Bodies Do the Thinking: Why Physical AI Changes the Intelligence Game
A mechanism-first reading of Physical AI as embodied feedback: where sensing, motion, learning, autonomy, and context become one business-critical control loop.
Cognaptus Insights
Industry thoughts, trends, and commentary.
Don’t Self-Sabotage Me Now: Rational Policy Gradients for Sane Multi-Agent Learning
A mechanism-first look at why adversarial multi-agent training breaks cooperative systems, and how Rational Policy Gradient turns sabotage into useful stress testing.
From Yarn to Code: What CrochetBench Reveals About AI’s Procedural Blind Spot
CrochetBench shows why multimodal AI can describe finished objects yet still fail at generating executable, structure-aware procedures.
Plans, Tokens, and Turing Dreams: Why LLMs Still Can’t Out-Plan a 15-Year-Old Classical Planner
A business-readable analysis of why frontier LLMs are getting better at formal planning, but still need symbolic validation before they touch real operations.
Safety in Numbers: Why Consensus Sampling Might Be the Most Underrated AI Safety Tool Yet
Consensus sampling reframes AI safety as probability-level agreement, offering a formal way to reduce undetectable output risks when inspection is not enough.
What We Don’t C: Why Latent Space Blind Spots Matter More Than Ever
A mechanism-first reading of WWDC, a latent flow-matching method for suppressing known signals so residual structure becomes easier to inspect.
When Heuristics Go Silent: How Random Walks Outsmart Breadth-First Search
A mechanism-first reading of when restarting random walks beat breadth-first search for escaping heuristic dead zones in AI planning.
Decoding Intelligence: When Spikes Meet Hyperdimensions
A comparison-based reading of SNN-HDC, a decoding method that trades one-hot outputs for binary hypervectors to reduce spike activity and estimated energy in neuromorphic vision.
Memory, Bias, and the Mind of Machines: How Agentic LLMs Mislearn
A practical reading of why agent memory can turn useful experience into brittle, biased behaviour when systems continuously rewrite their own lessons.
Parallel Worlds of Moderation: How LLM Simulations Are Stress-Testing Online Civility
COSMOS turns online moderation into a counterfactual simulation problem, showing why personalised interventions may reduce toxicity without the collateral damage of blunt bans.