Leopold Aschenbrenner, a former OpenAI researcher, plots AI capability over time in a graph that every marketing leader should see. The curve is steady from 2018 to about 2024, then it bends and goes vertical, and we are now at the bend.
The x-axis is time. The y-axis is effective compute, a proxy for how capable these systems are. The milestones are marked: GPT-2 wrote like a preschooler, GPT-3 like a schoolchild, GPT-4 like a smart high schooler. The next milestone on the chart is automated AI research. The confidence band after that point is so wide it fills the page.
The evidence that we have entered the early phase of this curve arrived over the past few months, and most of the marketing industry has yet to register what it means.
The MiniMax release
In March 2026, the Shanghai AI lab MiniMax released a model called M2.7. The release itself matters less than how they built it. The model participated in its own development: it identified its own performance gaps, planned changes, and ran improvement cycles inside an automated harness. MiniMax reports that the model handled 30 to 50 percent of the reinforcement learning team's daily research workflow, and that over 100 fully automated analyse-plan-modify cycles lifted internal evaluation scores by roughly 30 percent.
OpenAI has described GPT-5.3 Codex in similar terms: early checkpoints of the model were used to debug its own training and diagnose test results, and later checkpoints improved on the earlier ones.
Andrej Karpathy, former Tesla AI director and OpenAI founding member, open-sourced AutoResearch: a framework where you hand an AI agent a training script and let it propose changes, run experiments, evaluate, and iterate on its own, overnight, on a single GPU, without a human in the loop.
Each of these on its own would be a curiosity. Together they describe a pattern that every major lab is now reporting: AI is helping to build better AI.
Three consequences for marketing and brand teams
Most coverage of recursive self-improvement is written for AI researchers. The commercial implications land somewhere else entirely: on marketing, brand strategy, and analytics teams, and they are landing now.
1. Research cycles are collapsing
A traditional segmentation study costs north of 200,000 pounds and takes 12 to 16 weeks. The same is true of demand-space mapping, brand trackers, and competitive landscape work. Because this research is expensive and slow, it is infrequent, and the insight is often stale by the time it reaches a brief.
When AI systems improve at the rate we are now seeing, the unit economics of research change at the structural level.
I run systems today that do in hours what used to take weeks: competitor monitoring, contact enrichment, market signal scanning across hundreds of sources at once. They are production workflows that run every day. The models powering them are materially better than the ones I used three months ago.
Three months ago the models were worse, and three months from now they will be better again. If your planning cycle still assumes research is expensive and slow, it is built on an assumption that has already expired.
2. The strategy-execution gap is closing
The biggest waste in marketing is good strategy that moves too slowly from slide to market.
One element of Karpathy's framework transfers directly to marketing operations: autonomous experimentation. A system that designs an experiment, runs it, reads the results, adjusts, and runs again overnight without supervision is a repository you can download today.
Apply that pattern to campaign testing, pricing, creative optimisation, and media mix. A team that runs 50 experiments while a competitor runs 5 learns ten times faster, and in a market where the models themselves improve monthly, learning speed is the only durable advantage.
Strategy needs execution speed, and execution needs learning loops. Teams that connect the two, with strategy feeding automated experiments and results feeding back into strategy, will pull away, and the gap compounds.
3. Proprietary intelligence is becoming the durable advantage
Generic AI is a commodity. Your competitor has access to the same ChatGPT, Claude, and Gemini, answering the same questions from the same training data.
A model tuned on your customer interactions, your category data, your campaign history, and your competitive signals is a different asset. It is proprietary, and it no longer takes an ML team to build: an AI pointed at clean data can assemble it overnight, the same autonomous pattern the labs use on themselves.
Brands that start building proprietary intelligence now, even simple versions, accumulate an advantage a late mover cannot buy back by purchasing the same SaaS tool as everyone else. Your data is unique, and the intelligence layer on top of it should be too.
The next 18 months
Nobody knows exactly how fast this goes - the confidence band on Aschenbrenner's chart widens dramatically after the inflection. The direction, though, is settled, and every major lab reports the same pattern.
Three priorities follow for marketing and brand teams:
Build AI workflows now. The advantage compounds. A team that has run AI-assisted research, brief generation, and campaign analysis for 12 months operates at a different speed from a team starting next year, because the tools improve while you learn them.
Invest in your own data. Proprietary intelligence built on your category and your brand's signals is the differentiator, and it needs clean data and clear use cases more than it needs a data science department.
Treat AI as a system that learns. A tool does what you tell it, while a learning system gets better every time it runs at understanding your customers, your category, and your competitors.
The curve is steepening either way, and the teams that climb it early will be hard to catch.
We built NavigatorLab to apply this approach to a real category question: where growth lives in your market.
Sources and further reading
- Leopold Aschenbrenner, "Situational Awareness: The Decade Ahead" - situational-awareness.ai
- MiniMax, "M2.7: Early Echoes of Self-Evolution" (March 2026); coverage in VentureBeat
- Andrej Karpathy, AutoResearch - github.com/karpathy/autoresearch
- OpenAI, GPT-5.3 Codex release notes