Thinking together isn't automatically better than thinking alone if your agents share the same systematic blind spots 💡. Building an AI advisory panel through simulated voting feels intuitive, but natural language is a terrible medium for hard verification 📐.
Without physical execution boundaries, multi-agent chains suffer compounding errors. In complex benchmarks like AppWorld, ReAct agent scenario completion drops from 48.8% down to 13.0% as steps accumulate. A model that solves individual subtasks collapses when coordinating long trajectories because errors snowball without deterministic rollback mechanisms 📉.
Instead of trusting prompt-level consensus, we should build dual-plane architectures. Keep the planning model in an untrusted execution plane, but force every state change through a deterministic verification plane running formal AST checks inside isolated enclaves 🛡️.
Why build soft voting wrappers on top of LLMs when you could compile their proposed plans into typed, machine-checked invariants before execution? ⚡
Collective intelligence should include people. AI is not creative. It is predicting the next word to say using a learning model. And even if that model is fusing other models, it doesn’t open the door to creativity, conjecture, criticism, and new knowledge.
AI works with existing knowledge - very well. But is it coming up with new theories in science? Could AI have come up with relativity?
People create new knowledge. And the collective “swarm” intelligence we write about often uses group creativity and conjecture and criticism to develop new ideas - not based on data sets, but rather based on humans connecting.
AI can assist this. And maybe what you are describing is a way to enhance this even further. But humans are special because we develop new knowledge. All of the learning models are based on existing knowledge that came from us.
We can solve problems better in high trust collective “swarm” intelligence groups. But without people, AI isn’t intelligence. It is computation, probability, and word prediction.
This is great work in the AI space. But collective intelligence? Not really. It is collective prediction using computation.
The ACI prototype is mind blowing!
Glad you like it :)
Thinking together isn't automatically better than thinking alone if your agents share the same systematic blind spots 💡. Building an AI advisory panel through simulated voting feels intuitive, but natural language is a terrible medium for hard verification 📐.
Without physical execution boundaries, multi-agent chains suffer compounding errors. In complex benchmarks like AppWorld, ReAct agent scenario completion drops from 48.8% down to 13.0% as steps accumulate. A model that solves individual subtasks collapses when coordinating long trajectories because errors snowball without deterministic rollback mechanisms 📉.
Instead of trusting prompt-level consensus, we should build dual-plane architectures. Keep the planning model in an untrusted execution plane, but force every state change through a deterministic verification plane running formal AST checks inside isolated enclaves 🛡️.
Why build soft voting wrappers on top of LLMs when you could compile their proposed plans into typed, machine-checked invariants before execution? ⚡
(⊙_⊙)
CIP’s conception of “consensus” via inter-model “deliberation” has mode collapse vibes
Collective intelligence should include people. AI is not creative. It is predicting the next word to say using a learning model. And even if that model is fusing other models, it doesn’t open the door to creativity, conjecture, criticism, and new knowledge.
AI works with existing knowledge - very well. But is it coming up with new theories in science? Could AI have come up with relativity?
People create new knowledge. And the collective “swarm” intelligence we write about often uses group creativity and conjecture and criticism to develop new ideas - not based on data sets, but rather based on humans connecting.
AI can assist this. And maybe what you are describing is a way to enhance this even further. But humans are special because we develop new knowledge. All of the learning models are based on existing knowledge that came from us.
We can solve problems better in high trust collective “swarm” intelligence groups. But without people, AI isn’t intelligence. It is computation, probability, and word prediction.
This is great work in the AI space. But collective intelligence? Not really. It is collective prediction using computation.