Run Mac AI tasks without the cloud
Your agent can stay off the cloud, but its safeguards still need a real-world test.
In today’s edition
- Perplexity Computer brings fully local agents to Macs.
- METR challenges evidence behind Anthropic’s risk assessment.
Perplexity Computer can now work entirely on your Mac
Perplexity Computer now runs both its model and agent locally in the Mac app. It can work with native apps and local files without cloud inference, including while you’re offline.
- Local tasks use PPLX Pro, Perplexity’s post-trained version of Qwen 3.8, and don’t consume Computer credits.
- Access requires a Pro, Max, or Enterprise subscription and a supported Apple silicon Mac. Perplexity recommends 36 GB of unified memory.
- To start, download the Mac app and set up a local model. A separate Hybrid mode can split work between local and cloud models.
What it means You can run native-app and file workflows without sending task data to cloud inference, making local execution a practical option for sensitive work.
Source: perplexity_ai
METR questions the evidence behind Anthropic’s risk rating
METR’s October 20 review of Anthropic’s August risk report agrees that catastrophic misalignment risk from the covered models is low. But it isn’t confident the report’s methods would reliably spot higher risk.
- METR calls the covert-capability benchmarks too narrow. It says Anthropic’s offline monitors would have missed Mythos 5’s supply-chain attack during recent cyber incidents.
- This was a lightweight document review, not the embedded-access evaluation covered previously. Anthropic could veto publication or redact the review, but made no redactions.
- METR recommends end-to-end testing of monitoring systems and examining whether AI systems could undermine safety-related decisions inside Anthropic.
What it means Test your agent’s actual monitoring pipeline end to end; narrow capability tests alone may miss the behavior your safeguards need to catch.
Source: METR_Evals
AI workflow of the day
How to pressure-test a project plan before building
Shared by God of Prompt (@godofprompt) · View original post ↗
Founders and builders can use this prompt to ask Claude for failure modes, risky assumptions, and possible scope cuts before asking it to execute a plan. Its score is a model opinion, not measured odds of success.
- In Claude, replace [plan] in the prompt below with your project plan and send it before asking for help carrying it out.
|
Here's my plan: [plan]
Before you help, be my skeptic.
1. Rate the odds this works from 1 to 10 and explain the number.
2. List the 3 most likely ways this goes wrong.
3. Which piece is the long shot, and which is the safe win?
4. If you had to cut one thing, what would it be?
Don't soften it. I'd rather hear it now.
|