←Search results

Search “Anthropic”

By relevance · Total 1353 results Clear search

Monday

Tuesday

  3. Rohan Paul

    Anthropic in talks for a second $45 billion compute deal with Nscale, West Virginia reservation doubles to 920 MW

    Anthropic is in talks with Nscale for a second $45 billion compute agreement that would double its West Virginia reservation to 920 MW, adding another 460 MW building at the Monarch Compute Campus running Nvidia's Vera Rubin chips; Anthropic also holds a right of first refusal on the campus's expected third building. The first contract was signed in August, a six-year 460 MW lease with delivery starting at the end of 2027, averaging about $7.5 billion a year, or roughly $16.3 million per MW. Nscale's IPO filing lists the agreement with Anthropic at a value of up to about $44.6 billion, and the value of the second deal depends on whether the first Monarch building and its Vera Rubin racks can be powered on schedule by the end of 2027.

    Anthropic is in talks with Nscale for a second $45 billion compute agreement that would double its West Virginia reservation to 920 MW, adding another 460 MW building at the Monarch Compute Campus running Nvidia's Vera Rubin chips; Anthropic also holds a right of first refusal on the campus's expected third building. The first contract was signed in August, a six-year 460 MW lease with delivery starting at the end of 2027, averaging about $7.5 billion a year, or roughly $16.3 million per MW. Nsc

    Matched fieldsTitle

    #ai-news#Industry news#Anthropic#English text is being prepared.
  4. Rohan Paul

    OpenAI seeks to raise about $30 billion at a valuation of over $1.4 trillion, aiming to close before Anthropic's IPO

    According to Reuters, OpenAI is seeking to raise about $30 billion at a valuation of more than $1.4 trillion before Anthropic's IPO, a valuation at least 64% higher than the $852 billion it reached in March. The two companies are competing for overlapping long-term funds and sovereign wealth funds, and OpenAI hopes to lock in commitments before Anthropic files its S-1 and goes public in mid-November, to avoid direct comparison. Anthropic's third-quarter revenue grew 45% to $16.8 billion, a slowdown from more than doubling in the second quarter, leading some investors to argue for a valuation of $1.5 trillion rather than $2 trillion.

    According to Reuters, OpenAI is seeking to raise about $30 billion at a valuation of more than $1.4 trillion before Anthropic's IPO, a valuation at least 64% higher than the $852 billion it reached in March. The two companies are competing for overlapping long-term funds and sovereign wealth funds, and OpenAI hopes to lock in commitments before Anthropic files its S-1 and goes public in mid-November, to avoid direct comparison. Anthropic's third-quarter revenue grew 45% to $16.8 billion, a slowd

    Matched fieldsTitle

    #ai-news#Industry news#OpenAI#Anthropic
  7. Anthropic Alignment Science Blog

    Anthropic proposes GRAFT: localizing reward hacking to removable parameters rather than penalizing it

    Anthropic and other institutions propose GRAFT (Gradient Routing Adapter Fine-Tuning), which, when reward hacking is detected, updates only a dedicated FORGET adapter and removes it after training, thereby eliminating these behaviors. The method uses monitoring signals only to decide which parameters to update, and does not set rewards based on them, so the model is not reinforced for exploiting weaknesses in the monitor. In Countdown-Code RLVR experiments with Qwen3-8B, GRAFT prevented reward hacking while still enabling the model to learn the target task; the experiments were limited to small models and simple monitors.

    Anthropic and other institutions propose GRAFT (Gradient Routing Adapter Fine-Tuning), which, when reward hacking is detected, updates only a dedicated FORGET adapter and removes it after training, thereby eliminating these behaviors. The method uses monitoring signals only to decide which parameters to update, and does not set rewards based on them, so the model is not reinforced for exploiting weaknesses in the monitor. In Countdown-Code RLVR experiments with Qwen3-8B, GRAFT prevented reward h

    Matched fieldsTitle

    #company#Papers/Research#Safety / alignment#English text is being prepared.

Sunday