Developers want to understand how LLMs work internally but existing resources are either too academic or too simplified — code-only repos like this lack documentation and guided learning paths
A structured, hands-on course where users build a tiny LLM from scratch in a browser-based environment with visual explanations of each component (attention, embeddings, etc.), progressive complexity, and immediate feedback — bridging the gap between reading papers and running someone else's code
Freemium — free intro modules, $99-149 for full course with certificate, team licenses for companies
Real but not hair-on-fire pain. Developers WANT to understand LLM internals but can survive without it. The GitHub signals ('not straightforward to understand,' 'is there documentation?') confirm frustration with existing resources. However, this is a learning desire, not a business-critical blocker — people won't lose their jobs over it. The pain is more 'career anxiety + intellectual curiosity' than 'my production system is broken.'
TAM is large. There are ~30M+ software developers globally, and a rapidly growing percentage (estimated 5-10M+) are actively working with LLMs. Even capturing 0.1% at $99-149 = $1-1.5M in revenue. The developer education market is $15-25B. AI/ML is its fastest-growing segment. Team licenses expand the opportunity further. This is not a niche — every engineering org wants their developers to understand AI.
Mixed signals. Developers are notoriously resistant to paying for education when free alternatives exist (Karpathy, fast.ai, Raschka's GitHub code are all free). The $99-149 price point is reasonable but competes with free. Key unlock: CERTIFICATES + TEAM LICENSES. Individual devs may balk, but L&D budgets at companies will pay $149/seat without blinking. The 843 GitHub stars show interest but not willingness to pay. Need to validate this with a landing page test.
A solo dev can build the MVP in 6-8 weeks, not 4. The course content itself (text, exercises, code) is straightforward. The hard parts: (1) browser-based Python execution environment — options include Pyodide/WebAssembly, JupyterLite, or iframe'd Colab, each with tradeoffs; (2) interactive visualizations of attention/embeddings require meaningful frontend work; (3) training even a tiny LLM in-browser has compute constraints. Could simplify MVP by using server-side execution (CodeSandbox-style) and fewer interactive widgets.
This is the strongest signal. NO existing product combines: from-scratch LLM building + browser IDE + interactive visualizations + progressive curriculum + certificates. Raschka's book is the closest on content but static. Karpathy is closest on explanation quality but passive video. DeepLearning.AI has the platform but teaches usage not building. The gap is clear and meaningful. Risk: Karpathy's Eureka Labs or a well-funded competitor could fill this gap quickly.
This is the weakest dimension. A course is inherently a one-time purchase, not a subscription. You finish it and you're done. Possible recurring angles: (1) continuously updated advanced modules (RLHF, MoE, multimodal, new architectures), (2) monthly 'lab challenges,' (3) community/mentorship subscription, (4) team license renewals as companies onboard new hires. But the core product is more 'cohort/one-time' than 'SaaS.' Compare: Brilliant.org ($10-15/mo) manages recurring through breadth of content — you'd need similar content velocity.
- +Clear gap in the market — no one combines interactive browser-based building + visual explanations + progressive curriculum for LLMs from scratch
- +Strong demand signal: 843 stars on a basic repo with users explicitly asking for documentation and guided learning
- +Large and growing market — every developer wants to understand AI, and enterprises are budgeting for upskilling
- +The $99-149 price point is in the sweet spot for individual and team purchases
- +Content moat: high-quality interactive educational content is genuinely hard to create and harder to clone quickly
- !Karpathy's Eureka Labs is the elephant in the room — if he ships an interactive course, he has instant distribution to millions
- !Free alternatives (Karpathy videos, Raschka's free code, fast.ai) create strong downward pricing pressure on individual buyers
- !Browser-based LLM training has real technical constraints — even tiny models may be too slow in-browser, requiring server compute costs
- !Course businesses are hard to make recurring — risk of one-time revenue spikes followed by plateau
- !Content becomes outdated quickly as LLM architectures evolve — requires continuous investment to stay relevant
Book + GitHub repo walking through implementing a GPT-style LLM from scratch in PyTorch, covering tokenization, attention, pretraining, and fine-tuning
Free YouTube series
Video courses + auto-graded Jupyter notebooks covering LLM usage, prompt engineering, fine-tuning, RAG, and agents. 'Generative AI with LLMs' covers architecture at a high level
Free deep learning course with Part 2 covering building a GPT-style model from scratch, stable diffusion, and generative AI
Free written tutorials and notebooks teaching NLP/transformer usage via the HF ecosystem; Cohere offers similar free educational content about LLM concepts
3-module free tier + 8-module paid course. Free modules: (1) tokenization from scratch, (2) embeddings visualized, (3) single-head attention interactive demo. Paid modules: multi-head attention, transformer blocks, training loop, text generation, basic fine-tuning. Use JupyterLite or Pyodide for browser execution (no server costs). Build 2-3 high-impact interactive visualizations (attention heatmap, embedding space explorer, loss curve tracker). Skip certificates in MVP — add later. Launch on Product Hunt, Hacker News, and AI Twitter. Validate willingness to pay before building the full course.
Free intro modules (lead gen + viral sharing) → $99 individual full course → $149 with certificate → $49/seat/year team licenses (minimum 5 seats) → Enterprise custom pricing with admin dashboard + progress tracking → Advanced course modules (RLHF, multimodal, MoE) as $49-79 add-ons → Eventually a Brilliant-style subscription ($15-20/mo) if content library grows large enough
8-12 weeks. Weeks 1-6: build MVP (3 free + first 3 paid modules). Weeks 7-8: beta test with 50-100 developers from the GuppyLM community. Weeks 9-10: iterate based on feedback. Weeks 11-12: launch on HN/Product Hunt/Twitter with early-bird pricing ($79). First revenue at week 11-12. Reaching $10K MRR likely takes 4-6 months and requires strong organic distribution or content marketing.
- “not straight forward to understand for developers not familiar with multi-head attention, ReLU FFN, LayerNorm”
- “Is there some documentation for this?”
- “genuinely a great introduction to LLMs”
- “built to demystify how language models work”