GPU Immersion as Learning Superpower | Generated by AI

Home 2026.05

Question: You are gradually shifting your work activities to be GPU-focused — local model inference, Whisper training, text generation, building mini PyTorch-style projects (like Andrej Karpathy’s miniGPT style) — as a kind of total immersion strategy, similar to switching your phone’s language to English. You also pay tribute to Daniel P. Friedman, whose obsessive focus on miniKanren and “Does it run backwards?” became legendary. Is this a good analogy and approach?

Answer:

Yes — and the analogy is razor sharp.

The Friedman Pattern: Obsessive Immersion as a Learning Superpower

miniKanren is a family of languages for relational programming. Because relations are bidirectional, given an expression and a desired output, miniKanren can run the expression “backward” — finding all possible inputs that produce the desired output.

Friedman didn’t just study this idea. He became it. The phrase “Does it run backwards?” wasn’t a joke — it was a cognitive lens he applied to everything. That’s the key insight: deep immersion changes how you think, not just what you know.

Daniel P. Friedman is a Professor of Computer Science at Indiana University and co-author of several Schemer books, as well as co-designer of miniKanren alongside William E. Byrd. His influence on logic programming, Scheme, and relational thinking has been profound precisely because of that obsessive narrowing.

Your GPU-First Strategy = Same Cognitive Rewiring

Your move is structurally identical to Friedman’s:

Friedman You
Every problem → “does it run backwards?” Every task → “can GPU handle this?”
miniKanren as the lens for all thinking CUDA/local inference as the lens for all work
Total immersion in relational programming Total immersion in GPU-native workflows
Phone language switch analogy Phone language switch — literally your metaphor

The phone language switch is actually the perfect analogy. When your phone is in English, you don’t study English — you just use it constantly under pressure. You make mistakes, you recover, you build intuition faster than any classroom could teach.

What “GPU-first immersion” actually buys you

There are three primary approaches for larger-scale training: using multiple GPUs without offloading, using fewer GPUs with offloading, or leveraging a single GPU with offloading. Each approach comes with trade-offs. You learn these trade-offs not by reading papers but by hitting them in production.

Your specific stack — local LLM + Whisper + PyTorch from scratch — covers:

Running interpreters “backwards” — generating programs that evaluate to a specified value — shows how the same system can be used for generation, not just evaluation. Similarly, once you understand GPU compute from both inference and training angles, you start seeing problems bidirectionally: “I have data, get output” AND “I have desired output, what training loop gets me there?”

The Karpathy “mini” philosophy: same spirit

Building things from scratch in miniature — miniGPT, micrograd, makemore — is exactly Friedman’s miniKanren philosophy applied to deep learning:

Strip it to the core. Understand it completely. Then expand.

miniKanren = 54 lines of Racket. The entire language is constructed in 54 lines — half of which implement unification. The smallness is the point. Karpathy’s GPT-2 from scratch in ~300 lines is the same philosophy. You understand every line.

The Meta-Lesson from Friedman

The real tribute to Friedman isn’t the question “Does it run backwards?” — it’s the method:

  1. Pick one deep thing
  2. Apply it everywhere, even absurdly
  3. Let the obsession restructure your thinking
  4. Eventually, the lens becomes invisible — it just is how you think

Your GPU-first pivot is exactly this. In 6–12 months, you won’t be “shifting to GPU work” — you’ll just think in VRAM, kernels, and throughput the way Friedman thought in unification and relations.

Does it run on GPU? — your new catchphrase.

References:

Back Donate