Brief
DeepMind and UPenn detectors show reasoning for AI text
The reported detectors use sparse autoencoders to give human-readable reasons. The cached summary lacks version, code or availability details.
- Signal
- 39
- Facts
- 27
- Hype
- 9
classifiers
Topic
8 cards
Brief
Claude autoclassifier keeps blocking approved calendar deletions
A Reddit user says Sonnet 5.5 chat blocked agent calendar deletions even after the plan was approved. It is one user report with no official response.
- Signal
- 41
- Facts
- 23
- Hype
- 2
Brief
One user reports Claude security classifiers blocking blockchain prompts
About 8 refusals in 3 days, with no appeal path. It is a single anecdotal report.
- Signal
- 44
- Facts
- 16
- Hype
- 4
Brief
Frozen embeddings with logistic regression beat a local LLM on classification
A vendor blog reports 94.3% accuracy on banking77 at 118 ms on CPU, versus 68% for a zero-shot 14B model. The test used only 100 samples, so treat the gap with care.
- Signal
- 89
- Facts
- 62
- Hype
- 11
Brief
Decision model catches all 40 test prompt injections, but misses some
JobEmber tested loomcycle v1.109.0 on 78 texts at a 0.7 threshold. It caught every injection, but another model missed 9 of 40.
- Signal
- 77
- Facts
- 91
- Hype
- 8
Brief
Sentient launches an open leaderboard for model provenance
The benchmark tests whether models can prove where they came from. Nearly all ten tested methods, including Sentient's own, could be evaded by a determined attacker.
- Signal
- 43
- Facts
- 65
- Hype
- 26
Brief
Show HN: zero-shot classifier on scale-to-zero GPUs
One forward pass per call on RunPod scale-to-zero GPUs. The page could not be scraped, so the claims are unverified.
- Signal
- 30
- Facts
- 2
- Hype
- 21
Brief
Show HN: offline in-browser intent classifier using MiniLM
A roughly 23 MB WASM model with under 50 ms latency after warm-up. The claim that it is 100x lighter than Jev is not supported by the page.
- Signal
- 88
- Facts
- 67
- Hype
- 37