I’m currently working on a document parsing engine for a specific type of document. The inputs are usually PDFs. I’m able to get great structured output from both the latest Gemini Flash models and the latest Llama Scout models. The best latency I get with Gemini is about 5 seconds end to end. With llama hosted on groq it’s about 3 seconds.
My use case is latency constrained, so I’m exploring fine tuning / distilling to see if I can get latency sub second. I imagine these are the kinds of scenarios where it’s still worth it to fine-tune and distill.
My plan is to generate a lot of synthetic training data using more capable slower foundation models and use that to train the smaller model.
It’s really useful for generating synthetic data for search and recommendations that you can use to train a smaller / faster model. This is especially useful if you don’t have lots of click-through data or with cold start scenarios. There are some good articles that cover this, if you’re interested I’ll try to find them and share
- Conventional SGD: Fixed target (e.g. "make an exact replica of this butterfly image") and it follows greedy path to minimize the error
- Open Ended Search Process: No predetermined goal, explores based on what's "interesting" or novel. In Picbreeder, humans would see several generated images, pick the "interesting" ones, and the system would mutate/evolve from there. If you were evolving an image that looked like an egg and it mutated toward a teapot like shape, you could pivot and pursue that direction instead.
This is kinda the catch -- there is a human element here where individuals are choosing what's "interesting" to explore, it's not a pure algorithmic process. That said, yes, it does use a genetic algorithm (NEAT) under the hood, but I think what the authors are suggesting is that the key difference isn't whether it's genetic or gradient based optimization... they're getting at the difference in objective driven vs. open-ended search.
I think the main position / takeaway from the paper is that something about conventional SGD training produces these "fractured entangled representations" that work but are not well structured internally so they're hard to build on top of. They look at things like the curriculum / order things are learned in, objective search vs. open-ended search, etc...
Could a major opportunity to improve representation in deep learning be hiding in plain sight? Check out our new position paper: Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis. The idea stems from a little-known observation about networks trained to output a single image: when they are discovered through an unconventional open-ended search process, their representations are incredibly elegant and exhibit astonishing modular decomposition. In contrast, when SGD (successfully) learns to output the same image its underlying representation is fractured, entangled - an absolute mess!
This stark difference in the underlying representation of the same "good" output behavior carries deep lessons for deep learning. It shows you cannot judge a book by its cover - an LLM with all the right responses could similarly be a mess under the hood. But also, surprisingly, it shows us that it doesn't have to be this way! Without the unique examples in this paper that were discovered through open-ended search, we might assume neural representation has to be a mess. These results show that is clearly untrue. We can now imagine something better because we can actually see it is possible.
We give several reasons why this matters: generalization, creativity, and learning are all potentially impacted. The paper shows examples to back up these concerns, but in brief, there is a key insight: Representation is not only important for what you're able to do now, but for where you can go from there. The ability to imagine something new (and where your next step in weight space can bring you) depends entirely upon how you represent the world. Generalization, creativity, and learning itself depend upon this critical relationship. Notice the difference in appearance between the nearby images to the skull in weight space shown in the top-left and top-right image strips of the attached graphic. The difference in semantics is stark.
The insight that representation could be better opens up a lot of new paths and opportunities for investigation. It raises new urgency to understand the representation underlying foundation models and LLMs while exposing all kinds of novel avenues for potentially improving them, from making learning processes more open-ended to manipulating architectures and algorithms.
Don't mistake this paper as providing comfort for AI pessimists. By exposing a novel set of stark and explicit differences between conventional learning and something different, it can act as an accelerator of progress as opposed to a tool of pessimism. At the least, the discussion it provokes should be quite illuminating.
For anyone interested in NEAT, you’ll likely enjoy Ken’s later work on novelty search, open-endedness, etc.
His book “Why Greatness Cannot be Planned: The Myth of the Objectives” is one of the most perspective altering works I have ever read.
Very excited to see what he does next. He mentioned on Twitter a couple times his interest in representation learning and how objective based search affects this. Very interesting stuff
One downside of rolling it into an IRA is that you'll pay more in taxes to do backdoor roth IRA conversions every year. I think it's better to leave it in 401k if the fund options / fees are acceptable.
“Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis“
https://arxiv.org/abs/2505.11581