We have paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us. Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment. We care very deeply about AI safety. We believe the entire field will have to coordinate on shared safety standards, but will act unilaterally in the meantime. We expect confidence in safety to increasingly set the pace of AI progress. We are optimistic about the alignment work we are doing, and we remain committed to making frontier capabilities widely available.

Situational Awareness "purchased 'flex options,' a customized derivative, to amplify its AI trade." Flex options aren't what you might use as a pickup move at Surf Club Montauk. FLEX option terms are tailored to a specific buyer and seller. There's virtually no secondary market. Ooops

Apple argues that AI models cannot do math.. not even the grade school math. For years, labs like OpenAI and Google have bragged about near-perfect scores on benchmarks like GSM8K. They claimed AI had mastered logical problem-solving. Apple decided to test if that was true. They built a new benchmark called GSM-Symbolic. Instead of static questions, it uses templates to dynamically change names, numbers, and variables. The results exposed a devastating truth. When Apple changed the simple numbers inside a basic word problem, model accuracy plummeted. The AI wasn't solving the math. It was relying on pattern matching from its training data. It was guessing based on familiarity, not reasoning. Then they ran the test that exposed the illusion entirely. They added a single, irrelevant clause to a math problem. Just a sentence of background text that looked important, but had zero impact on the actual calculation. Every single frontier model, from ChaTGPT to Claude, suffered massive performance drops. Some crashed by up to 65%. Just by adding a piece of noise that a seven-year-old child could easily ignore. The conclusion is blunt. Current AI models do not possess genuine logical reasoning. They do not understand math. They are sophisticated mimicry engines replicating the shape of human logic without actually thinking. When the pattern is clean, they look like geniuses. The moment you introduce a minor twist, a variable change, or a distraction, the illusion shatters.

Serving inference is *not* high margin For demand to accrue to *your* model, you need to train the next gen. It’s only high margin until the next gen comes out Margin on GPT3 is useless once GPT4 is out Profitable via accounting tricks

AOC (Alexandria Ocasio-Cortez) just filed her annual financial disclosure. She reported: Up to $50,000 in savings Up to $15,000 in checking Up to $1,000 in a brokerage account Up to $15,000 in a 401(k) Up to $50,000 in student loan debt

With open model token prices rising, one can guess what is going on at Anthropic and ChatGPT. What happens when China floods the world with cheap Ascend chips to power the Chinese open models