Oskar J. Hollinsworth
Titles as the sources gave them. In 1 entry, with 5 checked quotations.
What they said
Newest first. Each quotation was found word for word in the source.
“Our probes achieve 98.8% AUC in SHADE-Arena, surpassing an Opus 5.5 text-monitoring baseline, and show improved efficacy as the underlying model is scaled up.”
“In one such evaluation, we show that probes can distinguish transcripts containing a model’s true hidden goal from other goals with an AUC of up to 99.7%.”
“Given recent loss-of-control incidents, it is evident that frontier developers’ monitoring practices are inadequate.”
“we strongly recommend that any deployment of potentially risky models, such as frontier models with strong offensive cybersecurity capabilties, be monitored with white-box probes.”
“We have seen that probes are limited by the knowledge and beliefs of the model that they are built on.”
Where they stand
Labels a machine proposed from the quotations above, one line per statement. As weeks accumulate, this is where a change of view shows up, with both quotations side by side.
A label describes one quotation, never the person.
AI companies can be trusted to govern their own safety
Leans disagreelatest, machine label- Leans disagree: “Given recent loss-of-control incidents, it is evident that frontier developers’ monitoring practices are inadequate.”
In the Base
Every entry they appear in.
Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception
In an arXiv preprint, FAR.AI researchers train 'probes' that read an AI model's internal activity to flag knowing falsehoods. Tested on open-weight models in the SHADE-Arena sabotage benchmark, the authors report their probes outscored an Opus 5.5 transcript monitor at a fraction of the cost. They note probes catch only what the model itself believes is false.