Keep pulling the thread on Akko Paoki.
OpenAI previously stated a goal of achieving AI with "research level intern capabilities" by September 2024.
OpenAI previously set a goal to develop a "fully automated AI researcher" by March 2028.
OpenAI's models have achieved the ability to solve problems at the level of the International Mathematical Olympiad (IMO), including a problem equivalent to "IMO problem six."
OpenAI is shifting its research focus from abstract benchmarks like mathematics to the practical application of its models in AI research, science, and other economically valuable activities.
Akko Paoki believes current AI models are now capable enough to materially change the economy and how work is done.
An internal OpenAI model, referred to as "GPT 5.2 Pro," has already generated minor but impactful research ideas that have been utilized by the team.
Akko Paoki predicts that AI models capable of working autonomously for several days to produce high-quality artifacts are "not very far" from being developed.
During the "first proof" mathematics challenge, an OpenAI model in training was able to solve novel research-level problems within approximately an hour of being prompted by OpenAI researcher James Lee.
The primary motivation for OpenAI's decision to hide the chain of thought in its O1 preview model was to prevent users from supervising it, which Akko Paoki believed would compromise its value for safety research.
Akko Paoki believes the AI industry must be prepared to make trade-offs and potentially slow down development based on findings from ongoing risk and alignment evaluations.
The majority of coding performed at OpenAI is done using its own AI coding tools, referred to as "codecs."
Akko Paoki predicts that AI systems will not be capable of autonomously solving high-level, open-ended problems like "solve alignment" within the current year.