In a ha-ha-only-serious way, I think this is a big part of the answer. DARPA is the only US funder that’s willing to pay for experienced staff. If your proposal needs a half-dozen seasoned scientists or engineers, they’ll pay for it.

In contrast, the NIH and NSF are very trainee-focused: a typical project consists of one professor loosely overseeing 1-3 “trainees”. We obviously need future scientists, but this model has a ton of drawbacks in terms of speed, quality, and even the nature of the project (which needs to let three people be somewhat-separately “evaluated”). It has also no room in it for any ICs more than four or five years post-PhD, which is bonkers given what the same organizations just paid to train them!

$3~4B is indeed quite big budget. However, it's dwarfed by big tech's R&D budget. FB alone spends $20B+ each year.

I guess one key difference is that unlike FB, DARPA money doesn't need to be confined to make quick and visible returns, thus they can focus on more long term problems.

[1] https://www.macrotrends.net/stocks/charts/FB/facebook/resear...

The Dream Machine: J. C. R. Licklider & The Revolution that Made Computing Personal, by Michael Waldrop[1], is one of the books out there that touches upon the early days of ARPA, what happened, and how computing & networking was jump started under by a far-seeing thrown-into-position bureaucrat who both dreamed big personally but more so, stood back & gave multiple far seeing groups of techies free reign to bring their dreams to life. A lot of other books I've read also speak similar stories, and I tend to believe them.

It's a near & dear book to me (although due for a re-read at this point), & I think many others. A history of many of the ideas circulating, and how many overlapped & reinforced on the way to building what ended up being the internet. There's more internet-centric accounts like Where Wizards Stay Up Late or Dealers of Lightning, but Dream Machines remains the favorite of many[2].

I just cant emphasize enough how much benefit I think there is to government granting good hard working big idea people, with not a lot of oversight or expectations. The counter though, is that these were relatively small organizations, all in all, and the goal is not to create jobs for life (as I feel like most other research institutions tend to do) but to fund projects, ideas, for a duration, and then move on, re-allocate funds.

[1] https://www.wired.com/2001/10/the-dream-machine-j-c-r-lickli...

[2] https://news.ycombinator.com/item?id=14119220

The kind of R&D that DARPA is doing and the kind of R&D that Facebook is reporting in their yearly report are very, _very_ different numbers. Every single tech company reports ongoing development work as "R&D," whereas much of the work that DARPA or other national labs do is exploratory and towards a lot of different projects.

For instance, at every company I've worked at, all of the infrastructure is under the R&D budget.

I'm gonna guess: tax evasion.

Like in the last two years, quite a lot of software developers in Poland suddenly started doing "R&D", because performing the necessary rituals to convince the tax man that your CRUD app is "research" let you drop your tax rate to 5%.

That's not the case in the US. There is no federal tax break in the US for "R&D" as far as I'm aware. Companies can deduct all expenses from their income. So there isn't even really a need for a tax break here. All salaries, rent, operational expenses are deducted from the income.

Moreover, many US corporations are effectively encouraged by tax law to go into debt (and always operated under debt), because having a zero effective income means you pay zero taxes. So that's what many companies -- they spend a little bit (or sometime a lot) more than their revenue, and thus report a loss each year (consequently paying zero taxes).

Yes, if you work in a large US company, they will tell you explicitly that they want to categorize as much work as possible as "research", for tax reasons. You may have to fill out questionnaires asking whether your work might fit any of a series of definitions of "research".

There's nothing secret about it.

Interestingly, the time is also classified by the ‘class’ of research, where class one research involve very basic improvements (like changing the speed at which you make something) that have a different tax bracket than class five research, which is closer to basic/applied research with no clear application in sight.

There's a bit of double-counting here too.

DARPA doesn't actually do much (any?) research on its own. Its role is to award contracts to "performers" that actually run the experiments, train the models, build the devices, etc.

In other words, a big chunk of (e.g.,) Teledyne's R&D budget is actually the same money as DARPA's R&D budget.

Not sure that's it. Consider the difference in motivation between trying to figure out how: to get people to click on advertisements; vs prevail when people are trying to kill you.

Source: was at BBN while DARPA pushed the Internet and created SIMNET. Did work for DARPA at own company later. Never met a wooly-haired, wild-eyed fellow nerd or VC with half the imagination of buzz-cut DARPA people.

3.5 billion for every DARPA project is a hilariously low number. DARPA projects are famous for having low budgets, completing under budget, and completing under time. It has an extremely worthwhile return on investment.

No US government research project aims to be under budget. Money is viewed as a finite expiring resource and there's every incentive to avoid your contractors from returning ceiling to the government.

I would not consider DARPA work to be low budget. DARPA programs are very typically in the $50 million range and program resources are split up between multiple performers. Sometimes larger program bets are made, especially in hot areas of research or programs where there are large integration costs.

A $50mil budget for a applied research project seems pretty low to me, especially when split up between multiple orgs.

I'd be keen to see a source on this if you have one.