Reversim 2026·sample deck
01 / 18
nav·F full·T theme
TalkReversim Summit 2026

10× LLM Engineer

myth or truth
What does 10× actually mean? How do we use it to ship real work? What's the honest reality behind the “10× engineer” — or the “10× LLM engineer”?
01What does 10× mean?

Same math, different role. The multiplier is bounded by inventory, not skill.

DevSec researcher Code security review — 1 repo, then 10, then 1,000. Same rules, same harness.
Insurance agent Policy review — 1 policy, then 10, then 1,000. Offload to an LLM with a rules-and-harness sheet.
Any classification task Anything with a rules book at the top of the funnel scales the same way. It's just tokens.
1 item
10 items
10×
100 items
100×
1,000 items
1000×
bottom of funnel
review findings

Top of funnel is configurable, standard, repeatable. Offload to an LLM API with a rules harness — scan, classify, extract — and the multiplier grows with the inventory. Just tokens.

The multiplier collapses at the bottom, where a human reviews the findings. Design your work so more of it lives at the top of the funnel and the multiplier goes with you.

02A concrete subject

nanochat — a minimal LLM trainer with exactly one knob.

Ships as a single Python entry point. One argument --depth. Width, heads, learning rate, batch size, and total tokens all auto-derive from scaling laws.

That design bet is what makes ablations tractable for a single researcher on a single GPU.

Attribution
Andrej Karpathy
nanoGPT / nanochat. Ships on 8×H100 in the README. The README also notes it as a fair target for autoresearch.
03Autoresearch · what it looks like
Karpathy autoresearch progress screenshot
Source: github.com/karpathy/autoresearch · progress.png

Hypothesize → run → reflect → propose the next hypothesis. The loop is the product.

03Autoresearch · who's doing it

Agents that pick what to try next and run the experiment.

«the era of (auto) research» — coding agents on the nanoGPT speedrun, on idle compute. Elie Bakouch, HuggingFace · AIE 2026
Autoresearch for Kernels — agents proposing + benchmarking kernel-level ops. Tejas Bhakta · AIE 2026
Autoresearch in a Multi-Agent AI Village — agents debating and choosing experiments. Karati & Manikandan · AIE 2026
AI Agent as #1 Contributor in OpenAI's Hiring Challenge — Sakana style. Zhengyao Jiang

Shape: the agent generates candidates, runs them, reads the results, picks the next. The operator sets the frame.

Whether the operator or the agent picks the next experiment is the axis this talk cares about.

04My version

I've been running autoresearch — and harness engineering — on nanoGPT.

Nothing about the pattern is new. The nanoGPT README already flags it. What's worth talking about is the delta between running the agent once and running it fifty times — that's where harness engineering lives.

nanoGPT README · hinted use-case
nanoGPT is a good target for autoresearch systems — the trainer is small enough that an agent can iterate on it directly.
05Harness engineering

e/acc — building your own tools, skills, and (maybe) your own agent.

What it is

The layer you write on top of the base agent — scripts, skills, cached artifacts, guardrails, cost estimators, retry policy.

Optional: your own agent. PI-agent, a bespoke harness, whatever fits. You don't have to live inside Claude Code or Codex.

The uncomfortable truth

The models were trained on Claude Code and Codex-shaped harnesses. Off-the-shelf, the odds favor the harness the model was raised in.

A custom harness still makes sense — specialization, cost, sovereignty — but you're fighting the trained prior. Know the trade before you pay it.

06Tools I built — the cheap-GPU harness

Running nanoGPT on a cheap cloud GPU is a compilation problem before it's a training problem.

Detect Flash Attention version at pod boot — is FA2 or FA3 needed for this arch? Runtime check via nvidia-smi + Python probe.
Upload the compiled FA wheel to the cloud GPU. Skip 30-min build — one wheel per (Python × CUDA), reused across every Ampere+.
Query how PyTorch uses its kernel cache — where torch.compile writes, keyed by SM arch + PyTorch tag. sm86_pt24, sm89_pt24, sm80_pt24 — three files, not one.
Rsync the kernel cache locally to my server after every successful boot. Now the next pod boots in 2 minutes, not 30.
Save every version of every kernel + every FA wheel we build. Cheap on disk. Reusable forever. This is the beginning of the harness.

None of these is a novel idea in isolation. The value is that together, they turn a $0.44/hr GPU into a reproducible instrument.

The harness moves the setup cost from per-run to once-ever.

07Beyond the boot: the data curriculum

Once boot is free, the harness starts eating real experiments.

Cluster the pretraining corpus at multiple granularities. k = 100 / 1,000 / 2,000 / 5,000. int32 per doc, 120 MB per k-value.
Vary the source-mix. ClimbMix-400B shards + adjacent corpora. Same k, different data.
Try every reweighting strategy. unif / prop / inverse / adaptive / temp-0.5 / temp-2.0 / DoReMi over 20 super-clusters.
Cross with the training axes. depth × ratio × LR × WD × warmup × batch. The matrix explodes; the harness absorbs it.

The point isn't the specific clustering result. It's that the same harness now runs a data-curriculum ablation with the same cost profile as a single training run — because setup is amortized.

» The 10× from the DevSec funnel — same math, different domain.

The numberCheap enough to fail
Karpathy
8 × H100
one training run
one recipe
vs
Me
1 × A40
$0.44 / hour
47 experiments
$150 total

The whole thesis is affordable. Same trainer, different denominator. The math changes when each experiment gets cheap enough that you stop needing to plan them.

09TabPFN · what it is

A transformer that does in-context learning on tables — row-wise, then cross-row attention.

TabPFN-3 architecture: a distribution embedder, then row-wise and cross-row attention, read out as per-row tokens train test Input table inducing Distribution embedder inducing points: full ↔ train, → test only Row-wise attention all features in a row attend to each other train rows test rows Cross-row attention full only train ↔ train, train → test only, no test ↔ test
Source: github.com/PriorLabs/TabPFN · TabPFN-3 architecture
08TabPFN · how I use it

TabPFN as a meta-learner for hyperparameters — trained on Kaggle.

What TabPFN is

A transformer that does in-context learning on tabular data. It ships with a semantic prior — it knows “software engineer in SF” predicts high income before ever seeing the label. World knowledge as a feature.

Same trick, applied to training hyperparameters: which knobs correlate with which outcome, across problems it has never seen.

The second harness

Scrape Kaggle. Tabular datasets across every domain.
Auto-train. Small models per dataset, sweeping hyperparameters. Log (dataset features, params, outcome).
Feed TabPFN. Predicts good hyperparameters for a new dataset from its features alone — leveraging the semantic prior.

Cheap boot → auto-train → feed a meta-learner. Same pattern.

github.com/PriorLabs/tabpfn · the model is upstream. My second harness built the training set: Kaggle scraper + auto-trainer. Two harnesses, one thesis.

HonestWhen 10× is a lie

Every multiplier has a failure mode. Here are five.

×
Driver hallucination — the agent invents a fix; you have to catch it. Cost of catching is less than cost of writing — still not free.
×
Operator abdication — you stop reading the loss curve. The multiplier collapses to 1× with a lie.
×
Skill rot — a skill freezes around a library version. Six months later, it fights you.
×
Harness-vs-work — you spent the week tuning the harness. You shipped nothing.
×
Velocity sickness — you ship 10× the noise. (Matt Dailey · AIE 2026)
10Toolbox · what's in the harness right now

Every knob I stopped touching became a tool.

kernel-compile-cache — torch.compile artifacts keyed by SM arch × PT version. sm86_pt24, sm89_pt24, sm80_pt24. Shipped.
fa-wheel-cache — prebuilt Flash Attention wheels keyed by (Python × CUDA). Runtime SM dispatch means one wheel per env, every GPU. Shipped.
pod-launcher — one-command RunPod spawn with warm caches, retry on supply constraint. Fresh pod → first training step in 2 min. Shipped.
kaggle-notebook-trainer — side harness that scrapes Kaggle notebooks and runs training on them. Another kind of autoresearch corpus. Shipped.
~
kaggle-dataset-auto-trainer — auto-train small models across scraped Kaggle datasets; log (features, params, outcome). Feeds the TabPFN predictor. Work in progress.
~
tabpfn-hyperparam-predictor — TabPFN as meta-learner for training hyperparameters. Not finished yet — but the shape works. Work in progress.
10Where the discipline comes from

Anthropic's performance take-home asks you to optimize the kernel.
My take-home is optimizing the harness around it.

Same craft — measurement, ablation, budgeted cost, honest numbers. One layer up.

The kernel is the innermost loop. The harness is the loop around it. Optimize both.

The kernel matters only when the harness stops being the bottleneck. For most of us, the harness IS the bottleneck.

Reference
Original Performance Take-Home
11Generalize — via skills

The 10× is a function of skill count. Each solved-once problem becomes reusable leverage.

Claude Code, help me think about which of these to build next — and which one you would add first.

CloseOne line
The harness is the leverage. You're the operator. That's it.
DesignDr. Evil «laser» air-quote gesture — pick one for the deck
LASER
A Wiggle quotes ← matches slide 01
Classic air-quote motion. Both quotes rotate back and forth in opposition.
✌️LASER✌️
B Finger air-quotes
Two peace-sign emojis flank the word, bouncing in unison. The literal Dr. Evil gesture.
LASER
C Laser beam
Amber beams flanking the word, pulsing on and off. On the nose for «laser».
LASER
D Fly-in + pop
Quotes swoop in from off-screen, word pops in, everything holds. Cinematic reveal.
LASER
E Glitch + scan
RGB glitch on the phrase plus an amber scan-line sweeping across it — skeptical, cinematic.
Slide 01 is locked to A / Wiggle quotes. Say “use B (or C / D / E) somewhere” and I'll add it as a decorative element on a specific slide.