See also No Revival for the Industrial Research Lab, Notes on Adam Marblestone’s Focused Research Organizations, and Is There a Translational Research Gap?.

More anecdotally, see Bus Factor 1, The Murder of Wilbur Wright and Highlights from the Emails on Golden Handcuffs.

All emphasis mine, some of these are several comments or emails grafted together, with minor omissions. Commentary at the bottom.

Adam Marblestone

One comment is actually I agree a lot of effort is going to “bridging the valley of death”.

I actually think FROs have a major, perhaps primary, purpose outside of that “translational” aspect. Even to develop tools for basic research, or public good — things that are never going to “graduate” to become startups —you may need to structure some efforts in a more “fast and focused”, team based, systems engineering heavy way than academia allows. Example: faster cheaper better brain mapping… not a startup, but needs a systems-focused non-academic team to get it up and running.

Thus, the Day One article which mentions “basic research” challenges that require FROs rather than “valley of death” may be a better motivation. The Twitter reply to Sarah is actually a kind of perversion of the main idea, therefore, massaging it to be more about valley of death per se than it need be.

…Picking the right problems is key, as is having metrics of progress but without letting that over determine the effort / prevent any pivoting on the way to the goal.

Many problems won’t be a good fit for them.

So yeah, I appreciate your pointing out that this is just one experiment that needs to be matched to specific rare problems that fit it, and not meant to replace the whole system.

That said, I think some of the problems they address — given that academia and startups are big enterprises already — could be “long poles in the tent” in some areas. In other words our R&D system is only as strong as the weakest element and if FROs can solve for some rare weak elements that could have a big impact overall.

Also there can be a notion of users, just scientific users not always VC grade giant commercial markets

Sarah Constantin

I think a handful of funds existing is good validation of the idea rather than evidence it’s already overdone. The gap is probably big even if these guys are doing everything right.

…btw, I should just say, I try a little to be right on Twitter, but I’ve never been interested in doing social science as an intellectual endeavor.  I’m glad some people are doing it, but I take my verbal activity way less seriously than that.

Nathan Taylor

Bell Labs was created as a defense against monopoly due to Bell risk of antitrust enforcement. As a public utility getting monopoly profits, Bell viewed itself as a public institution which could invest in basic R&D which might not impact the bottom line.  The clear parallel here is Google Lab moonshots. Again, a tech monopoly wants to showcase its virtues, and use its monopoly profits to invest in blue sky work. And Sun and XEROX PARC, just like Bell Labs, never really got return for their most important inventions. They became public goods, which new tech companies picked up and ran with. Or, given that basic research is so hard, the most common result is something like google glasses.

My point here is that Industrial Research Labs which we (historically) love are the product of tech monopolies tossing money around to make themselves feel good about their public service. And avoid risk of antitrust. They should arguably be funded by the company as a marketing expense. The people hired of course are good and sometimes produce incredible results.  But the benefits go to other companies than the ones who did the research. To use Bell Labs as a model we want to recreate without realizing it’s a product of monopoly profits, which had very little chance of benefiting the funding company itself, is to misunderstand what happened.

Bell Labs wasn’t murdered. It was a quirky exception vanity project which randomly produced large public benefits, while be of little or no use to Bell itself. It was never going to survive long, because it was a weird, quirky contingent exception. Trying to make more Bells Labs is a bad idea, because it’s not reproducible. Of course I may be wrong about this position about corporate research labs. And I’m exaggerating a bit to make the point. But I think this view is broadly defensible.

Where does that leave us? I think government funding for prizes (longitude and watches), or guaranteeing payment if tech pays off (mRNA vaccines) is a very sustainable way to go to push tech forward. Government leadership on picking where to invest, with large amounts of money going to companies, who only get paid if what they tried works. Otherwise the companies get nothing.

The result we want is technology with large public goods benefits. Well done government funding is a far better aligned way to create tech public goods than corporations.

Commentary
Adam’s “tent” metaphor is a useful alternative to thinking about one-dimensional gaps. Some technologies (including mRNA vaccines) require support over a path across the research space, so it’s not enough to have each individual discovery funding if you cannot do the “systems building” Adam talks about.

Sarah is maybe right about translational research, though given that Breakout Labs is a for-profit VC and 11 years old, I would expect it to scale up if the returns are actually good. Financial returns are an imperfect proxy for scientific ones, and as mentioned last time, 11 years might not be enough time to find out if the project was a success.

Nathan makes some good points, though I’m not sure Bell Labs was as quirky as he imagines. As I understand it, the lab was quite successful for multiple decades. At the very least, around 1940-1970.

As aggravating as it can be to see corporations waste money on vanity projects, there are reasons to avoid too much government funding, even when it’s well directed.

There’s some story in which:

• Bell Labs hires the world’s brightest minds, they do the best science anyone has ever done
• The government steps in and scales up the university system
• Universities now have the world’s brightest minds
• Universities are only accountable to bureaucratic funding, they become arbitrarily corrupt
• Meanwhile, it’s hard to launch a new private lab because the talent is tied up, and universities are already pumping out incremental research
• We end up in a stable equilibrium with no mechanism to fund “transformative research”

This is at least somewhat true, but I don’t know if it’s the most important truth.

I don’t know if “transformative research” is the right thing to aim for, if it is a thing at all, or if it’s even underfunded. Certainly, the NSF is eager to fund transformative research, whatever it is, so maybe it’s all fine.

On the other hand, I still get emails from people who say they stayed at Google for the paycheck, stability and prestige, even though they didn’t believe in the work. So certainly there is something wrong here, and some human talent is being squandered.

---

As a side note, this all leaves me with even less interest in Twitter. I thought the original thread was interesting and provocative, but found it less important the more I learned on my own. After contacting the people involved, Adam admits “The Twitter reply to Sarah is actually a kind of perversion of the main idea, therefore, massaging it to be more about valley of death per se than it need be.” and Sarah writes “I try a little to be right on Twitter, but I’ve never been interested in doing social science as an intellectual endeavor.  I’m glad some people are doing it, but I take my verbal activity way less seriously than that.”

Again, this is no one’s fault. Adam and Sarah are two of the smartest people I know, and neither is acting poorly here. They’re just doing what you’re supposed to do, which is to be relevant and engaging.

And still, neither of their stated aims is to use Twitter as a place for substantive and rigorous discourse. You may think that’s what you’re reading, but you are wrong. At best, it is a place to post links to other places.

See also Wake Up, You’ve Been Asleep for 50 Years, and Book Review: Why We’re Polarized.

Up until 1971, we were proposing and passing new amendments at a pretty steady rate:

The current drought is our longest since 1865, when a 62 year streak ended with the passage of the 13th amendment banning slavery.

There was a 27th amendment, completed in 1992, but it was proposed in 1789, and is a fairly minor issue relating to Congressional salaries.

More broadly, there’s even stagnation in terms of each amendment’s scope. Consider the section headings from Wikipedia:

This isn’t strictly chronological, but it’s close. We went from safeguards of liberty and justice (very important!) to banning slavery, discrimination and guaranteeing suffrage (also very important!) to minor issues related to processes and procedures.

The 25th amendment is just a clarification of succession rules in case the President dies, and the 27th is about congressional salaries. The 26th amendment lowers the voting age from 21 to 18, which is nice, but not nearly as big a deal as extending suffrage to women with the 19th.

And it’s not that no one is trying! There have been many proposed amendments since 1971, so it’s not for lack of want. Some of these are silly, but others are incredibly serious (repealing the penal exception clause from the Thirteenth Amendment [1]).

How does a constitutional amendment get passed? To play at Schoolhouse Rock for a second:

Proposal

• Get a 2/3rd vote in the Senate AND Get a 2/3rd vote in Congress, OR
• 2/3rd of state legislatures call for a national convention (this never happens)

Completion

• 3/4th of state legislatures, OR
• Ratifying conventions in 3/4ths of states (only happened for the 21st)

In short, the bar is very high! These rules are enshrined in Article V of the US constitution, and are unlikely to change.

From Wikipedia, here is a beautiful chart of Senate/House control:

The last time either party had a 66% majority was:

• House: 95th Congress (77 --79), Democrats had 292/435 seats, for a 67% majority.
• Senate: 89th Congress (65 – 67), Democrats had a 68% majority.

So how did the 26th Amendment get passed in 1971? It had overwhelming bipartisan support. The Senate voted 94 to 0, and the House voted 401 to 19.

The 24th (banning poll taxes), was a bit more controversial, but still won the Senate 77–16 and the House 295 to 86.

Our failure to propose and ratify a constitutional amendment since 1971 is not about rules, or about party majority. As far as I can tell, it’s just the result of polarization.

Scott Alexander’s review of Ezra Klein’s Why We’re Polarized opens:

In 1976, three years after Roe v. Wade, Democrats and Republicans were about equally likely to support abortion restrictions. That same year, a poll found that “only 54% of the electorate believed that the Republican Party was more conservative than the Democratic Party”; 30% thought there was no difference.

VoteView, Gallup and Vox (citing Vote View) provide some nice charts, all suggesting that yes, around 1970 – 1980, polarization took off:

I don’t know how seriously to take these. If you look more closely at the VoteView charts, you can see that I’ve cherry picked the extreme examples. In fact, on Regional/Social issues there actually seems to be a convergence starting around the same 1970 – 1980 period:

And of course, these are all important metrics, but don’t necessarily capture what you’re thinking about when you hear the word “polarization”. Scott’s post ends with a low confidence guess that Congress polarized in 1975 as a delayed reaction to 1960s civil rights, and normal people didn’t polarize until around 2000. He finishes by saying:

Every so often, people ask what an effective altruism of politics would look like. If you had some limited number of resources, and you wanted to improve (US) politics as much as possible so that the government made better decisions and better served its populace, what would you do? Why We’re Polarized and the rest of Klein’s oeuvre make a strong case that you would try to do something about polarization. Solve that, and a lot of the political pathologies of the past few decades disappear, and the country gets back on track.

On one hand, polarization seems obviously bad. You can’t get anything done without consensus, and we fall prey to zero-sum politics. I idolize the 1960s as much as anyone else, and it seems like some pieces of American were working then that don’t work anymore.

But also: Leopold Aschenbrenner’s description of European (particularly German) politics does not make the situation seem altogether enviable. Scott shares this graph from Vox, confirming that Germany really has seen a drop in polarization:

Low polarization is not all roses and harmony. As Leopold explains:

On the surface, German conventional wisdom decries the political divisions in the U.S.; it trumpets the supposed moral superiority of the German way over the American health care system or American foreign policy; it holds German democracy to be infinitely superior to American democracy (which, if you believe German media coverage, is on the verge of collapse and paralleled only by the Weimar Republic in 1933). But what this arrogance masks—and perhaps is deliberately intended to obscure—is the underlying reality of European “politics”: namely, that it is bereft of politics.

For the German voter has basically no say over his country’s fate. Sure, he may cast a vote in an election for parliament. But in the end, the same centrist parties seem to hold a majority in parliament, the same centrist parties form a coalition government, and the same party leaders remain in charge, making policy mostly through backroom deals rubber-stamped by the parliament. Besides relatively minor policy tweaks, the elections don’t seem to matter much.

I end up feeling lukewarm about all of this. I’m not sure why polarization happened, how clear cut it is, or if it’s even bad.

But I do know there’s a pressing need for progress, and many important human rights not yet enshrined in the constitution. Unless either party gains a 66% majority in both House and Senate simultaneously, our only hope is bipartisan agreement.

Appendix: We Should Ban Penal Labor

An alternative answer to this whole post is “we don’t pass amendments anymore because we already got all the important ones.”

I don’t buy it.

Penal labor is outside the scope of the main post, but really is horrendous. In short: the 13th Amendment banned slavery… unless you’ve committed a crime. From Wikipedia:

• “base pay being as low as 60 cents per day”
• “The state fire agency, California Department of Forestry and Fire Protection (CAL FIRE), mobilized over 11,000 firefighters in response, of which 1,500 were prisoners of minimum security conservation camps”
• “A wide variety of companies such as Whole Foods, McDonald’s, Target, IBM, Texas Instruments, Boeing, Nordstrom, Intel, Wal-Mart, Victoria’s Secret, Aramark, AT&T, BP, Starbucks, Microsoft, Nike, Honda, Macy’s and Sprint and many more actively participated in prison in-sourcing”

The pay isn’t $0, so maybe by some technical definition this isn’t slavery, but it is clearly wrong.

It’s hard to tell exactly how many prisoners are subject to penal labor. A Bureau of Justice Statistics report writes “More than half (54%) of all inmates held in facilities (88%) that operated work programs had work assignments at year end”

There are 2.3 million people in the criminal justice system, which multiplies out to 1.1 million in work programs. According to one estimate overall average salaries are around $0.14 to $0.63 an hour.

This is subject to a few caveats. If you have a job in a state-owned business, the estimate rises to $0.33 to $1.41 an hour, but those only constitute 6% of jobs. I also don’t think the average is weighted by population.

But none of that really matters. The wages are so incredibly low, and I feel confident condemning the practice.

This is much worse than “prisoners are poorly paid”. To take a slightly conspiratorial view, it is bad to let large companies profit from prison labor for the same reason it is bad to allow war profiteering. The more we benefit from an unfair system, the less incentive there will be to amend it.

---

[Edit 2021/02/17] It turns out this is not even conspiratorial, it is explicitly acknowledged. From Vox:

Harris’s office fought to release fewer prisoners, even after the US Supreme Court found that overcrowding in California prisons was so bad that it amounted to unconstitutional cruel and unusual punishment. At one point, her lawyers argued that the state couldn’t release some prisoners because it would deplete its pool for prison labor — but Harris quickly clarified that she was not aware her office was going with that argument until it was reported by media. [empahsis mine]

---

Obama’s famous A More Perfect Union speech opens:

The document they produced was eventually signed but ultimately unfinished. It was stained by this nation’s original sin of slavery, a question that divided the colonies and brought the convention to a stalemate until the founders chose to allow the slave trade to continue for at least 20 more years, and to leave any final resolution to future generations.

We are in the same position today. As great as the Thirteenth Amendment was, it also enshrines an exception to slavery. It is incumbent on us to fix it.

Or as Kanye succinctly tweeted: “the 13th Amendment is slavery in disguise”.

While I was busy explaining why the industrial research labs won’t return, an exciting thread unraveled on Twitter arguing about how to revive them anyway.

To kick things off, Activate CEO Ilan Gur argues that tech startups are the new Bell Labs, citing Moderna as proof. Then Sarah Constantin jumps in to explain that Moderna was just commercializing research from a university lab, and startups can’t do fundamental research because VCs won’t fund it.

Finally, Adam Marblestone chimes in to plug Focused Research Organizations as the missing gap between fundamental research in universities and applied science in startups.

This is all great, and highlights what I love about Twitter: smart people with diverse perspectives coming together in an ad-hoc way to discuss an important question, while making it accessible to lay audiences. [1]

I’m just not convinced it’s actually right.

Marblestone writes:

Academia’s big, so existence of fundamental research isn’t what we have to solve for (though can push more). Startups deploy well, so also not that, modulo VC risk/return params (1 of 3 proposals there). A gap is between those (another of 3 proposals highlighted).

This is a compelling narrative. Academic handles fundamental research, startups work on commercialization and deployment. What we’re missing is translational research to nurture pre-commercial research and bring it to fruition.

In fact, it’s so compelling that everyone else is already working on it. Off the top of my head, here are some mission statements to consider [4]:

• Cyclotron Road: “Our mission is to empower science innovators to advance their ideas from concept to viable first product, positioning them for broad societal impact in the long term.”
• ARPA-E: “advances high-potential, high-impact energy technologies that are too early for private-sector investment.”
• Sci-Founder: “help early career scientists start companies of their own.”

Not enough? Here’s the diagram from Actuate’s home page:

Okay, so maybe you’re convinced that the gap has been bridged, but it’s only a recent phenomenon, and only as a result of these brave pioneers.

Except here’s Breakout Labs in 2011: “Venture capital firms want research that can be quickly brought to market, and federal funding offers little room for risky, unproven ideas. We are jumping into this funding gap to energize innovative research.”

What’s going on here? Is translational research a great idea who’s time has come? An overcrowded space? A buzzword? A buzz-narrative?

Commenting on a draft of this post, Sarah writes: “I think a handful of funds existing is good validation of the idea rather than evidence it’s already overdone. The gap is probably big even if these guys are doing everything right.”

Similarly, Adam points out that the organizations listed represent a tiny share of all research funding. Gur’s thread mentions that translational research was 10% of Bell Labs’ budget. In contrast, these organizations sum up to around 1% of all federally funded research.

Maybe all of these organizations are good, and we just need to do even more. If that were true, we might expect Breakout Labs to have been an enormous success and raise/deploy more capital. But it’s always possible that while the idea in general is good, the specific implementation still matters, and Breakout Labs just happened to fail. It could also be that while 11 years sounds like a long time, it’s not quite enough to see the effects of what is explicitly pre-commercial.

Or maybe Sarah is right and venture capital, even with an explicit charter, just isn’t well suited to this type of investment.

---

Thanks to Sarah Constantin and Adam Marblestone for reviewing drafts of this post. Thanks also to Stephen Malina and rkris for their comments.

---

Appendix A: On the Basic/Translational/Applied Trichotomy

This taxonomy is worth keeping in mind, but it’s ultimately just one axis.

As Arora, Belenzon, et al. detail in section 5.1 of The Changing Structure of American Innovation, and as I summarized earlier, “inventions originating from large corporate labs are different”. Specifically, they are general purpose, practical, multi-disciplinary, and often resource intensive.

Along those lines, Google’s AI research has transformed the entire ecosystem. They:

• Pay extravagant salaries, driving the world’s best talent into AI
• Make high profile acquisitions, driving VC funding into AI
• Create Tensorflow, facilitating future AI research
• Develop TPUs, reducing the cost of compute, but only for particular applications
• Own one of the world’s largest datasets

None of this would be possible in startups or universities.

In this view, the gap is not just in “fundamental”, “applied”, or “translational” research, it’s in all the specific niches that industrial research labs could have funded, but can’t anymore. Different work is possible, and without those funding mechanisms, this work won’t happen.

That doesn’t mean the solution is a new government entity aimed at fostering “multi-disciplinary” research or whatever. We should just think about what particular aspects of Bell Labs were praiseworthy, and attempt to create the funding mechanism that will allow them to exist in some new form today.

---

After writing this, rkris and Stephen Malina told me about Spark, created at Berkeley, and pytorch, created by Soumith Chintala.

Soumith’s contributions to Pytorch seem to have occured mostly in 2016, at which point he already worked at Facebook. Pytorch has 46k stars on Github, compared to 153k on Tensorflow. Spark is at 29k.

Eyeballing the charts, the other early torch contributor as Yangqing Jia who was also at Facebook.

So I may have overstated the impossibility of creating a major ML framework in a university setting, but it does seem like Tensorflow is dominant, and Pytorch was heavily nurtured by Facebook.

Spark really was created at Berkeley. Aside from looking at the Github stars, I can’t really judge how influential it is compared to Tensorflow.

Stephen also notes that there are startups working on ASICs for AI.

Appendix B: Were mRNA vaccines invented by startups or universities?

Gur writes: “Startups are the most vibrant environment today for Bell Labs style intermixing of fundamental research and applied systems. And they are changing the world in Bell-Labs magnitude ways. Just look at @BioNTech_Group and @moderna_tx”

Sarah Constantin rebuts: “This is false. BioNTech and Moderna both commercialized/developed platforms based on innovations from academic labs.”

Who’s right?

In academia, Karikó was “issued an ultimatum, if she wanted to continue working with mRNA she would lose her prestigious faculty position, and face a substantial pay cut.” BioNTech licensed the work she had done with Drew Weissman, then hired her as Senior Vice President. She would go on to lead their work with Pfizer on the mRNA Covid vaccine. (Wired) The article adds that Karikó was considered “not of faculty quality” and that the Upenn admins laughed at her when she announced she was leaving. Another source adds that Karikó “spent the 1990s collecting rejections”.

For his part, Weisman still works in academia, though his lab is funded by BioNTech (MIT Technology Review)

And just to complete the holy academia/industry/government trinity, a third party organization claims Moderna "received approximately $20 million from the federal government in [DARPA] grants several years ago and the funds “likely” led to the creation of its vaccine technology. "

So who deserves the credit? Constantin is right that the innovation originally came from an academic lab, but this should come with the caveat that the university actively attempted to push out the lead researcher. It also seems likely that BioNTech would not have funded the research on their own, had they not seen Karikó publications.

All considered, I don’t walk away feeling like startups or universities are particularly good at fundamental research. Instead, it feels like a miracle that we ended up with commercialized mRNA vaccines at all.

It’s hard to know how much counterfactual impact government funding had, but if the DARPA grants were responsible for Moderna’s success, it’s a promising model for startups working on high-impact long-term technology, funded by someone other than VCs.

Footnotes

[1] It also highlights what I hate about Twitter. Gur provides no evidence for his claims, and is totally wrong on several of them. Constantin makes a vague complaint about VCs based on a single anecdote. [2] Adam’s tweets are confusing and difficult to parse. Everyone is talking past each other to further their own narrative. [3]

[2] Constantin writes:

A startup has to prove it can make money for investors within 5-10 years. The funding model’s not designed for long research projects… At a core level I think this is why my startup failed. I wanted to run a $1-5M experiment that nobody else was doing and seemed like the obvious first step in searching for anti-aging drugs. VC’s weren’t interested. You’re supposed to do that part in academia.

This might be true, but we should be very cautious of the argument “VCs wouldn’t fund my company working on X, so they must not be interested in the category X belongs to.” Per Thiel and Collison, startup failure is overdetermined. Maybe your startup couldn’t raise money because VCs are myopic, but there could be 100 other equally valid reasons.

[3] I’m being polite here, I really hate Twitter. Gur’s thread is filled with distracting gifs to the point of being nearly unreadable. The Constantin thread is over a dozen tweets long, with replies to individual subtweets that branch off, making the conversation impossible to follow.

I don’t know why anyone even tries to use this platform for substantial discourse. My best guess? They aren’t. Discourse was never the goal, just hot takes, dunks, and self-promotion.

I don’t say this to criticize the specific people involved, it’s all endemic to the platform, and not really anyone’s fault. Except for mine for reading the thread, and yours if you participate in this hellscape.

[4] I didn’t have to dig around to cherry-pick those quotes. It’s just the first line of each organization’s about page.