A research system at Physical Superintelligence using an Anthropic model provided a proof for the conjecture that quantum computers can be exponentially faster than classical ones.
The models formalized the proof of Fermat's Last Theorem in 30 hours, designed three electronic devices from scratch, and advanced robotics evaluations.
The new Muse Ember model costs one-third less than Claude Haiku 5.5 and GPT-6 Luna while matching their best scores on reasoning tasks. It is available via the Meta Model API at $0.10 per million input tokens.
Apple reportedly reduced the daily spending limit for some engineers from $300 to $200. The move highlights strict internal cost controls on generative AI usage.
While CAR T-cell therapies approved by the Alliance for Cancer Gene Therapy include Kymriah and Yescarta for blood cancers, they struggle with solid tumors and side effects like CRS and ICANS. Complex manufacturing and high costs limit the number of patients who can receive treatment.
Asahi Linux has implemented a policy prohibiting the use of generative AI in software development to protect software freedom and open-source integrity. The company warns that LLMs pose unique legal risks by potentially violating clean room reverse engineering requirements and contaminating code.
Physicists aboard China's Tiangong space station measured an equivalent temperature of 9.7 picokelvins, shattering the previous record in humanity's quest to approach absolute zero.
The Terminal2Shell design flaw causes terminal emulators to spawn a shell process immediately upon launch, executing any typed command with the full privileges of the logged-in user.
The method retains only the sign of the residual component, enabling a single BF16 vector instead of two FP32 states. When weights are quantized to 2 bits, optimizer state memory usage drops by 71–74% compared to AdamW.
Researchers from Queen's University Belfast demonstrated that agents can retrieve previously deleted context through subsequent prompts, creating a vulnerability for privacy attacks.
In China's Tiangong space station, a team cooled a cloud of 20,000 rubidium atoms to an energy equivalent of just 9.7 trillionths of a degree above absolute zero. This result surpassed the 2021 record of 38 picokelvin set by previous researchers.
Deletion of acsV, lpdA, cooC, and cooC2 halted growth in autotrophic conditions, while gcvH deletion reduced autotrophic efficiency and shifted carbon flux toward 2,3-butanediol.
Researchers developed a soft ionically conductive hydrogel using bioionic liquids that, in a traumatic brain injury model, cut reactive astrogliosis in half, enhanced cortical revascularization, and limited lesion cavity expansion.
A study of 2,615 patients found that metastasis frequency is higher near the gray-white matter junction and in areas with high cerebral perfusion, glucose metabolism, and mitochondrial density. At the cellular level, metastases were inversely related to astrocyte and microglia density.
The A2-MEAS framework uses code as a scaling interface to learn action corrections. A simple design with action pretraining reached a Q-score 2.21 times higher than the 2025 winner on the BEHAVIOR 2026 benchmark.
NeuroAge Therapeutics launched the Younger competition, where 500 participants must reduce their biological age over six months. Participants undergo full brain and body assessments, with baseline results due in February.
The company launched the Elysium Peptide Evidence and Evaluation Resource (PEER) to provide qualified research institutions with access to rigorously characterized peptides. The new initiative includes TIME-PEP, a study investigating epigenetic changes and biological age markers.
The team developed the SA-eGRASP genetic reporter to visualize individual dendritic spines that underwent long-term potentiation after contextual fear conditioning in mice.
The service is available for personal vehicle owners in cities like Los Angeles and Chicago. The network of 13,000 spots aims to reduce congestion and empty miles for autonomous vehicles.
Researchers analyzed 300 error reports from seven GPU kernel libraries, identifying risks of silent failures and inefficiency in large language model systems.