HRL Laboratories, LLC

AGI requires more than just scale. Our Neural Networks paper shows a >6x performance boost on a 3B LLM agent by teaching it metacognition. 🧠   #Problem: ⛔ AI agents using only the scaling paradigm (more data, bigger models) fail catastrophically in novel situations. We hypothesize a critical missing ingredient is metacognition—the ability to "think about our own thinking."   #Solution: 💡 Our MUSE (Metacognition for Unknown Situations and Environments) framework introduces two key metacognitive functions: (1) Self-Assessment: Learning to predict its own task success (2) Self-Regulation: Using that self-assessment to select the best strategy and adapt   #Insight: 🎯 Cognition and metacognition are fundamentally different. For example, an agent can be low-performing but highly metacognitive if it knows it’s likely to fail. Both skills need to be continually learned.   Facing novelty, MUSE switches its objective from maximizing reward to maximizing self-assessed competence. It runs what-if scenarios, vetoes paths it predicts will fail, and guides its search to strategies that leverage its strengths.   #Results: 📈 We tested MUSE against leading model-based RL (Dreamer-v3) and LLM (Reflexion): (1) On novel robotic tasks, MUSE: 70% success rate (20 episodes budget); Dreamer-v3: 0% (2) For 7B LLM agents, on novel text-based tasks, MUSE: >75-110% higher success rate than Reflexion (5 episodes budget) (3) Strongest effect on small models: a >6x boost on a 3B LLM agent   #Conclusion: ✅ Metacognition is critical in novel situations. It also makes smaller LLM agents dramatically more robust, facilitating adaptive AI that can learn on the fly and be deployed on the edge.   We need more natural intelligence concepts and brain-inspired architectures for AGI; scale alone likely won’t get us there. Thanks to my co-author Rodolfo Valiente and to HRL Laboratories, LLC for supporting this work.   #Paper: 👇 https://lnkd.in/gjdyCcib   #Metacognition #ContinualLearning #RL #LLM #AGI #EdgeAI #AgenticAI

We just wrapped up the latest Silicon Quantum Electronics Workshop (SiQEW) here in LA and I have to say, after three full days of talks, posters, and discussions, I’m more convinced than ever that semiconductor spin qubits (fast, tiny, and improving) are the technology to take quantum computing where it needs to go.    I opened on behalf of primary sponsor HRL Laboratories, LLC with a message about the importance of identifying shared challenges and of collaboration across the community. I highlighted HRL’s efforts on open-source control software (SpinQICK) and our collaboration with the LPS Qubit Collaboratory (including partners UCLA, U. Wisconsin, and MIT) to get HRL’s SLEDGE devices into more labs and make them more widely available for science.    I was encouraged to hear that same collaborative spirit echoed throughout the workshop. Many other talks featured shared foundry devices (e.g. from Intel Corporation and imec, to start-ups Diraq, Quantum Motion, to national labs, and to increasing numbers of university teams) as well as shared techniques and shared software. Beyond spin qubits, technologies for electrons on noble liquids and solids, i.e. from EeroQ Quantum Hardware and topological qubits, i.e. from Microsoft, also showed similar needs and progress. It was exciting to see the connections between these efforts at SiQEW, ranging from identifying common technical hurdles to common technical solutions in cryocontrol, fabrication, and measurement.  It is evident that the silicon quantum electronics community garners strength from its multiple diverse but nonetheless concerted technologies and practitioners.   But of course this post only nods to a small fraction of what the short workshop covered.  Across the board and from around the world – Si/SiGe, Ge/SiGe, nMOS, pMOS, implanted donor, STM donor, EO, LD, and hybrid – semiconductor spin qubits are making major strides and opening real pathways to practical utility scale.  I left feeling energized and motivated for us all to push even harder in these collaborative directions. SiQEW 2025 truly moved the field forward – many thanks to the sponsors and organizers who made it possible. I’m already looking ahead to SiQEW 2026 in South Korea!