Yeah, in essence. This is actually a pretty cool part of working in Lean. It's a somewhat normal convention to write something in a human readable way and then write a second optimized implementation with some kindness of correctness theorem connecting them. There was a whole open "competition" for writing a faster Lean kernel/proof checker that didn't sacrifice on soundness called Lean Kernel Arena. Fun reference point: https://kim-em.github.io/blog/2026-7-24-why-lean-is-faster-t...
Interesting that that's a feature of the water heater and not the tub. It has to travel through the plumbing all the way round-trip to the heater? You must have isolated recirculation plumbing to deal with the grey water, yeah?
This is pure escapism. Numbing yourself to the burden of being. Likely from a stress tolerance so shallow and under-developed that the slightest inconvenience or expectation feels like a crisis.
The truth is that after the brief initial relief of excusing yourself from being anything at all, you'll end up in a deep depression. It's a naive and immature fantasy.
Remove all electronics from your life, they are a constant nervous system stimulator that has prevented you from truly relaxing into a state of repair. Then introduce responsibilities into your life, start small. We are wired to be needed, wanted, useful, accountable, social. You're just stressed out.
Dave Beazley has a great talk about using Python built ins [0] for data analysis and other quick operations.
As a meta note, I've used many of these builtins over the years but, due to LLMs, have been using them less and less. Re-watching the video almost felt like watching bushcrafters make a chair using just a knife and saw...
That's silly to say when it can be fruitfully applied in so many situations. Any time you have noisy and sparse pairwise comparisons, you can think of them as one of the sides of the pair vouching for the other side. If you then solve PageRank for the entire graph, you get a somewhat principled global ranking of all items.
I used it recently to construct a top list of books I read a year based only on sloppy pairwise comparisons between them. I've also used it to judge the quality of other relevance algorithms while keeping the human input to a minimum.
I don't know of many alternatives that work better than PageRank under those conditions. Thurstone-type models require dense comparisons, and Elo doesn't fare very well when the comparisons are too noisy.
It's brutal, my son with a CS degree hasn't been able to get to single live interview ina year. A couple of screens has been it. I've reached out to my network, and the reply is always we're not hiring any juniors.
My heard hurts - i was stupid enough to think that SIMD was a CPU only thing - I don't understand why it would be ported to GPU - huge kudos to managing to surprise me
amusingly ive been working on ultra sparse llm inference/ training/ model design because nature loaths a dense graph/matrix and cause i think it shoukd be possible. i actually stood up a 20-25 percent faster than sota causal fast attention kernel yesterday, will be standing up cuda/metal/armv8 kernels too and thats gonna be fun.
i genuinely think these models should be like 0.1 percent sparse for same capabilities we associate with them today, but theres no sane way to do that with extent tools. i built the right core tech for that in 2014 when there wasnt a market, but now there is and the experimentation velocity is wild.
amusingly llms really have a hard time using my simple apis because its not in distribution array programs. but i literally stood up cpu custom memory format and micro kernel for dense causal attention in less than 24-36 hours and outperforms the equivalent fused ggml/llama cpp fast oath by like 20-25 percent
Much of math (or science) research has the strange quality of being mostly curiosity-driven, but having giant benefits that occasionally spin out to the public.
Some questions are more urgent and practical. My feeling is that the more directly practical a question is, the more likely the research community is to support AI usage in that question.
The annoying thing about recent AI advances is that they target questions on the wrong end of the spectrum: Erdos problems are exactly the sort of "useless" questions that people might answer purely for the love of the game. The sort of questions that a young person might cut their teeth on and gain confidence.
Solving questions like these automatically, I think, is not good for the long-term health of research. At least for the foreseeable future you still would like people to become interested and develop skills in these fields. These developments, and especially how they are presented, directly discourage that.
India has the problem with farming that the US is starting to have with AI. Farming in India is still far too labor intensive by world standards. 43% of workers still work in agriculture. [1] For the US, that number is under 2%. China is at 22% as of 2023, and dropping steadily.
This inefficient agricultural system is not by accident. It is supported by heavy subsidies. Attempts to cut the subsidies resulted in riots.[2] Trouble is ongoing.
Comments from someone who knows more about this than I do would help here.
The US and most of the EU went through that transition over several generations, and farming is still heavily subsidized in both areas. The transition happened faster in China, and a hukou system was put into place to prevent people from migrating from farms to cities faster than the cities could absorb them.
Looking at how countries coped with a fast transition from labor intensive agriculture to an urban society gives hints on how an AI transition may look. All the Asian countries that went from poor to rich in a generation did this, with different approaches. How that took place may provide more useful info than philosophy.
There's already an okay solution to supply-chain attacks against dependency managers like npm, PyPI, and Cargo: set them to only install package versions that are more than a few days old. The recent high-profile attacks were all caught and rolled back within a day, so doing this would have let you safely avoid the attacks. It really should be the default behavior. Let self-selected beta testers and security scanner companies try out the newest versions of packages for a day before you try them. Instructions: https://cooldowns.dev/
1. Drones are a relatively recent evolution but are really a continuation of asymmetric warfare that has been wildly successful post-1945. The US has been woefully unprepared for cheap, mass-produced drones that, as evidenced by these satellite images, are equivalent to high-precision missiles in terms of effectiveness but are substantially cheaper;
2. Censoring these images serves no military purpose. Iran, China and Russia (among others) have access to accurate satellite imagery so censoring these images really just belies a fear of public opinion. Any cost estimates of this war given by the administration (which tend to be $1-2B/day) don't seem to include repairing and replacing lost weapons, radars, facilities, aircraft and other base infrastructure. That's going to be billions more;
3. It seems clear that this war so ill-considered and the US was so unprepared that (IMHO) will go down as the biggest strategic blunder in US history as the US military and Gulf security guarantees have shown to be a paper tiger and there is no military out of this conflict short of the use of nuclear weapons;
4. Despite claims to the contrary, the US does not appear to have air superiority over Iran. The evidence for this is the continued use of missiles and other so-called "stand off" weapons (ie fired at range to avoid SAMs and anti-aircraft batteries);
5. Despite administration claims to the contrary, there are now desperate shortages of munitions for missile defences, Tomahawk missiles and various other missiles. Some of these had already been dseriously depleted in the 12 day War. This has made things substantially worse and it will likely take years to replenish supplies;
6. The future of Gulf bases and secruity guarantees is now unclear given it's now been demonstrated that the US can't protect them; and
7. I'm not sure the UAE (and Duabi in particular) ever recovers from this. The image that Dubai is some stable center for business and finance in the MIddle East has been shattered. Will the wealthy come back knowing the US can't protect Dubai? I honestly don't know. Dubai is a "wretched hive of scum and villainy" (to quote Star Wars). It's key in Iran evading sanctions, Russia evading sanctions and instrumental in the South Sudan genocide (ie there's a trade between UAE arms from the US and stolen South Sudanese gold from the RSF). The UAE has left OPEC. I honestly don't know if this will be a good or bad decision long-term.
There are 3 players in this war and they all have very different goals. Israel wants to wreck Iran. The US wants out. Iran simply needs to survive. I'm not sure where we go from here. To back down, the US would need to split with Israel and that's a pill likely too difficult to swallow given that Israel is the only reason we're in this war at all.
Looming over all this is the upcoming summit between the US and China, currently set for next week. It's already been delayed once because of this war. Having this situation unresolved is going to greatly weaken the American negotiating position. The US may well want to delay it again. If so, (IMHO) China may well cancel it entirely.
I also liked the one on lambda calculus. I hope one day we will be able to find interpretation of what it actually means for PLUS Times Plus. Maybe this is how we will explore nonstandard arithmetic.
Edit: Oh wait, no, I was thinking of the Drew's Campfire double pendulum video. That video was extra interesting because the creator is not a typical content producer. He just has a few videos without any views, then dropped what might be one of the best videos of all time, and then went back to his technical videos.
More on what astronauts found “objectionable” and “distasteful” with Apollo's system, from the PDF linked in the OP (1):
"In general, the Apollo waste management system worked satisfactorily from an engineering standpoint. From the point of view of crew acceptance, however, the system must be given poor marks. The principal problem with both the urine and fecal collection systems was the fact that these required more manipulation than crewmen were used to in the Earth environment and were, as a consequence, found to be objectionable. The urine receptacle assembly represented an attempt to preclude crew handling of urine specimens but, because urine spills were frequent, the objective of “sanitizing” the process was thwarted.
The fecal collection system presented an even more distasteful set of problems. The collection process required a great deal of skill to preclude escape of feces from the collection bag and consequent soiling of the crew, their clothing, or cabin surfaces. The fecal collection process was, moreover, extremely time consuming because of the level of difficulty involved with use of the system. An Apollo 7 astronaut estimated the time required to correctly accomplish the process at 45 minutes.* Good placement of fecal bags was difficult to attain; this was further complicated by the fact that the flap at the back of the constant wear garment created an opening that was too small for easy placement of the bags.** As was noted earlier, kneading of the bags was required for dispersal of the germicide.
*Entry in the log of Apollo 7 by Astronaut Walter Cunningham.
**The configuration of the constant wear garments on later Apollo missions were modified to correct
this problem."
This is indeed a great quote (one of many gems from Sir Tony) but I think the context that follows it is also an essential insight:
> The first method is far more difficult. It demands the same skill,
devotion, insight, and even inspiration as the discovery of the simple
physical laws which underlie the complex phenomena of nature. It
also requires a willingness to accept objectives which are limited
by physical, logical, and technological constraints, and to accept a
compromise when conflicting objectives cannot be met. No committee
will ever do this until it is too late.
I'm a occasional hobbyist maker and i've used Autodesk Fusion, Solid Edge, OpenSCAD and other niche parametric programs, but always felt FreeCAD was too complex. But I really wanted it to work for me because it's FOSS and 100% offline. So with the new FreeCAD 1.1 RC I found an hour long tutorial and dove in.
(1.1 is supposedly much easier to work with)
After doing the tut I can say that 1.1 is very nice, i can uninstall Fusion and Solid Edge finally :)
> The transformer architectures powering current LLMs are strictly feed-forward.
This is true in a specific contextual sense (each token that an LLM produces is from a feed-forward pass). But untrue for more than a year with reasoning models, who feed their produced tokens back as inputs, and whose tuning effectively rewards it for doing this skillfully.
Heck, it was untrue before that as well, any time an LLM responded with more than one token.
> A [March] 2025 survey by the Association for the Advancement of Artificial Intelligence (AAAI), surveying 475 AI researchers, found that 76% believe scaling up current AI approaches to achieve AGI is "unlikely" or "very unlikely" to succeed.
I dunno. This survey publication was from nearly a year ago, so the survey itself is probably more than a year old. That puts us at Sonnet 3.7. The gap between that and present day is tremendous.
I am not skilled enough to say this tactfully, but: expert opinions can be the slowest to update on the news that their specific domain may have, in hindsight, have been the wrong horse. It's the quote about it being difficult to believe something that your income requires to be false, but instead of income it can be your whole legacy or self concept. Way worse.
> My take is that research taste is going to rely heavily on the short-duration cognitive primitives that the ARC highlights but the METR metric does not capture.
I don't have an opinion on this, but I'd like to hear more about this take.
Not go to all “ackchually” but modern GPUs can render in many other ways than rasterising triangles, and they can absolutely draw a cylinder without any tessellation involved. You can use the analytical ray tracing formula, or signed distance fields for a practical way to easily build complex scenes purely with maths: https://iquilezles.org/articles/distfunctions/
Now of course triangles are usually the most practical way to render objects but it just bugs me when someone says something like “Every smooth surface you've ever seen on a screen was actually tiny flat triangles” when it’s patently false, ray tracing a sphere is pretty much the Hello World of computer graphics and no triangles are involved.
Edit: for CADs, direct ray tracing of NURBS surfaces on the GPU exists and lets you render smooth objects with no triangles involved whatsoever, although I’m not sure if any mainstream software uses that method.
> For the chess problem we propose the estimate number_of_typical_games ~ typical_number_of_options_per_movetypical_number_of_moves_per_game. This equation is subjective, in that it isn’t yet justified beyond our opinion that it might be a good estimate.
This applies to most if not all games. In our paper "A googolplex of Go games" [1], we write
"Estimates on the number of ‘practical’ n × n games take the form b^l where b and l are estimates on the number of choices per turn (branching factor) and game length, respectively. A reasonable and minimally-arbitrary
upper bound sets b = l = n^2, while for a lower bound, values of b = n and l = (2/3)n^2 seem both reasonable and not too arbitrary. This gives us bounds for the ill-defined number P19 of ‘practical’ 19x19 games of
10^306 < P19 < 10^924
Wikipedia’s page on Game complexity[5] combines a somewhat high estimate of b = 250 with an unreasonably low estime of l = 150 to arrive at a not unreasonable 10^360 games."
> Our final estimate was that it is plausible that there are on the order of 10^151 possible short games of chess.
I'm curious how many arbitrary length games are possible.
Of course the length is limited to 17697 plies [3] due to Fide's 75-move rule. But constructing a huge class of games in which every one is probably legal remains a large challenge; much larger than in Go where move legality is much easier to determine.
The main result of our paper is on arbitrarily long Go games, of which we prove there are over 10^10^100.
While useful it needs a big red warning to potential leakers. If they were personally served documents (such as via email, while logged in, etc) there really isn't much that can be done to ascertain the safety of leaking it. It's not even safe if there are two or more leakers and they "compare notes" to try and "clean" something for release.
The watermark can even be contained in the wording itself (multiple versions of sentences, word choice etc stores the entropy). The only moderately safe thing to leak would be a pure text full paraphrasing of the material. But that wouldn't inspire much trust as a source.
Super cool to read but can someone eli5 what Gaussian splatting is (and/or radiance fields?) specifically to how the article talks about it finally being "mature enough"? What's changed that this is now possible?
For most daily needs: chef's knife, pairing knife, serated/bread knife. Possibly useful 'extras': kitchen shears, petty/utility, boning, slicing/carving. They do not recommend sets.
I will beat loudly on the "Attention is a reinvention of Kernel Smoothing" drum until it is common knowledge. It looks like Cosma Schalizi's fantastic website is down for now, so here's a archive link to his essential reading on this topic [0].
If you're interested in machine learning at all and not very strong regarding kernel methods I highly recommending taking a deep dive. Such a huge amount of ML can be framed through the lens of kernel methods (and things like Gaussian Processes will become much easier to understand).
For the particular case of the 5 delimiters '\n', '.', '?', '!', and ';', it just happens to be so that you can do this as a single shuffle instruction, replacing the explicit lookup table.
You can do this whenever `c & 0x0F` is unique for the set of characters you're looking for.
