jethkl

I generally agree, though direction, intuition, and domain knowledge are still relevant. Your breadth-first-search framing feels right, but you still need a sense of which paths are worth following, and you need to know when to trust the results.

I’ve been doing more math as a hobby in the past few weeks — working on lesser-known conjectures and exploring proofs of hard theorems — than I could have managed over the previous several years. It’s an exciting time.

if you tell me someone just stumbles onto and old coin...

I found a bill from the Weimar hyperinflation era. Its face value was several billion (Milliarden). Its only value was as a curiosity.

I don't think his Usonian concepts have had much impact on society.

The word Usonain has vanished, but the style's influence has not.

In 2024, I spent $750,000 on a 1200 sqft rancher built in 1962.

The Jacobs First House [1] in Madison, WI was the first Usonian house; it is credited with many features that became common in the mid-century ranches of the 50's and 60's. Stewart Hicks has a good deep dive [2] into Wright's influence on 20th century architecture.

[1] https://en.wikipedia.org/wiki/Herbert_and_Katherine_Jacobs_F...

[2] https://www.youtube.com/watch?v=ZXyiK-zVKsE

Wasserstein distance (Earth Mover’s Distance) measures how far apart two distributions are — the ‘work’ needed to reshape one pile of dirt into another. The concept extends to multiple distributions via a linear program, which under mild conditions can be solved with a linear-time greedy algorithm [1]. It’s an active research area with applications in clustering, computing Wasserstein barycenters (averaging distributions), and large-scale machine learning.

[1] https://en.wikipedia.org/wiki/Earth_mover's_distance#More_th...

There is an analogue of the CLT for extreme values. The Fisher–Tippett–Gnedenko theorem is the extreme-values analogue of the CLT: if the properly normalized maximum of an i.i.d. sample converges, it must be Gumbel, Fréchet, or Weibull—unified as the Generalized Extreme Value distribution. Unlike the CLT, whose assumptions (in my experience) rarely hold in practice, this result is extremely general and underpins methods like wavelet thresholding and signal denoising—easy to demonstrate with a quick simulation.

Are they financially screwed

I’ll share an anecdote I witnessed in my extended family–it was horrible. When US Air went bankrupt, employees with decades of service, expecting high five-figure to low six-figure annual income, learned they would get roughly $0.20 on the dollar. For many who were entering retirement, the impact was life-changing, and the stress and disruption it caused could well be argued as life-shortening.

PBGC did take over, but that did not solve the problem.

I believe the root cause was mismanagement of the pension, with the bankruptcy merely exposing this. But I wouldn’t be surprised if, at every opportunity during the bankruptcy process, changes were made that eroded the program’s health.

https://www.thestreet.com/investing/stocks/us-airways-pilots...

This is an instance where Conway's Law applies: state and county systems were kept separate so that maintenance and repairs crews wouldn’t accidentally duplicate work. https://en.wikipedia.org/wiki/Conway%27s_law

There are many physical scale models of the solar system around the world, many walkable, some bikeable: https://en.wikipedia.org/wiki/Solar_System_model

I've seen several, Planet Trek in Wisconsin is a good bikeable one with high quality signage. The sun is downtown, the moon is the size of a peach pit, Pluto is ~20 miles away.

how helpful was ai for this? The paper is light on details, but it says the agent was used to generate a kind of seed set (rank-1 bilinear products) that were then fed into the subsequent steps. Evidently this idea succeeded. Curious if anyone here has insight into if this is a common technique, how this agent's output would compare to random or a simple heuristic that attempts the same. Also interested to see how the training objective gets defined since the final task is a couple of steps downstream from what the agent generates.

My hope is that a lib like this one or similar could rally mindshare and become integrated as the new standard, and adopted by the wider developer community. In near term, it comes down to trade-offs. I see no decision that works for all use cases. Dependencies introduce ticking time bombs, stdlibs should be correct and intuitive, but at least when not they are usually well tested and maintained, but when stdlib don't meet urgent production needs you have to do something.

Link to Tom Scott & Computerphile from 10y ago on tz madness. https://www.youtube.com/watch?v=-5wpm-gesOY

you can reverse the playback, all the physics of billiards bouncing around works equally well in either time direction.

If the balls start clustered together, they will scatter over time about the box and the replay of their paths has only one plausible time direction.

It is extremely unlikely that all the billiards will, simply by chance at some point in the future, collect together so they are contained within a very small volume. this shows that there is asymmetry in which direction time flows.

Sharing a simple thought experiment that was shared with me years ago that explains (to me at least) why this is an interesting question. Imagine a billiard ball with nonzero velocity bouncing around an enclosed box. When the ball encounters a side of the box, it bounces off elastically. A replay of this ball's path over time is equally plausible if the replay were run forward or reversed. The preceding is also true if one imagines 2 or 3 balls, with the only difference being that the balls may also bounce off each other elastically. Even in this scenario, reversibility of playback holds no matter the configuration of the balls: they could all start clustered or be scattered and the replay would be plausible when played in either time direction. But this is no longer true when the box (now much larger) contains millions of billiard balls. If the balls start clustered together, they will scatter over time about the box and the replay of their paths has only one plausible time direction. This is because it is extremely unlikely that all the billiards will, simply by chance at some point in the future, collect together so they are contained within a very small volume. To summarize, in the "few scenario" we can plausibly reverse time but in the "many scenario", we cannot. The only difference between scenarios is the number of balls in the box, which suggests that time is an emergent property. layer8's answer elsewhere in this thread says the same, but more succinctly.

I thought the article would discuss physics of flight, but it's about physics of iridescence. Cool, but not what I expected. I've seen butterflies fly in a stiff wind. It seems impossible even when you watch it. I was hoping for some insight into that.

your comment introduced me to the term permaculture, looks interesting. Do you leverage numerical optimization - linear programming or integer programming for example - in your work?

That's a clarifying response: I've applied (non)convex programming (LP, QP, MIP, etc) to all the above and a few more, all of which i'd classify as classical applications. Less traditional applications -- I'd like to explore these more --- include data envelopment analysis, which provides a framework for assessing the efficiency of processes based on 1 or more input metrics, and several ideas in papers published at NeurIPS and other conferences that integrate LPs into neural networks in various ways, including AUC maximization. I've also worked on first order methods to solve LPs, and while I'd like to continue in that direction, the area is very crowded with very good existing tools, new and emerging tools that are also often good, and very strong teams building on all of the above. One of the biggest challenges that I see in the OR space is that it requires human expertise to leverage the technology.

LPs (and MIPs) are used for scheduling, planning, and logistics. These are some of the earliest applications of LPs, dating to the 1930's and 1940's, and these applications are still relevant. The wikipedia page has a good overview of the history and utility of linear programming. https://en.wikipedia.org/wiki/Linear_programming

Thank you! -- and thank you to the others who shared!

An extremely useful law to remember....

Would you be willing to provide an example where you applied Little's Law?

Martin Gutmann's recent talk ( https://www.youtube.com/watch?v=b0Z9IpTVfUg ) supports your point. While Shackleton is celebrated for overcoming disaster, Admunsen reached the south pole by avoiding disaster in the first place. Yet despite Admunsen being more effective, he is not celebrated to the same degree (at least in the US) as Shackleton.

You are missing the joke. The correct line would read:

  isHexadecimal("HADC0FFEE"); // false

The engineering hall statue! Thank you for posting your comment. I thought it was an abstract sculpture trying look "engineery". But it's a realistic representation of a cultural wound, and it is a warning.

https://engineering.wisc.edu/wp-content/uploads/2022/10/Kids...

This is such a prescient analysis from over 30 years ago, and he delivers so well - extemporaneous on a white board. The only error I see is the timing of when the market opportunity for NeXT would match their product: he predicted 1993. The world got there - pretty much as he predicted - but it took a long time.

It's great to see research in this field, I know there is opportunity here, and I hope to someday benefit from progress. But I skimmed the paper, and it doesn't appear solve a problem that I have. From the practical standpoint, what I want from a time series tool includes: 1) a small set of simple levers that I can review and tune 2) short training time for any input sets of size O(10k) to O(100k) (this covers seconds/day, minutes/week, hours/year) 3) the process of train + forecast can run fine on CPUs -- not GPUs with low memory overhead 4) decent out of the box performance that basically passes the sniff test and 5) a simple way to include regressors. I've enough experience to have learned to be wary of fully automated tuning, benchmark performance metrics, elaborate models, etc.

One amazing feature of Lustron houses is that their enameled exterior has aged amazingly well. I live near several of them, and all of them, without exception, appear to have survived 7 decades in a harsh northern climate. The enamel appears to be colorfast, does not chip or stain, and the material must be nearly indestructible. It isn't obvious from the photos, but the roofs are also enameled, and I believe all the houses near me still have the original roofs. I'm guessing the individual panels are basically impossible to replace or repair. They have survived well in a very tough environment.

Better realtime data may improve actionability, and the big audience reach may influence the overall narrative. Independent measurements are extremely important to have for science and data validation. But monitoring methane has been going on for a few years. NASA even has a portal:

https://methane.jpl.nasa.gov

https://www.esa.int/Applications/Observing_the_Earth/Coperni...

a defense against collisions like this is through CTEs that select a minimal set of columns, with column names suitably selected and standardized:

  CTE_A AS (SELECT ... comment as comment_a from A...),
  CTE_B AS (SELECT ... comment as comment_b from B...)

The article omits the "natural join". A natural join automatically joins tables on columns having the same name, and the columns in the resulting output appear exactly once. Natural joins often simplify queries, and they automatically adapt to schema changes in source tables. Natural joins also encourage standardized naming conventions, which has longer-term benefits.

the bug is the diy csv loader. use instead the csv module or its equivalent. My experience parsing csv's (how hard can it be?) often end with nested lists of iterators, complex logic to deal with edge cases, and other things that fail on painful-to-diagnose corner cases - exactly like what the author describes.

Good question, my thoughts: 1) to apply an LP or MILP to a practical problem that a business cares about requires a rare mix: one or a small number of people to have domain knowledge, knowledge about LPs/MILPs, and a good fit to the problem at hand. 2) the types of problem where LPs/MILPs reduce business expense is (currently) different from the spaces where ML has found success. This could change, but LPs have been applied extensively to strategic/logistics/planning applications, which aren't as approachable as applications of chatGPT, xgboost, etc. 3) LPs - since they are convex and provide global solutions - don't naturally support a kaggle-type competition that pits individuals and teams against each other. 4) there exist good open source solvers and very nice APIs (scipy, cvxpy, cvxopt, etc for example) but also high-cost and high-performance commercial solvers that businesses do pay for (Gurobi, CPLEX, etc).

Cool! Obligatory https://xkcd.com/195/

Its locality properties provide a practical alternative to the Hilbert curve for this type of map. Also interesting that it is loopy.