Personal research journal documenting insights, discoveries, and learnings across technology domains
Claude 3.5 Sonnet demonstrates 40% higher task completion in agentic scenarios compared to GPT-4, suggesting model-specific agent architectures.
Circle STARKs with Mersenne31 fields achieve 10x proving speedup, potentially enabling real-time ZK proofs for complex computations.
Willow demonstrates genuine QEC milestone where errors decrease with scale, but practical quantum advantage remains 10-15 years away.
Showing 10 notes
Claude 3.5 Sonnet demonstrates 40% higher task completion in agentic scenarios compared to GPT-4, suggesting model-specific agent architectures.
Circle STARKs with Mersenne31 fields achieve 10x proving speedup, potentially enabling real-time ZK proofs for complex computations.
Willow demonstrates genuine QEC milestone where errors decrease with scale, but practical quantum advantage remains 10-15 years away.
AlphaFold 3 struggles with disordered proteins, allosteric sites, and dynamics - use as hypothesis generator, not definitive answer.
Private mempools paradoxically increase builder centralization - encrypted mempools with threshold decryption may be better solution.
ML-KEM hybrid TLS is ~2x slower with 34x larger ciphertexts, but completes in <50ms - ecosystem compatibility is the real challenge.
BCI security regulations inadequate - need new framework combining medical device, critical infrastructure, and neural-specific protections.
LLMs exhibit "functional reasoning" - achieving reasoning-like outcomes through different mechanisms. Important for AI safety considerations.
Base editing efficiency 3-4x lower in clinical cell types vs lab standard HEK293T - delivery optimization is critical bottleneck.
QML shows no practical advantage yet - promising areas are quantum kernels, sampling, and chemistry ML in 5-10 year horizon.
Have feedback on these research notes? Want to discuss ideas or share your own insights? I'd love to hear from fellow researchers and curious minds.