The Technical Stack: What You Must Know to Start in AGI Research

Why Tooling Fluency Is a Prerequisite, Not a Differentiator

This is the first article in a new series — AGI Researcher Foundations — that goes deeper than the Researcher's Field Guide and Day in the Life articles already on this site. Those covered papers, labs, career paths, and daily rhythms. This series covers the actual knowledge base underneath all of that: the technical stack, the math, the core deep learning concepts, and the specialized foundations (generalization, RL, interpretability, alignment, world models, systems, and methodology) that every one of the researcher profiles in this series' Field Guide had to master before their work became notable.

This first article is deliberately the least glamorous: the programming languages, frameworks, compute options, and tooling every researcher uses daily. It's tempting to skip straight to "the interesting ideas" — generalization, alignment, world models — but every lab covered in the Research Frontier Map assumes this layer is already solid before a candidate is evaluated on ideas at all. Fluency here is table stakes, not a differentiator — but the absence of it is an instant disqualifier, which is exactly why it deserves a full article of its own.

Programming Foundations

Every profile in this series' Field Guide — from Sutskever to Neel Nanda to John Jumper — built their technical credibility on top of a small, consistent set of programming skills. None of them are exotic.

Core ML Frameworks: PyTorch vs. JAX

The framework question is less a technical debate than a signal of which lab culture you're entering. PyTorch remains the dominant framework by raw volume — well over 70% of recent arXiv ML papers are implemented in it — and is the default at OpenAI, Meta, Mistral, and Cohere. JAX has a smaller but concentrated footprint, with notable adoption at Google DeepMind, Anthropic, Apple, and xAI, particularly for TPU-heavy training workloads where JAX's functional, compiler-first design (via XLA) gives more direct control over device placement and parallelism.

The practical takeaway: learn PyTorch first — it's the common denominator across nearly every course and paper reproduction in this article — and pick up JAX specifically if you're targeting a DeepMind/Anthropic-style, TPU-centric research role, per the lab-specific patterns covered in this series' Field Guide.

The Research Engineering Stack

Beyond the modeling framework itself, a specific supporting stack shows up across nearly every research team covered in this series' "Inside an AGI Startup Team" articles.

For distributed training specifically, three approaches compete: DeepSpeed's ZeRO optimizer (used historically for models like BLOOM and MT-530B), NVIDIA's Megatron-LM for tensor and pipeline parallelism (often wrapped by DeepSpeed, NeMo, or vLLM's training paths), and PyTorch's own native FSDP2, which shipped in PyTorch 2.6 and increasingly competes directly with DeepSpeed's ZeRO-3 stage for teams that want to stay within the core PyTorch ecosystem rather than adding an external dependency. For inference and serving, vLLM has become the dominant engine via its PagedAttention design and continuous batching, and has expanded into RL training infrastructure as well — see its 2026 "vime" RL framework announcement.

Compute Literacy

Understanding what "training on 8xH100" actually means — cost, memory constraints, interconnect bandwidth — is itself a research skill, not just an operations detail. For someone starting out, several real, current options exist at very different price and reliability points.

Pricing sourced from Jarvis Labs' 2026 cloud GPU provider comparison and Spheron Network's RunPod vs. Lambda Labs comparison; confirm current rates directly before committing to a provider, since GPU spot pricing shifts frequently.

Data Engineering Basics

Data literacy is as important to research outcomes as modeling skill, a point emphasized throughout this series' coverage of Sutskever's own generalization thesis in the SSI investigation — data quality, not just model architecture, increasingly determines what a model can and cannot do.

Evaluation & Debugging Tooling

Eval methodology has become its own specialized discipline, as covered in the Frontier Map's "unclaimed territory" analysis. Two frameworks anchor this space, though their status has shifted recently:

AI-Assisted Research Tooling

As documented in this series' Day in the Life article, AI coding tools are no longer optional convenience — they're part of the baseline stack. Claude Code and GitHub Copilot are used daily for infrastructure and tooling code; Goodfire's Silico platform (covered in the Lane 3 team article) shows the more advanced end of this trend, where an AI agent plans and runs its own research experiments. For a newcomer, the practical skill is knowing when to lean on AI-assisted coding for boilerplate and infrastructure, versus when a research result specifically requires your own from-scratch implementation to build real understanding — a distinction covered directly in the paper-reproduction scenario below.

Key Papers to Read First

These aren't the deep specialized papers covered in later articles in this series (those belong to the generalization, RL, interpretability, and alignment foundations articles) — these are the papers that establish the shared vocabulary the rest of the field assumes you already have.

Courses to Complete

These are real, currently accessible courses — not a generic "learn ML" reading list. Where a course's current-year status couldn't be independently confirmed, that's flagged directly rather than presented as certain.

Real Scenario Walkthroughs

Self-Assessment Checklist

Each of these eight items gets its own deep dive — the why, the how, and sourced references and videos — in Article 2: The Self-Assessment Deep Dive.

What's Missing From This Article

Where This Series Goes Next

This article deliberately stayed at the tooling layer. Article 2 in this series takes the Self-Assessment Checklist above and goes deep on each item — the why behind it, a concrete how-to path with sourced tutorials and repos, and videos for each — before Article 3 moves to the mathematical foundations underneath everything covered here (linear algebra, probability, optimization, information theory). From there, the series builds through core deep learning concepts, generalization and learning theory, RL foundations, interpretability, alignment, world models, systems, and research methodology, before a final capstone article ties every foundational concept back into one unified map, cross-linked to this series' Frontier Map and Field Guide.

🎥 Recommended Videos

🧭 Closing — The Stack Is Learnable, the Judgment Isn't Yet