01 / Pacing
The pause is a feature.
Every stage ends at a decision point: continue, ask a follow-up, or turn the explanation into a revision card.
ContinueAsk whyMake a card
The Liminal Method01 / 06
The summary is where we begin.
Choose a hard ML or SWE concept. LiminalML takes it beyond the overview through intuition, derivation, implementation, interview probes, and unaided retrieval.
02full sessions$0no account157indexed topics
live study thread / attention mechanismssaved
stage 03 / 06
The Math
Derivation
Why divide by √dₖ?
Var(q · k) ≈ dₖ. Without scaling, softmax saturates and the useful gradient disappears.
Attention(Q, K, V) = softmax(QKᵀ / √dₖ)V
every term motivated3 of 6
The tutor pauses here. You decide when to continue.
Ask a follow-up, or move forward when it clicks.
Classical MLDeep LearningReinforcement LearningTraining EngineeringSystems and MLOpsFrontendBackendSystem DesignUI/UXCS FundamentalsClassical MLDeep LearningReinforcement LearningTraining EngineeringSystems and MLOpsFrontendBackendSystem DesignUI/UXCS Fundamentals
Built for active recall
A study partner with a point of view.
Chat is optimized to keep answering. LiminalML is designed to stop, let you think, and make you produce the next step yourself.
02 / Implementation
def attention(q, k, v, mask):
scores = q @ k.transpose(-2, -1)
scores /= q.size(-1) ** 0.5
scores.masked_fill_(~mask, -inf)
return scores.softmax(-1) @ v test passes / gradients finite
03 / Memory
Leave. Come back. Keep going.
Every thread saves at the exact stage where you stopped. No repeated setup and no lost follow-up questions.
Attention mechanismstoday
KV cache3d ago
04 / Your context
Your work becomes the test.
Add a resume only if you want to. Stage 6 can then ask you to defend choices you actually shipped.
grounded retrieval
Your retrieval system had a 30ms budget. Which attention layer would you remove first, and what quality loss would you expect?
The six-stage loop01 / 06
Orient
See the whole idea before the machinery.
Start with the problem, the constraints, and the mental model that makes every later equation feel inevitable.
Leaves you with a durable mental model.
The real interface
Don't take the landing page's word for it.
Run a scripted Attention Mechanisms session here. It uses the exact rendering, pauses, math, code, and diagrams of the live product without spending a model call.
Demo session · Attention Mechanisms
Scripted replay · no model callsTry it right here
Watch a full session run.
A recorded replay of a real six-stage session, including the exact pauses where the tutor stops and waits, plus two follow-ups you can ask. No account, no model calls.
6 stages · the same math, code, and diagrams a live session renders · ~3 minutes
Six stages, in order, never compressed. The pause after each one is the point.
Read one first, no sign-up
17 concepts written in the session format.
Deep LearningAttention MechanismsHow scaled dot-product and multi-head attention work; the soft key-value lookup at the heart of every Transformer; with the math, runnable PyTorch, and calibrated interview questions.9 min · 6 stagesDeep LearningTransformer ArchitectureThe Transformer block from the ground up; self-attention plus a position-wise feed-forward network, residuals and LayerNorm, and the encoder/decoder configurations; with the math, PyTorch, and calibrated interview questions.9 min · 6 stagesDeep LearningBackpropagationBackpropagation as reverse-mode autodiff; the chain rule over the computational graph, the gradients for a linear layer and ReLU, and why gradients vanish; with a runnable manual backward pass.9 min · 6 stagesDeep LearningBatch NormalizationWhat Batch Norm normalizes and why, the critical train-vs-inference distinction, BN vs. Layer Norm, with the math and a from-scratch PyTorch implementation.8 min · 6 stagesOptimizationGradient Descent (SGD, Momentum, Adam)SGD, momentum, and Adam explained; the update rules, why mini-batching wins, Adam's bias correction, and when plain SGD generalizes better; with from-scratch implementations.8 min · 6 stagesLLMsKV CacheHow the KV cache makes autoregressive LLM decoding affordable; what it stores and why reuse is valid, the memory cost, why decoding is memory-bandwidth-bound, and how MQA/GQA shrink it; with code.8 min · 6 stages
Two tracks, one method
Choose the depth. Keep the structure.
Interview prep is the sharpest use case, but the product is built for any engineer who wants a concept to survive beyond the tab they learned it in.
ML / Research71 topics
From classical models to training systems.
Transformers, optimization, RLHF, distributed training, serving, and the mathematics underneath modern systems.
Software engineering86 topics
From browser internals to distributed systems.
Rendering, caching, databases, data structures, backend architecture, and production tradeoffs.
Pricing
Try it before you exist to us.
Only Pro takes a card. The trial does not charge for seven days and cancels in one click.
Guestno account
$0
2 full sessions, then sign in
Two complete six-stage sessions
All 157 topics, all 17 concept pages
No email, no card, no profile
No practice lab, nothing saved between visits
Freesign in
$0
8 sessions a month
8 sessions a month, no card
Saved threads you can resume
5 practice-lab code reviews a month
Revision cards
Stage 6 grounded in your resume, not generic examples
Pro7-day trial
$9 / month
Unlimited sessions
Unlimited review sessions
Unlimited practice-lab code reviews
Everything in Free
New topics ship to Pro first
Request topics directly
Start before you sign up
Pick the concept you keep almost understanding.
Leave able to derive it, implement it, and explain where it fails.
Start a deep study session2 sessions without an account · 8 / month when you sign in · unlimited on Pro