rationale:

Current State and Strict Resolution Criteria As of October 27, 2026, there are zero qualifying AI predictions that have been experimentally tested. The resolution criteria impose a narrow funnel: discoveries must be genuinely new particles, phases, or physical effects; the central idea must come from a general-purpose AI; the prediction must be publicly credited to the AI before measurement; and independent experimenters must confirm it. Recent high-profile AI physics outputs fail these strict tests. The GPT-5.2 Pro single-minus gluon amplitude conjecture GPT-5.2 derives a new result in theoretical physics Single-minus gluon tree amplitudes are nonzero and the 4D Yang-Mills and QCD existence proofs Working with academic communities to build scientific ... AI makes discoveries in mathematics, computer science ... are theoretical or mathematical results with nothing to measure. OpenAI's GPT-5 science case studies focus on derivations rather than new testable effects Early experiments in accelerating science with GPT-5, and recent evaluations note that autonomous agents still struggle to select high-level research questions Assessing AI agents at autonomous theoretical physics .... Specialist machine-learning discoveries of new superconductors Machine Learning-Guided Discovery ofKagome ... and dusty-plasma physics AI just discovered new physics in the fourth state of matter rely on purpose-built neural networks or involve excluded categories like new compounds. Additionally, the AIMS agent's work on Wigner-crystal melting explained an effect after the fact rather than predicting it AIMS: an AI experimentalist turns uncertainty into quantum ... AIMS: an AI experimentalist turns uncertainty into quantum ..., and AI-designed optics experiments rely on optimization algorithms rather than general-purpose systems AI suggests new physics experiments that could outperform ... AI suggests new physics experiments that could outperform ....

Pathways to Experimental Confirmation The most plausible route to accumulating confirmed predictions before the end of 2030 runs through tabletop platforms: quantum simulators, cold atoms, photonics, and condensed matter. Fundamental particle physics faces severe timeline constraints, such as the ATLAS detector completing Run 3 in June 2026 with beams not expected to return until 2030 ATLAS enters the high-luminosity era. In contrast, accessible condensed-matter and quantum-device effects can be realized much faster. Experimental latency in these fields ranges from months, as seen with Floquet time crystals https://arxiv.org/abs/1603.08001 https://arxiv.org/abs/1609.08684, to several years, as was the case for the fabrication challenges of magic-angle graphene A new twist on graphene: an interview with Pablo Jarillo .... Given the roughly four-year window remaining, predictions generated between 2027 and 2028 in these faster-cycle domains are the most credible candidates for timely confirmation.

Accelerating Capabilities and Research Generation There is a clear trajectory toward scalable, AI-generated research proposals. Systems like AI-Mandel have demonstrated high-volume generation of experimental blueprints, even if they currently remain simulated https://arxiv.org/html/2511.11752v1 https://arxiv.org/html/2511.11752v1. Furthermore, AI agents are increasingly capable of operating real hardware, such as calibrating 108 of 112 qubits in under five hours Vibe Calibration: Autonomous Bring-up of a 112-Qubit .... The precedent set by mathematics—where arXiv papers crediting an AI surged from about 9 in June to 50 in September 2026—suggests that theorists will increasingly adopt and publicly credit AI systems for their theoretical groundwork. If academic theory groups begin generating testable predictions in bulk for quantum devices or cold atoms, a healthy pool of candidates will emerge by 2028.

Forecast Rationale and Uncertainties The forecast distribution balances the strict resolution criteria against the rapid acceleration in AI capabilities, yielding a median estimate of 3 confirmed predictions with a substantial right tail. I expect the first confirmed prediction will likely arrive well before the 2030 deadline . Together with the faster experimental cycles available in tabletop physics, this supports a 25th percentile of 1 rather than zero: at least one qualifying success is likely even if the broader pipeline develops slowly. The 10th percentile remains at zero because attribution norms could remain conservative, AI output could remain heavily skewed toward pure theory, or the required independent replication and field acceptance could take too long. Conversely, the 75th percentile of 6 and the 90th percentile extending to 15 confirmations capture scenarios where AI agents become routine collaborators and automated lab environments drastically compress the time from a public prediction to experimental validation. These outcomes allow for substantial scaling without assuming that an early first success removes the strict attribution and experimental bottlenecks for subsequent predictions.