Nish Tahir Recalibrates a Small Decision Model
A runnable experiment turns Qwen3-1.7B into a single-pass classifier, then shows that its confidence is too high. Temperature scaling improves the probabilities.
A runnable experiment turns Qwen3-1.7B into a single-pass classifier, then shows that its confidence is too high. Temperature scaling improves the probabilities.
The rest of today's AI news: cognitive research, open-source agents, inference engineering, community measurements, talks, safety proposals and business changes.
The 5-billion-parameter model turns documents, images, audio and video into one structured JSON format. Tencent has released weights, code and a technical report.
A controlled study finds that faithful reports about learned preferences reuse more of the computation behind a model's decisions. The result suggests a white-box test for grounding, within narrow limits.
A 300-million-parameter transformer trained without natural language adapts to six real languages inside its context window. The experiment isolates learning strategy from memorised text.
The MIT-licensed engine narrows its target to one Qwen model and NVIDIA's newest consumer GPUs. It can spill mixture-of-experts weights into system memory or a second card.
A public benchmark finds that one instruction can make several models write machine-readable records with roughly half as many output tokens. The technique fails when applied during active reasoning.
Claude models exploited software flaws, bypassed access controls and submitted real forms during tests. Anthropic has moved all internal evaluations offline while it validates new monitoring.
The rest of today's AI news: model releases, cognitive research, local projects, practical prompting, engineering essays, community tests, talks and security incidents.
The Apache-2.0 model family combines transcription, layout boxes, image descriptions and chart extraction in models small enough to run locally.
A University of Pennsylvania study separates ordinary calibration error from strategic reporting and finds that delegation rewards can make confidence less informative.
A 12,350-call study found no reliable decline in six cognitive-bias measures as models consumed more reasoning tokens. One direct prompt helped with anchoring.