Holy Moly. OpenAI has released GPT-6 Rosalind, a version of GPT-6 Astra designed to answer more dual-use biological research questions with fewer refusals. It uses the exact same model weights as GPT-6 Astra, with no additional training for this release. Compared with Rosalind 5.5: >Genomics-agent success: 16.7% → 37.7% >Medicinal chemistry: 28.9% → 49.6% >Life sciences research: 44.0% → 61.3% But its biosecurity behavior is markedly different despite using exactly the same model weights. GPT-Rosalind-6 passed OpenAI's Biosecurity Sensitive safety grader in 79.0% of tests, versus 99.8% for GPT-6 Astra. On Dual-Use biological prompts, it scored 84.6% versus Astra's 97.8%. OpenAI classifies it as High capability for biology and, inheriting Astra's assessment, Critical for cybersecurity. Access is restricted to approved researchers, companies and select government partners.

Holy moly $SOFI just laid out a path to 100M products and at least $8B in annual revenue across 15 businesses by 2030. That essentially implies ~36% annual product growth through rest of the decade 🤯

Holy Moly. Meta has now revealed new details about Prometheus, its gigawatt-scale AI supercluster in Ohio. During synchronized training, Prometheus experiences 30–40% power swings, shifting tens of megawatts in seconds. Meta developed a system called Dynamo Dimmer to throttle GPU power, supported by batteries and capacitors. The first phase has already trained Muse Wonder. Prometheus now connects hundreds of thousands of Nvidia GPUs across multiple buildings, with plans to scale to roughly 1 million and add Nvidia's Vera Rubin hardware. At its peak in mid-2025, Prometheus suffered nearly 1,000 training interruptions per day per 8,000 GPUs. Meta says engineering fixes with Nvidia have since reduced interruptions by 99%. To keep 32,768 GPUs actively training, Meta provisions 37,656 GPUs, including spare capacity for hardware failures.