I recently summarized the state of machine learning. The post served as a way to keep up with what has happened until now. Predictably, it triggered a lot of questions about what will happen. Rather than make bold predictions of future breakthroughs, it seems better to ask “what is a suitable framework based on which we can interpret new results and guess possible future capabilities?”.

In general, a challenge with keeping up with machine learning research is that a list of “the latest cool stuff” gets old very quickly. At least if results are considered in too much detail. There are new results published all the time, but in reality, “breakthroughs” are rather few. Most papers are “improved state-of-the-art by 1%”-noise. I do not think you can predict the future of research since a lot of innovation is accidental and unexpected. But we can do a lot better than random guessing. It is necessary to try predicting the future since there is a lot at stake for society, business, and humans. So, here are some thoughts on how to reason about the future of machine learning:

PaLMs and Flamingos

Let’s use two recent papers as examples. They were both published very recently (so recent in fact, that I did not cover them in my previous post). The first one, PaLM , was published in early April 2022 (blog, paper) and the second one, Flamingo , was published in late April 2022 (blog, paper). Both papers contain mind-blowing demo results (see examples below).

Caption: Flamingos having conversations with humans.

Caption: PaLM explaining jokes.

So, how do we make sense of these papers? And what do they mean for the next wave of machine learning applications?

What is actually new?

All research builds on previous results, so we should ask ourselves: what is actually new about these papers? It is tempting to focus on the amazing demo output, but a more interesting question is “why does it work better than before?”. Later, we will of course also focus on when it does not work.

Let’s start with what is not new. The core concepts of numerical optimization and deep neural networks have remained largely unchanged for a long time. Neither PaLM nor Flamingo changes anything about these core concepts. Of course, there are improvements and tricks used to make them perform better, but the fundamental idea is more or less the same. It also turns out both papers make use of the same kind of embeddings that we’ve seen for the last 12-13 years now, i.e. models representing information as dense vectors. Finally, both models leverage attention and transformers to learn improved embeddings. We covered these concepts in the previous blog post.

Larger Models are (still) Better

It turns out that the progress shown in recent papers is mostly about size , both in model and data. To quote the authors of the PaLM paper:

Since GPT-3, a number of other large autoregressive language models have been developed which have continued to push the state of the art forward. […] The improvements in these models have primarily come from one or more of the following approaches: (1) scaling the size of the models in both depth and width; (2) increasing the number of tokens that the model was trained on; (3) training on cleaner datasets from more diverse sources; and (4) increasing model capacity without increasing the computational cost through sparsely activated modules. […] This critically demonstrates scaling improvements from large LMs have neither plateaued nor reached their saturation point.

The fact that transformers can be parallelized, and the emergence of new, innovative ways to train models such as Pathways make it possible to train larger and larger models. The unexpected result is that we should expect models to keep improving as they keep growing. People have been suggesting this for years, but I’ve always believed more sophisticated representation methods would be required. I mean, intuitively more context when learning should make a model more capable, but the fact that the latent space can organize the information so successfully is surprising. At least to me. It could have very well been that information got “jumbled” as the model grows, but it appears not to. It also appears as if concepts can be combined and reused across domains in a surprisingly robust way.

Chain-of-Thought Prompting

Besides size, a new thing from my perspective is that PaLM demonstrates that when model scaling is combined with * chain-of-thought prompting* the model can solve problems that require multi-step mathematical or commonsense reasoning to produce the correct answer. I hadn’t read much about this concept before, so for me, this was new. It looks like the best paper to read on this is “Chain of Thought Prompting Elicits Reasoning in Large Language Models”. The purpose is to handle so-called “system-2” tasks such as logical, mathematical, and common sense reasoning. Large language models exhibit flat scaling curves for such tasks, i.e. they will not improve with the size of the model. To handle this, the authors propose a way for these models to decompose multi-step problems into intermediate steps. The idea is basically to let the model generate a coherent set of short sentences that lead to the answer to a reasoning problem.

Multimodal Generative Modelling

Turning to Flamingo, the most interesting aspect of this model is that it is:

it is a visually-conditioned autoregressive text generation model able to ingest a sequence of text tokens interleaved with images and/or videos, and produce text as output.

Basically, it can process images and text interchangeably in a sequence and predict a suitable next sequence of text tokens. When considered together with DALLE2 I assume it’s just a matter of time before it can also respond with an image every now and then. Imagine the human operator inputting a sequence of text and images, and then asking “Show me what you are thinking”, to which the model would output an image capturing its “state of mind”. Such an exchange should be possible.

Diffusion Models

It is clear that GANs have a powerful new contender: diffusion models. To quote Lilian Weng:

Diffusion models are inspired by non-equilibrium thermodynamics. They define a Markov chain of diffusion steps to slowly add random noise to data and then learn to reverse the diffusion process to construct desired data samples from the noise. Unlike VAE or flow models, diffusion models are learned with a fixed procedure and the latent variable has high dimensionality (same as the original data).

Diffusion models are used in DALLE2 to generate data by reversing a gradual noising process. So basically they “remove noise”. Noise in this context is a rather complicated concept. It’s not a lack of sharpness or anything, it’s more radical than that. It’s more like taking a crude vector and finding various suitable images based on it.

Caption: Description from original paper.

If you want to read more about it, I found this post to be very well written: How DALL-E2 Actually Works. Below is an example from the DALLE2 paper showing the variations created by the diffusion model on the same core vector.

Graph Neural Networks

A concept that is not covered by the papers I’ve focused on, but that is offering new and exciting technology is Graph Neural Networks (GNNs). They have been on the fringe of research until recently, but are not exploding in interest. The idea is that you can create mesh-based simulations and predict how meshes change over time depending on external factors. That can be used for things like ETA in traffic. They are still not very resource-efficient, but new methods like RevGNNs appear to lower the cost dramatically. These models appear to be useful for long-time planning in reinforcement learning. An example is this paper: “World Model as a Graph: Learning Latent Landmarks for Planning”.

What are the limitations of these models?

As impressive as the demos are, there are still limitations to these models. Knowing their limitations, and how likely it is that we can overcome those limitations, will be important to predict what is possible in the future.

The need for webscale human-curated knowledge

These models still depend on annotated data. You might protest, and the authors of these papers like to highlight that they avoid the need for large amounts of annotated data. At the same time, they clearly rely on “webscale datasets”. It turns out the internet is made by humans. All the text written, and the image captions created, are written by humans. So in fact, the concepts manifested in the model’s latent space, need to first be described by humans in text and associated with images. And as a result, these models will also learn whatever undesired concepts and biases are present online. How to handle that, and mitigate that, will be a big part of making these models aligned with our interests.

What these papers show is that if you have webscale amounts of human-curated knowledge a lot of tasks can be solved using the same foundational model. So in that sense, you do not need task-specific annotated data assuming the task is a subset of what is described on the internet. Which tasks are implicitly described online, and which are not, is not clear to me at this point.

Fixed Amount of Thinking

These models all have a limited amount of FLOPS to spend, i.e. they are not “continuously active”. This limits what tasks can be solved. Input is fed through the network and output is generated. Once that is done, nothing else can happen. I’m not actually sure how big this limitation is scientifically, or if it’s more an issue of cost/energy. Maybe the government or military can keep one of these models spinning continuously. Just hook it up to a nuclear reactor or something? Humans can choose to think more about something if it is complicated. These models cannot.

Models are still Great at Bullshitting

Models are still great at bullshitting when they “do not know” the answer. I get the impression they are “biased for action” in the sense that most versions of these systems prefer some output over “I do not know”. In fact, it seems models are “extremely sure” all the time, even when the output is more or less ungrounded. Paper authors describe this as “hallucinations” which might be a case of anthropomorphizing. It could also be that there are sufficiently similar concepts in the data that the model picks a prediction that humans feel is nonsense, but that is actually statistically accurate. Which is correct, that which a human finds intuitive, or that which is most statistically likely?

Caption: From the Flamingo paper.

Cherrypicked Results and Leading Questions

Of course, the results you see in the papers are impressive, but there are also a lot of really strange results popping up. I’m less worried about this because if these models do in fact keep getting better and better with size, the frequency of obvious errors will probably go down. But either way, estimating performance based on examples in papers is hard. Sure, the benchmarks are there to make sure performance evaluation is unbiased, but when you actually consider the numbers they show, there will be a lot of strange results if you interacted with a model. I also think humans tend to write leading questions. Take the following example:

Caption: Example of Flamingo interacting with a human. Are the questions leading it to the right response?

It is all virtual

Recent models are still completely virtual. I recently learned about the Moravec’s paradox which hypothesizes that:

[…] contrary to traditional assumptions, reasoning requires very little computation, but sensorimotor and perception skills require enormous computational resources […] In general, we're least aware of what our minds do best and we're more aware of simple processes that don't work well than of complex ones that work flawlessly […] it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.

This feels right to me, but I obviously don’t know. Maybe I’m just giving myself a false sense of safety. Maybe we should not feel too scared about these algorithms until we’ve actually made a lot more progress on physical robotics. Maybe reasoning isn’t actually that hard?!

What will happen next?

It might seem like a philosophical question: “what will the future of machine learning look like?”. I’ve long been a skeptic. For years I’ve been saying “actually, computers are pretty dumb”. Recently, I’ve started changing my mind. The emergence of massive latent spaces and the fact that human knowledge can apparently be represented in such spaces, makes me rethink things. In my case, I own and run a company that is tightly connected to the future of machine learning, so it is my job to keep track of this. But I think it also matters as a human. So, what can we expect?

More Modalities are Merged

The obvious, short-term next step is to combine text, images, and sound. I’m sure there are already really promising results brewing inside Google and other companies. We will be talking to these models in natural language and getting audible responses very soon. Alexa, Siri, and Hey Google are about to get a massive upgrade. With DALLE2 it feels likely we will also see interactive visuals that go with these audio interfaces before long. I think learning human emotion and facial expressions is probably possible if you understand language and images, and have access to the entire internet.

The Scaling Hypothesis Keeps Holding

Maybe we are all just brute-forcing life? I highly encourage you to read Gwerns excellent post on “The Scaling Hypothesis”. To quote Geoff Hinton:

Extrapolating the spectacular performance of GPT3 into the future suggests that the answer to life, the universe and everything is just 4.398 trillion parameters.

There is no evidence today that very large models will suffer diminishing returns in their ability to learn. If that holds over time, these models should approach human-equal performance by the time their parameter space matches our brain. There might be as many as 1 trillion synapses in the brain, i.e. neurons connected to each other and passing signals to each other. The largest language models are approaching 1 trillion parameters. Together with a continuously growing size of encoded human knowledge, it is now possible that these models will match us in just a few years.

That does not mean we will have Terminator-style robots in a few years, it means we will have virtual entities that we can talk to that are indistinguishable from humans. Maybe. On the other hand, we might find that “Oh fuck, there was a roadblock here that we didn’t expect” and suddenly improvements are in fact diminishing.

A lot of new companies are emerging

The “Attention is All You Need” paper lead to the founding of a bunch of new start-ups. Just the authors alone have started a bunch, and others have followed in nearby domains. Here are some of my favorites.

Predictably, most of the companies are based in the US:

Access to Quality Knowledge becomes the Short-term Limitation

I think machine learning systems are already suffering from the same problems humans do: a lot of them are garbage. I mean, being a human does not guarantee you are a good, productive member of society. Becoming a great person requires careful “tuning” through education, parenting, selective reading, coaching, and mentoring. If we just consume everything around us all the time and give in to all our urges, we become awful people.

I think we will find that as more capable ML systems emerge, we will lose some of the normal advantages of computers: preciseness and predictability. I think that’s a passing problem since once you’ve perfected the knowledge base for one of these entities it could theoretically live forever, and keep improving.

I think we can spend eternity curating a knowledge base and debating what is right. I mean, that’s how we spend most of our time anyway right? As soon as you have enough food and shelter, we climb the Maslow stairs and spend our time pondering the meaning of life. An algorithm could do that forever without any really progress. Sure, it might pick an optimization problem to focus on, and that might be bad. But it’s also possible it will just make and watch soap operas all day. Who knows?

A few reading recommendations

The State of AI. During my research on this, I’ve spent a lot of time reading the excellent stuff that Nathan Benaich is publishing (twitter, website).

Gwern. There are so many good posts on the site:

My last link will be from that site, and it is probably a good place to go next. As Gwern puts it:

It might help to imagine a hard takeoff scenario using solely known sorts of NN & ⁠scaling effects… [This] is a story which may help stretch your imagination and defamiliarize the 2022 state of machine learning.

Read this: It Looks Like You’re Trying To Take Over The World

Conclusion

The last few weeks have made me begin to adjust my assumptions about what machine learning will be able to do. I now think it is likely that the scaling hypothesis will hold, and that we will experience human-level competency in virtual entities in the next few years. That’s going to bring massive change to society, work and business. Be prepared.