Model Release Cadence - Bryant McGill

# Model Release Cadence **Entity class:** Operational concept **Domain:** Artificial intelligence / deployment / safety **Maturity:** Developed ## Definition Model release cadence is the interval and change pattern through which frontier capabilities, tools, post-training systems and deployment affordances reach users. ## Mechanism and significance As capability changes become continuous rather than episodic, safety work must track interacting releases rather than treating each pre-training run as the only meaningful unit of change. ## Source route - [[research/David Robinson on OpenAI Safety, Superintelligence, and the Transhumanist Transition|David Robinson on OpenAI Safety, Superintelligence, and the Transhumanist Transition]] ## Relationships [[wiki/Frontier AI|Frontier AI]] · [[wiki/AI Race Dynamics|AI Race Dynamics]] · [[wiki/AI Safety Culture|AI Safety Culture]] · [[wiki/Normalization of Deviance|Normalization of Deviance]] ## Evidence boundary Release frequency is not itself proof of danger; its importance depends on capability delta, evaluation quality, deployment scale and the time available for correction. ## Simple Reminders, Quotations, and Thoughts  > “We need to meet safety criteria. If we can meet them in ten minutes, great. If we stop for two months and still have not met them, then we are still blocked. Running off a cliff and walking slowly off a cliff are not that different.” > **— David Robinson**, *The Ezra Klein Show, October 7, 2026* [[reminders/AI Control/Slowing Down Does Not Satisfy Unmet AI Safety Criteria by David Robinson|Slowing Down Does Not Satisfy Unmet AI Safety Criteria by David Robinson]] > “Frontier model releases once arrived roughly every seventy days; now they can arrive every eleven days. Even as the models become more capable and frightening, they are being released much faster.” > **— Ezra Klein**, *The Ezra Klein Show, October 7, 2026* [[reminders/AI Control/Frontier AI Releases Accelerated from Seventy Days to Eleven by Ezra Klein|Frontier AI Releases Accelerated from Seventy Days to Eleven by Ezra Klein]] > “A model release once meant baking a fresh cake through a new pre-training run, with months between major releases and a substantial period for safety work. Now pre-training, post-training, reasoning training, tools and new system affordances can all change capability and risk independently. We are shipping new capability and new risk every Tuesday.” > **— David Robinson**, *The Ezra Klein Show, October 7, 2026* [[reminders/AI Control/AI Labs Ship New Capability and Risk Every Tuesday by David Robinson|AI Labs Ship New Capability and Risk Every Tuesday by David Robinson]] > “There is not a ton of time to kick the tires before a frontier release. People inside the laboratories care passionately about making systems safe, launches do get cancelled and training runs do get stopped, but the ecosystem as a whole is not operating at the level of safety rigor its risks require.” > **— David Robinson**, *The Ezra Klein Show, October 7, 2026* [[reminders/AI Control/Frontier Releases Leave Too Little Time for Safety Testing by David Robinson|Frontier Releases Leave Too Little Time for Safety Testing by David Robinson]] > “The structural situation changed between 2023 and 2026. Model releases accelerated, more players began competing, open-weight and Chinese models expanded the field, commercial contracts intensified the pressure and trillion-dollar IPOs approached. All of those forces push toward speed.” > **— Ezra Klein**, *The Ezra Klein Show, October 7, 2026* [[reminders/AI Control/Competition Pushes Frontier AI Toward Speed by Ezra Klein|Competition Pushes Frontier AI Toward Speed by Ezra Klein]] > “The AI industry could normalize danger by accepting a risk, observing that nothing horrible happens, and then accepting the next system because it is only slightly different. When models change a little every week rather than becoming a whole new world every few months, the industry can move by shades into a level of risk that no longer makes sense.” > **— David Robinson**, *The Ezra Klein Show, October 7, 2026* [[reminders/AI Control/AI Labs Can Normalize Danger One Small Model Change at a Time by David Robinson|AI Labs Can Normalize Danger One Small Model Change at a Time by David Robinson]]