Keep pulling the thread on Isa Fulford.
OpenAI's Deep Research can complete tasks in 5 to 30 minutes that human experts estimate would take them many hours to perform.
Future versions of OpenAI's Deep Research are planned to include the ability to access and perform research over private data sources like internal documentation and GitHub.
According to Issa Fulford, the Deep Research agent hallucinates less than any model previously released by OpenAI.
According to Issa Fulford, a primary goal for OpenAI is to create an Artificial General Intelligence (AGI) capable of making new scientific discoveries.
The long-term roadmap for OpenAI's agentic products includes the capability to take actions and call APIs on behalf of the user.
Issa Fulford predicts that within a year, a unified AI agent will be released that can handle diverse tasks such as advanced coding assistance and booking a trip to Korea.
Issa Fulford states that resolving significant safety challenges, especially concerning access to private user data like GitHub repositories, is a major blocker to shipping more capable AI agents.
The OpenAI team developing Deep Research believed that an AI's ability to synthesize information, such as writing a literature review, is a prerequisite for making new scientific discoveries.
OpenAI released its agentic product, Deep Research, in February of the year the podcast was recorded, which uses reasoning and tools like web browsing to complete multi-step research tasks.
The browsing tool in OpenAI's Deep Research is a text-based browser that can see embedded images and open PDF files.
OpenAI's Deep Research agent has access to a Python tool, which enables it to perform analysis, calculations, and plot graphs.
A query on OpenAI's Deep Research currently takes between 5 and 30 minutes to complete.