June 6, 2026
Brainstorm or brainworm?
Human-Like Neural Nets by Catapulting
Scientists pitch a wild shortcut to human-like AI and the comments are absolutely not buying it
TLDR: A new proposal claims super-sized AI trained in a very unusual way could become more human-like, safer, and better at generalizing from less information. Commenters were fascinated but deeply skeptical, arguing the idea makes giant promises without proof and wildly underestimates how much “training” evolution gave the human brain.
A bold new AI proposal just dropped, and the internet responded with the digital equivalent of spitting out its coffee. The idea: maybe today’s artificial intelligence is smart in the wrong way, and the fix is to train absurdly huge neural networks with very aggressive settings on smaller, carefully chosen piles of data so they suddenly “click” into more human-like understanding. The author suggests this could make AI learn more like people do: with less data, better common sense, and fewer bizarre mistakes. In plain English, it’s a pitch for a shortcut to smarter, safer AI.
But in the comment section? Total cage match. Skeptics called the core claim “all vibes,” with one reader basically saying, show me the biology notes, because this sounds made up. Another flatly rejected the fantasy that some hidden math trick can magically replace missing training data. And the most brutal reality check of all? A commenter reminded everyone that the human brain didn’t just appear out of nowhere—it was “trained” by evolution over billions of years, using, you know, the entire planet as compute. Casual.
Not everyone was purely mocking, though. One community member tried to calmly summarize the argument, translating the giant-brain theory into something almost understandable. Still, the vibe was clear: fascinating idea, huge claims, tiny patience. The unspoken meme of the thread was basically: bro invented one weird trick for consciousness. If the proposal wanted curiosity, it got it. If it wanted trust, the comments said: not so fast.
Key Points
- •The article proposes that human-like generalization in AI could emerge from training extremely overparameterized neural networks with very high learning rates and strong regularization on small, curated datasets.
- •It frames the difference between current large language models and human brains as a bias-variance tradeoff, suggesting LLMs minimize variance while brains may minimize bias.
- •The article claims this training regime could produce a 'catapulted LLM' with stronger generalization, adversarial robustness, improved economics, cloning resistance, and AI safety benefits.
- •It suggests testing the hypothesis by training multi-trillion-parameter models for relatively few steps with high cyclical learning-rate schedules and evaluating them on hard benchmarks such as arithmetic and small-image classification.
- •The article positions the proposal as a response to unresolved anomalies in AI, including why current models rely on Chinchilla-style scaling while humans appear to learn from much less data and possibly less compute.