Learning Engineering with AI: Four Fast Study Methods

Learning a hard topic used to take weeks: muffled lecture recordings, slide piles, and endless reading lists. Engineer Aleks Gornik says that routine has changed; AI pulls him out of stuck points within minutes. The secret is not outsourcing the work to AI but turning ideas into visuals, cutting noise from signal , and reaching the best resource fast. This article gathers the four methods from his 12-minute talk plus the evidence behind them.

The narrator graduated in electrical engineering and interned across AI, electronics, and software as a student. He now works on a technical team at a startup training large language models, so learning hard things is literally his day job. Throughout the talk he draws on his own student years: as an intern in 2023 he asked a free chatbot to explain topics like he was five, the moment that sparked his fascination with AI.

Turn ideas into visuals

The first method is rendering abstract ideas visual. As a kid he watched the 3Blue1Brown channel and was enchanted by the animation tool its host built in Python. That tool now lives openly on GitHub: the MIT-licensed project called Manim , an animation engine for explanatory math videos with tens of thousands of stars. Work once reserved for mathematics now applies to engineering topics too.

The examples are concrete: a mechanical engineering student with no design idea asks the Claude assistant for an interactive presentation with three robotics project pitches plus a rotatable 3D design. On the electrical side, the narrator recalls his confusion wiring an STM32 board to an ST-LINK debugger for power measurements; with a weak datasheet, he uploads photos of his boards and gets a step-by-step wiring board generated. Electronics forums confirm such hookups genuinely confuse people: a StackExchange thread viewed over 17,000 times explains that the 3.3-volt line on the STDC14 header is a reference voltage, not a supply, and that some pins may stay unwired for debug-only use. His side note matters too: paying a bit more for stronger model tiers visibly improves 3D output.

The second method is triaging what to read with AI. A large part of his job is scanning research papers; grasping one small tweak added to a transformer block makes reading everything impossible. University looks the same: the lecturer recommends five books of a thousand pages each, one from the 1970s. The fix is handing the syllabus and reading list to the assistant and asking which book and which chapters actually pay off for the exams ahead. A well-written prompt returns two or three decisive chapters instead of the whole book, and the same filter decides which tutorial problems deserve solving.

Unstick yourself with twenty percent

The same triage works for papers. Asking the assistant which of five papers found on social networks is worth reading and most legitimate shrinks two hours of skimming to fifteen minutes. A university study from Argentina shows prompt design is decisive here: according to findings published on arXiv, the first attempt scored a zero pass rate, while a second prompt version with evidence rules lifted the median score to 88 out of 100. Broader evidence exists too: in a randomized controlled trial published in Nature, students working with an AI tutor learned more in less time than peers in active-learning classes, and appeared more engaged and motivated. Both studies teach the same lesson: disciplined prompts deliver, casual use does not.

The third method is finding the twenty percent that unblocks you. The idea dates to 2023: the same explain-like-I-am-five trick from his intern days carried into his thesis work. Simulating lung airflow with graph neural networks, he drowned in deep-learning detail: forward passes, activation functions, matrices, endless architectures. The way out was shrinking the problem: focus on the smallest knowledge needed to run a training loop and push a few epochs of data through. Learning to move data across a network with PyTorch meant postponing the mathematics of backpropagation; getting moving never required knowing everything.

Hunt resources with an agent swarm

The fourth method is letting AI agents find the best resources. Back in computer-aided design for work after avoiding it at school, he set a swarm of agents combing the internet loose on free templates and datasets. The result stunned him: the NASA 3D Resources archive publishes downloadable and printable designs of mission hardware, including the Perseverance Mars rover, free for anyone. Building your own Mars rover at home is genuinely possible. The Claude documentation confirms this usage pattern: when side tasks would bloat the main conversation, a dedicated agent works in its own context and returns only the summary. The critical warning stands unchanged: click every link and verify each one, because AI still hallucinates.

Science explains why these four methods work. Under the forty-year-old cognitive theory of multimedia learning , the mind processes verbal and visual information through separate channels, working memory is limited, and lasting learning happens through selecting, organizing, and integrating; a recent review published by Springer lists 15 evidence-based teaching principles derived from them. Visualization and triage match those principles exactly: only needed knowledge enters limited memory, the rest stays out.

In short, the recipe has four principles: visualize ideas, triage readings, focus on the unblocking twenty percent , and let agents hunt the best resources. The resume template and one-to-one coaching pitches occupy only a small slice of the talk; the real value sits in copyable prompts. For a student facing exam week, the first step is small: feed tomorrow's syllabus to the assistant and ask which two chapters truly matter. The rest is repeating the same loop at every stuck point.