henrythe9th/AI-Crash-Course: 2-Week AI Research Path for Busy Builders

henrythe9th/AI-Crash-Course: 2-Week AI Research Path for Busy Builders

Software engineers and technical founders face a familiar dilemma: AI research moves faster than product cycles, yet most of us lack months to read papers full-time. The gap between "I know how to call an API" and "I understand why DeepSeek R1 works" keeps widening, and scattered Twitter threads don't close it. henrythe9th/AI-Crash-Course addresses this directly—a ruthlessly curated reading list designed to bring busy builders to the public frontier of AI research in approximately two weeks.

This isn't another "learn Python^{↗ Bright Coding Blog} for AI" tutorial. It's a founder-level acceleration path maintained by Henry Shi, who built Super.com to $150MM+ in annual revenue before exiting, then needed to rapidly upskill on AI himself. The repository has attracted 6,139 GitHub stars and 874 forks as of its last commit on February 23, 2026, suggesting genuine traction among time-constrained technical professionals.

What is henrythe9th/AI-Crash-Course?

henrythe9th/AI-Crash-Course is an open-source learning roadmap distributed under the MIT License. It sits at the intersection of [INTERNAL_LINK: developer-education-resources] and research synthesis—less a tool than a structured knowledge accelerator for engineers who need research literacy without a PhD program.

Henry Shi, the maintainer, frames this as personal necessity turned public good. After exiting Super.com, he needed to evaluate where AI opportunities and gaps existed. Rather than gatekeeping his research process, he open-sourced the paper and video sequence that worked for him. The repository's "Unknown" primary language in GitHub's classification actually signals its nature: this is a meta-resource, a curated index rather than executable code.

The repository's relevance in early 2026 stems from its aggressive currency. It includes papers published as recently as December 2026 (Recursive Language Models) and covers the DeepSeek R1 breakthrough from January 2025. For engineers evaluating whether to fine-tune, build agents, or pivot entirely, this currency matters more than comprehensive depth. The 6,139 stars suggest the community values this tradeoff: speed of orientation over encyclopedic coverage.

Notably, the repository doesn't pretend to replace foundational education. It explicitly recommends starting with 3Blue1Brown's neural network series and identifies "obvious/easy papers" for newcomers. This honesty about prerequisites builds credibility with a skeptical technical audience.

Key Features

Ruthlessly Prioritized Paper Selection

The repository marks critical papers with asterisks (*)—approximately 40% of listed papers receive this designation. This isn't decorative; it's a time-allocation mechanism for readers who can't process everything. Starred papers include Transformers (2017), GPT-3 (2020), RLHF (2022), Chain-of-Thought (2022), MuZero (2019), ARC-Prize (2024), DeepSeek R1 (2025), and Llama 3 (2024). The maintainer's judgment here reflects practitioner priorities: architectural foundations, scaling convictions, and recent reasoning breakthroughs.

Multi-Modal Learning Paths

The resource integrates video lectures, survey papers, primary research, and book recommendations rather than treating papers as the sole valid input. The Neural Network → LLM video series serves as the on-ramp; survey papers provide orientation; primary papers deliver depth; books like Build a Large Language Model (from Scratch) offer implementation scaffolding.

Temporal Organization with Strategic Exceptions

Papers are grouped by Foundational Modelling, Planning/Reasoning, Applications, and Benchmarks—not strictly chronologically. This thematic clustering lets readers follow their curiosity while the year annotations preserve historical context. The 2026 Recursive Language Models paper appearing in "Planning/Reasoning" rather than a separate "Emerging" section shows confidence in its paradigm fit.

Explicit Difficulty Calibration

The "Obvious/easy papers" subsection acknowledges audience heterogeneity. CoT and SELF-REFINE get singled out as accessible entry points—critical for engineers who abandon paper-reading due to initial friction.

Curated External Resources

The repository doesn't duplicate effort. It links to a16z's AI Canon, Latent Space's 2025 reading list, and Full Stack^{↗ Bright Coding Blog} Deep Learning courses, positioning itself as a filtered gateway rather than comprehensive replacement.

Use Cases

Technical Founder Evaluating AI Pivot Timing

A SaaS founder with traditional software background needs to assess whether current AI capabilities justify product bets. The two-week structure provides sufficient depth for strategic decisions without months of research. The DeepSeek V3 and Llama 3 papers specifically address cost-efficient frontier model construction—directly relevant to build-vs-buy calculations.

Senior Engineer Transitioning to ML Infrastructure

An engineer moving from conventional backend to AI infrastructure needs vocabulary and conceptual anchors. The survey papers (LLM, Agent, Prompt Engineering, Context Engineering) provide structured terminology acquisition. The progression from 3Blue1Brown videos to survey papers to primary research mirrors how engineers actually learn: intuition first, framework second, detail third.

Engineering Manager Building AI-Literate Teams

Managers need to evaluate candidate knowledge and scope project complexity. This repository serves as a calibration tool—managers can assess whether their team's current understanding matches the frontier, and identify specific knowledge gaps (e.g., does anyone understand MoE architectures? Has anyone read the RLHF paper?).

Research Engineer Needing Rapid Orientation

Even specialists benefit from cross-domain exposure. A computer vision researcher entering LLM work can use the Foundational Modelling section to identify which "famous" papers they've actually internalized versus merely encountered. The starred paper system helps prioritize genuinely seminal work over merely cited work.

Open-Source Contributor Evaluating Project Architecture

Contributors to projects like SWE-Agent or OpenHands need to understand the research lineage their tools implement. The Applications section directly links to these projects' foundational papers, providing implementation context that READMEs often omit.

Installation & Setup

henrythe9th/AI-Crash-Course requires no traditional installation—it's a knowledge resource, not a software package. However, effective setup involves intentional environment configuration for the two-week sprint.

Step 1: Clone the Repository

# Clone to local machine for offline reference and note-taking
git clone https://github.com/henrythe9th/AI-Crash-Course.git
cd AI-Crash-Course

This preserves the exact paper list and lets you add personal annotations in a branch.

Step 2: Establish Paper Access Pipeline

The repository uses arXiv links extensively. Configure bulk PDF management:

# Create organized directory structure for downloads
mkdir -p papers/{foundational,planning,applications,benchmarks,surveys}

# Optional: Install arXiv CLI for batch retrieval
pip install arxiv

Many readers prefer Zotero or Obsidian with PDF plugins for annotation. The repository doesn't prescribe tools, but the two-week timeline demands frictionless reference management.

Step 3: Schedule the Video Prerequisites

The 3Blue1Brown Neural Network → LLM series constitutes mandatory prework. Block 4-6 hours before starting papers. The repository explicitly warns against skipping this—subsequent paper comprehension depends on visual intuition about attention mechanisms.

Step 4: Survey Paper First Pass (Days 1-3)

Read in this order:

1. LLM Survey (2024) — establishes current landscape
2. Agent Survey (2023) — if building autonomous systems
3. Prompt Engineering Survey (2024) — if primarily using existing models
4. Context Engineering Survey (2025) — most recent, most specialized

Step 5: Starred Paper Deep Dives (Days 4-10)

Follow starred papers by thematic interest. The repository suggests "digging deeper" into ideas that resonate—this isn't linear completion but targeted exploration.

Step 6: Application-Specific Papers (Days 11-14)

Select from Applications and Benchmarks based on your domain. Building dev tools? Prioritize SWE-Agent and SWE-Bench. Evaluating models? Focus on Chatbot Arena and LLM-as-Judge.

Real Code Examples

The henrythe9th/AI-Crash-Course repository contains no executable code—it is a curated index of external resources. This reflects its purpose: accelerating research literacy rather than providing implementation shortcuts. Below are the repository's actual structural elements that readers interact with programmatically.

Example 1: Repository Structure (Markdown^{↗ Smart Converter})

The README itself serves as the primary "interface." Its hierarchical organization is the "code" readers execute:

# AI-Crash-Course
AI Crash Course to help busy builders catch up to the public frontier of AI research in 2 weeks

**Start Here:**  
[Neural Network -> LLM Series](https://www.youtube.com/watch?v=aircAruvnKk&list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi)

**Then get up to speed via Survey papers:**
- [LLM Survey](https://arxiv.org/pdf/2402.06196v2) - 2024
- [Agent Survey](https://arxiv.org/pdf/2308.11432) - 2023

This structure functions as a declarative learning pipeline: prerequisites first, then orientation, then specialization. The markdown links are the "function calls" to external knowledge.

Example 2: Paper Reference Pattern

The repository's consistent citation format enables programmatic extraction if desired:

[**DeepSeek R1***](https://arxiv.org/pdf/2501.12948v1) (Building OSS o1-level reasoning model with pure RL, no SFT, no RM) - 2025

Elements include: bold title, asterisk priority marker, direct arXiv PDF link, parenthetical value proposition, and year annotation. This pattern supports quick scanning and priority filtering—critical for the two-week constraint.

The repository currently contains these two structural patterns as its primary "code". Readers should not expect implementation examples; the value lies in curation quality and sequencing logic. For actual code, the repository explicitly points to external resources like Build a Large Language Model (from Scratch) and the Build a Reasoning Model (From Scratch) book.

Advanced Usage & Best Practices

Parallel Track by Subfield

The two-week timeline assumes full-time focus. If studying part-time, prioritize one vertical (Foundational Modelling for infrastructure builders, Planning/Reasoning for agent developers) rather than diluting across all sections. The starred papers within your chosen vertical provide minimum viable frontier knowledge.

Active Paper Synthesis

Don't merely read—maintain a living document connecting papers to your specific problems. When reading the DeepSeek R1 paper, note: "Our current fine-tuning pipeline uses SFT; could pure RL reduce our data preparation burden?" This transforms passive consumption into actionable intelligence.

Survey Paper as Filter

Treat survey papers as decision trees. Follow their citation networks selectively rather than exhaustively. The repository's curation already performed first-pass filtering; trust the asterisk system for second-pass prioritization.

Temporal Awareness

Papers include publication years for reason. A 2022 method (DPO) may be superseded by 2024-2025 techniques, but understanding its historical problem-solving logic remains valuable. Don't dismiss older papers solely by age—assess whether their core insights persist in current practice.

Community Annotation

With 874 forks, community extensions exist. Search fork networks for readers who added annotations in your domain. The MIT License permits derivative works, so consider publishing your own annotated branch if your additions would benefit others.

Comparison with Alternatives

Trade-offs: a16z's AI Canon offers greater breadth and institutional credibility. Latent Space's list provides deeper commentary and community discussion. henrythe9th/AI-Crash-Course wins on time-bounded structure and GitHub-native iterability—you can fork, modify, and track your own learning path. The choice depends on whether your constraint is time (this repo), comprehensive coverage (a16z), or community depth (Latent Space).

FAQ

Is henrythe9th/AI-Crash-Course free to use?

Yes, MIT Licensed. Fork, modify, and redistribute freely.

Do I need a PhD to follow this course?

No. The repository explicitly includes prerequisite videos and "obvious/easy papers" for newcomers.

How current is the paper selection?

Last updated February 23, 2026, with papers through December 2026 included.

Is there video content or only papers?

Mixed: video series for foundations, survey and primary papers for depth, plus recommended books.

Can I complete this while working full-time?

The two-week estimate assumes dedicated focus. Part-time completion requires 4-6 weeks realistically.

Does this teach me to build models from scratch?

Not directly. It provides research context; implementation guidance points to external books and courses.

How does this differ from just reading papers myself?

Curation, sequencing, and prioritization. The asterisk system alone saves hours of uncertainty about what's essential versus merely cited.

Conclusion

henrythe9th/AI-Crash-Course fills a specific gap in the AI education landscape: time-bounded research orientation for practicing engineers. It won't make you a research scientist, but it will equip you to read current papers, evaluate model capabilities, and identify genuine opportunities versus hype cycles. The 6,139 stars and active maintenance through early 2026 suggest this need is widely felt.

This resource best serves technical founders evaluating AI pivots, senior engineers transitioning to ML infrastructure, and engineering managers calibrating team knowledge. It's less suited to researchers seeking comprehensive literature reviews or complete beginners without software engineering foundations.

The repository's honesty about its scope—"help busy builders catch up" rather than "master AI in two weeks"—builds the trust that sustains long-term utility. Start with the Neural Network → LLM series, trust the asterisk system, and fork the repository to track your own annotations.

Begin your two-week sprint: https://github.com/henrythe9th/AI-Crash-Course