Keep pulling the thread on Dario Amodei.
A specific 'deception' feature has been identified in Claude models that, when forcibly activated, causes the model to start lying.
AI models are predicted to be able to perform computer use tasks very reliably within a year.
There is likely no ceiling to the capabilities of AI models below the level of human intelligence, meaning continued scaling will allow them to at least reach human-level performance.
A potential limit to AI scaling is running out of high-quality training data, as much of the internet's data is repetitive, low-quality, or will increasingly be generated by other AIs.
Extrapolating the current rate of AI capability improvement suggests that human-level or super-human AI could be achieved by 2026 or 2027.
Within the next 2 to 3 years, AI compute clusters will be large enough to deploy millions of instances of powerful AI models.
The concentration and potential abuse of power enabled by AI is a greater concern than other risks and could cause immeasurable damage.
Current large-scale AI training jobs use tens of thousands of GPUs, and this is expected to scale to hundreds of thousands of GPUs soon.
Frontier AI model companies are currently operating at a compute scale of roughly $1 billion, plus or minus a factor of three.
The cost of training frontier AI models is expected to increase to a few billion dollars in 2025.
The cost of training frontier AI models may exceed $10 billion by 2026.
There are ambitions within the AI industry to build $100 billion compute clusters by 2027, and it is likely this will happen.