Everyone's obsessed with new training tokens... It doesn't need to be more knowledgeable, it just needs to practice more. Ask any student: practice is synthetic data.
Overfitting can be caused by a lot of different things. Having an over abundance of one kind of data in a training set is one of those causes.
It’s why many pre-processing steps for image training pipelines will add copies of images at weird rotations, amounts of blur, and different cropping.
> The more concepts the model manages to grok, the more nonlinear its capabilities will be
These kind of hand wavey statements like “practice,” “grok,” and “nonlinear its capabilities will be” are not very constructive as they don’t have solid meaning wrt language models.
So earlier when I was referring to compounding bias in synthetic data I was referring to a bias that gets trained on over and over and over again.
These kind of hand wavey statements like “practice,” “grok,” and “nonlinear its capabilities will be” are not very constructive as they don’t have solid meaning wrt language models.
So, here's my hypothesis, as someone who is adjacent ML but haven't trained DNNs directly:
We don't understand how they work, because we didn't build them. They built themselves.
At face value this can be seen as an almost spiritual position, but I am not a religious person and I don't think there's any magic involved. Unlike traditional models, the behavior of DNNs is based on random changes that failed up. We can reason about their structure, but only loosely about their functionality. When they get better at drawing, it isn't because we taught them to draw. When they get better at reasoning, it isn't because the engineers were better philosophers. Given this, there will not be a direct correlation between inputs and capabilities, but some arrangements do work better than others.
If this is the case, high order capabilities should continue to increase with training cycles, as long as they are performed in ways that don't interfere with what has been successfully learned. People lamented the loss of capability that GPT 4 suffered as they increased safety. I think Anthropic has avoided this by choosing a less damaging way to tune a well performing model.
> We don't understand how they work, because we didn't build them. They built themselves.
We do understand how they work, we did build them.
The mathematical foundation of these models are sound. The statistics behind them are well understood.
What we don’t exactly know is which parameters correspond to what results as it’s different across models.
We work backwards to see which parts of the network seem to relate to what outcomes.
> When they get better at drawing, it isn't because we taught them to draw. When they get better at reasoning, it isn't because the engineers were better philosophers.
Isn’t this the exact opposite of reality?
They get better at drawing because we improve their datasets, topologies, and their training methods and in doing so, teach them to draw.
They get better at reasoning because the engineers and data scientists building training sets do get better at philosophy.
They study what reasoning is and apply those learnings to the datasets and training methods.
> We do understand how they work, we did build them. The mathematical foundation of these models are sound. The statistics behind them are well understood.
We don't understand how they work in the sense that we can't extract the algorithms they're using to accomplish the interesting/valuable "intellectual" labor they're doing. i.e. we cannot take GPT-4 and write human-legible code that faithfully represents the "heavy lifting" GPT-4 does when it writes code (or pick any other task you might ask it to do).
That inability makes it difficult to reliably predict when they'll fail, how to improve them in specific ways, etc.
The only way in which we "understand" them is that we understand the training process which created them (and even that's limited to reproducible open-source models), which is about as accurate as saying that we "understand" human cognition because we know about evolution. In reality, we understand very little about human cognition, certainly not enough to reliably reproduce it in silico or intervene on it without a bunch of very expensive (and failure-prone) trial-and-error.
The thing that you are handwaving away as just "which parameters correspond to what results" is precisely the important, the inexorable thing which defines the phenomena, and it is exactly the thing which we don't have access to, and which we did not and could not design, plan or engineer, but which emerged
And who will tell the model whether its practice results are correct or not? Students practice against external evaluators, it’s not a self-contained system.
We know that unique data improves performance.
These LLM systems are not students…
Also, which students graduate and are immediately experts in their fields? Almost none.
It takes years of practice in unique, often one-off, situations after graduation for most people to develop the intuition needed for a given field.