455 karma · joined May 10, 2020
fastneutron235 -at- gmail -dot- com
In all seriousness, the MIT admissions reps I interacted with at the time were responsive, friendly, and more transparent with me about the process than you'd expect. I actually did ask them for feedback once the cycle was over, and it basically came down to the concern that I would be too focused on doing "cool stuff" and not enough on core academics, which was probably evidenced by my HS transcript. I was a solid A- student, but you could tell my priorities were extracurricular.
Hindsight being 20/20 two decades on, they weren't far off the mark on that feedback, but it never wound up being a problem.
In aggregate, this is how you wind up with stub functions and narrative descriptions rather than full working implementations. The RLHF is optimizing for correctness within some constrained token count.
The stress we put ourselves through in high school when applying to colleges was not worth it, and I can’t it being any better these days. As trite as the advice from my elders seemed at the time, your choice of undergrad is not necessary a limitation if you’re determined enough to get to a particular career destination.
Either way, MIT is being overly smug about it with articles like this. The admissions director at the time even brought it up in an interview with the associated press that fall [1]. Things turned out fine for me, and I’m glad that whole undergraduate admissions rat race is in the past.
1. https://mitadmissions.org/blogs/entry/marilee_jones_in_the_n...
I want to echo this point, because it gets talked about in the popular media like it's just another kind of garden variety crude oil, when it is anything but. For the curious, there are independent sellers (on ebay and elsewhere) who sell specimens of the stuff, along with samples of other kinds of raw fossil fuel and energy minerals. It's very helpful for demos and discussions like this one.
This stuff is almost literally road tar mixed with sand, almost like asphalt. It's difficult to break up by hand when cold, and when warm it has a tarry, putty-like texture. Contrast this with light-sweet crude, which is a pale yellow, gasoline-smelling liquid. Once you have a feel for these things, it doesn't take a leap of the imagination to grasp that the latter is going to take a lot less effort and energy to turn into useful products than the former!
Now, technology has advanced to the point where all the low hanging fruit are gone in terms of performance optimization, and we’re better able to determine how close to the “cliff” we are, with tighter uncertainty bounds, than ever before with probabilistic design methods. By this logic, we should be able to squeeze out margin that was never needed (or never there) in the first place.
No safety professional would trade away margin you say, but the engineer is often between a rock and a hard place when a few percent traded away in margin translates to multiple millions of dollars in cost reduction or profit. Is it unethical to remove margin if your Bayesian UQ calculations are saying you’re still safe, if not safer than you were under the other methodology?
This tension is going to keep on building as AI-enabled solutions start penetrating more and more into traditional engineering fields, with as-yet unknown consequences.
Some back of the envelope math to illustrate why:
- The decay energy of Ni-63 is 66 keV, but since the energy is split with an irrecoverable neutrino, the average recoverable energy per decay is more like 17 keV.
- The typical power consumption of an iPhone is about 10 watts, give or take.
- Converting (1eV = 1.6e-19 J, 1W = 1J/s), we get 2.72e-15 Watts per decay.
- Assuming 10 Watts for your phone, we get 3.68e15 decays/second, 3.68e15 Bq, or *99364 CURIES* of activity
- The US NRC exempt quantity for Ni-63 is currently 1e-5 Curies (10 micro-Curies) (10CFR30.71, Sched. B)
For reference, the activity of the Am-241 in a typical smoke detector is on the order of 0.1 MICRO Curies. A betavoltaic battery for a smartphone would need a full 20 orders of more activity per unit, 18 orders of magnitude higher than the current regulatory limits, and we're talking about a product that nearly every person in the country owns. The item would be physically warm to the touch, could never be turned off, and would need piles of paperwork to track the origin, transfer of custody, and proper disposal.
Maybe this makes sense for remotely-deployed military equipment and other niche applications where the user would go thorough the necessary licensing steps, but for consumer products, forget it. Absolute fiction.
While it may be true that the absolute number of people involved is small compared to the wider tech industry, long-term code reproducibility is critically important for engineering-heavy industries with long-lived, high consequence physical assets. Think aerospace, nuclear, petrochemical, electrical, and other civil infrastructure. If something in the environment changes and you need to re-analyze the performance of that asset to ensure safety & reliability under the new conditions, the first thing you do is re-run the benchmark cases to verify the answers didn't change from the last time you ran it. If they did, it throws your entire basis for design into question unless you can conclusively resolve the inconsistency.
Disagree on this one. In my experience, the computational kernels of our newer scientific & engineering codes (Fortran or C++) are almost never invoked as standalone applications, and are instead embedded within much more complex workflows that are orchestrated with a Python API or Java application. The string handling, data containers, and support for standard file formats in these higher level languages are indispensable at this level, and add negligible overhead compared to the computational expense of the main solvers. Yes, these things are more ephemeral, but they are also more interchangeable. Confining them to a higher abstraction layer lets you more cleanly decouple this stuff from the essential numerical parts of the code.
The historical lack of support for these basic amenities in Fortran (and the C-family to a lesser extent) has resulted in the need to maintain innumerable ad-hoc I/O formats, home-rolled nonstandard parsers, and other bizarre, Rube Goldberg-like abuses of the language. Since the original developers of these legacy codes were scientists and engineers first, and developers second, these constructs inevitably wind up being tightly coupled to a larger, organization-spanning workflow whose input-output mapping has remained bitwise identical for three decades, and must continue to do so for just as many more.
Higher-level languages like Python have greatly alleviated this nightmare, at least on a go-forward basis.
With quantum error correction, we're not talking a $25B facility with many competing hands in the pie. Instead, we're talking about improving the spatial and timing precision of RF, optics, and photonic hardware. Unlike fusion, where basically everything needs to be a custom job, there are established industries around these quantum-adjacent technologies with decades of accumulated know-how and plenty of low-hanging fruit to get most of, if not all of the way there.
Even if/when these things are working on a useful scale, quantum computing is going to be more comparable to special-purpose accelerators like GPUs, or perhaps akin having a re-programmable physics lab-on-a-chip. It will be advantageous for the subset of problems that can be cast in the form of Schrodinger's equation, of which there are many. For example, you can use the quantum computer to generate high-fidelity solutions of the electronic structure of some classically intractable molecule, and use those solutions as the training set for a classical AI model that then goes and performs the bulk of the computation on whatever it was you were trying to do.
While I think it’s probably fair for people with English as a second language to use LLMs as a writing assistant (for example), there definitely needs to be some kind of author disclosure statement at the very least. This could be similar to how more papers are now including contribution statements when there are many coauthors.
These are incredibly expensive even on today’s hardware. If you look through some of the unclassified ASCI reports from the early 2000s, 3D calculations of this equation set were implied to be leadership-class computations. At the time of the Cray, it must’ve been coarse-grid 1D as the standard, with 2D as the dream.
The fidelity would almost certainly be super low compared to modern FEA software, but it would be a fun exercise to try.
If private capital wants to fund deeptech, more power to them. If they want forego due diligence and fund tomato computers, that’s (literally) their business.
If it’s such a travesty that VCs are making bets on tech like this, how then do you fund long term R&D projects with a high risk of failure? Isn’t that the point?
Joking aside, these slides are excellent! Is there an associated video or course that they were a part of?