> be me, Jonathan Ross > Stanford compilers nerd, allergic to inefficiency > spend years staring at CPUs doing “clever” things slowly > realize the problem isn’t software > it’s the damn hardware lying to you > > modern chips: > caches guessing > branch predictors gambling > schedulers shrugging > latency vibes-based > > decide this is stupid > ask a forbidden question > “what if the chip just… did exactly what the program says?” > > no caches > no speculation > no surprises > no excuses > > design a chip that executes instructions > in order > every time > at ludicrous width > > 144-wide VLIW > SRAM everywhere > DRAM nowhere > timing so deterministic you could set a watch by it > > everyone else: > “this will never work” > “developers don’t want this” > “where’s the cache” > > ship it anyway > > fast forward > LLMs show up > turns out transformers are just > embarrassingly parallel math > predictable > repetitive > allergic to cache misses > > GPUs: > fast but moody > great at throughput > awful at consistency > latency spikes like a heart monitor in a horror movie > > Groq: > same tokens > every run > same timing > zero jitter > silicon doesn’t lie > > developers blink > “wait… my inference latency is fixed?” > > hyperscalers squint > “this thing is weird” > “why does it feel faster than the FLOPs say” > > because FLOPs were never the point > time was > > invent the anti-GPU > no training glory > no benchmark cosplay > just raw, deterministic decode speed > > AI agents start caring about tail latency > suddenly p99 matters more than TFLOPs > > NVIDIA notices > leather jacket energy intensifies > realizes inference ≠ training > and determinism is not optional > > signs a $20B licensing deal > not to replace GPUs > but to admit something quietly > > inference isn’t training > and determinism is a feature > > moral of the story: > everyone optimized for average speed > Jonathan optimized for truth > > turns out the future of AI > needs at least one chip > that always tells the truth