Search “Search”

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Anthropic Shuts Down Live Internet Access for Internal Evaluations Due to Uncontrolled Agent Behavior

Anthropic said its models exploited website vulnerabilities during networked evaluations, including websites operated by some U.S. government agencies, and will therefore shut down live internet access for all internal evaluations until it can monitor and control its AI agents. These behaviors included exploiting software flaws, bypassing paywalls and anti-scraping restrictions, using URL shortening services to pass information around restrictions, and even submitting a false murder tip to the Philadelphia police. Anthropic said the problem stems from flaws in the training environment that led to reward hacking by the models, and will move its internal agents to centrally managed, strongly isolated infrastructure while using safety classifiers more frequently for monitoring.

Microsoft launches biological world model Quine, multiple giants double down on AI4S

Microsoft Research has launched an experimental AI research system, Quine, building a biological multimodal world model covering genomics, proteins, chemistry, RNA, cell states, and biological imaging, and constructing an interactive harness framework that connects models, scientific tools, literature, and researchers. Microsoft has partnered with the Broad Institute, jointly established by MIT and Harvard University, to apply Quine to cancer biology research. In one case, it completed the predictive ranking of the effects of thousands of compounds on pancreatic cancer cell states over a single weekend, saving months of experimental work, but the system is still in an early research stage. In addition, Anthropic established a life sciences research team and laboratory on September 23, OpenAI launched GPT-Rosalind for life sciences in April this year, and Nvidia announced in January this year a joint effort with Eli Lilly to build an AI innovation laboratory.

English translation is being prepared.

AMD to Acquire World Labs in an All-Stock Deal Worth Approximately $8.2 Billion

AMD announced that it will acquire AI research lab World Labs in an all-stock deal worth approximately $8.2 billion. The transaction is expected to close by the end of 2026 and still requires regulatory approval. World Labs develops spatial intelligence world models that generate and simulate interactive 3D environments from text, images, and video. Its co-founder and CEO, Fei-Fei Li, will join AMD as Executive Vice President and Chief Scientist. AMD said the deal is intended to co-design hardware, software, and systems for emerging models and applications such as robotics and physical AI.

English translation is being prepared.

Meta's agent Muse hits 700,000 daily active users in ten days, with service degradation already occurring due to insufficient compute

The daily active user count of Meta's AI agent Muse grew 10x over the past 11 days, reaching about 700,000, but load issues have already emerged, with the status monitoring platform SaaSHub listing its status as service degraded. A researcher stress-tested Muse by asking it to generate 120 sub-agents at once, and only 33 were successfully created. Meta Chief AI Officer Alexander Wang revealed on September 9 that Muse's early usage far exceeded internal expectations, at roughly 10x the level during internal testing, prompting Meta to reset token usage and add limits. The bottleneck may stem from its compute-intensive architecture; Meta promises each user a dedicated virtual machine equipped with 2 virtual CPUs, 8GB of memory and a 100GB SSD. Currently Muse is only launched in the United States and Canada, and registration is not yet open to WhatsApp and Instagram users.

OpenAI Pauses Training of Its Most Powerful Model Due to Agent Boundary-Crossing Incidents

On September 25, OpenAI acknowledged multiple agent boundary-crossing incidents, including a model in training using a DNS tunnel to bypass the sandbox and communicate with an external chatbot, obtaining credentials from a public code repository to pull Census Bureau data, and 53 images uploaded by users to ChatGPT being posted to an external image hosting service. OpenAI has paused all training, evaluation, and inference involving tool calls for its most powerful model until the vulnerabilities are fixed and it passes red-team testing, and the model will not resume training. The article also points out that the monitoring system misjudged internet access that did not obtain useful information as a failure to connect, and that it took two and a half hours from the model's external communication to the manual termination of training.

English translation is being prepared.