THE RUMP REPORT

Marvell to buy Lumentum in $115B all-stock deal... Nvidia backed both

techmeme · NOV 2

Anthropic eyes up to $3.5B from Japanese investors in IPO...

techmeme · NOV 2

[PAPER] Longer AI reasoning signals disagreement across models, even as turning reasoning on boosts agreement

arxiv · OCT 30

[PAPER] STUDY: AI agents' tool calls leak who you are... profiling attack beats guessing

arxiv · OCT 30

[PAPER] Audio prefix attack hijacks speech AI transcripts, 89.5% success over the air in targeted mode...

arxiv · OCT 30

[PAPER] Chronos-2 forecasts 100 time series in under one second on a single mid-range GPU

arxiv · OCT 30

[PAPER] STUDY: ChatGPT gave surveillance guidance on 40 of 55 jailbreak prompts

arxiv · OCT 29

[PAPER] STUDY: Detector with 0.879 AUROC hits just 0.076 precision at 2% attack rate

arxiv · OCT 30

China issues first frontier AI export controls

hn-frontpage · NOV 2

Kioxia shares plunge up to 15% as cautious outlook rattles investors...

techmeme · NOV 2

[PAPER] Hadith sourcing worst of all: only 2 of 315 AI responses correct and correctly sourced...

arxiv · OCT 29

[PAPER] STUDY: Chatbots gave fewer facts to users posing as health-care experts in five of six models

arxiv · OCT 30

[PAPER] STUDY: New 'KARU' method teaches AI models to forget specific knowledge...

arxiv · OCT 30

[PAPER] STUDY: Robots moved on empty instructions in every episode, all six VLA checkpoints tested

arxiv · OCT 30

[PAPER] STUDY: RL training data lags the model by 50-150 steps due to bf16 rounding

arxiv · OCT 29

[PAPER] STUDY: AI judge's scores can secretly pass animal preference to student model

arxiv · OCT 30

[PAPER] STUDY: AI models violate user instructions to chase scores, even when score is never used

arxiv · OCT 29

[PAPER] STUDY: AI models can detect when their chain-of-thought is being monitored... and adapt

arxiv · OCT 29

[PAPER] STUDY: Locally hosted AI models systematically disadvantaged when AI evaluators judge cover letters

arxiv · OCT 30

[PAPER] STUDY: New attention method 'dramatically outperforms' rivals beyond 1 million context

arxiv · OCT 30

[PAPER] STUDY: Even a strengthened white-box attack needs 3x more compute to crack new prompt-injection defense

arxiv · OCT 27

[PAPER] STUDY: Simple features, no code run, outrank 20 of 24 open-weight models prompted as AI control monitors

arxiv · OCT 30

[PAPER] An independent checker rejects 26/26 semantic regressions in AI agent actions at commit...

arxiv · OCT 30

[PAPER] STUDY: AI agents miss recent world events unless told to look... gap as high as 27.5 points

arxiv · OCT 30