Interview drills

Do these out loud, timed, recorded. A silent "I could do that" is not practice.

Implementation (90 min energy, 3×/week from P3 on)

Build the muscle CodeSignal actually hits: multi-part problems that grow.

Rotate:

In-memory database: SET/GET/DEL, then TTL, then SCAN, then transactions. Each part extends the last. Speed matters.
Stack-sampling profiler → Chrome-trace events. Parse messy input, nest stacks, emit JSON.
Tiny continuous batcher: requests arrive with different prompt lengths; you emit batches under a token budget.
KV cache allocator: pages of tokens, allocate/free, fragmentation, prefix reuse.
Toy all-reduce: simulate ranks, ring vs tree, count bytes on the wire.
Correctness harness: two implementations of softmax / GEMM, compare, print the first mismatch with dumps (this is the job).
Multi-thread crawler or pipeline with backpressure (reported Anthropic-like flavor).

Rules: no LLM in the timed window if you want transfer to the rounds that ban it. Look up syntax. Do not look up the design.

Blank page. Camera on. If you peek, you failed.

Attention: every tensor shape, FLOPs, bytes, why /sqrt(d).
Online softmax: the m,l recurrence, why it is equivalent.
FlashAttention IO complexity vs standard.
KV cache bytes for a named config (you pick GQA 70B-class: 80 layers, 8 kv heads, 128 dim, 8k, batch 16, fp16).
Backprop through one attention head.
Arithmetic intensity of decode GEMM vs prefill GEMM.
Collective volumes for tensor-parallel MLP (column + row).
Why Pre-LN + RMSNorm is stabler than Post-LN at depth.
RoPE: rotate pairs of dims, why it gives relative positions.
Mixed precision: where you keep fp32 (master weights, loss, accumulators) and why.

Prompt: "Design an inference API for a frontier chat model."

They will push on:

Prefill vs decode (different utilization, different kernels, different batching)
Variable-length requests
When to flush a batch vs wait
KV memory as the real capacity limit
Paging / prefix cache
Streaming tokens to the client
Tail latency vs throughput
Autoscaling: queue depth weighted by expected tokens, not GPU util
Multi-tenant fairness / credits (one reported question was GPU scheduling with credits)
Failures: stuck decode, accelerator XID, partial stream
Heterogeneous accelerators (GPU/TPU/Trainium) if staff-level

A passing answer names the bottleneck of the current design before the interviewer asks.

Have a real view, written once, then thrown away so you speak it:

What Constitutional AI does, step by step, and one failure mode (sycophancy, over-refusal, principle conflict).
A capability you would not want scaled without a corresponding control.
A time you picked the simple thing that worked over a clever one.
Why this lab, not "a frontier lab."
Something you changed your mind about after reading their work.

If your answers could be pasted onto an OpenAI or xAI application unchanged, rewrite them.

Draft weekly from P1, in the Greenhouse shape:

The most impressive low-level or performance thing I've done: …

Keep it to one paragraph + URL. Update it when the artifact earns it. This is the application.