Source item

What I read this week... 1) Meta will cut roughly 600 roles across its AI division, which has sparked confusion about whether the company is pulling back from artificial intelligence just as the rest of Big Tech doubles down. The details tell a different story. The layoffs primarily affected Meta’s legacy Facebook Artificial Intelligence Research (FAIR) and infrastructure teams, which were consolidated under the new team, Meta Superintelligence Labs, that Mark assembled in June for $14.3B. Meta Superintelligence Labs is comprised of four groups: TBD Lab: The team managing Meta’s large language models, led by Alexandr Wang (prev. CEO of Scale AI). FAIR: A team focused on longer term research for AI led by Rob Fergus (Meta’s Director of AI Research). Products and Applied Research: The consumer integration team led by Nat Friedman (prev. CEO of Github). MSL Infra: The team for infrastructure to sustain artificial intelligence models that is led by Aparna Ramani (Meta’s VP of engineering) In contrast, hiring continues aggressively within its TBD Lab, which builds Meta’s next-generation foundation models and agent systems. Executives described the reorganization as a move toward smaller and faster execution from teams that can operate more like startups within the company. “By reducing the size of our team, fewer conversations will be required to make a decision,” - Alexandr Wang, CAIO 2) Starcloud is preparing to deploy AI-data-center capacity in orbit this November using NVIDIA H100 GPUs and solar-powered systems that exploit the unique advantages of space. The project aims to solve three key bottlenecks in current AI infrastructure. 1. Energy consumption. Training large models today demands tens of megawatts, and terrestrial grids plus cooling constraints create bottlenecks. Starcloud flips that dynamic by using abundant sunlight without having to worry about weather dynamics. 2. Cooling and heat-management. On Earth, data centers rely on fresh water for cooling through evaporation towers. In orbit, the vacuum of space becomes an infinite heat sink where heat is dissipated through large deployable cooling panels. 3. Data locality and latency for Earth-observation and satellite workloads. Starcloud plans early use-cases around processing sensor data (optical, hyperspectral, SAR) in orbit rather than down-linking terabytes of raw data to Earth. The key applications that will be unlocked by the above are analysis of Earth observation data, which could inform crop type detection and better predict local weather. The satellite will also speed up wildfire detection and distress-signal response when emergencies pop up. 3) Google announced a breakthrough quantum algorithm this week using its new Willow processor. The company demonstrated a “verifiable quantum advantage” with an algorithm called Quantum Echoes that ran 13,000 times faster than one of the world’s top classical supercomputers. The research was published in Nature and improves upon their 2019 study that only sampled bitstrings from a highly chaotic quantum state of qubits. The new experiment modeled how information spreads in a quantum system using a 105-qubit superconducting chip with some of the highest fidelity rates ever recorded. This shift from “faster” to “verified” is what makes the Willow milestone distinct. It establishes a baseline for scaling future quantum processors that are both powerful and testable by other advanced quantum chips. Simulating molecules and materials is one of the earliest real applications for quantum systems, with potential breakthroughs in battery chemistry, drug design, and clean energy catalysts. Links to sources and other articles/X I read are on substack: https://t.co/Kx1AwDTeBO