lambdatronics

Location: Seattle, WA, USA

Remote: Yes

Willing to relocate: Yes

Technologies: Python, Matplotlib, Scipy, Numpy, sk-learn, Keras, Tensorflow, Julia, Fortran, SQL, git, slurm, Linux, MS Office, analog circuit design, optical systems design, high-energy laser operation, optical spectroscopy, photonics, sensors, interferometry, polarimetry, plasma physics, ultra-high vacuum, X-ray detection, Bayesian inference, probability theory, statistics, Monte Carlo methods, information theory, calculus, differential equations, linear algebra

Résumé/CV: https://drive.google.com/file/d/1KJUpr03_t5CqS-u_HAr3P6iWJzb...

Email: [email protected]

I have a PhD in experimental physics + 8 years of experience, include 4 years at startups. I can take complex electro-optical hardware projects from concept through design, construction, testing, operation, data analysis, comparison with simulations, and visualization/presentation of results. I have run physics simulations on HPC clusters. I routinely use the scientific Python stack for data analysis and visualization.

I am a self-starter comfortable with minimal supervision, and working with small cross-functional teams. I am excited to learn new technologies, on the job or in my own time. For instance, I taught myself about relational algebra and SQL because I wanted to explore probabilistic databases as an approach to statistical learning.

>already a fairly radically super-human speech center >We're asking a speech center to do an awful lot of tasks that a speech center is just not able to do

Exactly!

>We need more parts.

Yeah, imagine what happens once we get the whole thing wired up...

>a way to generalize the compute graph as a learnable parameter.

Agreed. Seems analogous with how human mental processes are used to solve the kind of problems we'd like LLMs to solve (going beyond "language processing" which transformers do well, to actual reasoning which they can only mimic). Although you risk it becoming a Turing machine by giving it flow control & then training is a problem as you say. Perhaps not intractable though.

>What would it even mean for a particle to have an "action" (whatever that is) that is not minimized?

The particle doesn't have an action. The trajectory of a particle is what the action is defined in terms of. One way to think of it would be "it's a measure of how much the trajectory deviates from the one dictated by Newton's equations." Pretty much like what you said: "I can always find a measure that something else is always a minimum of."

About what a trajectory with non-minimal action would look like: it would be an arbitrary violation of the equations of motion for the system (ex: free particle moving in a zigzag instead of a straight line at constant velocity). Moving in a straight line at constant velocity is what Newtonian mechanics prescribes, and that trajectory will minimize action for the corresponding Hamiltonian.

Cool! I had been meaning to try my hand at a to-do app ever since I read "Getting Things Done" a few years ago, and I finally spent a couple weekends on it this month. The tree structure is 100% the way to go, and it's surprisingly uncommon as you note. It's been a fun & educational project so far. It's only SQLite file & some Python functions, no polish whatsoever. Probably will never get to the level of what you have, so hat's off!

I see what you did there with the Dr. Horrible reference! :D

IIRC sodium ion batteries have traditionally had trouble with rapid degradation due to the large size of the sodium ion, which disrupts the carbon electrode via swelling during intercalation. So I would be worried about the number of cycles they can last.

Thanks stuart8ol for all the detailed answers in this thread. I learned several new things today (use of compressors for fuel cells to increase power density, logic of airplane sizes). This is the stuff I come to HN for! Also cool to see you using H3X motors - I remember their launch here, I was excited for them.

Good luck decarbonizing aviation!

> But the trade off should not be "if there is any risk at all, then no"

I'm a kitesurfer and I'm with you 100%. I'm glad kiting is not banned in most places. Kiters die every year -- I knew someone who was killed, and I've had a couple close shaves. In fact, today I was kiting and witnessed a kiter get rescued by the Coast Guard. It's understandable that society would want to have some say due to the externalities (loss of productive members of society, cost of rescues, risk to bystanders, etc).

IMO it's not the case that 'no risk is acceptable' -- it's just that the risk tolerance for various things seems really arbitrary & out of proportion to reward. For instance, if we didn't accept any risks as a society, we'd ban alcohol & tobacco along with kitesurfing, skydiving and motorcycles, but we don't.

The paper says that this is only half of a solution: "Our investigation here has focussed [sic] on the fundamental Li-NRR performance at the cathode. Further developments towards a complete ammonia electrosynthesis system will require investigations of appropriate anode reactions while eliminating sacrificial solvent oxidation. A feasible initial strategy is to couple the Li-NRR with the H2 oxidation reaction, which has already been demonstrated but requires improvements in stability and activity. A more-desirable anode process is H2O oxidation, which presents larger challenges because of the potential interference of water with the Li-mediated process and vice versa."

This system involves ethanol as a sacrificial hydrogen donor: "The amount of ammonia produced in the 96 h experiments (3.9 ± 0.1 mmol) was around four times higher than the amount of ethanol present (1 mmol), indicating that it is not a completely sacrificial reactant but can also operate as a proton carrier."

First you want to displace natural gas for fertilizer production. But yes, if the energy efficiency is good enough and the electrolyzer costs are very low, it would make more sense than electrolyzing hydrogen and then running that through Haber-Bosch. Remains to be seen if either of those criteria can be met.

Probably not just yet. I calculated downthread that the productivity is something like 16x less per unit area than a hydrogen electrolyzer, so that would need to be improved to make it cost-effective probably. Also they don't mention the energy efficiency, only the "current efficiency" so I would assume the energy efficiency is also poor. Sounds like there's much to be done still.

Probably refers to "Coulombic efficiency." Ie, it takes 4 electron transfers to turn 2 H2 + N -> NH4, so that gives you a conversion factor between Coulombs of electrons (1 Amp of current is 1 Coulomb per second) and number of NH4 molecules produced.

Yeah, someone would have to get access to the paper to see if they state the energy efficiency. I assume that b/c they don't mention it, it is abysmal. There's pressure to put good results into the abstract.

For hydrogen electrolysis they typically quote around 1 A/cm^2 current. One Coulomb is ~6e18 charges, whereas one mole is 6e23 molecules, so that makes about 1e-5 H atoms per cm per second. Of course making one molecule of ammonia needs 4 H atoms, so it works out to something like 16x lower productivity. I assume it's not competitive as is.

Only if there's enough wind shear between the aircraft and kite altitudes... kinda like dynamic soaring [0] but without the need for the aircraft to jump back and forth between the two regions. You'd probably need to use a sailplane or something to get high efficiency flight at low speeds. I think it would be super dangerous, b/c in order to access the strongest differential wind speed, the craft would need to be very low to the ground (just a few meters).

[0] https://en.wikipedia.org/wiki/Dynamic_soaring#Manned_aircraf...

No-moving-parts is probably not realistic - the core would be much hotter than the radiating surface due to the thermal resistance of the shielding. More likely, you would have a working fluid to transfer the heat. Most nuclear reactors operate at much lower temperatures where TPV wouldn't be efficient or cost-effective. It's certainly possible to go much higher. Nuclear-thermal rocket propulsion tests ran with exhaust temperatures up to ~2200 C [0]. Whatever fluid is used, you would need to avoid radioactivity in it, b/c that would probably degrade the TPV. Also you would probably want to avoid having heat exchangers because each one incurs a temperature drop. So helium would fit that bill. That's what the "high temperature gas reactors" use [1]. IDK if helium-compatible plumbing/pumping could be made to work at >2000 C though.

Edit: BTW, there are radioisotope thermoelectric generators used for space applications primarily, but they are not true nuclear reactors - they produce short-range radiation that doesn't require much shielding. Nuclear reactors produce neutrons and gammas that require thick shielding.

[0] https://en.wikipedia.org/wiki/Project_Rover#Kiwi_A

[1] https://en.wikipedia.org/wiki/High-temperature_gas_reactor

Yeah, tungsten is far too expensive. Sounds like more realistically you would have molten silicon in graphite plumbing [0]. This article claims the self-discharge rate could be made 1%/day.

There's been talk of doing solar concentration -> hot object -> thermophotovoltaic converter. IIRC concentrated solar already exceeds 40% efficiency so it doesn't make sense to add the extra step, unless you are using the solar concentration to "recharge" a heat storage system.

I buy the argument that thermophotovoltaics can become cheaper on a $/kW basis than comparably-efficient fluid/mechanical heat engines. The power per unit area is intrinsically orders of magnitude higher than for direct solar, so even if these cells are pricier than regular solar cells, they have a fighting chance. Also, both the power density and efficiency increase with temperature, and in principle the operating temperature can be higher than that of a turbine (since the materials don't have to simultaneously withstand crazy mechanical stresses and reactive chemical environment).

[0] https://pubs.rsc.org/en/content/articlelanding/2019/EE/C8EE0...

Just to amplify on your point about lithium: the tritium production function is critical. Every fusion neutron needs to produce more than one tritium atom on average, so that the reactor is sustainable (there are inevitable losses & tritium also decays radioactively) or even making excess tritium (to bootstrap other reactors). This is challenging b/c even in the best case each neutron can produce maybe 2 tritium atoms, so there's not much margin. The lithium needs to comprise most of the material surrounding the plasma, limiting the fraction that's available for other functions (structural supports, heat shielding, cooling, plasma control & heating systems, sensors, etc).

Not at the moment. IMO the prospects aren't good, due to the low efficiency (high energy cost) and high equipment cost. There are start-ups working on it (ex: Prometheus Fuels [0], Twelve [1], Synhelion [2], honorable mention to Terraform Industries [3] which is targeting methane).

Everything that can be electrified, will be electrified, because it's more efficient. It seems like shipping & aviation are probably the hold-outs.

I used to think this was the way, especially for balancing the grid, but the more I read [4-7], the more it looks like batteries will be used for fluctuations <1 day, and demand adaptation for longer periods. Some of that adaptive demand may wind up being used for hydrocarbon synthesis, but I don't expect it to compete with fossil fuels for a long time, if ever.

[0] https://prometheusfuels.com/

[1] https://www.twelve.co/

[2] https://synhelion.com/

[3] https://www.terraformindustries.com/

[4] https://doi.org/10.1016/j.adapen.2021.100051

[5] https://doi.org/10.1016/j.joule.2020.07.007

[6] https://doi.org/10.1016/j.joule.2019.11.009

[7] https://www.liebreich.com/the-clean-hydrogen-ladder-now-upda...

"Information theory and statistical mechanics" by E.T. Jaynes [0]. Starting from a Bayesian/information-theoretic perspective, he derived statistical mechanics in just a few pages. This approach finally made stat-mech 'click' for me.

It's a paradigm shift. Entropy (in the Bayesian sense) is fundamentally a subjective quantity. Maximizing entropy is just minimizing the assumed information. When dealing with large complex systems over long timescales, the only pieces of information one is justified in assuming are the values of those quantities that are conserved globally by the dynamics (eg, total energy, total mass, etc). The "thermodynamic entropy" S is just the maximum entropy, given a certain set of conserved quantities (an "ensemble") -- it is therefore more or less objective, modulo the conserved quantities.

The second law of thermodynamics is just the information processing inequality: you don't have any more information about the future state of the system than you do about the present state. If anything, you have less, because if you have any information about the current state of non-conserved quantities, that information is not valid at other times (assuming you don't have complete information & the ability to fully simulate the dynamics). From this perspective, entropy does not generate the "arrow of time": the argument is symmetric in time.

Another from Jaynes worthy of mention: "Prior probabilities" [1] discusses the use of group theory to derive "non-informative" prior distributions by considering the set of transformations that result in equivalent inference scenarios. This is resolved another question I had when studying stat mech: "How come we start with an assumption of uniform probability in phase space as expressed in (x,p) coordinates -- if we transformed to a different coordinate system, it would look like a non-uniform distribution!" The answer is that we assume Galilean invariance. The concept applies outside stat mech as well.

Plenty of other interesting stuff [2].

[0] https://bayes.wustl.edu/etj/articles/theory.1.pdf [1] https://bayes.wustl.edu/etj/articles/prior.pdf [2] https://bayes.wustl.edu/etj/node1.html

ReBCO still isn't practical for grid transmission AFAIK. It needs to be cooled to something like 20 K to achieve high current. That's ~4x warmer than traditional superconductors (~4.5 K) but it's still not "high temperature" by anyone else's standards.

ReBCO is a huge boon for magnetic-confinement fusion. The comment from their CEO that "In fact, from a physicist’s standpoint, our machines look kind of boring [...] we could get extremely high performance through the brute force of the magnetic field” hits the nail on the head.

Experiments like DIII-D have spent the last 20-30 years trying to eek out a few dozen percent more performance by fine-tuning the shape of the plasma & the current distribution inside it, etc. Some of the configurations they've tested are right on the edge of instability, requiring high-speed control systems to walk up to (but not over!) the edge of an invisible cliff in parameter space. If we learned anything from Three-Mile-Island and Chernobyl, we do not want that kind of twitchiness in a nuclear context. Sure, the worst-case fallout isn't as bad, but the multi-billion dollar reactor could be crippled for months/years, depending on how things play out. Boring is good, in a nuclear context.

The downside to the use of brute force is literally the forces involved. The original design (a class project) estimated that the structural steel required to hold the coils of a reactor together would run almost $5 billion & account for ~80% of the total cost of the core. (See Table 11 of [0]) This is for a power plant that would deliver ~ 250MW of electricity. The total plant cost would probably be $10-15B, which is not economical. However, they say "there likely exists a better economic optimization of magnetic field strength versus mass for a full power plant." Most likely, that involves reduced field but larger size reactors, which might be slightly more expensive but significantly more powerful. (IIRC, the latest designs for ARC are inching up toward 4 meters major radius.)

I doubt this approach can reach economic breakeven, but I think they have a good shot at exceeding energy breakeven. OTOH, I'm a physicist rather than an engineer, and I work for a competitor, so take my opinion with a grain of salt.

[0] https://library.psfc.mit.edu/catalog/reports/2010/15ja/15ja0...

I toured a multi-million-dollar electron microscope at ORNL. They had a piece of tissue paper taped to the side. I asked about it, and they said that it was an airflow monitor. It turned out to be more sensitive than some very expensive sensors, according to the tour guide.

That's a great way to put it. ("properly functioning state of consciousness") Because some folks would argue that the it isn't functioning. That brings us to the question of "What does it mean to be properly functioning?" which isn't by any means settled, although "experiencing 24/7 visual snow" certainly isn't it.

I've worked on two different plasma devices where the engineers used a different coordinate system from the scientists.

>But points of sail don't have any handedness, correct? I.e. they're mirrored on both sides of the wind.

The right-of-way rules in sailboat racing depend on who is on 'port tack' or 'starboard tack' (ie, they differentiate those mirrored halves). There's nothing physically special, though.

>To think deeply about meta science in any official capacity as part of PhD training requires you to go out of your way.

Yup. Not just life sciences, this applies to physics as well.

>I also imagine that even in less expensive/slow fields there is an increasingly unapproachable body of prior knowledge to learn, and people don't want to be trainees forever.

This is where AI might come in handy, for compressing the time it takes to ingest the prior work by providing tailored educational material, or even obviating the need for it by augmenting our intelligence (esp. when it comes to recall/search). Otherwise, we'll eventually be in the situation of Schlemiel the painter [0].

[0] https://en.wikipedia.org/wiki/Joel_Spolsky#Schlemiel_the_Pai...

I wouldn't say it is useless. But, there's an extent to which "I can explain it to you, but I can't understand it to you." There's an interesting school of thought saying that in order to understand an idea, it requires creative act similar to that of the individual who originated an idea. (B/c verbal communication is very limited.) For instance, in order to comprehend the idea of Newton's Laws, one needs to have in some sense the capability to have come up with them, because teaching is really just prompting the pupil to reinvent the concept in their own mind.

IMO, this is why "natural language processing" with LLMs will not amount to "natural language understanding." Natural language explanations are targeted at intelligences capable of forming concepts based on IRL experience (naive physics, etc). The goal of teaching is to induce the formation of the concepts, not to directly transmit the concepts.