We all want a better world. A cure to disease, a pill that will let us live forever, the ability to teleport to anywhere on earth, free energy, no crime, infinite food, etc etc. But these things require scientific progress. The faster science advances, the closer we get to these goals. So its natural to ask what are the drivers of the rate of scientific progress. What are the causes of its lack thereof? When is progress quick? When is it slow? And why? Once we understand this a bit more we can hopefully strengthen the former, and eliminate the latter.
A common excercise in the military is an After Action Review (AAR). Its 4 questions. What was the goal, what was the intended actions, what actually happened, and what can we learn for next time.
What if we did this for the last 1000 years of scientific discovery. How much time and progress was lost because of a suboptimical policy of mankind. E.g. poor leadership, allocation of resources, and guidance by "the experts" (aka the old guards of academia that refuse to update their priors).
Lets double click into the latter as an example. As Max Planck said: “A new scientific truth does not triumph by convincing its opponents, but rather because its opponents eventually die.”
Scientist funerals literally reset priors; empirical evidence shows outsider citations surge 8-10 % when an eminent scientist dies! Thats crazy.
How true this is this today with the shift in trends of AI from certificates, global optimality, and rigorous proofs to empericism, scaling laws, and "trying things to see what works". Starting at Stanford in 2012, in the Computer Vision lab of all places, I watched in real time as Fei Fei Li went from: "Deep Learning doesn't work", to "Deep Learning is interesting for a narrow set of problems" to "Deep Learning is only useful for Computer Vision but cant do everything" in the span of 2 years. Which to her and Chris Manning's credit, were "relatively early" on the DL train. But even now, 13 years after AlexNet, Stanford professors are still shunning DL. A famous Optimization professor actually said to me a few months ago: "Francois, people are talking about using Deep Learning for end-to-end self driving cars and planes. This is absolutely crazy. This is solved. Its called Real-time Convex Optimization." Did not have the will to argue that if its so solved, why am I driving my car to work still. And last I checked humans (a deep learning system) are still the only ones the FAA let fly planes.
And this has a direct negative effect on scientific progress. The oldest quartile's inability to update priors (in expectation) results in a large filter lag for scientific discovery. Research is led and funded by those people or people who listen to those people. And those people have no first hand experience in training deep learning models. They should have the George Washington humilty to say, hmm maybe I should step aside and let someone else run with this now. Or maybe we should do as Mckinsey and Deloitte do, and mandate 65 yo retirement ages or earlier. Maybe end tenure?
Counterfactuals are really hard to run without a giant simulator, but having run a small company for 10 years, I can attest to the fact that our progress did slow bc of my poor decision making and resource allocation. And I would argue that I was much better than most CEOs now having seen what an "experienced" CEO did in my place (is killing the company). Poor leadership is seldomly detected in real time and sometimes never detected at all. Take the CEO of every major retailer over the last 30 years that watched a bald finance guy, non-retailer, walk into their industry and walk out the wealthiest guy in the world 20 years later. Man. Talk about regret. But does anyone point the finger at Doug McMillon or Rodney McMullen or Bob Miller or Brian Cornell? How do these people still have jobs? How can they ever work again after losing this bad to Jeff Bezos for this long. Its befuddling. The regret in the case of retail is easily measured by the difference in market caps from 1995 of these organizations to today, compared to the difference in market cap of amazon to today. Man that is a lot of regret.
Its not as simple to do this for PIs, grant approvers, etc, but it can be done. Everyone who missed deep learning and is still pushing GOFAI should be forced to early retirement. They are just increasing drag coefficient. Especially if you still are not a Deep Learning believer after a decade of insane results.
Having just done my first regret optimization proof I want to map our current scientific discovery engine to an optimization procedure to help think about what increases noise in scientific discovery, what decreases noise, what slows descent, and what increases it. Lets see how I do.
Let πt be the policy, our sociotechnical “solver” consisting of grant agencies, corporate labs, review boards, and cultural norms that decides how we allocate resources (e.g. what ideas get funded and how much).
Let xt be the state of humanity's technology in year
The update rule is xt+1 = πt(xt).
Cumulative regret after T years is RT = ∑t=1T[πt(xt) − πt(x*)], where πt(xt) quantifies the state of humanity's technology at time t, and πt(x*) is the best state of humanity's technology possible given existing resources in T time. i.e. Now knowing what is possible (GPUs, iphones, nuclear energy, GPS, vaccines, ChatGPT, Google, Amazon, etc) could we have achieved it faster / earlier? Why could we not have invented all of them 10 years earlier, 100 years earlier, or even 1000 years earlier? Does there exist a better π than our current one? Very likely. Regret will allow us to compare π's and we want to choose the best π.
Bias (term ∝ diameter) and Variance (term ∝ η·σ2): chasing the wrong vector, Lamarckism (experience is inherited) vs. Darwinism (genetics are inherited), Tesla vs. Edison, and Good old fashion AI (GOFAI) vs. Deep Learning. Entrenched dogma strictly increases Regret and just slows us down or even pushes us in the wrong direction. Some can see x* but the old guards stop us from getting there to save their ego. This causes either bias or variance, not sure which in this anology. I guess its a function of consistentcy of intransigence of the oldest quartile. lol.. But in either case, High-Bias or High-variance gradients thrash the search, slowing or (in the case of AI winter) stopping convergence.
Step-size η: Reviewer #2 vibes, funding winters, religion, anti-STEM protests (like what is happening in k-12 education, arguing that "math is racist") all slow down learning rate. And the Apollo missions, NASA in general, industrial revolutions, capitalism, venture capital, silicon valley, DARPA, (unfortunately) WW2, etc. all improved the rate of world's innovation and the learning rate.
It may be fair to argue that the path that was taken WAS optimal, and we had to study Lamarckism to get to Darwinism, which may be true in some cases, you need to get to A to get to B. I am not a biologist, but I am an AI engineer, and in the case of the AI winter, I can not agree, there was no reason to go back to statistical methods and convexity in lieu of deep learning. That was strictly a pause of progress without logic. That slowed down mankind for at least a decade.
Actually a better example is Black-Scholes. MIT rocket scientists published "The Pricing of Options and Corporate Liabilities" in 1973. They win the Nobel Price in 1993. And still very few are using it! They start their own hedge fund LTCM in 1994. Now almost all options traders use it. But it took 20 years to adopt. Why not 2 years. Why not 2 days. Its due to a limitation of our thinking. We had to wrap our heads around it. And that takes too much time. Are we that dense? Why not go faster?
A Danish king gifted Brahe the island of Hven plus taxes to run a private observatory; Kepler mined that data to discover his laws, paving the way for Newtonian gravity.
A corporate lab with slack capital birthed the bit and information theory in 1948, no immediate product needed .
Quality-control in beer required small-sample inference; the Student-t test emerged inside industry, not academia.
AT&T Bell Labs funded the first convolutional-net hardware for zip-code reading, long before ImageNet fame.
DARPA cuts and the Lighthill report froze U.S./UK AI budgets, flipping the progress vector for 10-20 years. Time wasted!
| Wave of “winter” | Key decision-maker | What they did | Impact | |
|---|---|---|---|---|
| 1966 (U.S.) | John R. Pierce – Chair, ALPAC panel (Bell Labs) | Led the ALPAC report declaring machine-translation progress “disappointing,” recommending withdrawal of large MT system funding. | U.S. agencies (DoD, NSF, NIH) cut most MT grants, ushering in the first AI winter. | NOTE: The hilarious irony that the Transformer was discovered on the machine translation task! lol |
| 1973 (U.K.) | Sir James Lighthill – Author, Lighthill Report | Argued AI had not met its promises; British Science Research Council used the report to terminate nearly all university AI funding. | Many U.K. AI labs closed or shrank; senior talent emigrated, deepening the chill. | |
| 1987–1988 (U.S.) | Jack Schwarz – Director, DARPA IPTO | Slashed the Strategic Computing Initiative’s AI budget, calling expert systems “clever programming” unworthy of further investment. | Lisp-machine vendors collapsed, venture funding dried up, triggering the second U.S. AI winter. |
Each of these individuals had direct authority (as advisory chairs or budget controllers) and their decisions rapidly dried up research and commercial investment in AI, leading to prolonged periods of reduced AI activity.
Similar to AT&T Bell Labs, DeepMind funded by Google, discovered some of the most critical AI research of the past decade.
While Google Brain inventory the transformer, Google was under anti-monopoly scrutiny so had zero appetite for commercialization and it was OpenAI that commercialized it.
Empericism over mathematical rigor: eliminate requirements for papers to have mathematical rigor. Results are all that matters. Does it work or not. Thats all. No more reviewer #2 BS.
Industry discovers much more than academia per capita / per dollar spent: It took Bell Labs to invent Information Theory, not every stats/prob/math department. It took Google Brain to invent the transofmrer. Its much more often the case that industry invents vs. academia. And its definitely the case that they commercialize.
Invention without commercialization does not improve humanity: a paper on arxiv is not the goal. That is a step along the way. The goal is to improve the quality of life for mankind. Perhaps this is why industry is better at research that academia. They are closer to the feedback signal and have a stronger incentive to solve it vs. publish an epsilon paper.
Cut tenure: emulate the beneficial “funeral regularizer” without actual funerals; NO MORE TENURE! Grant cycles must retire stale priors.
Compute vouchers & open benchmarks inspire innovation: ImageNet was such a good example of this. The dataset inspires new invention. This cheapens experiments.
Democratic / Prediction-market-guided grants: just a crazy idea.. there is wisdom in the crowds. Polymarket works for politics. Could it work for research directions? Replacing "the experts" with "crowd-source priors" to align the gradient direction early may be a new way to fund research and get the BS out of academia. It worked for wikipedia. It works for StackOverflow. Why can't it replace PI's, Grant Decision Makers, and conference reviewers.
Cross-sector sabbaticals: rotate academics ↔ startups ↔ government every 5 years; share priors, lower bias. The mindset difference between gov, academia, and startups is stark. Perhaps mixing things up will assist in the flow of ideas.
Cross-speciality sabbaticals: Same argument for disciplines. Speaking with people in Physics, Statistics, Finance, and Medicine is like going back in the stone age. They have so little understanding of what Deep Learning is, and how beneficial it can be for their discipine. The more mixing things up the better.
Looking back on the last 1000 years and what empowered and what inhibited scientific discovery can help guide us to improving scientific discovery per unit time, per dollar, per capita. My conclusion is the famous charlie munger quote "Show me the incentive, I will show you the outcome". Academia, the experts with tenure, reviewer #2 all have no incentive to take big swings, go against the grain, and improve mankind. The tallest blade of grass is the first to be cut. And what if you are wrong? You are out! So just be quite, be happy, get invited to the white house, and never shake the tree that hard. Vs. industry who has clear incentive to shake things up. Startups who definitely have incentive to improve the status quo. These are all better systems than academia. If I had to bet on who will invent the future, a tenured professor or a startup, I would bet on the latter everyday. With this in mind, forget getting rid of tenure, what would really happen if we got rid of academic research entirely. Should we shift funding from academia to startups directly? Hmm.. I guess this is already happening bc every professor at Stanford has their own startup or two so they can innovate without the pressures and disincentives of academia.