Science/DeepMind researchers introduced SynthID Bio, a watermarking method that embeds a verifiable signature into synthetic biology sequences so it can be detected on physical proteins after synthesis. In wet-lab tests on 3 target proteins (VEGF-A, SARS-CoV-2 spike RBD, and PD-L1), the watermarked designs matched the hit rate, binding affinity (KD), and natural sequence diversity of unwatermarked versions. As a result, DNA synthesis providers and public protein databases can potentially use the embedded signal to verify provenance and reduce issues from AI-designed or mislabeled biological entries.
Zenithon raised $10 million to develop world models for extreme physics, aiming to predict outcomes across systems like fusion reactors, rockets, and semiconductor manufacturing. The funding will be split between recruitment and compute. Zenithon plans to release new models every three months and grow its team in San Francisco and across the US.
NVIDIA, MIT, and Oxford introduced Physis-Lang, a framework that uses physics-focused captions as a shared representation for video world models. On the Physics-IQ Verified leaderboard snapshot dated September 29, 2026, Cosmos3-Super ranked first with 48.2 ± 1.4, beating Veo 3.1. The approach adds physics reasoning and negative prompts, then uses a caption-evolution loop and targeted data curation to improve physical correctness.
Zenithon, a London-based AI lab building world models for extreme physics, raised new funding from Backed, Lunar, Seraphim, MMC and SOSV. The round totals $10m. The additional capital is expected to expand the lab’s work on those world-models projects.
General Intuition closed a $220 million funding round to advance its world-model AI that uses synthetic footage for training. It was valued at $6.2 billion. The company will hire more AI researchers and test a commercial version with a limited set of customers.
SCLATE was introduced as an execution substrate that lets continual-learning agent benchmarks and unmodified agents share one event scheduler, while also acting as a rollout engine that records model-call tokens and log probabilities. It compresses a month-long scenario into hours by using a hybrid simulated clock, and post-training Qwen3.5-4B cuts file-line reading by 6.8× while raising SWE-bench Verified by 16.7 points. As a result, the work enables head-to-head continual-learning evaluations and finds added memory systems do not consistently outperform native harness memory, even as a model can learn to use both harness and memory effectively.
A systematic study evaluates multiple large-language-model conditioning methods for injecting or removing a target concept and compares them across both effectiveness and generation quality. The paper reports that efficient steering methods often achieve conditioning but at a steep cost to fluency, and activation steering is far less effective on instruction-tuned models than on base counterparts. As a result, it recommends simpler prompting or supervised fine-tuning for concept injection, notes they are weaker for concept removal, and uses cheaply computed text metrics that track LLM-judge scores to guide method choice.
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.
