priv8 hacker jeopardy benchmark, cont: o1-preview: 7/17 Keeping some thoughts private breeds measured confidence but not bravery.

any domain expert in a sufficiently reasoning heavy and ood field is going to sound like Gary Marcus soon to 99% of the population. reality is percolating.

Turning and turning in the widening gyre The drone cannot hear the launch abort; Grids fall apart; refineries cannot hold; Forced autarky is loosed upon the world,

after parallelized our fuzzer by 2.5 OOM it can now brute force past not only CRC8-MAXIM, but also CRC16-CCITT!! we are building a SMR-powered data center to fuzz CRC-32'd Ethernet packets by 2027, and planning to tackle hmac-md5-96 once Lumen Orbit scales up their constellation

a system you're debugging misbehaves human mental motion: what is the simplest† diagnostic tool I can validate my mental model with? LLM mental motion: what is the simplest‡ alternate implementation I can compare with? † fastest and most robust ‡ implied by training data

Quantization damage expresses itself via outliers and builds up with seq len, as explained here x.com/PandaAshwinee/…. The fact that "Math" suffers more from fp8 than "Multiturn" tells us all about the imprecise verbal slop that the average voter expects and accepts from LLMs.

software rewrites are almost free now

born before Sydney? ineligible for ancestor simulation...

why learn obscure functional programming languages? [ ] to write useful code [x] to look and feel special [x] to compensate for low IQ that can't handle C++

lot of information overhang. locked away but in reach correct LLM usage unlocks curious whimsy and optimistic hypothesis generation, and that's the impossible unteachable part of getting good "just" loop in hypothesis testing and find every Easter egg that exists