(Images created via OpenAI’s image generation technology)
“We’re in the cusp of a system that can do new science.”
That line on page 3 of Openai’s latest “preparation framework” (version 2 updated April 15, 2025) illustrates a potential paradigm shift in the R&D ecosystem.
Going forward, the framework addresses the possibility that AI can “recursively self-improve.” The resulting “major acceleration of AI R&D speed” warns that new features and risks can be introduced rapidly. This acceleration could outweigh current safety measures, render surveillance “inadequate” and explicitly flag the danger of “maintaining human control” on the AI system itself.
Speaking at the Goldman Sachs event (released on YouTube on April 11th), Openai CFO Sarah Friar reinforced this view, saying that the model has already “invented something novel in their field,” and that it merely reflects “extending it” existing knowledge. Friar also focuses on a rapid approach to artificial general information (AGI), suggesting that “we might be there.”
Friar said that CEO Sam Altman’s view (AGI) – it is “immediate” for AI to handle the work of the most valuable human beings while acknowledging the ongoing debate, let alone its feasibility. This suggests that researchers move from AI as a tool to AI, as researchers may be closer than many people understand, and early examples have emerged potentially in areas such as software development.
https://www.youtube.com/watch?v=2kzqm_bue7e
Major R&D institutions are actively building “autonomous research” capabilities. For example, national laboratories such as Argonne and Oak Ridge are developing “autonomous driving labs” specifically designed for materials science and chemistry. Los Alamos is also working with Openai
Test the inference model for energy and national security applications of Venado supercomputers.
Generally, National Labs investigates AI use and undertakes core research tasks. Iterate towards hypotheses (often via optimization strategies), multi-step experiment design, robot execution control, real-time analysis of results, and discovery goals that significantly reduce human intervention within a given operation. Requiring human surveillance for validation and strategic orientation – perhaps working at a “level 3” or a new “level 4” of research autonomy – such initiatives show that AI will move beyond passive data analysis to directly participate in the scientific discovery process. As seen in the recent DOE “1,000 Scientists AI JAM,” it involves empowering researchers directly. This massive collaboration brought together around 1,500 scientists from multiple national labs, including Argonne, to test advanced AI inference models from companies like Openai and humanity, and tested real-world scientific problems. Researchers specifically investigated the possibilities of enhancing tasks such as hypothesis generation and experimental automation.
Currently, developers have a variety of views on the possibilities of Genai-enabled tools, but a similar migration is currently underway in software development. Today’s AI often acts as an assistant, but technology is rapidly improving software games, especially in popular languages ranging from JavaScript to Python. Openai’s model shows a major advancement in “getting closer to the human level” with key benchmarks. This supports potential monks who are described as “agent software engineers.” This is AI that allows you to “go out and work independently” and can be built, tested, documented, and more. This evolution into more autonomous capabilities could completely reconstruct the field.
Openai’s 5-level AI
Maturity Framework
Openai reportedly benchmarks advances towards artificial general information (AGI) using an internal five-level framework. This structure was discussed within the company in mid-2024 and later reported by outlets like Bloomberg, and outlines a clear stage in AI functionality.
Level 1: Chatbot/Conversation AI: A system that is proficient in natural languages like chatgpt. Level 2: Inferentialist: AI capable of solving basic problems comparable to highly educated people. At this level, the model can also demonstrate new inference skills without external tools. Level 3: Agent: An autonomous AI system that manages complex tasks on behalf of users and allows you to make decisions over a long period of time. Level 4: Innovators: AI can greatly contribute to creativity and discovery by generating new ideas, supporting inventions, or driving breakthroughs. Level 5: Organization: The pinnacle phase where AI can potentially manage and operate complex functions across the organization, potentially exceeding human efficiency.
Generally, National Labs investigates AI use and undertakes core research tasks. Iterate towards hypotheses (often via optimization strategies), multi-step experiment design, robot execution control, real-time analysis of results, and discovery goals that significantly reduce human intervention within a given operation. Requiring human surveillance for validation and strategic orientation – perhaps working at a “level 3” or a new “level 4” of research autonomy – such initiatives show that AI will move beyond passive data analysis to directly participate in the scientific discovery process. As seen in the recent DOE “1,000 Scientists AI JAM,” it involves empowering researchers directly. This massive collaboration brought together around 1,500 scientists from multiple national labs, including Argonne, to test advanced AI inference models from companies like Openai and humanity, and tested real-world scientific problems. Researchers specifically investigated the possibilities of enhancing tasks such as hypothesis generation and experimental automation.
Currently, developers have a variety of views on the possibilities of Genai-enabled tools, but a similar migration is currently underway in software development. Today’s AI often acts as an assistant, but technology is rapidly improving software games, especially in popular languages ranging from JavaScript to Python. Openai’s model shows a major advancement in “getting closer to the human level” with key benchmarks. This supports potential monks who are described as “agent software engineers.” This is AI that allows you to “go out and work independently” and can be built, tested, documented, and more. This evolution into more autonomous capabilities could completely reconstruct the field.
