By November, Ben had learned that university life could fill every hour from breakfast to dinner without finishing a single assignment. He was a first-year computing student. Classes occupied his days, coursework ate his evenings, and his laundry was developing a separate life under the desk. He could spend an hour searching for an explanation of a programming problem and come back with more browser tabs than answers.
Then, after OpenAI released ChatGPT on November 30, 2022, a classmate sent him the link. Ben pasted in a question about a loop he could not understand. The reply explained the steps in plain English, and he could ask another question without starting his search again. For that task, it felt better than digging through search results. He leaned back in his chair, delighted that someone in the room seemed to know what was going on.
Over the following weeks, Ben used the chatbot whenever a lecture left him behind. It helped him read unfamiliar code and turn a messy idea into a plan. Then it recommended a library function he could not find in the documentation. Another answer explained a different problem from the one he had asked. The friendly voice stayed calm through all of this, which Ben found increasingly rude.
He corrected an answer and watched the next reply improve. Later, in a fresh conversation without that correction, the same misunderstanding returned. Ben stared at the screen. Once Ben understood a loop, he could usually use it again without somebody repeating the explanation. He sometimes forgot the lecture by Friday, admittedly, but that seemed like a different department's problem.
That evening he opened a new notebook and wrote "My own study AI" at the top. He pictured a small helper that learned his course a little each day, kept its corrections and stopped inventing functions. He already knew some Python. Surely the remaining work was mostly connecting things, choosing a pleasant name and finding somewhere to put the settings button. His first search returned neural networks, linear algebra and a training guide long enough to qualify as additional coursework.
Ben took the guide to Maya, a tutor in the computing lab. She opened a small training example and pointed to the model's prediction. Beside it were rows and columns of numbers, called matrices. Ben changed one of the weights, the adjustable values the model used, and ran the example again. He compared the prediction with the old one. Maya showed him how linear algebra combined the numbers, with other functions helping the model learn more complex patterns. Ben now had something to poke at, along with several new words to look up.
Ben paused the example after a prediction and compared it with the expected answer. The code measured the error, then worked backward to find how small changes to each weight would affect it. An update step adjusted the weights using those results. Maya made him circle that step in the PyTorch training explanation. Running a trained model to produce a reply was different. Ordinary inference used the weights already learned, without performing those training updates for every conversation.
Then he asked why every large-model guide seemed interested in expensive graphics cards. Maya explained that GPUs could perform many suitable calculations in parallel. A larger model and more training data meant more work, and training also needed memory for temporary calculations and the extra values used to update weights. The memory used during training could therefore exceed what was needed to run the same model for replies. Ben's laptop had not betrayed him. He had given it a job without checking the job description.
He also learned why model weights were kept in working memory to avoid repeated trips to slower storage. He had assumed every model had to be copied into system RAM all at once. Maya showed him how some software could load parts or share work across devices, sometimes at the cost of speed. That could help it fit, but the training work would still be there. Ben stayed with the small example. When Ben found that the BLOOM model card listed 384 training GPUs, he quietly closed the shopping tab he had opened for a laptop stand.
By December, the question had changed. Ben no longer expected to recreate a large chatbot over a weekend, but he could not see why a machine needed so much preparation. He could study something in the evening, sleep and sometimes understand it better the next day. Why could his model not do that? He arrived at the lab carrying this question and a coffee large enough to suggest that he was not fully committed to the sleep part.
Maya asked him whether he had started university without knowing any language, numbers or useful facts about the world. Ben considered his last mathematics test and decided the question was unfair. Still, he had years of experience to connect each lesson to. Learning a new loop used ideas he already knew about counting and order. Training a large language model from random weights was a much broader task than adding one lesson to an already experienced student.
Sleep could help the brain strengthen and reorganize memories, Maya explained, but it did not turn every late-night reading session into understanding. Maya showed him a human sleep experiment that found signs of earlier learning in brain activity during sleep. That supported a role for memory processing during sleep, not a promise that Ben could sleep through the reading itself. He remembered a morning when the lecture still made no sense and accepted this qualification with personal disappointment.
They returned to the small training exercise rather than the giant model guides. A modest experiment could be far smaller than a large pretraining run. Training time depended on the model, data, hardware and goal, so days or months were not a rule for every AI project. A released model could receive more training, too. Switching off Ben's laptop overnight did not make that training happen. His computer's version of a good night's sleep was simply being switched off.
During the holiday break, Ben found the number that seemed to settle everything. The human brain ran on roughly twenty watts, while large AI systems used racks of computing equipment. He took this as a strong argument that his own head was excellent value, even after allowing for snacks. He wrote "20 W" in his notebook and began preparing a complaint against the entire computer industry.
Before he could finish, Maya asked what the number actually measured. Maya opened the energy audit by Levy and Calvert and pointed to what the twenty-watt figure covered. It was a rough budget for the whole brain, based on energy from glucose, rather than the cost of learning one fact. The brain used energy to maintain living cells and communicate as well as process information. Ben had been treating a whole organ's running budget like a bill for one completed assignment.
He tried the arithmetic anyway, this time with units. At a constant twenty watts, twenty-four hours would use 480 watt-hours, or 0.48 kilowatt-hours. Ben wrote the units beside the result and underlined the constant-power assumption. Comparing it with a training cluster still required matching the task and what each measurement included. It also left out the rest of his body, which was currently making a persuasive request for lunch.
The brain's efficiency remained impressive. The comparison simply could not tell Ben how many times better it was at training the same model, because nobody in his notebook had performed that matched test. Nor had he become an adult learner overnight from random starting values. He had grown and learned for years. Ben crossed out the efficiency ratio he had been about to calculate and replaced it with "Different jobs, different starting points".
Back in the lab after the break, Ben asked whether he could skip the giant training job and update a small model every evening. Maya said this was a real research problem, not an idea waiting for a student to invent it. Continual learning meant learning from new experiences while trying to retain earlier abilities. Ben briefly mourned the loss of his discovery, then felt much better about having something useful to read.
The difficulty appeared when they talked through what an update changed. Knowledge was spread across many learned values, and adjusting them for a new task could damage performance on an old one. Researchers called a severe version catastrophic forgetting. Maya showed him research on protecting important weights and reusing past examples. Both approaches tried to preserve earlier abilities as the model learned something new. These were methods with tested settings, not guarantees that any model could absorb any lesson safely.
Ben understood the appeal of forgetting when he looked at his exam timetable, but he wanted his helper to keep the useful parts. Maya asked whether the helper should trust a user's confident but false rule, or learn from somebody's private notes. A system that changed from every conversation would need choices about which examples to trust, which changes to keep and how to test old abilities again. Fast updates alone would not answer those questions.
That changed how Ben saw his earlier correction in the chat. The correction had supplied useful context for a reply. It had not demonstrated a lasting change to the model's weights. A product might also save information outside those weights and provide it later, much as Ben could open his notebook. Looking better informed in a conversation and undergoing further training were different ways a system could change its answer.
Ben eventually returned to the first page of his notebook. Under "My own study AI", he added a more useful description. The helper should answer questions from course material he supplied and show him where the answer came from. It did not need to learn the whole internet again because he had missed a lecture. He could build around an existing model, provide relevant notes and check whether the resulting answer matched them.
He read about retrieval-augmented generation, which combined a model with information found in a separate collection. Updating the collection could make new material available without retraining the model's weights. It could still retrieve the wrong passage or misread the right one, so Ben kept the notes beside the answer. He also learned about LoRA, which trained extra adjustable values while keeping the original weights fixed. That was a possible adaptation method, with its own data and hardware needs, rather than a bedtime setting for his computer.
For his first attempt, Ben asked which file format a coursework submission needed. The reply matched the PDF rule in his notes. When he asked when the tutor's office opened, however, the helper supplied a time the notes never mentioned. Ben marked that answer wrong and added an instruction to say when the material did not contain an answer. On the next try, the helper admitted that the opening time was missing. He kept both questions in his notebook to check again after later changes.
That night, Ben wrote down the test for his next session. He would choose a small set of notes, ask a question with a known answer and check the result against the page. The laptop would not wake up knowing more unless some actual update process ran. Ben might retain more of what he had practised, although his laundry remained beyond help. He closed his laptop, left the notebook open on the desk and went to bed.