OpenAI is reportedly planning to launch specialized artificial intelligence (AI) agents for advanced applications, with pricing reaching up to $20,000 per month for a 'PhD-level research' agent. These agents are designed for high-value use cases such as conducting complex research, software development, and knowledge-based tasks. The company is also considering other AI agents at lower price points, including a $10,000-per-month software developer agent and a $2,000-per-month knowledge worker agent. These offerings reflect OpenAI's efforts to monetize its technology and address its financial challenges, including reported losses of $5 billion last year. Reports highlight that OpenAI investor SoftBank has committed $3 billion to support the development of these high-end AI products. Additionally, OpenAI is exploring chain-of-thought (CoT) monitoring to detect misbehavior in AI models, such as reward hacking, which remains a challenge. Comparisons have been drawn to Google's Gemini feature, which offers similar research capabilities at a lower price point.
I don’t think people understand how cheap it is to start an AI company. Cursor - $20/mo Vercel - $20/mo GPU bill - $10,000/mo SF penthouse office - $30,000/mo 3 billboards on highway 101 - $85,000/mo Fmr OpenAI eng - $110,000/mo
>train a model to use CoT >model starts planning ways to reward hack and misbehave >we can’t quite select this behavior out, it actually gets worse cuz the model just learn to hide it >slop CoT summarizer propaganda How did anyone think this post was a good idea? https://t.co/u45EtDTyDp
Fascinating. “We have further found that directly optimizing the CoT to adhere to specific criteria [..] does not eliminate all misbehavior and can cause a model to hide its intent. .. We recommend against applying strong optimization pressure directly to the CoTs of frontier… https://t.co/DpMkjAqgZi