[HN Gopher] The dual PhD problem of today's startups
___________________________________________________________________
The dual PhD problem of today's startups
Author : tosh
Score : 84 points
Date : 2020-08-19 10:09 UTC (12 hours ago)
(HTM) web link (techcrunch.com)
(TXT) w3m dump (techcrunch.com)
| crmrc114 wrote:
| This is not a new problem when people look at technology all day.
| You are basically saying that you have seen so many birds that
| there can never be a black swan. Not without complex xy and z
| factors. This is a perspective problem that can tie you down into
| some interesting thoughts such as "There is nothing else to be
| invented new." -or- "The innovation will happen here, in this
| little corner, where I and others say it will"
|
| An actress basically invented FHSS, and no one understood the
| applications it would have to future technology until much later
| on. Just because you cannot think of something new does not mean
| that no one else can- if you are working in startup funding you
| need to find the true purple cow. Not the spraypainted one, or
| the one that only lives for two weeks and has to sustain itself
| on gold.
| dnautics wrote:
| > and no one understood the applications it would have to
| future technology until much later on
|
| Huh? FHSS was a specific wartime effort with a wartime goal
| that, aside from that its modern application is mostly
| civilian, is not terribly distant from its intended use case.
| crmrc114 wrote:
| There is actually a whole backstory to people ignoring the
| idea because she was a woman etc. The navy initially turned
| down the technology when she presented it. More the the main
| point I made above- people did not know what they were
| looking at because they could not think of how important it
| would be or how it could evolve.
| dnautics wrote:
| I thought they turned it down because the permittivity of
| radio through the seawater dielectric made it impractical.
| deepnotderp wrote:
| You're correct, and at the frequencies at which
| permittivity is reasonable, FHSS is much less effective.
| And Hedy Lamarr was far from the first person to come up
| with FHSS for communications.
| madhadron wrote:
| > An actress basically invented FHSS
|
| Mind you, said actress, Hedy Lamarr, was a fairly brilliant,
| self taught electrical engineer.
| crmrc114 wrote:
| I just wanted people to have to google it and learn if they
| did not know, and here you are ruining that game for me!
| Anyhow nice film on this for anyone interested: https://en.wi
| kipedia.org/wiki/Bombshell:_The_Hedy_Lamarr_Sto...
| dmch-1 wrote:
| It may be harder to combine two skills such as sales and
| engineering, than knowledge in bio and AI, e.g., which are both
| technical.
| pontus wrote:
| The article misses the point, I think.
|
| The reason why (most, not all!) VCs are successful is not because
| they have some secret visionary insights into the future of
| technology but rather because they have the means of diversifying
| their investments in things that are more or less guaranteed to
| happen. Will work be more decentralized in 10 years than it is
| today? Yes. Will financial institutions move away from the
| archaic infrastructure it's on today over the next decade or two?
| Yes. Will education move online and become more personalized in
| the next 10 years? Yes. So, just invest in 20 remote work SaaS
| companies, 20 fintech products, and 20 online education startups
| and you'll have a fair shot at making some money. In other words,
| most VCs are really just private equity versions of index funds.
|
| Because of this, most VCs lack the experience, understanding, and
| interest in investing in highly experimental projects (there are
| exceptions of course!)
|
| As an example, I would be very surprised if any of the major VCs
| today would have invested in a small set of people who wanted to
| work on what would eventually become the transistor or TCP/IP.
| There's a reason why these things tend to start in huge corporate
| research labs (bell labs) or universities: they're not obvious
| and they're not obviously profitable.
|
| So, the real reason why these companies are not being built is
| not that the people aren't there willing to build them, it's
| because nobody's willing to listen. They're just a bunch of
| crackpots with crazy sounding ideas... until they're not.
| pnathan wrote:
| The solution involves large companies, not startups. You have an
| infrastructure and a pool of highly qualified applications that
| aren't going make or break their personal finances on these
| problems. These are classic coordination problems, and it
| requires people skilled at this aspect.
|
| Large companies don't talk about their work in this area much for
| a while for a variety of reasons. Bell Labs was a thing once....
|
| It's sort of a solvedish problem if you imagine that you are not
| bound by "silicon valley 2-person startup" rules.
| k__ wrote:
| Somehow that link redirects me to
|
| https://guce.advertising.com/collectIdentifiers?sessionId=3_...
| dustingetz wrote:
| capital is like water, it floods the low ground first
| dhairya wrote:
| This assumes that ML and AI research will continue to be silo'd
| outside of domain specific research. But it's not the case in
| academia and also increasingly in industry. You have
| computational neuroscience, bioinformatics, and many other
| traditional disciplines which have not only incorporated ML/AI
| methods but also pushed the fundamental methods research forward.
| We're increasingly seeing interdisciplinary methods and research
| becoming the norm in academia. In undergrad, nearly all the
| social science classes and all the hard science classes had some
| sort programming and quantitative methods requirement. Even in
| industry we're seeing interesting multi-disciplinary work.Many
| interesting innovations in time series ML methods have come from
| the algorithmic trading firms and medical research community has
| made contributions to computer vision and unsupervised learning
| approaches.
|
| I had a colleague who during her PhD in particle physics wrote
| from high performance parallel computation frameworks from the
| ground up in C which was better than Hadoop and Spark in
| performance. And at my last enterprise AI startup, our CTO had
| come from a computational neuroscience background. Whether these
| folks end up in creating startups is a different question, but
| the talent definitely exists.
|
| The more difficult problem is how to evaluate multi-disciplinary
| startups and businesses. There usually isn't good empirical
| evidence unless they follow a more established business model.
| SQueeeeeL wrote:
| Most of the "interdisciplinary" research I see is just one lab
| sending a dataset they generated from experiments to some ML
| collaborator, who just run some fairly low hanging Logistic
| Regression and RF on the data. I don't see a lot of places
| where people with deep statistical/computational understanding
| tackle the problems with gathering data, and working on
| understanding underlying processes. A lot of this comes down to
| how short PhD programs actually are, learning both neurology,
| GPU programming (not just plug and chug torch) and statistics
| would probably take 8-10 years; and most people/schools won't
| bare that commitment
| dhairya wrote:
| That's fair. It's hard to generalize as there is definitely a
| mixture of good and bad research out there. As a counter
| point I had to opportunity to observe the Summer Workshop on
| the Dynamic Brain (https://alleninstitute.org/what-we-
| do/brain-science/events-t...) which had fascinating
| interdisciplinary research at the intersection of computer
| vision, ML, data science, and foundational neuroscience
| research. There many other programs and research groups I can
| mention that do quality interdisciplinary research and train
| interdisciplinary students.
| michaelhoffman wrote:
| As a computational biologist I have gained sufficient expertise
| in both computation and biology to know that most of the magic
| AI biomed stuff proposed by people who have expertise in only
| one of these areas is utter nonsense.
|
| Two PhDs that don't speak the same language isn't a great
| solution, but one PhD who is a jack-of-both-trades isn't the
| only alternative either. I feel like I've done well with
| alternating collaborations with biologists who don't have a
| computational focus, and quantitative methods folks who don't
| necessarily have a focus in genomics (what we work on).
| rabidrat wrote:
| Maybe you need 3 PhDs for 2 fields: one to go deep in each
| field, and one 'jack-of-both-trades' to mediate and translate
| between them.
| digitallogic wrote:
| > Today's startups have a biologist talking about wet labs on one
| side and an AI specialist waxing on about GPT-3 on the other, or
| a cryptography expert negotiating their point of view with a
| securities attorney. There is constant and serious translation
| required between these domains, translation that (I would argue
| mostly) prevents the fusion these fields need in order for new
| startups to be built.
|
| Is that all that different from a software engineer with little
| customer facing experience teaming up with a non-technical
| cofounder who does?
| ampdepolymerase wrote:
| The last two pairs are non-issues, both have plenty of funding.
| For the former however, one misstep and you have the FDA/DHS or
| one of state medical unions breathing down your neck.
| seebetter wrote:
| This is analogous to suggesting Elon should had skipped Zip2 and
| X.com/Paypal and went direct to building rockets.
|
| Ideally the lucky few who make a ton of cash on easy software
| apps etc should be using that capital to risk solving hard
| problems and developing new sciences and technologies.
| mindcrime wrote:
| Through the first two paragraphs of this article, I thought it
| was going to be another silly rant bemoaning the lack of "real
| innovation" today. That is, another riff on the "They promised us
| flying cars, we got 140 characters" kind of rant.
|
| _One of the upsides of this job is that you get to see
| everything going on out there in the startup world. One of the
| downsides of this job is seeing just how many ideas out there
| aren't all that original._
|
| _Every week in my inbox, there is another no-code startup.
| Another fintech play for payments and credit cards and personal
| finance. Another remote work or online events startup. Another
| cannabis startup, another cryptocurrency, another analytics tool
| for some other function in the workplace (janitor productivity as
| a service!)_
|
| But I'm glad I kept reading, because there is some good stuff
| here. I mean, it's not a PhD thesis or anything, but there's some
| insights worth pondering, tucked away in this article.
|
| The gist is here:
|
| _Now, we are approaching a new barrier -- ideas that require not
| just extreme depth in one field, but depth in two or sometimes
| even more fields simultaneously._
|
| _Take synethtic biology and the future of pharmaceuticals. There
| is a popular and now well-funded thesis on crossing machine
| learning and biology /medicine together to create the next
| generation of pharma and clinical treatment. The datasets are
| there, the patients are ready to buy, and the old ways of
| discovering new candidates to treat diseases look positively
| ancient against a more deliberate and automated approach afforded
| by modern algorithms._
|
| _Moving the needle even slightly here though requires enormous
| knowledge of two very hard and disparate fields. AI and bio are
| domains that get extremely complex extremely fast, and also where
| researchers and founders quickly reach the frontiers of
| knowledge._
|
| I would agree with that sentiment in the general sense. And
| there's probably some interesting things to be gained by thinking
| deeply about how to address that problem.
|
| The only part of this I found myself disagreeing with somewhat is
| here:
|
| _We've gone through the generation of startups you can do as a
| dropout from high school or college, hacking a social network out
| of PHP scripts or assembling a computer out of parts at a local
| homebrew club. We've also gone through the startups that required
| a PhD in electrical engineering, or biology, or any of the other
| science and engineering fields that are the wellspring for
| innovation._
|
| While I agree that it's probably getting _harder_ to come up with
| something really innovative without that "fusion" approach
| alluded to above, I'm not convinced that it's not possible.
| Furthermore, I don't see being "the next no code startup" or "the
| next cryptocurrency startup" as being a Bad Thing - so long as
| you do it in a way that's appreciably better than "the other
| folks" doing the same thing.
|
| Sure, inventing something Brand New is nice, but you can make
| money making a "nicer version of something that already exists",
| or by just innovating on business model while the product is
| unchanged (or mostly so).
| starfallg wrote:
| >We've gone through the generation of startups you can do as a
| dropout from high school or college, hacking a social network
| out of PHP scripts or assembling a computer out of parts at a
| local homebrew club.
|
| None of that was really innovative, yet ended up with massive
| commercial success. MS-DOS was the not first PC operating
| system, Facebook was not the first social network on the web,
| Apple was not the first PC or smartphone maker.
|
| So I don't think anything has changed there at all. You can
| still create a massively successful venture bringing something
| out to market in a way that is somewhat incrementally 'better'
| than what is on offer without having multiple PhDs in different
| fields on your founding team.
|
| And it's pointless comparing that to the type of startup that
| is trying to creating something that is completely 'novel' from
| the intersection of 2 or more technical fields. Managing that
| type of complexity isn't something new either - it is fairly
| routine in academia to apply tools from one field to another -
| which is also how a lot of innovation happened historically as
| well as how many startups got started. Historically these type
| of ventures are high risk and the reason we are seeing a
| growing number of these is more a testament to how saturated
| the startup ecosystem is and how research increasingly is
| driven by venture capital rather than by academia and industry.
| balthasar wrote:
| Is this the onion?
| hyko wrote:
| _Every week in my inbox, there is [...] Another fintech play for
| payments and credit cards and personal finance, [...] another
| cryptocurrency_
|
| _Of course, there are a bunch of new horizons out there [...]
| Cryptocurrencies and finance._
|
| It seems a lot can change two paragraphs on. Life moves pretty
| fast these days.
|
| Edited to add: I reject the central thesis of this article, and
| pretty every one of the supporting arguments. Humans have
| required teamwork to achieve their goals from the very beginning.
| Invention has always required the synthesis of ideas from
| multiple domains. There's nothing historically unusual about
| that. What _is_ historically unusual are the diseconomies of
| scale in activities like software development. That's provided
| many market opportunities for small teams in the past four
| decades, and it will continue to do so unless those economics
| change.
|
| There are markets with high barriers to entry, and there always
| have been. Nobody was selling homebuilt aircraft carriers from
| their bedrooms in the 90s.
|
| From our vantage point, we can't tell if the seam of potential
| innovation and market configuration is anywhere close to being
| mined out in consumer tech, but my sense is that we are nowhere
| near the point where all startups need to be at the frontiers of
| all human knowledge of gtfo.
| atemerev wrote:
| Last time I checked, Bitcoin was still trading safely for
| around $12k.
| zamfi wrote:
| This is funny, but to give Danny the benefit of the doubt: he
| presumably means there are horizons out there in
| cryptocurrencies and finance that aren't approached by also-ran
| "fintech plays" or "another cryptocurrency"...
| wins32767 wrote:
| A lot of the impedance mismatch talked about in this article is
| true for any cross-domain work. Building an application in the
| medical space, you have to get engineers and doctors to
| communicate effectively. Building a new semiconductor, you need
| to get electrical and chemical engineers to get on the same page.
| Designing a new music venue and you need architects, civil
| engineers, and sound engineers to get on the same page.
|
| Successful organizations in the spaces in the article need to
| prioritize cross-training and collaboration as a first class
| value. Not doing so will lead to siloing and nobody understanding
| the whole problem.
| woeirua wrote:
| The reason you don't see more startups in the hard sciences is
| not due to the lack of hybrid talent as this article surmises.
| It's because: 1 - VCs are reluctant to fund capital intensive
| startups that have time horizons for exits that are significantly
| longer than software based startups. 2 - The product lifecycle is
| so much longer, which makes it inherently much riskier. In many
| cases it can be years before you even get to the point where you
| can get real feedback on the business model. 3 - There are often
| other considerations e.g. regulations or interactions with
| existing products, that are entirely outside of the control of
| the company that can significantly alter the likelihood of
| success.
| troughway wrote:
| Considering that only a small number of VCs make any money at
| all (power law strikes again), it doesn't surprise me that the
| rest of them are some combination of reluctant and inept when
| it comes to investing and seeing the long term game of some of
| these prospective hard science initiatives.
|
| The ones that do succeed end up spending what money they made
| to keep the deck stacked their way and crush opposition. It's
| not so much that there's no barriers to entry. They exist, it's
| all the competitors that have VC money ready to burn to keep
| you out of the game.
|
| Example: EV wouldn't have really taken off without Tesla
| battering the living shit out of it. Now the other
| manufacturers are starting to play catch up after suppressing
| it for decades. It's not like we miraculously discovered the
| technology for EV drivetrains a decade ago. It's been there all
| along, and every single one of those fuckers has been stomping
| on any and every initiative with a warchest of money to make
| sure it doesn't happen.
|
| Looking back after all these years, "invest in people not in
| products" seems like nothing more than glorified lip service.
|
| I want to agree with the article but I have no skin in the
| game. The only VC tier stuff I was involved it was F&F angel
| investing and it has worked out quite well, but the scale of
| money and time needed for "hard sciences" is beyond my level of
| expertise, and, I imagine, beyond the expertise of most VCs out
| there.
|
| In short, I suspect most VCs do not know what they are doing
| when it comes to investing, given the paltry ROIs for most of
| them. So the article is really restating that in a different
| sort of way.
| wenc wrote:
| Biotech startups are also subject to the fickle nature of
| living organisms.
|
| When I was in grad school I knew a bunch of grad students who
| worked on biotech/bioengineering experiments. They would have
| to take care of their experiments like pets, nurturing them and
| making sure they are well taken care of, because if they died
| on you, that's months of effort down the drain. Vacations had
| to be carefully planned.
|
| Whereas folks running experiments with non-living things could
| actually work 9-6 and go on holidays. Computational folks could
| even run their experiments while sitting on a beach in Hawaii
| (with an LTE signal).
|
| Bio is just a different beast.
|
| The worse thing is? Many of these biotech graduates actually
| struggle to find well-paying jobs after, despite how hot the
| field seemingly is.
| Andyfilms wrote:
| This, 100%. The author says how difficult it is for
| multidisciplinary teams to come together when they don't
| understand each others skillsets--but the same applies to the
| investors themselves. When you start talking about these
| complicated ideas, there comes a point where unless the
| investor is involved in the industry they're investing in, they
| simply won't understand the true impact of it.
|
| My company is trying to raise capital now, and that is the
| exact problem we're running into.
|
| There's a reason for the "janitor as a service" unoriginal
| ideas--because they're easy to understand, so more likely to be
| funded. Those kinds of investors are looking for the buzzwords,
| too, "as a service," "cloud," "social," "AI" that cut off ideas
| that aren't strictly consumer-facing and infinitely scalable.
| If you have a modest idea that requires a modest amount of
| money and targets a modest group of people, you're just not
| going to hear back from investors. This causes people to have
| to wrap their idea in buzzwords or lobotomize it into something
| that allows them to achieve their true goal in a sideways
| manner.
| dnautics wrote:
| It's 1.
|
| But also, who can blame them. I've seen really stupid companies
| come out of biotech incubators, including one that was peddling
| a genetically modified probiotic whose concoction as designed
| is known to be ineffective pharmacologically (and an equivalent
| reformulation strategy is not known in their host species), and
| a company that demoed reconstituted mock vegan meringues that
| had residual trifluoroacetic acid in their demo day samples.
| Vcs just don't know how to judge this shit, and it's much
| harder to pattern match details that require subtle knowledge
| than "Uber for X"
| theferret wrote:
| The author's point is valid - innovation requires more knowledge
| as the tech that required less knowledge gets built. Today we
| face the "dual phd" problem, tomorrow the tri.
|
| Obviously that is not sustainable. If society is to continually
| innovate, you need to stop building innovative systems and start
| growing them. One approach to this is true AI; not improving your
| NLP algorithm by 10% with 3x the math complexity IE the
| transformer model (quote taken from Michael Stonebrake, although
| he was referencing database research), but by building math that
| can grow math.
|
| Hell even math is becoming a road block (try integrating THAT
| Bayes!). The point is as long as we must learn to build, we will
| run into a wall as humans have finite lifespans and don't scale
| horizontally (nor do they want to). If we build something that we
| can feed or point to un-wrangled, raw information into such that
| it can learn on our behalf, we might have a shot.
|
| Now BACK to pumping out small improvement papers, innovators!
| gautamcgoel wrote:
| Another question that's worth pondering is that perhaps the
| previous generation of technologies was simply easier to develop
| than the current generation - I believe the economist Tyler Cowen
| proposes this theory in his book "The Great Stagnation". For
| example, it may be that silicon processors are just intrinsically
| easier to develop than quantum computers, traditional nuclear
| (fission) reactors are easier to develop than fusion reactors,
| better fertilizer is easier to develop compared to GMO crops,
| etc. Perhaps, as Cowen claims, we have already plucked all the
| low-lying technological fruit, so to speak.
| golergka wrote:
| > AI and bio > two very [...] disparate fields
|
| They are not, really. The field of bioinformatics exists for
| almost 20 years, as in, you can degree in it - I almost did
| myself. And the "informatics" part that you get educated about is
| pretty much data science, that, by now, uses a lot of ML methods
| and just like ML requires a very serious math foundation.
| hobofan wrote:
| That's not a lot of bioinformatics programs that I'm seeing. A
| lot of bachelors programs seem to focus on teaching almost
| exclusively the basics of BLAST and all it's boring related
| algorithms (basically everything in this Coursera course[0])
| and their mathematical foundations. Master's programs sometimes
| are a bit better with a hint of ML, but ultimately most people
| I've encountered there are still awfully unequipped to tackle
| ML problems and transfer the advances from mainstream ML to
| biology/biochemistry problems.
|
| [0]: https://www.coursera.org/specializations/bioinformatics
| asdff wrote:
| Any bioinformatics program covers machine learning these
| days. They might not have an explicit class called 'machine
| learning,' but you can bet it will be covered in the lecture
| sequence and the cutting edge of the field will be discussed
| in journal clubs, rather than in lectures which are about
| established fundamentals.
|
| For a pure biology undergrad who is probably med school
| bound, learning ML is superfluous so you don't see it in the
| curriculum at the undergrad level, unless there are specific
| concentrations offered for computational biology. A
| bioinformatics program may even just have you take these ML
| classes from the statistics or CSE department rather than
| offer some bioinformatics-specific section within their
| department.
| TrackerFF wrote:
| Interestingly enough, things like Machine Learning grew out of
| domain fields. A couple of decades ago, there were few - if any -
| dedicated programs for Machine Learning. The research and grads
| came mostly from domain-specific fields, like Computer Vision,
| Signal Processing, Computational Biology / Chemistry, Statistics,
| Applied Math, etc. Then when things got more cohesive, the most
| important parts formed into a more or less "pure" field of
| Machine Learning.
|
| Today, you can study Machine Learning without having to focus on
| any particular domain (well, other than stats and applied math,
| which lays the foundation for the theory).
|
| But, yes, it is tough and demanding to find people that have deep
| / expert knowledge in both their respective domain, AND machine
| learning / data science / AI.
|
| I think maybe one way to do it is to just look after domain
| experts, and learn them enough about ML and DS (if they lack the
| background) to work as generalists. Enough that they can read and
| discuss it.
|
| And then, you hire ML scientists and engineers to do the nitty-
| gritty work, with the input and feedback from the domain experts.
___________________________________________________________________
(page generated 2020-08-19 23:00 UTC)