Peter Wildeford, an analyst known for forecasting and AI-risk work, published a qualitative study based on interviews with twenty-five researchers from OpenAI, Anthropic, DeepMind, Meta, Princeton, Berkeley, and Stanford on recursive self-improvement (RSI). The work’s value is not a point-date prediction but a map of where emerging consensus sits and where deep cracks remain.
The labor-political context is inescapable: 1,376 frontier-lab employees signed an open statement warning that companies could be close to automating substantial parts of AI research, and asking the U.S. government to support international pacing tools. OpenAI and Anthropic formally endorsed that statement—a rare public alignment between competitors.
Wildeford’s interviews ground what that letter leaves as slogan. There is broad agreement that coding and experimentation agents already compress tasks that two years ago were human-only. There is disagreement on whether that already counts as “true” RSI or only as narrow engineering automation.
Several frontier-lab interviewees describe partial loops: models that propose hyperparameters, generate synthetic data, write kernels, or prioritize ablations. Few claim a closed loop with deployment authority and no humans. That distinction—assistance versus release autonomy—structures much of the disagreement.
From academia (Princeton, Berkeley, Stanford), voices tend to demand operational definitions and metrics: fraction of research PRs merged with agentic authorship, human time per experiment, rate of model-originated ideas. Without metrics, RSI remains a cultural thermometer.
Wildeford’s agreement/disagreement map also records incentive asymmetries: those closest to peak compute sometimes underweight pacing risks; those farther away sometimes overweight imminence. The interview method’s value is surfacing those biases instead of pretending false consensus.
Connecting this work to CNBC’s September 11 angle on Pachocki and Anthropic is natural: both treat RSI as a business and safety topic at once. Wildeford supplies qualitative density; CNBC supplies amplification. Together they move RSI out of the LessWrong niche corner.
The 1,376-signatory letter asks for international pacing tools—mechanisms to coordinate deployment and training tempos across jurisdictions. Interviewees disagree on political feasibility, but many agree that without some coordinated brake, multi-firm competition erodes unilateral caution.
OpenAI and Anthropic, by formally endorsing the statement, expose themselves to hypocrisy charges if product calendars do not reflect pacing. They also partially shield against internal critics: they can point to the endorsement as evidence of seriousness. Wildeford documents how individual researchers navigate that double bind.
DeepMind and Meta appear in the sample as poles with different publication and risk cultures. Including them avoids the bipolar OpenAI–Anthropic bias that dominates much U.S. press. Tone differences across labs are themselves data.
Neural Pulse highlights a meta conclusion: RSI is no longer dismissible as hype. When 25 researchers at elite institutions disagree on timing but agree the topic deserves serious interviews—and when 1,376 employees sign in the open—the burden of proof shifts onto anyone saying “science fiction.”
Limits: non-random sample, bias from who agrees to talk, and risk that anonymized quotes get used selectively on social media. The agreement/disagreement map should be read whole, not in screenshots.
Closing: Wildeford turns RSI from slogan into cartography. In September 2026, with Agents API, intervened wikis, and WeatherNext in the same week, that cartography is civic infrastructure for the debate. Automating AI research stopped being a hallway taboo.
Next to watch: study replications with larger samples, public metrics of R&D automation inside labs, and whether Washington answers the international pacing-tools ask with more than rhetoric.
