Training a Large Language Model
Do you understand how a Large Language Model (LLM) is created?
What we call AI today is not designed or programmed in any traditional way. When you ask a LLM to answer a question, that question is not generated by any programming. Instead, the LLM is "trained" by giving it all available text on the Internet. The LLM simply "learns" how that text is organized and stores them in a huge table of weight values. Those weight value are completely undecipherable by any human directly. These values are collected by "self-organization" which is a collection of information that it learns on its own.
DNA gathers information in the same way: it is presented with trillions of life/death situations and it learns what does and does not work in order to stay alive. Science deniers claim that information can not be accumulated this way, however AI would never work if that was true.
Claude says this on the topic:
Training an LLM and evolution of DNA are fascinating examples of information accumulation through iterative selection — but they differ deeply in mechanism, timescale, and what kind of "knowledge" gets encoded. Here's a structured comparison: What they share
Both processes involve a system that gets better at producing adaptive outputs by being exposed to a large volume of examples and feedback. In each case, "learning" isn't consciously directed — it emerges from the aggregate pressure of many, many trials. And both encode information in a compressed, distributed form: DNA doesn't store a blueprint for a liver cell in one place, and an LLM's knowledge of French grammar isn't stored in one layer or set of weights.
The feedback mechanism
This is the sharpest difference. Evolution is driven by differential reproduction — organisms that happen to be better suited to their environment leave more offspring, so their DNA propagates. There's no gradient, no explicit error signal. It's selection operating blindly on random variation (mutations, recombination). LLM training is gradient descent — a mathematically precise error signal flows backward through the network, nudging billions of parameters in the direction that reduces prediction error. It's not random; it's directed. The contrast is something like the difference between wandering a landscape until you stumble into lower and lower valleys (evolution) versus having a GPS that always tells you which direction is downhill (backpropagation).
The "data" being learned from
DNA encodes information about ancestral environments and selection pressures — what pathogens existed, what food sources were available, what predators were present. This information is smeared across deep time, millions of generations. An LLM trains on human linguistic output — text that is itself already a distillation of human thought, culture, and recorded knowledge. So LLMs are learning from a kind of second-order signal: not raw environmental pressure, but the accumulated outputs of minds that were themselves shaped by evolution.
Timescale and iteration
Evolution operates over thousands to millions of generations, each lasting years to centuries. LLM training happens in weeks on hardware, iterating over hundreds of billions of tokens in a single run. Evolution can't "restart" — it's a continuous process with no clear beginning. LLM training has discrete runs with checkpoints and restarts.
Generalization vs. specialization
DNA is remarkably plastic within a species but conservative across deep lineages — the basic machinery (ribosomes, ATP synthesis, DNA replication) is nearly universal because it's been locked in by billions of years of selection. LLMs generalize across tasks in a way that has no clean evolutionary analog — a model trained on text can reason about code, logic, and poetry simultaneously, which is more like a cognitive architecture than a genome.
What gets "forgotten"
In evolution, information that's no longer selected for degrades — genes accumulate mutations and become pseudogenes. In LLMs, training data that's underrepresented doesn't get "forgotten" so much as underweighted; the model may simply not learn it well. But there's no mechanism analogous to extinction, where an entire lineage of information disappears permanently. The deepest similarity might be philosophical: both are processes where structure in the environment gets internalized into structure in a substrate — whether that substrate is a genome or a neural network. Neither "understands" what it's learning in any intentional sense; both just reflect the statistical regularities of what they were exposed to, filtered through a selection pressure.