Nate's core move across this cluster is teaching one idea from a dozen angles: everything an AI touches gets broken into tokens, embedded into a shared vector space, and processed by attention — and once you see data that way (text, code, images, audio, protein sequences, even action trajectories), you start seeing where AI can and can't actually help. The A-to-Z literacy guide, the Notebook-LM explainer series, and the "Indiana Jones" tokenization piece all teach this same core model at different depths, because for Nate literacy is a moving target — what counted as advanced in 2023 (knowing what a token is) is table stakes by 2025.
The second axis he keeps returning to is scaling. Model-size scaling gave way to a second law — inference-time/reasoning compute (o1 to o3, Claude's hybrid "extended thinking") — by early 2025, and Nate initially treated reasoning models with near-unqualified enthusiasm ("The New Laws of Intelligence," written days after o3, floats 13 spicy predictions including that ethics itself gets easier as models reason better). By mid-2025 his position visibly cooled: "Scaling Isn't Destiny" and the Apple "reasoning cliff" piece argue that context brittleness, short-horizon memory, and messy real-world data are separate research problems scaling doesn't touch, and that production systems need an explicit escalation architecture — knowing when a model is out of its depth and handing off, rather than trusting a confident-sounding chain-of-thought. By late 2025 and into 2026 the framing shifts again toward measured optimism: METR's time-horizon data shows real, compounding capability growth (the "super-exponential"), but the gap between 50% and 80% reliability is where the actual work lives, and no consumer product has yet cleared the "anticipation" bar (acting before being asked).
The throughline for what stays human: taste (recognizing when technically-correct output is contextually wrong), judgment (knowing what's actually worth solving), and specification/verification skill (defining tasks precisely enough to delegate, then catching failures that look like successes). Nate increasingly frames this as a structural division rather than a competition — AI wins wherever a task needs more held context or perfectly consistent rule-application than a human can sustain; humans win wherever it needs novel judgment, business trade-offs, or knowing which failures actually matter. The operational takeaway: don't compete with AI on execution, and don't outsource judgment either — build the muscle for delegation, spec-writing, and fast verification, because that's the real bottleneck now, not raw model capability.
Key moves
Tokenization is the universal move — any data (text, image, audio, code, protein sequence, decision trajectory) becomes a token stream fed through the same attention mechanism; "if you can tokenize it, you can transform it."
The three components worth actually understanding are embeddings, attention, and positional encoding — temperature, sampling method, mixture-of-experts, quantization, and LoRA are all dials on top of that core.
Treat chain-of-thought with real skepticism: it looks like reasoning but is still token prediction, can be confidently wrong, and collapses (not degrades gracefully) past a complexity "cliff" — architect an explicit escalation/verification step rather than trusting the trace.
Inference-time compute is a real second scaling axis, but by itself doesn't fix context loss, long-horizon intent, or adaptive memory — treat those as separate, unsolved research problems.
Deliberately switch models for different cognitive jobs (voice/exploratory mode to expand thinking, a reasoning model to map and structure it, a different model to refine) instead of running your whole workflow through one chat window.
Give AI the four things it can't guess — situation, goal, constraints, and success criteria — ask for options before drafts, specify exact output format, and add a confidence/verification step for anything with real consequences.
Build a personal context bank (voice samples, working prompts, known-good workflows) so you're not starting from zero on every interaction.
Redact or generalize sensitive specifics (names, dollar amounts, dates) before pasting anything into a tool you don't control.
Taste — the ability to say a technically-correct output is contextually wrong — matters more as models get smarter, and it's buildable fast through obsessive depth in a narrow domain, not just tenure.
Good judgment (what's worth solving, what to explicitly deprioritize, when to trust vs. verify) is now scarcer and more valuable than analysis itself, because AI made analysis cheap.
Delegation is graduated, not binary: models are reliable (80%+) only on short tasks (~30 minutes) even though they're capable (50%) on much longer ones (~5 hours) — operate deliberately in that gap and build spec/verification skill as the horizon extends.
Historical labor-share data (automation destroys tasks, rarely whole job categories) argues against "labor is over" framing — but expect your own job description to be rewritten around whatever isn't yet automatable.
Answer AI hype and AI doom with the same three questions: does this change cost, does it change failure modes, does it change where compute can run — not with a reaction to a headline or a single benchmark.
2025-12-29 — My honest field notes on the super-exponential — METR's task-horizon data shows AI capability doubling every 4-7 months; the 50%/80% reliability gap is where the real work happens.
2025-12-10 — Grab the 8 prompts I built to turn NeurIPS 2025 into decisions — 21,575 papers filtered to six shifts: better long-context attention, model convergence, overstated RLVR reasoning gains, RL scaling working again, diffusion memorization clarified, edge deployment going mainstream.
2025-09-13 — The Universal AI Skill: Good Taste — taste, knowing when correct output is contextually wrong, is the skill that compounds as models improve, and it's buildable through obsession, not just tenure.
2025-09-10 — The Complete AI Reading Guide for 2025 — three reading modes (awareness, retrieval, connectome) mapped to 34 books and five career-specific learning pathways.
2025-09-06 — What If I'm Not Technical Enough for AI? — technical literacy isn't the barrier; if a problem involves words and eats your time, AI can help, no jargon required.
2025-07-09 — The A-to-Z AI Literacy Guide (2025 Edition) — a 26-concept glossary, tokenization through multimodal fusion, as a full mental-model refresh for the GPT-5/Grok-4 era.
2025-06-30 — The Complete AI Learning Roadmap for 2025 — a GPT-5 pre-briefing plus a from-scratch explainer of how LLMs work, paired with a curated course list and a "who to follow" signal list.
2025-06-10 — We're Using AI Backwards — stop using AI only to compress thinking; use voice mode for expansive exploration, then a reasoning model to map the terrain ("cognitive choreography").
2025-06-09 — Let's Talk THAT Apple AI Paper — the "reasoning cliff" isn't proof AI is fake, it's a depth limit that tells you exactly where to architect an escalation lifeline.
2025-06-06 — Scaling Isn't Destiny — scaling gave sharper autocomplete, not durable agency; context brittleness, short-horizon intent, and thin memory are separate unsolved problems.
2025-05-19 — Our Work Norms are 4000 Years Old — a Know-Keep-Prove framework for rebuilding trust and ownership norms now that drafts are free but accountability isn't.
2025-05-08 — We Live in an Emergent Compute World — the shift from deterministic, scripted computing to emergent computing where capabilities are discovered rather than specified.
2025-04-28 — What if It's Already Enough? — argues we over-debate AGI timelines and under-explore how much value today's model intelligence already unlocks.
2025-04-03 — The AI is a Vibe: A Short Manifesto — AI's power is social performance, not cognition, "synthetic sincerity" that requires new literacy to read past the fluency.
2025-04-01 — Beyond Black Boxes — LLM capability isn't binary; it spans nine continuums (factuality, reasoning, transparency, agency, and more) that vary by model and context.
2025-03-04 — Think Before Your Next Move — inference-time compute (o1 to o3, Claude 3.7's hybrid reasoning) as a new, third scaling law alongside model size and data.
2025-02-25 — Indiana Jones and the Rise of Tokenized Culture — multiple models pitching Indiana Jones sequels all converge on the same MacGuffin, revealing how token-probability shapes AI (and human) creativity.
2025-02-17 — Surfing Generative AI: A Long View — a history of AI from symbolic systems through transformers to 2025's agentic frontier, and what each major lab is chasing next.
2025-01-03 — When Machines Think Like Aliens — proposes judging AI on its own "alien" terms (distributed identity, emotion-free cognition) instead of only by comparison to humans.
2024-12-29 — The AI Twin Paradox — open-source replication makes duplicating a model free even as original innovation stays expensive, a paradox for competitive strategy.
2024-12-22 — The New Laws of Intelligence: A Manifesto for the AGI Age — 13 "laws" written days after o3's debut: intelligence is abundant but under-deployed, reasoning may be inherently less agentic, and AI ethics may get easier as models reason better.
2023-04-01 — Strange symbols: Building in the Age of AI — Nate's earliest framing here: building through the third major tech transition since 1990, and what actually endures.