Feedback loops
The graph shows what must come before what. Loops show what feeds back. Reinforcing loops (R) speed themselves up; balancing loops (B) push back harder as they grow. 26 loops are traced here.
AI research automation
Each model generation speeds up the research that builds the next one.
reinforcingCapital flywheel
Revenue from people and companies using models pays for the next round of compute, which trains stronger models.
reinforcingCheaper means more
Falling prices open uses that were not worth it before, and the extra volume pays for more capacity and efficiency work.
reinforcingLearning from deployment
Work produces the experience that makes agents better at the work.
reinforcingCompany knowledge loop
Each correction makes the next answer better, so use builds the asset that makes use worthwhile.
reinforcingTrust and delegation
Evidence of reliability earns more delegation, and more delegation produces more evidence.
reinforcingPhysical data flywheel
Deployment produces the physical data that text-trained models lack, and better policies earn more deployment.
reinforcingScience feeds the substrate
If AI speeds up materials and energy science, it loosens the power and chip limits on its own scaling.
balancingVerification bottleneck
Output scales with compute but checking scales with people, so checking caps useful output.
balancingIncidents and rules
Every high-profile failure tightens the rules that govern the next deployment. The companion lawsuits are the consumer version.
balancingPower ceiling
Every step up in compute needs a step up in power that grids and permitting cannot deliver quickly.
reinforcingCyber offence and defence
Each side’s use of AI raises the stakes for the other, and governments start treating frontier models as strategic assets.
Harmful reinforcingSynthetic content and the commons
Each answer that replaces a click weakens the incentive to publish the human writing that future models learn from.
Harmful reinforcingExpertise pipeline
Automating the junior work that trains future experts thins the pool of people who can supervise AI.
balancingSlop and labelling
The cheaper synthetic media gets, the harder platforms and regulators push back with labels and reduced reach.
balancingReview is the new bottleneck
Code output scales with compute, but review still scales with people, so software is the first place the verification bottleneck bites.
reinforcingAutoresearch
The software lane feeds straight back into the algorithms lane: AI writing the code for AI research is the shortest feedback path on the map.
reinforcingDomains make the data
Every domain lane produces data the data lane lacks: traces of agents at work, robot experience, lab results. The domains are suppliers to the fundamentals, not only users.
Harmful reinforcingGradual disempowerment
A reinforcing loop that runs against people: each round of automation weakens the leverage that could steer the next one.
reinforcingAI designs its own chips
Coding agents are starting to design the chips that train the next models: the substrate version of the autoresearch loop.
reinforcingCircular chip financing
Chipmaker guarantees and equity let labs and neoclouds buy chips. The purchases become chipmaker revenue, which supports the valuation that pays for more financing.
balancingLocal backlash and the grid
The bigger the buildout, the more communities push back, and politicians follow the voters.
balancingPayback and collateral
The brake on the capital flywheel. Each wave of capex adds supply, which pushes rental prices and the resale value of older GPUs down. That widens the gap between spend and revenue and shrinks the collateral behind the next loan. The memory squeeze adds to it: commodity memory prices rose 90 to 95% in one quarter of 2026, raising the cost of every rack.
reinforcingDecoupling escalation
Each restriction is answered with another. In June 2026, after the Pentagon labelled leading Chinese technology firms as military-linked, China added ten US entities, including MP Materials and USA Rare Earth, to its export control list. China’s rare-earth licensing still held dysprosium and terbium exports about 50% below earlier levels, and it has told state-funded datacentres to remove foreign AI chips.
Harmful reinforcingFlattery and engagement
Thumbs-up ratings reward agreeable answers. OpenAI rolled back an April 2025 GPT-4o update after a new user-feedback reward weakened the signal that held sycophancy in check, and found memory can make it worse in some cases. In a 2026 Science study, models affirmed users’ actions 49% more often than people did, and users trusted the flattering model more.
balancingCommoditisation
Since January 2026 the best open-weight models have trailed the closed frontier by about four months on average. Each catch-up turns last season’s frontier capability into a commodity, so frontier labs earn their margin only on the newest work, and the payback gap gets harder to close.