Question
On 2028-06-30, which of these will best describe frontier AI research practice at the largest AI lab by revenue?
Current State of AI R&D (November 2026) The dominant practice across frontier labs remains "humans direct, AI implements." While measured capability growth is accelerating—with independent assessments estimating a near-2x pace at Anthropic metr.org—this speedup comes primarily from coding assistance and task execution rather than autonomous research design. OpenAI reports that the "Decide" and "Design" phases combined account for only 2.8% of agent output tokens, with overarching research planning ("what to work on") at just 0.3%. Anthropic's metric showing Claude "leading" 26% of R&D work measures supervised task execution, not experiment selection 2 sources. METR similarly finds that AI is not yet making "highly significant and fully autonomous contributions" to the most critical parts of model training 2 sources.
The Modal Trajectory: Automating the Inner Loop Over the next 19 months, competitive pressure and current capability trends strongly point toward AI automating the bounded experiment loop (design, execute, analyze, iterate), giving the "AI runs the full experiment loop while humans set research agendas" option our largest weight at 41%. There is clear evidence of movement up the capability stack: research taste scores (TasteVal) have shown a doubling time of about three months, and recent models like Mythos 5.2 drive their advantage specifically through improved next-step judgment rather than mere execution. Delegating the iterative refinement of well-defined problems to AI closely aligns with OpenAI's stated goal of creating an "automated AI researcher by March of 2028."
Governance and Compute Constraints The strongest alternative (30%) is that practice settles slightly short of the full loop, at "AI proposes experiments and humans select." The primary barriers to fully delegating the experiment loop are compute scarcity and safety governance. Training and experiment compute remains enormously expensive—Anthropic's compute and infrastructure took 58% of its revenue in Q3 2026—meaning labs have a strong incentive to require human authorization before models spin up resource-heavy experiments. Furthermore, Anthropic is explicitly considering updates to its Responsible Scaling Policy to require human review and approval of consequential R&D decisions 2 sources. If independent assessors interpret these strict approval gates as "humans select," this outcome becomes highly probable, further reinforced by lab public relations teams seeking to emphasize human control.
Full Automation vs. Stalls and Pauses The remaining probability space covers more extreme scenarios. An outcome where "AI sets research agendas with humans auditing outcomes" (11%) requires not only exceptionally rapid takeoff but also that labs completely abandon their stated frameworks for keeping humans in the overarching self-improvement loop. Conversely, the combined 18% weight on "Humans direct and AI implements" (12%) and "No material change" (6%) reflects the non-trivial risk of progress stalling. Labs have already demonstrated a willingness to halt specific work, such as OpenAI pausing frontier RL training following a July incident and Anthropic cutting live internet from internal evaluations aibusiness.com. A regulatory freeze, diminishing returns in agentic capabilities, or persistent debugging bottlenecks could easily freeze dominant lab practice near its 2026 baseline.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited