THE COLLECTION
EVERY SIGNAL, SORTED.
Search the whole shelf by topic, tool, term, or plain old curiosity.
Australia, University of Sydney, Deep Learning, Reproducibility, Riddled Basins, Neural Networks, Model Evaluation, Initialization, Training Runs, Rare Cases, Deployment Artifacts and Research Methods
A one-bit nudge sent AI training elsewhere. Ship the ensemble receipt
A study of roughly 500,000 neural networks found that absurdly tiny changes can send training toward different internal solutions. The useful response is not panic. It is to stop treating one lucky run as the whole product.
OpenAI, Decisions API, Application Routing, Agent Control, Model Evaluation, Calibration, Human Review and Operational Accountability
OpenAI made tiny AI decisions fast and cheap. The threshold needs a receipt
A new API turns text, images and structured data into bounded probabilities, choices and scores at very low latency and cost. That makes the model easier to put in the hot path, which is exactly where teams need to record the threshold, the action and what happened next.
South Korea, Sovereign AI, Foundation Models, Public Investment, Model Evaluation, Open Models, Benchmarks, Usability, Expert Review, Compute and Accountability
Korea picked three sovereign model teams. Publish the scorecard, not just the winners
South Korea advanced Upstage, SK Telecom and LG AI Research after a 100-point evaluation that mixed benchmarks, expert review and public usability. Motif led a major external benchmark and still missed the cut, which is exactly why the public needs the complete scoring receipt.
New Zealand, Conversational AI, Turing Test, Model Evaluation, Experimental Design, Human Judgment, Replication, Philosophy of Mind and Benchmark Limits
Thirty-six people spotted GPT-4-Turbo. The five-minute Turing Test was the shortcut
University of Canterbury researchers let interrogators keep talking until they felt ready to decide. Most took longer than five minutes, and 36 of 37 identified GPT-4-Turbo. The result challenges a famous shortcut, not every current model or the meaning of intelligence itself.
France, Time-Series Forecasting, Open Models, Probabilistic Prediction, Industrial Operations, Energy, Supply Chains, CPU Inference and Model Evaluation
Mistral made four small forecasting models. The baseline still gets a chair
Timestral turns histories, related measurements and known future events into probabilistic forecasts on ordinary CPUs. Mistral reports leading public benchmark results. The useful deployment test is whether one of its four sizes beats a plain baseline on the decision that actually costs money.
United Kingdom, Canada, AI Security, Model Evaluations, International Standards and Public Assurance
Britain and Canada renewed AI security cooperation. The shared tests need publishable results
The two governments want interoperable standards, closer information sharing and coordinated evaluations of frontier AI. That is useful plumbing. Now they need a public layer that shows what was tested, what was learned and what outsiders can reproduce.
Germany, Europe, Open Weights, Bilingual AI, Sovereign AI, Benchmarking and Local Deployment
Aleph Alpha opened Kolibri's weights, but kept the benchmark home field
Germany now has a downloadable bilingual reasoning model with a permissive weight license and unusually detailed training notes. The next test is whether outside teams can reproduce the performance claims on tasks the model maker did not design.
Open-Weight Models, Sparse Attention, Coding Agents and Reproducible Evaluation
NaiveAI opened a 309-billion-parameter model, but the benchmark still belongs to its maker
Naive-N0.5-Flash exposes weights, inference code and a sparse route through a very large model. That is a useful release. The speed and capability scoreboard is still company-run, hardware-hungry and waiting for independent reproduction.
AI Security and Public Oversight
Europe’s cyber agency got access to two frontier models
ENISA is testing Anthropic’s Mythos 5 and OpenAI’s GPT-6 Astra. That sounds procedural, but regulators cannot examine a machine they are not allowed to touch.