Background & Context§
In an era where artificial intelligence is often treated as a black box, a growing movement advocates for a return to first principles. AI by Hand, the research publication of By Hand Research, founded by Prof. Tom Yeh, aims to demystify AI by dissecting models at the mathematical and algorithmic level. The platform offers subscribers articles and live seminars that encourage learners to trace computations by hand—spreadsheets, pen and paper, and all. This approach resonates with a philosophical tenet often attributed to physicist Richard Feynman: "What I cannot create, I do not understand."
The project arrives amid a broader industry push toward interpretability and transparency, as regulators and enterprises demand accountability in AI systems. By focusing on the foundational mathematics—rather than high-level abstractions—AI by Hand positions itself as an educational tool that bridges the gap between theory and practice. Its trending status on Hacker News, where the discussion drew both praise and criticism, signals a genuine appetite for hands-on learning in a field increasingly dominated by plug-and-play APIs.
The News: What Happened Exactly§
AI by Hand is the latest initiative from By Hand Research, led by Prof. Tom Yeh, a professor known for his work on AI education and model interpretability. The publication's core premise is to make the internal workings of neural networks—transformers, attention mechanisms, and backpropagation—tangible through manual calculation. Subscribers receive free articles and participate in live seminars, while paid members gain access to a full research library containing detailed, step-by-step breakdowns of complex models.
However, the initial user experience has been a point of contention. Upon visiting the website, users are greeted with a subscription gate that obscures the actual content. One Hacker News commenter noted, "I'm not sure what this is. Click past the subscribe page to get a bunch of links to descriptions of articles I could read if I was a subscriber?" This friction was echoed by another user who lamented, "I'd say bad UX design. It may or may not be something good behind the door. But why do I enter?" The library itself is partially locked, and even the accessible portions lack visual aids. A commenter observed, "half the posts are locked- other half have no visualizations whatsoever," while another expressed disappointment after watching the intro video: "All I saw were numbers in an Excel sheet. That doesn't make this fun."
Despite these UX criticisms, the underlying content has found an appreciative audience. Several HN users shared similar educational projects, such as a GitHub repository that traces through training a GPT-2 model from scratch using NumPy, and another that builds LLMs from the ground up. These parallels suggest that the "by hand" methodology resonates with a segment of the AI community that values deep, structural understanding over surface-level familiarity. The philosophy is succinctly captured in the tagline of one such project: "What I cannot create, I do not understand."
The initial reception highlights a tension between the rigor of manual computation and the desire for engaging, accessible learning experiences. While AI by Hand may excel in mathematical depth, its delivery leaves room for improvement—a critical consideration for any educational platform in the fast-paced AI landscape.
Historical Parallels & Similar Incidents§
AI by Hand is not the first attempt to teach AI through manual computation. One notable predecessor is Karpathy's micrograd, a tiny autograd engine that demonstrates backpropagation in just over 100 lines of code. Micrograd became a pedagogical staple because it distills the core mechanics of neural network training into a digestible, hackable package. Unlike AI by Hand, micrograd is open-source and immediately interactive, allowing learners to run and modify it in real time. The contrast is stark: micrograd's hands-on, code-optional approach lowers the barrier to entry, whereas AI by Hand's paywalled, spreadsheet-centric style may deter casual learners.
Another antecedent is the "Deep Learning: A Visual Approach" book by Andrew Glassner, published by No Starch Press. This book takes a highly visual, intuition-first approach to explaining deep learning concepts, using diagrams and illustrations to demystify backpropagation, convolutions, and transformers. Its success underscores a market demand for accessible AI education—a demand that AI by Hand's text-heavy, spreadsheet-driven method may not fully satisfy. However, AI by Hand's focus on manual calculation is more extreme than Glassner's, offering a level of granularity that appeals to researchers and engineers seeking to verify model internals down to the arithmetic.
A more recent parallel is the LLM-from-Scratch GitHub repository, which guides learners through building a large language model from zero, including pretraining. This project gained traction on Hacker News as a response to the increasing abstraction of AI development. Like AI by Hand, it stresses foundational understanding, but it does so through code rather than manual math, making it more accessible to practitioners who prefer programming over pen-and-paper calculations.
The lesson from these parallels is that the "by hand" philosophy has merit but must be balanced with usability. The academic rigor of AI by Hand is commendable, but the negative feedback on its UX points to a broader issue: even the most complex topics can be taught in engaging ways. The success of micrograd and visual books suggests that interactivity and clear visualizations are key to effective learning. AI by Hand could learn from these predecessors by offering more visual aids, interactive examples, and possibly a freemium model that allows users to preview content before subscribing.
As the AI field matures, the demand for interpretability and deep understanding will only grow. AI by Hand addresses a real need, but its current execution risks alienating the very audience it aims to educate. By drawing on the lessons of micrograd and visual approaches, it has the potential to refine its methodology and become a valuable resource for AI education. Whether it will adapt remains to be seen, but the conversation it has sparked is already a step toward better, more transparent AI learning.
Conclusion§
AI by Hand's launch has ignited a debate on how best to teach AI foundations in a world dominated by high-level APIs. Its manual, math-driven approach is a refreshing counterpoint to black-box models, and its alignment with the philosophy "What I cannot create, I do not understand" resonates with a dedicated cohort of learners. However, the platform's poor UX and lack of visual engagement have drawn sharp criticism, and its initial reception underscores the challenge of balancing rigor with reach. As AI continues to evolve, initiatives like AI by Hand will likely play an increasingly important role, but they must learn from successful predecessors to truly democratize understanding. The T: the community's response to AI by Hand may well shape the future of AI education—one hand-calculated step at a time.