Independent alignment researcher
Garrett Baker
Karma: 3,150
There is evidence that transformers are not in fact even implicitly, internally, optimized for reducing global prediction error (except insofar as comp-mech says they must in order to do well on the task they are optimized for).
Do transformers “think ahead” during inference at a given position? It is known transformers prepare information in the hidden states of the forward pass at t that is then used in future forward passes t+τ. We posit two explanations for this phenomenon: pre-caching, in which off-diagonal gradient terms present in training result in the model computing features at t irrelevant to the present inference task but useful for the future, and breadcrumbs, in which features most relevant to time step t are already the same as those that would most benefit inference at time t+τ. We test these hypotheses by training language models without propagating gradients to past timesteps, a scheme we formalize as myopic training. In a synthetic data setting, we find clear evidence for pre-caching. In the autoregressive language modeling setting, our experiments are more suggestive of the breadcrumbs hypothesis.
Hi John,
thank you for sharing the job postings. We’re starting something really exciting, and as research leads on the team, we—Paul Lessard and Bruno Gavranović - thought wed provide clarifications.
Symbolica was not started to improve ML using category theory. Instead, Symbolica was founded ~2 years ago, with its 2M seed funding round aimed at tackling the problem of symbolic reasoning, but at the time, its path to getting there wasn’t via categorical deep learning (CDL). The original plan was to use hypergraph rewriting as means of doing learning more efficiently. That approach however was eventually shown unviable.
Symbolica’s pivot to CDL started about five months ago. Bruno had just finished his Ph.D. thesis laying the foundations for the topic and we reoriented much of the organization towards this research direction. In particular, we began: a) refining a roadmap to develop and apply CDL, and b) writing a position paper, in collaboration with with researchers at Google DeepMind which you’ve cited below.
Over these last few months, it has become clear that our hunches about applicability are actually exciting and viable research directions. We’ve made fantastic progress, even doing some of the research we planned to advocate for in the aforementioned position paper. Really, we discovered just how much Taking Categories Seriously gives you in the field of Deep Learning.
Many advances in DL are about creating models which identify robust and general patterns in data (see the Transformers/Attention mechanism, for instance). In many ways this is exactly what CT is about: it is an indispensable tool for many scientists, including ourselves, to understand the world around us: to find robust patterns in data, but also to communicate, verify, and explain our reasoning.
At the same time, the research engineering team of Symbolica has made significant, independent, and concrete progress implementing a particular deep learning model that operates on text data, but not in an autoregressive manner as most GPT-style models do.
These developments were key signals to Vinod and other investors, leading to the closing of the 31M funding round.
We are now developing a research programme merging the two, leveraging insights from theories of structure, e.g. categorical algebra, as means of formalising the process by which we find structure in data. This has twofold consequence: pushing models to identify more robust patterns in data, but also interpretable and verifiable ones.
In summary:
a) The push to apply category theory was not based on a singular whim, as the the post might suggest,
but that instead
b) Symbolica is developing a serious research programme devoted to applying category theory to deep learning, not merely hiring category theorists
All of this is to add extra context for evaluating the company, its team, and our direction, which does not come across in the recently published tech articles.
We strongly encourage interested parties to look at all of the job ads, which we’ve tailored to particular roles. Roughly, in the CDL team, we’re looking for either
1) expertise in category theory, and a strong interest in deep learning, or
2) expertise in deep learning, and a strong interest in category theory.
at all levels of seniority.
Happy to answer any other questions/thoughts.
Bruno Gavranović,
Paul Lessard
From The Guns of August
Old Field Marshal Moltke in 1890 foretold that the next war might last seven years—or thirty—because the resources of a modern state were so great it would not know itself to be beaten after a single military defeat and would not give up [...] It went against human nature, however—and the nature of General Staffs—to follow through the logic of his own prophecy. Amorphous and without limits, the concept of a long war could not be scientifically planned for as could the orthodox, predictable, and simple solution of decisive battle and a short war. The younger Moltke was already Chief of Staff when he made his prophecy, but neither he nor his Staff, nor the Staff of any other country, ever made any effort to plan for a long war. Besides the two Moltkes, one dead and the other infirm of purpose, some military strategists in other countries glimpsed the possibility of prolonged war, but all preferred to believe, along with the bankers and industrialists, that because of the dislocation of economic life a general European war could not last longer than three or four months. One constant among the elements of 1914—as of any era—was the disposition of everyone on all sides not to prepare for the harder alternative, not to act upon what they suspected to be true.
But such people are very obvious. You just give them a FizzBuzz test! This is why we have interviews, and work-trials.
This style of argument proves too much. Why not see this dynamic with all jobs and products ever?
I don’t think the bitter lesson strictly applies here. Since they’re doing learning, and the bitter lesson says “learning and search is all that is good”, I think they’re in the clear, as long as what they do is compute scalable.
(this is different from saying there aren’t other reasons an ignorant person (a word I like more than outside view in this context since it doesn’t hide the lack of knowledge) may use to conclude they won’t succeed)
Sounds right. It would be interesting to see how extremely unconvincing you can get the prompts and still see the same behavior.
Also, ideally you would have a procedure for which its impossible for you to have gamed. Like, a problem right now is your could have tried a bunch of different prompts for each value, and then chosen prompts which cause the results you want, and never reported the prompts which don’t cause the results you want.
The main concern I have with this is whether its robust to different prompts probing for the same value. I can see a scenario where the model is reacting to how convincing the prompt sounds rather than high level features of it.
Is this coming from deep knowledge about Symbolica’s method, or just on outside view considerations like “usually people trying to think too big-brained end up failing when it comes to AI”.
I will tell you the received wisdom from my friends’ experience, and their friends’ experience with polyphasic sleep: It is in theory doable, but often in practice a disaster because you end up getting less sleep than the minimal theoretical requirements given by polyphasic sleep.
If on polyphasic sleep you are sufficiently undisciplined that you end up racking up 40 hours of sleep debt, this wisdom would say you should probably stop doing polyphasic sleep. And instead of biohacking your way out, just have a few nights of normal sleep.
My intention was not to dismiss or downplay the importance of various values, but instead to clarify our values by making careful distinctions. It is reasonable to critique my language for being too dry, detached, and academic when these are serious topics with real-world stakes. But to the extent you’re claiming that I am actually trying to dismiss the value of happiness and friendships, that was simply not part of the post.
I can’t (and didn’t) speak to your intention, but I can speak of the results, which are that you do in fact down-play the importance of values such as love, laughter, happiness, fun, family, and friendship in favor of values like the maximization of pleasure, preference-satisfaction, and short-term increases in wealth & life-spans. I can tell because you talk of the latter, but not of the former.
And regardless of your intention you do also dismiss their long-term value, by decrying those who hold their long-term value utmost as “speciesist”.
This view seems implicit in your dismissal of “human species preservationism”. If instead you described that view as “the moral view that values love, laughter, happiness, fun, family, and friends”, I’m sure Aysja would be less alarmed by your rhetoric (but perhaps more horrified you’re willing to so casually throw away such values).
As it is, you’re ready to casually throw away such values, without even acknowledging what you’re throwing away, lumping it all unreflectively as “speciesism”, which I do think is rhetorically cause for alarm.
My apologies
The question is not how big the universe under various theories is, but how complicated the equations describing that theory are.
Otherwise, we’d reject the so-called “galactic” theory of star formation, in favor of the 2d projection theory, which states that the night sky only appears to have far distant galaxies, but is instead the result of a relatively complicated (wrt to newtonian mechanics) cellular automata projected onto our 2d sky. You see, the galactic theory requires 6 parameters to describe each object, and posits an enormously large number of objects, while the 2d projection theory requires but 4 parameters, and assumes an exponentially smaller number of particles, making it a more efficient compression of our observations.
I’ve usually heard the justification for favoring Everett over pilot wave theory is on simplicity terms. We can explain everything we need in terms of just wave functions interacting with other wave functions, why also add in particles to the mix too? You get more complicated equations (so I’m told), with a greater number of types of objects, and a less elegant theory, for what? More intuitive metaphysics? Bah!
Though the real test is experimental, as you know. I don’t think there’s any experiments which separate out the two hypotheses, so it really is still up in the air which actually is a better description of our universe.
I have a bit of a different prescription than you do: Instead of aiming to make the community saner, aim to make yourself saner, and especially in ways as de-correlated from the rest of the community. Which often means staying far away from community drama, talking with more people who think very differently than most in the community, following strings of logic in strange & un-intuitive directions, asking yourself whether claims are actually true when they’re made in proportion to how confident community members seem to be in such claims (people are most confident when they’re most wrong, for groupthink, tails come apart, and un-analyzed assumptions reasons), and learning a lot.
A kind of put on your own mask before others’ sort of approach.
People have criticized Eliezer for taking time to write fan fiction and indulge in polyamorous orgies, but notice that he hasn’t burned out, despite worrying about AI for decades.
Not really relevant to your overall point, but I in fact think Eliezer has burnt out. He doesn’t really work on alignment anymore as far as I know.
Do these things happen automatically as a consequence of trying to be rational, or did just someone accidentally build the Bay Area community on top of an ancient Indian burial ground?
As someone “on the ground” in the Bay Area, my first guess would be that the EA and rationality community here (and they are mostly a single community here) is very insular. Many have zero friends they meet up with regularly who aren’t rationalists or EAs.
A recipe for insane cults in my book.
My guess is that we (re)perceive our perception as a meta-modality different from ordinary modalities like vision, hearing, etc, and that causes the illusion. It’s plausible that being raised in a WEIRD culture contributes to that inclination.
This seems exceedingly unlikely. Virtually every culture has a conception of “soul” which they are confused about, and ascribe supernatural non-materialist properties to.
I wonder if everyone excited is just engaging by filling out the form rather than publicly commenting.