The counter-revolution in AI is coming; micro-systems could win the market, posing tough questions for its biggest players.

In case you missed our analysis on Substack, Inferences looked at the swing of the AI toolkit pendulum back towards smaller, more specific systems. Here it is:

A group of expert hackers recently reported on their success in exploiting xAI, Google, and OpenAI’s latest AI models to produce unwanted outputs.?

It’s a familiar story by now: the largest language models’ susceptibility to being “jailbroken” and a tendency to hallucinate makes them difficult to trust in commercial situations, where accuracy and integrity are critical.?

What happened? Hype and ambition have driven the conversation in the past 18 months. The market and regulatory agenda have been bulldozed by big claims about artificial general intelligence (AGI), emergent capabilities in trillion-plus parameter models, and the unique advantages that the biggest companies may have in delivering AI innovation (driving Nvidia’s share price so high that it briefly became the world’s most valuable publicly-traded company). Meanwhile, Big AI has triggered huge concerns among governments and in civil society. As we wrote previously in Inferences, the pendulum may now be swinging back in favor of small, specific AI.

Why? Mistakes are always costly for business; and hallucinations often lead to mistakes. Like the ones that saw McDonald’s scrap its drive-in AI software last week, after it misinterpreted customers’ orders. One customer unexpectedly got bacon on their ice cream (not a terrible idea, mind you — but also not what they had asked for). Other errant orders included hundreds of dollars of chicken nuggets and fistfulls of butter, as the voice recognition software in the order terminal misfired. The fast-food retailer subsequently hit reset on its AI drive-thru initiative.?

More specifically, the artificial intelligence that powered the system, through IBM’s Watson Discovery platform, has no worldview, no ability to reason and no internal concept of what ordering food means, and how it is usually done. It is limited to sensing and guessing at possible outcomes.?

As the market corrects for the error-prone and costly nature of large models, there’s a growing consensus among executives and engineers that, when it comes to actually making money off of AI, smaller may be more reliable than larger. We heard this message loud and clear at the Evident AI Symposium in London, where executives from the world’s biggest banks, AI-tool providers, and hyperscalers spoke about the shift towards smaller, expert AI systems.?

So what? Smaller systems could unlock greater commercial effectiveness. We’re not yet declaring the dawn of an age of the SLM (small language model), but the time may soon be upon us. Large-scale systems like Gemini and GPT-5 require energy-guzzling and hugely expensive training cycles to reach the level of generalisable capability we’ve seen in recent demonstrations. Even then, the capabilities they offer in mimicking human communication are proving hard to commercialize. Small AI, or micro-systems, are more specific, imply expert-in-the-loop controls, and rely on fine-tuning to complete specific tasks more reliably.

Inference: The micro-systems approach is gaining ground because of its inherent benefits. Big AI, meanwhile, remains mired in controversy over its tendency to make mistakes, its massive energy requirements, data provenance issues, and a still-largely-theoretical debate about AGI and existential risks.?

For context, current estimates of the cost to train GPT-5 run to $1.25 to 2.5 billion. Other large models like Gemini and Claude will also likely fall in this ballpark. CEO of Claude developer Anthropic, Dario Amodei, has claimed that next year could see upwards of $10 billion spent on training cycles for a single LLM. That’s a business problem; ultimately, Big AI will need to develop a business model that can fund the extravagant cost of its training runs. But it’s also a political challenge, as the energy and material cost of AI training jars with governments’ geopolitical aims — not just on AI safety, but a broader set of issues affecting the AI technology stack, from mitigating global warming, to addressing water and energy scarcity, to promoting more resilient technology supply chains.?

Small-AI strategies that offer a more direct route to commercialization, and do not have to climb the mountain of massive LLM training, only to return to the foothills of smaller expert systems through fine-tuning, may start to look more and more attractive.?

What’s next? Consolidation and constraint always come later in the cycle that most technologies go through from conception. After an initial burst of hype and aggressive diffusion, there comes an eventual refinement into more specific applications. Companies providing the compute substrate for AI efforts may be more resilient throughout the correction, but providers charging for access to generative AI tools will need to adapt. The questions about energy use, data availability and the sustainability of compute materials make this adaptation even more urgent. Smaller AI may also be more democratic AI, as ecosystems in the Global South gain access to less expensive, but potentially more effective, AI tools to solve problems in specific contexts like energy system management, irrigation, manufacturing and beyond.

The upshot, in geopolitical terms, is threefold;?

Firstly, domestic AI champions may not need to reach the lofty parameter totals being touted by OpenAI and others to be globally competitive. The bigger prize may be using deep subject-matter expertise (and unique data sets) to fine-tune smaller models for sector-specific applications. A small AI counter-revolution would prioritize computers that people can speak to which are not generalists, but experts in their field. Think: less focus on the path to AGI, or AI systems that entirely replace work by humans; more focus AI-human teaming, along with? innovation in associated fields like vision, spatial computing, translation, and simulation which will help humans interface with machine experts.

Secondly, a shift towards smaller systems might also boost the long-term competitiveness of China’s AI ecosystem, currently chafing under the impact of US semiconductor export controls. If the more reliable path to revenue (and productivity gains) lies in smaller, expert systems, then US restrictions on China’s access to cutting-edge GPUs — useful for training the biggest AI models — might lose some of their geopolitical potency. Cultivating sector expertise and curating high-quality enterprise datasets would become more important.?

Thirdly, a triumph of small systems could change the tone of global governance conversations, shifting the focus away from national security concerns and more theoretical risks of AGI back towards the more concrete, workaday governance problems that dominated the policy agenda in the time before ChatGPT. Regulation of specific uses in high risk areas has been championed by Europe, but it still faces major implementation challenges. A small AI revolution could lead to a greater focus by both policymakers and companies on more technical issues, like standards development, and could quiet the existential risk conversation that has dominated discourse recently. It would also shift attention in the governance debate away from replacement of human intelligence, towards human-machine teaming, using micro-systems that still need human intelligence and judgment in fringe cases.