#数据/训练

Our paper on Subliminal Learning was just published in Nature! Last July we released our preprint. It showed that LLMs can transmit traits (e.g. liking owls) through data that is unrelated to that trait (numbers that appear meaningless). What’s new?🧵 https://t.co/Iiv9sgjJki

this is INSANE! the world’s first colosseum built for ENHANCED athletes. every competitor FULLY enhanced. every dose tracked. every compound clinically approved.​​​​​​​​​​​​​​​​ $1.2B backing before a single event. built from scratch. who’s not tuning in to watch humans push past every limit we thought existed? Las Vegas. May 24

Demand for human experts training AI is exploding. Handshake’s gross annualized AI training revenue rose to nearly $1B from $550M in January and from just $5M to $10M a year ago. https://t.co/li5aLNZDeC

Stanford published a 423 pages long 2026 AI index report. Model progress is not flattening, with top systems now hitting or beating human baselines on some hard coding, science, and math tests, while some older benchmarks are getting less useful because frontier models are bunching together near the top. The U.S.-China gap at the model frontier is now very small, with the U.S. still ahead in top model count and private investment, while China leads in papers, citations, patent volume, and much of the industrial deployment base. AI adoption is moving unusually fast, reaching about 53% global population-level use in 3 years, with strong business uptake and a large estimated consumer surplus even when tools are free. Responsible AI is lagging capability, with documented AI incidents rising to 362, weak safety reporting, and real tradeoffs where improving one safety dimension can hurt another metric. The hardware story is just as important as the model story, because AI now depends on a fragile stack centered on data centers, huge energy use, and a chip supply chain heavily tied to TSMC. The labor signal is mixed, because AI is producing real productivity gains in fields like support and coding while some entry-level software jobs are already weakening. The most practical near-term win may be medicine, where AI note-taking tools are reducing clinician paperwork and burnout, even though strong real-world clinical evidence is still thin for many systems.

Robotic data is insanely expensive and brutal to collect. It’s the only thing holding back general-purpose robots right now. Figure CEO @adcock_brett : "If we could get a pile of data in the helix stack, we would solve general robotics right now." https://t.co/YWpA20gdn5

Indian factory workers wear head-mounted cameras to capture data for training robotics AI models. This image captures a blunt truth about robotics: teaching a machine to move in the real world is still painfully expensive. What looks dystopian at first is also a clue about the bottleneck. Robots do not learn useful physical behavior from internet-scale text the way language models do. They need embodied data: hands reaching, wrists turning, objects slipping, fabric folding, tools resisting, people recovering from small mistakes in real time. That data is rare because reality is slow, messy, and costly. A robot fleet is expensive to buy, expensive to maintain, hard to supervise, and dangerous to scale in uncontrolled settings. Even teleoperation is costly, because every minute of human-guided movement requires hardware, operators, calibration, and failure recovery. So companies go looking for the cheapest possible proxy for physical intelligence. First-person video from factory workers is not the same as robot action data, but it can still be valuable because it captures sequencing, posture, bimanual coordination, and the micro-adjustments that make real work look easy. The frontier in robotics is not just better models. It is better pipelines for collecting reality itself. That is why warehouses, factories, kitchens, and repair benches matter so much: they are dense environments of repeated contact with the physical world, which is exactly what robots lack. The unsettling part is that this turns human labor into training infrastructure twice over, first as work, then as data. And until embodied data becomes cheaper to gather than human motion is to record, robotics will keep learning from workers before it fully replaces them.

Do all reasoning tokens matter equally? We study the functional importance of reasoning tokens implicitly encoded in LLMs. Work w/ my cool advisor @dilekhakkanitur at @convai_uiuc; to appear in ACL Main 2026. #ACL2026 Paper: https://arxiv.org/pdf/2601.03066 Demo: https://greedy-token-pruner.github.io https://t.co/1ZUpLhQmPj

Platform Engineer - Benchmark Lead ARC Prize Foundation is hiring a senior engineer to build our benchmark platform * Expand ARC-AGI-3 * Own ARC-AGI-4 * Lay the foundations for ARC-AGI-5 Come build the benchmark that defines progress toward AGI $7.5K referral bonus https://t.co/hEm9NafkQP

http://x.com/i/article/2041371538482761728

Here is a quick start script including the setup, technical details, and a candid look at where Kinetic excels versus its current limitations 🪡 https://github.com/jigyasa-grover/kinetic-finetuning-on-cloud-tpu

Fine-Tuning Gemma 2B on PubMedQA: Building a Medical Q&A Assistant with LoRA, Keras Kinetic, and Cloud TPU https://kuanhoong.medium.com/fine-tuning-gemma-2b-on-pubmedqa-building-a-medical-q-a-assistant-with-lora-keras-kinetic-and-0e6eaf566834 #TPUSprint https://t.co/h3Olhekb79