SubQ LLM 12M-token context and Google's $3.2B Anthropic bet

Researchers discovered that LLMs consistently generate the same fictional name pairs, like Elena Vasquez and Marcus Chen, and embed them across thousands of AI‑crafted documents. These correlated priors act as model‑specific fingerprints, even producing fake scholarly articles with real DOIs, exposing a new avenue for misinformation and provenance tracking.

The authors model multi‑agent LLM deliberation as a closed‑loop dynamical system where each agent carries a hidden internal belief, a “anchor”, that persists across rounds. When an anchor sits far from the initial opinions, it can drag the group’s confidence beyond the convex hull of starting answers, amplifying errors instead of correcting them. Designers of LLM‑based deliberation must therefore account for anchoring effects to avoid misleading consensus.

Subquadratic unveiled SubQ, the first commercial LLM that abandons the quadratic attention bottleneck. Its 12‑million‑token context window claims 52× faster inference at 1 M tokens and up to 1 000× compute savings at full length, potentially slashing costs for long‑context applications. If validated, it could render many RAG pipelines obsolete.

Google has staked $3.2 billion on the Lake Mariner data‑center in western New York, where it will lease thousands of its Tensor Processing Units to Anthropic. By mirroring Nvidia’s model of financial guarantees and compute‑as‑a‑service, Alphabet is turning its custom chips into a commercial product line that could reshape AI infrastructure markets.

Midjourney’s new medical division unveils an “Ultrasonic CT” that scans the entire body in about 60 seconds using sound and water, not radiation or magnets. The company plans 50,000 units globally and a spa‑style health center in San Francisco, positioning the device as a faster, cheaper alternative to MRI.

DeepSeek demanded that any investor in its $7.4 billion fundraising round sign a no‑poaching clause, barring them from hiring its staff or encouraging spin‑outs. The stipulation highlights how fierce the talent war has become among China’s AI giants, with firms scrambling to lock down scarce expertise.

MosaicLeaks shows that agents that blend private documents with web searches can reconstruct confidential facts from query logs, leaking up to 34% of information. The proposed privacy‑aware RL training, PA‑DR, raises correct multi‑hop reasoning to 58.7% while dropping full‑information leakage to under 10%.