeed Venture Partners, joined by backers including Sequoia, Kleiner Perkins, and Goldman Sachs. The company plans to deploy more computing infrastructure for AI research, especially “new generalist models,” and it will use the funding to support custom large language model development and broader legal benchmarking via LAB.
Hyper-τ-bench evaluated whether AI developer agents can build customer-service agents from business materials, and Sierra reports that the best autonomous setup still fell below 25% test success.
Alibaba.com’s Accio team released CommerceAgentBench, an open-source benchmark that grades AI agents by end-to-end commercial outcomes rather than model reasoning. The benchmark includes 107 tasks built from real e-commerce work and the best model completed 61.7% of them. The results shift deployments toward “precision delegation,” where businesses hand off only the workflows with high pass rates and keep humans checking the low-performing exception areas.
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.
The article reviews OpenAI’s GPT-6 Astra and explains how looped transformers (recurrent depth) could relate to claims about hiding reasoning traces. GPT-6 Astra is reported to score 99.9% on the ARC-AGI-3 benchmark. It argues that recent improvements are tied mainly to computer-use and RL-with-verifiable-rewards training, while the looped-transformer idea is positioned as an architectural detail to evaluate alongside those reasoning-trace claims.
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.
