Brad retweeted
🚨 @AnthropicAI: AI “MAY BE” capable of self-replicating!
Now read the actual research:
• No fully autonomous rogue replicator demonstrated.
• Cyber exploitation deliberately simplified.
• Real-world replication remains unproven.
So we’re turning controlled experiments and speculative projections into alarming headlines now?
AI safety deserves rigorous science, not fear mongering driven marketing dressed up as a breakthrough.
anthropic.com/research/self-…
This is how you reduce your Jev calls while maintaining accuracy:
I have some doubts here. You can build a decision graph with zero Jev nodes — just rules over the state — and get behavior that looks every bit as intelligent.
We love how this doomo doomonstrates real-time intelligence and what can be doone with code + AI!
~10 calls/sec = ~$7/hour
Very underutilized design pattern no one is talking about (yet) for Jev-like models:
cascade.
A small model answers everything. A big model answers only what the small one is unsure about. In my runs the big one handled 13% of questions and the pair matched it alone — 90 vs 91, 145 vs 144.
It works because the confidence is a real probability read from the logits, not a token the model typed.
Cascade engine implemented in my repo (pls ⭐️ thanks)
github.com/bradAGI/ruling
Jev's answers are public for 256 judgments, so I ran the same rows through an open model on my laptop.
Jev: 238. Qwen3.6-35B-A3B, untrained: 231.
Not a significant difference.
The part I care about: anything above read-only runs inside a Seatbelt profile, so the folder you picked is the only place a write lands.
macOS refuses it. The agent is not being asked to behave.