Hacker News Reader

Respectfully I don't think you're characterization of early 20th century and today's research is accurate at all.

There is a stupendous amount of new science being carried out. Science that we'll look back on 80 years later (just as we look back on the science of the 1920s) and recognize how impactful that science has been. In fact, I would imagine that the science that is being worked on today will prove far more impactful than the research of the 1920s.

Think about the seminal technologies that are still in their infancy! Deep learning in it's current form (AKA the form that actually works) is ~10 years old. That's it. And in those ~10 years we've gone from cute theory to being able to convincingly emulate video, translate languages, solve dictation, and generate near-human quality text. Then there's CRISPR, which could potentially cure 1000s of diseases and dramatically improve diagnostic tools, and is similarly ~20 years old. There are many thousands of other groundbreaking technologies that are being worked on as we speak.

Contemporary scientists are conducting this research.

Deep learning has nothing to do with "science" and the last algorithmic advances that enabled it happened rather more than 10 years ago: the neocognitron, in 1979, the Long-Short Term Memory recurrent neural nets in 1995, backpropagation (for neural network training) in 1986, etc.

In general, all the interesting work that enabled today's deep learning boom happened towards the end of the 20th century and recent advances are primarily owed to increases in computational power and availability of data sets.

Says not me, but Geoff Hinton:

Geoffrey Hinton: I think it’s mainly because of the amount of computation and the amount of data now around but it’s also partly because there have been some technical improvements in the algorithms. Particularly in the algorithms for doing unsupervised learning where you’re not told what the right answer is but the main thing is the computation and the amount of data.

The algorithms we had in the old days would have worked perfectly well if computers had been a million times faster and datasets had been a million times bigger but if we’d said that thirty years ago people would have just laughed.

http://techjaw.com/2015/06/07/geoffrey-hinton-deep-learning-...

Google published "Attention is all you need" in 2017, which introduced the Transformer Deep Learning Architecture. This is the general framework by which GPT-3 was trained, and is increasingly replacing most other architectures in NLP tasks.

So.. I guess I disagree strongly with your premise? Look, it takes some time for these ideas to propagate to non-experts. In 20 years a hacker-news poster will likely post the same comment, referencing the 2017 Transformer paper to enforce the idea that all the _good_ science in DL was done years ago.

You're mistaking the hot new fad of Transformers for "new science" and the field of NLP for a scientific field. Transformers have only taken over NLP because that entire field of research has run out of new ideas, and has been desperately grasping for straws in the last twenty years. Consequently, NLP researchers have been jumping on this bandwagon or that ever since I can remember. I enquired about a PhD in NLP as soon as I finished my degree, in 2011, but by 2014, when I finished my Master's, I knew that was the last thing I wanted to study: a subject riddled with sloppy practices, flooded with shaky empirical results, without any real theory to guide it, without even any useful benchmarks or credible metrics that could justify the empirical direction. That is why the GPT-3s and the BERTs and all those data- and power-hungry monstrosities reign supreme: because NLP is built like a house of cards, with no scientific foundation whatsoever, so that anybody can claim anything they like and there is almost nobody left to challenge even the most outlandish claims.

>> In 20 years a hacker-news poster will likely post the same comment, referencing the 2017 Transformer paper to enforce the idea that all the _good_ science in DL was done years ago.

On the contrary, my expectation is that in 20 years from now people will point to Transformers and laugh and joke about how misguided people were "back then" to think that just building larger neural nets would lead to AGI, as they do now for earlier AI approaches.

What I'm not convinced is that you are any different than the "non-expert" in your comment, to whom "it takes some time" for ideas to propagate. You are holding up Transformers as some kind of scientific breakthrough but that is only the most obvious conclusion to draw from the overhyping of the approach on social media and the lay press.

I work in the field and transformers have literally transformed science.

Can you please clarify which "field" is that and how do you "work" in it?

the field is the application of machine learning to biology (specifically drug design, but also the wider area of computational biology in all its forms). Here's some of my prior work: https://arxiv.org/abs/1502.02072 however that doesn't use transformers. The DeepMind work on protein structure prediction does, and it represents one of the greatest technical achievements of the modern era.

Transformers basically took anything that was "sequence learning" (large fractions of computational biology are) and made it work 100X better over night.

Thank you for clarifying. I see that the paper has six authors two of whom declare equal contribution. Could you please further clarify what was your personal contribution to the paper (to the extent that this doesn't unecessarily expose your personal information)?

Regarding Transformers, the results on protein structure prediction are certainly impressive, but they are not enough to justify your comment that "transformers have literally transformed science". At most we can tell that they have considerably improved protein structure prediction. Perhaps your comment was an exaggeration?

Edit: I can't find the paper on the ICML 2015 website, although I can see that it has a few hundred citations, as a preprint, which is not uncommon these days. But, could you point to some published work you contributed to?

I really do believe that the attempt to industrialize scientific research has failed. Science cannot be done at scale. It comes from singularly focused projects lead by pure inquiry. I don't regret my PhD because I enjoyed achieving it, but I see it as a waste of time.

There's an analogy in Software Engineering. Programmers are akin to Artisans or Craftsmen, but the industry wants and needs them to be something more like little factories. I don't think that approach makes good software, even if it's a good product. Just as I don't think the system we have created in academia doesn't produce good research, even though it makes good careers.

I don't want to point a finger, but the same time period has seen explosive growth of the life sciences. Doing good science on "squishy stuff" is just plain hard, for two reasons: First, isolating a single cause and effect, is almost always impossible. Everything is multivariate and noisy. Second, many of the things we're urgently interested in learning involve human health and welfare, and there's probably a widespread feeling that we don't have a century or two to wait for solid answers.

My day job involves designing measurement equipment, some of which is used in life science. I absolutely care about knowing whether a measurement is any good or not, to the point where it's not just a paycheck, but a true passion. And I work with people in actual research who care just as much, using my stuff to do front-line work.

There are certainly problems, and I don't wish to defend the "publish or perish" system, but I don't think a cynical view of the causes is a complete or even accurate picture of what's going on.

Mind-blowing discoveries and realizations that fundamentally change how we view the universe and to some extent even life itself in the late 20th century, off the top of my head:

* Cosmology. Most of the cosmological theories (e.g., Big Bang) didn't develop until after 1950, and the Hubble Space Telescope has advanced the field considerably.

* Supersymmetry/superstring theory. I hesitate to put this here because I think this is a wrong turn for theoretical physics, but it is a very well-known model that developed in the 1970s, with the superstring theory revolution occurring in the 1970s.

* Central dogma of molecular biology. Proposed in the late 1950s, this states that basically DNA -> RNA -> protein, and nothing goes back the other way. It has also taken a massive beating in recent decades, as we've discovered there's quite a few ways that information gets passed around in cells that don't follow this path (e.g., epigenetics, and it's an epic acerbic word match if you want to restrict epigenetic only to things like methylation of DNA or include any other heritable non-genetic traits as epigenetic).

* Plate tectonics. This became accepted geological theory only around the late 1960s.

This is just a list of pure science discoveries I can think of off the list of my head. Stretch to applied science advancements, and there's several that have caused massive revisions to fields--organic synthesis techniques, NMR spectroscopy, radiocarbon dating, genetic sequencing are all things that came to mind while compiling the above list. If you switch over to math/CS side of things, you can also consider that the development of RSA quite literally opened up a new world of what can be done with computers, or the Cook-Levin Theorem that really kicks off computability theory (and leads to the asking of the question P=NP). Pretty much everything about CS only develops in the late 20th and early 21st century.

Hell, that last comment prompted me to remember something else that rather fundamentally changes how we view the universe. It turns out that it's far easier to teach a computer how to play chess than speak with the intelligence of a three-year-old, never mind an adult. What can be more mind-blowing than discovering that we can't make a more effective learner than a toddler?

FWIW, most studies of academic problems sort of pinpoint the [late?] 90s as a change point era for various reasons.

I also have a hunch that replication problems are not equal everywhere. I suspect that biomedicine and the social sciences are worst off (although I think CS ML might also have some issues).

I'm not agreeing or disagreeing with the OP, or you, but the 1950's and '60s should be "mid-20th century", not "late". I don't know if there's a commonly accepted definition of "early", "middle" and "late" that does not involve waving hands.

I assume we have a whole lot more scientists than we have back then, both in total and per capita.

Still, if scientific progress is in any way real and not cyclical (as in Kuhn), then science should be getting harder. It stands to reason that if there are fewer and fewer secrets to find, finding them (or randomly stumbling upon them) will be a lot rarer.

We tend to assume that there is an infinite amount of knowledge to be gained of the natural world, but it's not necessarily so. It's especially not evident that our tools are evolving as fast as they need to, or even that they can evolve much further.

I'm not saying "everything has been found", more like "all progress makes further progress require more effort, by definition".

I agree with this. I’ve been coming to similar conclusions myself.

If you have not, you should read The Idea Factory.

https://www.amazon.com/Idea-Factory-Great-American-Innovatio...

We owe most of our "modern world" to the hard science done at Bell Labs in the 1900s, before the corporation was singularly focused on quarterly profits.