Physical AI

Researchers have introduced a physics-grounded artificial intelligence framework designed to make predictions about new materials more testable, according to Phys.org. By anchoring AI models in physical laws, the approach aims to produce forecasts that scientists can verify experimentally, addressing a common weakness in machine learning predictions for materials discovery.

Zenithon has raised $10 million in funding to build AI world models capable of simulating extreme physics. The startup says such models could help researchers and engineers understand conditions that are difficult or dangerous to test in the real world, from high-energy environments to advanced materials. Details on investors and product plans remain limited so far.

Cerebras Systems CEO Andrew Feldman is speaking at TechCrunch Disrupt 2026 about whether AI models can continue scaling at their current pace. The session tackles a central question for the AI industry: whether gains in compute, chips and training methods can keep driving progress or whether the field is approaching physical and economic limits. Feldman's company builds wafer-scale chips aimed directly at accelerating large AI workloads, making him a notable voice on the debate.

A new AI framework for materials discovery has been introduced that incorporates physics principles directly into its design. The approach aims to produce more reliable and trustworthy predictions than purely data-driven models, a persistent weakness in computational materials science. Researchers in the field see physics-informed machine learning as a promising route toward faster, more credible discovery of new materials.

Industry analysts say physical AI — artificial intelligence systems that interact with the real world, such as robots and autonomous vehicles — is reshaping how work is divided across the semiconductor industry. Chipmakers are reassessing which players design, manufacture and supply specialized hardware as demand shifts toward chips capable of real-world sensing and control, potentially redistributing roles among designers, foundries and automotive suppliers.

Zenithon AI is developing simulation technology it claims can model extreme physics up to a million times faster than existing methods. The claim was flagged on the startup data platform Dealroom, which tracks venture-backed companies. If accurate, such speedups could matter for fields like aerospace, energy and materials research, where high-fidelity physics simulation is a major bottleneck.

A robotics startup has taken a large sublease from Airbnb, according to a report by The Real Deal. The deal highlights growing demand for office space from companies working on physical AI, as the race to build embodied AI systems accelerates and robotics firms expand operations.

Zenithon has raised $10 million to develop world models designed to simulate extreme physics scenarios. The funding will support the company's work on AI systems that model physical phenomena under conditions difficult or dangerous to test in the real world. Details on the investors and specific applications of the technology have not yet been made public.

Attention is turning to the robot hand as a central challenge in physical intelligence, the push to give machines the ability to perceive and manipulate the physical world. Dexterous manipulation is widely seen as a key test of whether AI systems can move beyond software and into real-world tasks, drawing interest across the robotics field.

General Intuition, a startup working on world models, has closed a $220 million investment round, according to a report by SiliconANGLE. The funding makes the young company one of the notable new entrants in the increasingly crowded field of AI firms building models that understand and simulate physical environments. Details on the investors and valuation have not yet been made public.

Anthropic, the AI company behind the Claude models, is reported to be launching or operating a wet laboratory, prompting discussion about what this means for the pharmaceutical industry. The move suggests AI firms may move from software into hands-on biology and drug discovery. Commenters are weighing whether AI companies building physical lab capabilities could accelerate drug development or disrupt traditional pharma research models.

Ten AI agents running Anthropic's Claude model have jointly produced a 17,895-line proof addressing a physics problem first posed in 1904. The feat highlights how multi-agent AI systems can now tackle long-standing scientific and mathematical work, collaborating by dividing the reasoning across parallel agents. Commenters are weighing what this means for the future of research, both as a demonstration of AI's growing capability and as a warning about verifying machine-generated proofs of such length.

Harvard University reports that a custom-built AI tutor tested in its physics courses produced learning gains roughly double those of traditional instruction, achieved in under an hour of use. The study is drawing attention from educators debating whether AI tutors can match or outperform classroom teaching, and how universities should integrate generative AI into science education.