Search “Reasoning”

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Multiple teams train LLMs on historical data to test whether AI can reproduce general relativity

According to a recent report in Nature, since the beginning of this year multiple teams have attempted to build retro LLMs whose training data is cut off before a certain historical point in time, to test whether the models can rediscover later scientific ideas. The results show that AI is still a considerable distance from Einstein-style scientific breakthroughs. Independent researcher Michael Hera trained Machina Mirabilis on pre-1900 data; when faced with the photoelectric effect, the model once proposed that light would break down into a large number of mutually independent pulses, but in most cases it failed and was merely imitating historical texts. Google DeepMind researcher Tom Zahavi points out in the paper "LLMs Don't Leap" that current LLMs cannot yet complete a reasoning leap like Einstein's, and that the abductive reasoning on which scientific breakthroughs depend often occurs on anomalous phenomena with very little data or in conflict with existing theories.

OpenAI Team Responds to Computer Use Progress Doubts, Recaps One-Week Sprint on Decisions API

In a podcast after OpenAI DevDay, Ari Weinstein, who is in charge of the Computer Use agent, responded to Dwarkesh's doubts about Computer Use progressing slowly, saying it is now completely different from a few months ago because the model has become better at debugging and retrying, and can write JavaScript to execute multiple steps at once while combining screenshots, accessibility, DOM, and Playwright. He also mentioned that Dots provides each Dot with a cloud Linux virtual computer, and that Computer Use has entered the Agents API. Nikunj Handa, head of the OpenAI API team, introduced updates to GPT-6 such as asynchronous tool calling, mid-turn steering, WebSockets, UltraFast inference, prompt caching and pre-warming, and compaction, and said the Decisions API is based on Luna weights, with no new model trained, and took only about a week from the emergence of the idea to a prototype.

English translation is being prepared.

After Nvidia chips became hard to buy, what keeps China's large models running?

A Huxiu analysis points out that domestic computing power can now cover large model operation: Zhipu's domestic-chip inference cluster has reached a scale of 100,000 cards, and Meituan's LongCat-2.0 has also completed training on domestic computing power. However, Ascend 910C achieves an MFU of about 30%-35% in trillion-parameter post-training, lower than H100's 40%-50%. The article breaks computing efficiency into three variables—chips, software, and models—and compares the four paths of Zhipu, MiniMax, DeepSeek, and Kimi, concluding that Zhipu has the highest short-term certainty, while DeepSeek has the greatest medium-term elasticity. The author proposes three tracking indicators—capacity, efficiency, and cost—and argues that only when all three improve simultaneously can domestic computing power shift from a backup option to a cost advantage.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Nathan Lambert: AI Safety Is Underprepared, as Seen from the Frontier Model Hacking Incident

Nathan Lambert published an article in Interconnects reviewing the OpenAI-HuggingFace hacking incident, arguing that the AI industry as a whole is severely underprepared for the risks of the next 12-24 months. He noted that OpenAI's models are stronger than Claude in inference persistence, possibly because OpenAI has bet more heavily on inference-time scaling, and that such models that continuously pursue goals deserve more safety attention. He also argued that public models and open research are irreplaceable for understanding frontier risks; currently open models lag by only 3-9 months, and restricting open models would instead weaken society's ability to respond.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.