Studying what follows.
We research the material and societal consequences of running AI at scale: the energy these systems consume, the hardware reshaping how they run, and what they do to the integrity of what we see and read. Most of what is said about AI is said to sell something. We advise the organisations that need more than that.
The more consequential question is direction. The prevailing use of AI replaces human thinking. We think the higher-value target is developing human learning, not automating it. AI-driven coaching produces gains in human capability, particularly where the bottleneck has always been feedback.
We are a small, deeply analytical research group.
Datacenter operators report efficiency gains. The aggregate numbers tell a different story.
A Canadian startup is etching neural networks directly into silicon. The implications go further than the benchmarks suggest.
A physicist used AI to extend the frontier of theoretical research. Experienced developers used the same generation of tools and got slower. AI does not replace expertise. It amplifies it. The danger is what happens when people skip the struggle that builds it in the first place.
Most AI strategies optimise for replacing human expertise. The higher-value target is accelerating how people develop expertise. When judgment is what makes AI useful, the real bottleneck is learning, not automation.
The prevailing use of AI, across every domain, replaces human thinking rather than developing it. Learning is the counter-case: the one place we can prove the alternative works. The bottleneck has always been feedback, not information.
Generative video is improving faster than the models built to detect it. The more resilient approach works upstream: cryptographic attestation at the hardware level, proving footage was optically captured rather than computationally produced.
Social media companies deploy algorithms that shape the information diet of billions with no mandatory proof of competency. AI widens the gap further. Every previous technology that could cause this kind of harm eventually triggered licensing. The only question is when.
Almost every worry about AI in the classroom is a worry about the machine being wrong. Set those aside. Assume it is right, fast, and getting righter. That is the version that should trouble us most, and it is the one the tools are racing toward.
Do it the way you're supposed to and you get the answer everyone gets. The interesting discoveries live in the deviation, and that is precisely the part a prediction engine is built to erase. AI makes variations cheap to generate. Judging which deviation matters is still a human task.
Inference Research grew out of a background that spans experimental physics and climate science. The thread connecting these fields is the same one that runs through this work: complex systems behave in ways that reward careful observation over confident prediction.
The research group exists because the conversation about AI's real-world impact needs more rigour and fewer press releases. We bring a physical-sciences perspective to questions that are too often framed as purely technical or purely political.
We're always interested in hard problems.