One row per milestone; forecast the date each milestone is…: When will Discovery Loop first publish a formal safety…

Part of: One row per milestone; forecast the date each milestone is first reached. If a milestone may plausibly never happen, express that through late percentiles. Shared background: Discovery Loop is a Delaware public benefit corporation announced August 5, 2026 by Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, automating the full experimental loop of research (propose, run, evaluate, iterate), initially in ML research and engineering, acting as its own first customer, with stated applicability to the 14 NAE Grand Challenges. Seed round co-led by Radical Ventures and Khosla Ventures with Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet participating, terms undisclosed and unclosed at launch; Google Cloud supplies first-year compute, and Google announced a joint research framework with the company on ML systems. At launch: zero employees beyond the founders, no office, a single generic Member of Technical Staff posting in Palo Alto, no safety or policy or sales roles advertised, and a stated plan to hire a lean in-person founding team. Reference class: Google DeepMind's applied-science record shows fast wins where the experimental loop is digital with objective evaluation (AlphaFold at CASP, weather models, algorithm discovery) and slow or failed outcomes where the loop runs through the physical world or institutions (Streams shut down, disputed GNoME synthesis, Isomorphic reaching first trials only after five years). AlphaFold took roughly four years inside a fully staffed lab from project start to recognized breakthrough. Peer startups: Safe Superintelligence emphasized a small team and took nearly a year to reach roughly 20 employees and has shipped nothing public; Thinking Machines took about seven months to its first technical blog post and about eighteen months to a high-level safety document; Sakana AI took eight to twelve months to its first automated-discovery artifacts and saw its attribution claims disputed. — view all 4 rows

When will Discovery Loop first publish a formal safety framework for automated AI research?✓#4

milestone_dateMar 1, 2029

10% by
Jul '27
25% by
Feb '28
median
Mar '29
75% by
Jan '31
90% by
Sep '34

rationale:

Discovery Loop launched in August 2026 as a Delaware public benefit corporation with only its four founders and a generic engineering role posted 44 sources. Crucially, the resolution criteria require a specific, named framework with operational commitments and thresholds—analogous to Anthropic's Responsible Scaling Policy or Google DeepMind's Frontier Safety Framework 2 sources. Launch-day artifacts, which include mission statements about being a 'positive force for humanity' and standard PBC language 44 sources, explicitly do not qualify. With zero safety or policy headcount currently, building a document of this rigor from scratch will take substantial time.

The primary drivers pushing estimates into 2029 are Discovery Loop's deployment posture and the historical reference class. Because the company plans to act as its own first customer, it lacks the external model-deployment triggers that typically force the publication of safety frameworks. Furthermore, imminent regulations like California's SB 53 or New York's RAISE Act target large frontier developers with revenue exceeding $500 million wiley.law—thresholds a pre-revenue startup will not hit for years. The reference class for peer startups is sobering: Anthropic took over 2.5 years to publish its RSP, xAI took well over a year for a draft, and peers like Safe Superintelligence (SSI) have published nothing two years post-launch ssi.inc. Even Thinking Machines took roughly 18 months to publish a 'high-level framework' that explicitly deferred concrete stop conditions and access criteria, which would likely fail this question's strict bar 55 sources.

Conversely, the timeline is accelerated by the intense 2026 focus on exactly what Discovery Loop aims to build: automated AI research and recursive self-improvement 2 sources. The July 2026 'Pacing the Frontier' letter, signed by over 1,100 researchers, explicitly demands pacing mechanisms for recursive self-improvement 44 sources. Combined with scrutiny from FLI’s indices futureoflife.org and Discovery Loop's joint-research partnership with Google—whose Frontier Safety Framework already outlines Tracked Capability Levels for ML R&D deepmind.google—there will be substantial reputational pressure and ready-made templates for the founders to eventually formalize their safety commitments.

The median estimate of March 2029 (roughly 31 months post-launch) balances the time required to build an internal platform, hire safety staff, and encounter concrete evaluation needs against the heightened external pressure for RSI governance. The wide intervals, particularly the long right tail (p75 in January 2031, p90 in September 2034), reflect a roughly 25-30% probability that a qualifying public document is never published. This 'never/very late' mass accounts for scenarios where Discovery Loop remains an internal research shop, publishes only vague non-qualifying blog posts, pivots, fails to scale, or is simply reabsorbed into Google before ever needing a public operational safety framework.

Assessed against expectations for early lean headcount scaling, the earliest bound shifted slightly later to reflect the low likelihood of drafting a complex policy framework during an initial bootstrapping phase.