lambdatronics

>HN equivalent of physics or "real" engineering

Where's that? :D

Have a look at this study for a tokamak-based system: "The cost of the fusion-specific technologies accounts for approximately 75% of the total direct costs of constructing the plant" [0] They also conclude that "The model levelized cost of electricity was found in the amount of ∼160 $/MWh." Note that 57% of that cost is due to depreciation of the hardware, and another 23% is replacement costs for materials damaged by the neutron radiation.

Another source says: "Setting the capital cost of the fusion-specific technology to zero results in a LCOE of 72 $/MWh, which borders current prices for renewables, but is clearly physically implausible."[1]

[0] https://www.sciencedirect.com/science/article/pii/S036054421... [1] https://arxiv.org/abs/2101.05727

Working on nuclear fusion. Debatable whether that's ultimately the best way to tackle climate change given how well solar is doing these days, but hey...

For me, it was approaching it from the information-theoretic perspective. E. T. Jaynes' paper was what made it all click for me: https://bayes.wustl.edu/etj/articles/theory.1.pdf

Edit: but that was only after I had grokked Shannon's paper on information theory, which I felt was pretty intuitive.

Interesting to get some fusion nerd specs on the Trenta machine: 10 MJ stored energy, ~100us shot duration, merged plasma temperature ~ 1keV (pre-compression), 10 minute shot cycle, multiple MA coils in the compression region. If the illustration is correct, the compression coils are extremely thick. That's probably necessary to (1) withstand the magnetic pressure from the 10-20 T magnetic fields they claim to create (that would be about 400-1600 bar) (2) reduce resistive losses to enable the ~95% energy recovery they claim.

Just going back and look at GPT-2's output, it's amazing how much better this system is. It still doesn't "understand" anything, but the coherence of what it spits out has gone up drastically.

https://thegradient.pub/gpt2-and-the-nature-of-intelligence/

There's always General Fusion.

Uhhh... it's definitely being used for "stockpile stewardship." The fusion crowd went splitsies with the weapons folks to get this funded in the first place.

The bigger the target, the more energy is needed to compress it. So it would require more laser energy to get to 100 MJ or 1 GJ, I think. But maybe only a few times more powerful. (Pity they didn't build in some head room!)

I think Jevon's paradox is also hitting to some extent, coupled with the ecosystem getting bigger, so a typical project may have more transitive dependencies behind it. The functionality is going up, and the speed of compilation is going up, but the total latency is staying similar or increasing, in my experience.

I would say there have been a handful of important milestones this year, but this I would consider a breakthrough. Most of the other stuff is overhyped for sure.

Yeah, either heavy-ion beams or electrically-pumped excimer lasers seems like the path forward for the driver. Higher efficiency, higher repetition rate, possibly more robust. They also need to do away with holraums and switch to direct drive, to reduce target cost, ease alignment issues, and increase energy efficiency.

I don't hold out much hope for a practical, economical reactor from inertial confinement, but it's certainly exciting to see them achieve ignition & scientific breakeven, even if it's 10 years behind schedule. The one nice thing about ICF is that the energy gain shoots up dramatically once you cross the ignition threshold. That means they're arguably closer than tokamaks, even though both concepts need ~100x the demonstrated gain to get from where they are now to a workable reactor. (Ie, tokamaks have hit Q~0.3, need to get Q~30, vs ICF that has hit Q~1, needs Q~100).

Last time, they got something like 80% return on the laser energy input, now it's over 100% apparently. And, they had trouble repeating that last record, so people were questioning how meaningful it was if it couldn't be repeated. Now they've been able to repeat it & improve on it.

Yeah. Super interesting concepts I hadn't come across before, but it's hard to see the motivation for the design choices, in light of the stated goal (which is itself a bit 'out there'). OTOH, it's in general a hard problem to predict the emergent characteristics of the ecosystem of a programming language just from the underlying design choices.

Humans can think; humans are machines; therefore machines can think.

Edsger Dijkstra: "The question of whether Machines Can Think (…) is about as relevant as the question of whether Submarines Can Swim."

>we need to pair generative models with verification systems >code and math are easier to verify

I would love to see a two-stage pipeline using a LLM to convert natural language specifications into formal specifications for something like Dafny, and then follow up with another model like AlphaZero that would generate code & assertions to help the verifier. This seems like something that a major group like DeepMind or OpenAI could pull off in a few years.

Agreed. You can be a decent manager with mediocre tech skills provided you work really hard on the soft-skills part, but greatness only comes from excellence in both soft & 'hard' skills.

"Managers are hubs of communication, the better they communicate across these sphere boundaries, the more people they can communicate with, the more data they have, which, consequently leads to better decision making."

"As a manager, your job is that of a bullshit umbrella. You need to decide what crap your team needs to deal with and what crap can be ignored." https://randsinrepose.com/archives/managementese/

Leaders are shock-absorbers: https://www.edbatista.com/2022/10/leader-as-shock-absorber.h...

Leaders need to ensure alignment: "[Manager's] goals do not automatically become their employees' goals. On the contrary, much or most of a manager's job is to align these goals." http://yosefk.com/blog/people-can-read-their-managers-mind.h...

I like it. Some restaurants hand out little pagers, this could substitute for those.

“A technology that works 70% of the time is a bad one.” https://missiveapp.com/blog/life-and-death-of-read-tracking

With enough dimensions, everything is on the Pareto frontier: https://erikbern.com/2016/10/25/pareto-efficiency.htm

> people believe is true, but cannot build upon because they don't fully understand it.

OK, maybe it doesn't help people build on the proof, but does it help computers build on the proof? Why does it have to come back to human understanding? Maybe it's not possible right now for computers to generalize proof strategies, but I wouldn't bet against it becoming possible down the road.

Interesting thought. One could do Monte Carlo simulations to estimate the probability distribution of time-to-finish. I considered something like this [0] to explain the empirical log-normal distribution of project completion times [1], although I hadn't thought about including sequencing & bottlenecks. Probabilistic programming [2] is the discipline concerned with solving this type of problem - some tooling might exist, although probably w/o a nice UI.

[0] https://heteroskedasticblog.wordpress.com/2021/12/04/softwar...

[1] https://erikbern.com/2019/04/15/why-software-projects-take-l...

[2] https://en.wikipedia.org/wiki/Probabilistic_programming

From the introduction: "On the other end of the spectrum, formats such as HDF5 and BLZ address problems with large data sets and distributed computing, but don’t really address the metadata needs of an interchange format. ASDF aims to exist in the same middle ground that made FITS so successful, by being a hybrid text and binary format: containing human editable metadata for interchange, and raw binary data that is fast to load and use. Unlike FITS, the metadata is highly structured and is designed up-front for extensibility." [0]

Frankly, I feel like making the metadata of a scientific data structure human-editable is something of a mis-feature, or at best a non-feature. I use metadata in HDF5 files as a form of provenance tracking and I'd rather there be some friction to editing it.

[0] https://asdf-standard.readthedocs.io/en/1.0.3/intro.html

Good question. I think historical designs supported the fuel elements on material structures that ended up absorbing a large fraction of the fragments, leading to inefficiency & overheating the fuel. Here's an outside-of-the-box idea, using a dusty plasma cloud. [0] In this case, I'd be concerned about a criticality accident if the fuel dust containment failed & all the dust landed in a pile.

[0] www.rbsp.info/rbs/PDF/aiaa05.pdf

The power density is much higher than traditional photovoltaics (more in line with concentrating photovoltaic power), so in theory the cost should be low, although one probably needs more sophisticated cells to achieve reasonable efficiencies. On the other hand, since there are no moving parts, the temperature can be very high, which increases the maximum theoretical efficiency. (Carnot limit applies.)

The major issues I can see are (1) achieving sufficient efficiency at a reasonable cost, by recycling the low-energy photons instead of absorbing them as heat (2) avoiding radiation damage to this sophisticated device - maybe this isn't too hard with a liquid salt heat exchanger?

I loved the grade-school math textbook series by Harold Jacobs:

- Mathematics: A Human Endeavor - Elementary Algebra

They are peppered with cartoons (especially B.C.) and have lots of real-world motivation. I definitely recommend them for your kids if you have the choice.

Yeah. This blog post [0] has a nice graph showing that at continued exponential growth of energy consumption, we have less than a few hundred years until the planet would become uninhabitable simply due to the amount of additional heat. In particular, in about 400 years, the energy demand would equal the total solar energy flux hitting the planet.

[0] https://dothemath.ucsd.edu/2011/07/galactic-scale-energy/