How modern AI models, agents and infrastructure work.
Goal-driven loops, tool calls & verification — how an AI agent turns a goal into safe, checked action.
Backpropagation, attention & inference — how neural networks and transformer models actually compute.
Train a real feedforward classifier in your browser and trace every forward pass and backprop update by hand.
Run a real miniature transformer through prefill and autoregressive decoding, with temperature/top-k/top-p sampling and a live KV cache.
Open up a real trained transformer block and inspect every operation — embedding, attention, residual and feed-forward — on one token.
Bits, gates & algorithms — from binary and logic gates through CPU cycles, memory, data structures, threads and client-server computing.
Assemble and run real RV32I-style programs on a five-stage pipeline, with forwarding, caching and a cycle-by-cycle trace.
Sample sensors, run a firmware scheduler and drive a motor or servo on a simulated MCU, with 11 injectable electrical faults.
An eight-bit register, place values and a ripple carry chain — unsigned carry versus signed overflow.
Six gates on a digital trainer — input switches, output lamp and an automatically cycling truth table.
Fetch, decode and execute a LOAD, ADD, STORE and HALT program on a simple accumulator machine.
Register, direct-mapped cache and RAM — locality, conflict thrashing and streaming access traces.
Call-stack frames beside a heap arena — clean unwinding, leaked blocks and a use-after-free access.
Array, linked list, stack, queue and binary search tree — one operation trace each, with comparison counts.
Bubble, insertion and selection sort on eight tagged values — comparisons, swaps and inversions.
Two threads share one counter — scheduler slices, a lost-update race and a mutex-protected critical section.
Clients, network delay, a FIFO queue and one worker — queueing latency, server errors and request timeouts.
Containers, clusters & scale — how modern applications are packaged, scheduled, balanced and replicated across pools of servers.
Namespaces and cgroups on a shared kernel — CPU quota throttling, memory limits and a modeled OOM kill.
Two guest OSes on one host — weighted vCPU scheduling and memory admission control under a hypervisor.
Pod scheduling across three worker nodes — resource requests decide placement, and a node failure evicts pods.
Round robin vs least-outstanding routing across three uneven backends, with queue limits and failure detection delay.
A utilization-based controller reacting to a demand surge — startup delay, replica ceiling and downscale stabilization.
Leader log and two followers — async vs two-copy quorum acknowledgment, replay lag and network partitions.
Invocations meet a bounded environment pool — cold starts, warm reuse, idle expiry and throttling.
Trace packets across public and private subnets — internet gateway, NAT egress, routes and security-group rules.
Edge caches on a geographic map — TTL, hits and misses, stale versions and origin outage.
Competing consumers vs fan-out pub/sub, with acknowledgments, redelivery and repeated side effects.