If I understand correctly, they already exclude AI overviews for under-18 users in the EU to comply with local laws.
2,143 karma · joined October 30, 2013
It's been some time since I've thought about Van Eck phreaking: https://en.wikipedia.org/wiki/Van_Eck_phreaking
>Each Party acknowledges and agrees that a breach or threatened breach by LICENSEE of any of its obligations under this Agreement would cause the LICENSOR irreparable harm for which monetary damages would not be an adequate remedy and agrees that, in the event of such breach or threatened breach, the LICENSOR will be entitled to equitable relief, including a restraining order, an injunction, specific performance, and any other relief that may be available from any court, without any requirement to post a bond or other security, or to prove actual damages or that monetary damages are not an adequate remedy. Such remedies are not exclusive and are in addition to all other remedies that may be available at law, in equity, or otherwise
>LICENSEE shall use the Database and Data for the Permitted Use only and shall not disclose, release, distribute, or deliver the Database or Data, any portion of the Data or Database, or any other information derived from the Data or Database to any third party without LICENSOR's prior written consent.
* xhigh: 62 score at $14.83 per task for Fable 5.2, vs 50 score at $10.85 per task for 5.1
* high: 62 score at $9.08 per task for 5.2 vs 47 at $7.54 per task for 5.1
* medium: 56 at $3.80 per task for 5.2 vs 43 for $5.56 for 5.1
* low: 54 at $3.10 per task for 5.2 vs 39 at $4.37 per task for 5.1
At medium and low the scores are much higher for Fable 5.2, while actually costing less per task than 5.1
Something like Orca's Agent Dashboard, or something else? https://www.onorca.dev/docs/model/agents-sessions
> [Before writing this, Astra thought for 1,034 tokens. That thinking is hidden. The API returns only its 5 headings: Reconstructing the active map / Mapping maze corridors / Preparing exact maze map / Reconstructing row groupings / Checking channel patterns]
> According to a draft report reviewed by The Washington Post, Cathedral used AI agents to find weaknesses in business software used by Anthropic for administrative and sales tasks. The researchers found ways to access internal data, including meeting transcripts and Slack messages, after running their agents for less than 12 hours.
>You can see the computer age everywhere but in the productivity statistics. Robert Solow (1987)
> Prohibited: high-performance detection technologies used to counter cross-border data transmission security risks, including deep-synthesis forgery detection (266401J); AI cyber offensive capability technologies, such as malicious code generation and mutation, automated mass generation and distributed large-scale deployment of anti-detection variants (266402J); large models trained using China’s Core Data (266501J); and high-performance AI agents designed for highly sensitive scenarios, such as critical information infrastructure, national defense and core government affairs (266502J).
I imagine by this point the models already have the plugin code in their training corpus, plus plenty of RL data from users interacting with the plugins.