Deep Research, AI Takeoff and the Limits of Early Reactions

In February 2025 I look at reactions to OpenAI’s newly released deep research. Researchers and commentators describe its ability to search and synthesize sources, while some see it as evidence of rapidly accelerating AI progress. These reactions range from personal experiences to predictions, and they do not by themselves establish a general intelligence breakthrough.

I cite Stability AI founder Emad Mostaque’s prediction that machines will soon perform most digital knowledge work faster, better and more cheaply. He describes an intelligence takeoff. That is a forecast about future capability and economic impact, not a measured result from the deep research launch.

I connect this view to Leopold Aschenbrenner’s scenario of recursive self-improvement: an AI discovers useful knowledge, applies it to improve itself and repeats the process. I argue that we are approaching such a scenario, but the examples discussed here do not demonstrate an autonomous, sustained cycle of AI improving its own successor.

I mention an example discussed by Simon Willison in which a model was prompted to suggest a code optimization that reportedly made a specific operation twice as fast. A human-directed optimization of one workload is interesting, but it is not proof that a model has autonomously improved its own overall capabilities.

Technology columnist Kevin Roose compares reported scores on Humanity’s Last Exam, with deep research reaching 26.6% on the cited configuration shortly after a lower leading result. These are benchmark-specific results with particular tools and testing conditions, not a score for all human knowledge or a reliable forecast of a takeoff.

I credit DeepSeek’s R1 paper with making reasoning-model techniques more visible, and highlight Nathan Lambert’s observation that OpenAI trained deep research using reinforcement learning on browsing and reasoning tasks. I see a combination of reasoning and tools as important, although describing it as a proven recipe for AGI goes beyond what these examples establish.

Ethan Mollick’s essay [“The End of Search, the Beginning of Research”](https://www.oneusefulthing.org/p/the-end-of-search-the-beginning-of) argues that reasoning systems and autonomous research agents are converging. In his examples, he judges the resulting reports to be of different levels of quality. Those are his evaluations of particular outputs, not a guarantee that every report matches expert work.

I describe the model as the reasoning component and tools as the way an agent can act: it can search the web, write and execute code, or consult other data sources. The combination can produce a synthesized report, but the quality still depends on the sources, the task and the model’s interpretation.

The video shows a GPQA Diamond chart comparing model scores over time with human expert reference lines. The chart illustrates progress on a defined set of difficult science questions. It does not establish that a model is a general-purpose PhD, can independently conduct all research, or has exceeded experts across their work.

I interpret the graph as evidence for a hard takeoff and imagine many agents discovering AI research and improving themselves in parallel. That is my extrapolation. A rising benchmark score alone cannot establish recursive self-improvement, its rate or whether a hard takeoff is occurring.

I mention Sam Altman’s suggestion that inexpensive compute could produce substantially more value for a user. That is an illustration of possible value rather than a controlled economic measurement. Actual value and cost will differ with the task, output quality and review required.

Biomedical scientist Derya Unutmaz describes favorable early experiences with deep research and says two reports on cancer cases were impressive. This is a small anecdote and his professional assessment, not a clinical validation study or proof of reliable medical advice.

I also recount an OpenAI employee’s account of using deep research to investigate a breast-cancer treatment question for his wife and checking the information independently. The report is an example of information gathering, not evidence that the tool should decide treatment. Patients should discuss clinical choices with qualified clinicians and verify cited sources.

In a launch clip, Altman estimates that the system can perform a single-digit percentage of economically valuable tasks. He explicitly presents that as his estimate. I discuss its possible economic scale, but the percentage is not established by a comprehensive measurement of work across the economy.

Altman distinguishes synthesizing existing information from a future system that invents new knowledge. I think the direction could be transformative; the statement is a prediction, not proof that the product has achieved autonomous scientific discovery or recursive improvement.

Google staff react to the launch by pointing out that Gemini had already offered a product called Deep Research. I find the name overlap striking. Their reactions support the observation that a similarly named product existed earlier; they do not establish that OpenAI copied Google’s implementation.

Finally, I note Altman’s contemporaneous comment that deep research was not the additional announcement he had teased for o3-mini. That was a February 2025 snapshot of product announcements, not a statement about current release plans.

OpenAI’s own [launch material](https://cdn.openai.com/API/docs/deep_research_blog.pdf) notes that deep research can hallucinate, make incorrect inferences and struggle to distinguish reliable sources from rumors. Those limits matter especially for medical and other high-stakes research, where claims and citations need independent review.