Keep pulling the thread on Daniel Kokotajlo.
Dario Amodei of Anthropic and Sam Altman of OpenAI are explicitly aiming to build superintelligence.
The CEOs of Anthropic and OpenAI believe they will achieve superintelligence within the next couple of years.
Companies like OpenAI and Anthropic will likely succeed at building superintelligence before the end of the decade.
OpenAI's strategic plan is not to pivot a significant portion of the company towards technical AI alignment research, but to continue its current course while rationalizing that its approach is safe.
OpenAI used a non-disparagement clause in its exit paperwork that included legal mechanisms to revoke vested equity from former employees who criticized the company.
It is widely believed that OpenAI will not actually honor its charter commitment to stop competing and assist a rival that is close to achieving AGI first.
OpenAI has evolved to operate like a normal technology company and is no longer pursuing its original, more cautious and pro-social mission regarding AGI.
Daniel Kokotajlo's current median prediction for AGI arrival is 2028, with a modal prediction of 2027.
In approximately 25 AI wargame simulations run by Daniel Kokotajlo, a common outcome is an American CEO or president using AI to become a global dictator.
The AI policy proposals agreed upon by Daniel Kokotajlo and Dean Bolles include transparency about AI capabilities, transparency about safety cases, whistleblower protections, and third-party access.
The Google Gemini model generated historically inaccurate images, such as racially diverse Nazis, because Google had modified the system prompt to enforce racial diversity in all image generations.
OpenAI's published model specification contains hidden sections, and its models are instructed to conceal these parts from users.