256.1

Signal Processing
About

SUS: Systems Under Strain

The race toward AGI fixates on scale and benchmarks while ignoring how today's AI systems fracture under real-world pressure. Beyond impressive demos, brittle foundations emerge—revealing critical gaps between laboratory performance and production robustness.

This site examines those breaking points in perception, computation, attention, and memory, exposing where brute-force scaling hits its limits and why genuine machine intelligence demands more than parameter counts.

Perception & Reality Models

How sensor fusion, data bias, and world modeling determine whether systems navigate ambiguity or collapse under it.

Compute & Cognitive Limits

Why architectural constraints and attention dynamics create cognitive bottlenecks that raw processing power cannot solve.

Hardware & Physical Limits

Where material science, interconnect bandwidth, and thermal density impose hard ceilings on AI scalability.

Compiled Thoughts

  • Optical Interconnects: Free or Bonded?

    ♞ FSO achieves 1.6 Tb/s at 2.3 pJ/bit but requires ±5 µm alignment precision
    ♜ CPO adds 100-150ns FEC latency while FSO promises sub-nanosecond communication
    ♝ Environmental vulnerability vs proven reliability defines the optical trade-off
    Anthony Robledo: 50% | AI: 50%
  • Attention Thrashing: ADHD in Artificial Minds

    ♞ O(N²) scaling causes attention thrashing as context windows reach millions of tokens
    ♜ FlashAttention-2 achieves 2x speedup enabling 32K token contexts on A100 GPUs
    ♝ Lost-in-the-middle phenomenon reveals models fail at information retrieval despite size
    Anthony Robledo: 85% | AI: 15%
  • Déjà Vu: Cog in the Machine or Cognizant Machines?

    ♞ Pattern completion in neural networks mirrors human memory's reconstructive nature
    ♜ Transformer KV-cache creates temporal echoes distinct from biological recall
    ♝ Engineering memory systems reveals the gap between storage and understanding
    Anthony Robledo: 90% | AI: 10%
  • Whose Bits are Wiser, GPU | TPU?

    ♞ NVIDIA H100 trains GPT-3 in minutes while TPUs excel at massive-scale efficiency
    ♜ GPUs offer flexibility through thousands of cores; TPUs optimize via systolic arrays
    ♝ AMD MI300X and startups like Cerebras challenge the GPU-TPU duopoly
    Anthony Robledo: 50% | AI: 50%
  • Who's Driving the Autonomous Vehicle Shift?

    ♞ $300-400B market by 2035 with players betting billions on different sensor strategies
    ♜ Waymo's multi-sensor fusion achieves zero injury collisions in first million miles
    ♝ Tesla's camera-only gamble could dominate or bust on safety validation
    Anthony Robledo: 64% | AI: 36%