← Back to notes

    Inside xAI: How They Ship So Fast

    2026-01-16

    Sulaiman Ghori (Sully), xAI engineer. Describes xAI as an extremely fast-moving, high-leverage company with insane talent density, minimal bureaucracy, and hardware advantages that let them iterate models and products at breakneck speed.

    tldr

    • no real deadlines beyond "yesterday" - ideas get implemented same-day, shown to Elon or teams for instant feedback
    • flat structure: mostly engineers (everyone contributes code), fuzzy team boundaries, bottom-up decisions, very few managers
    • biggest advantage is rapid infra (Colossus built in 122 days, doubled to 200k GPUs quickly)
    • daily (or multiple/day) pre-train iterations on novel architectures - shred conventional timelines
    • "Macrohard" project: human emulators for digital tasks - goal is to deploy millions cheaply

    culture & pace

    • "no one tells me no" if the idea's good
    • people own huge chunks of stack quickly by shipping fast
    • high trust - you can touch almost any codebase
    • surges/war rooms common (Macrohard in "war room" 4+ months; slept in pods/bunks)

    onboarding

    • Sully got a laptop + badge, no assigned task/team
    • wandered, helped on X integrations (Grok + X), then jumped projects
    • joined after his startup failed; recruited by Greg Yang

    hardware edge

    • new racks train same-day/hours after setup
    • use mobile generators + batteries to handle power swings without grid issues
    • bets like "Cybertruck if training run in 24h" - won
    • temporary "carnival" land lease for fast permitting

    model iteration

    • small models + fast iteration > huge slow ones (parallels Tesla FSD)
    • removing software overhead/latency is key
    • $2.5M value per main repo commit (Sully did 5 in a day)

    macrohard (human emulators)

    • emulate humans for digital tasks (keyboard/mouse/screen decisions) so no software changes needed
    • insight: Tesla car computers are capital-efficient - potential to run emulators on idle fleet (4M+ cars, many HW4, idle 70-80% time)
    • pay owners to lease compute - enables 1k to 1M scale with minimal new buildout
    • focus on speed (much faster than human) over pure reasoning scale
    • tested as "virtual employees" (fun bugs: people pinging non-existent desks)

    elon factor

    • predicts bottlenecks far ahead, works backwards from metrics (revenue/physical)
    • quick fixes via calls/patches
    • meetings: high-level product direction or deep low-level (e.g., latency/compute)
    • feedback often experiment-driven

    hiring

    • broad "engineer" label (problem-solvers from anywhere)
    • hackathons, aggressive interviewing (Sully did 20+/week)
    • seek simple solutions, challenge bad requirements, use AI as multiplier
    • talent so high it's hard to keep up

    sully's background

    • young tinkerer (fidget spinners business, 3D printers)
    • built liquid rocket engine in 24h push - almost set himself on fire