Atomic Razor

At Least It is a Plan of Sorts

I have been approved by my household CFO to get a Nvidia Project DIGITS personal AI supercomputer. This will supplement the Nvidia Jetson Orin Nano Super Developer Kit. I am hoping that the Jetson will arrive before May as it will allow me to get started. Supposedly, the Project DIGITS machine should be available in May. We shall see. But obviously I cannot spend all my time doomscrolling One of the ways I'm going to navigate the coming Time of Troubles is using the Project DIGITS and Jetson machines to research into

• program synthesis
• reasoning models
• Foundation Models
• CUDA (originally Compute Unified Device Architecture, a Nvidia proprietary parallel computing platform and API that allows software to use certain types of GPUs) for accelerated general-purpose processing)
• Deep Reinforcement Learning
• Cellular Automata
• Artificial life models
• simulation of physical systems

Obviously, there's a lot of other things that I want to do as well as we discussed here both recently and over the years and decades. But the thing I need to do it position myself with enough knowledge to get a job where I can do this sort of thing during my working day and get paid for it. I am hoping that at some point in the year or two I can get to join a company that is doing things with the proverbial data centre of Nobel Prize winners. I know, I know, but I have to have some hope. Perhaps then things can really get started. 

The Galaxiads

Look at this beauty! 

By ESO - http://www.eso.org/public/images/potw1330a/, CC BY 4.0, https://commons.wikimedia.org/w/index.php?curid=27491069

To quote the AAB description of this image

Spiral galaxies are usually very aesthetically appealing objects, and never more so than when they appear face-on. And this image is a particularly splendid example: it is the grand design spiral galaxy Messier 100, located in the southern part of the constellation of Coma Berenices, and lying about 55 million light-years from Earth. While Messier 100 shows very well defined spiral arms, it also displays the faintest of bar-like structures in the centre, which classifies this as type SAB. Although it is not easily spotted in the image, scientists have been able to confirm the bar’s existence by observing it in other wavelengths. This very detailed image shows the main features expected in a galaxy of this type: huge clouds of hydrogen gas, glowing in red patches when they re-emit the energy absorbed from newly born, massive stars; the uniform brightness of older, yellowish stars near the centre; and black shreds of dust weaving through the arms of the galaxy. Messier 100 is one of the brightest members of the Virgo Cluster, which is the closest cluster of galaxies to our galaxy, the Milky Way, containing over 2000 galaxies, including spirals, ellipticals, and irregulars. This picture is a combination of images from the FORS instrument on ESO’s Very Large Telescope at Paranal Observatory in Chile, taken with red (R), green (V) and blue (B) filters.

Which is a rather wonderful piece of writing. 

But this wasn't the galaxy I saw on NASA APoD the other day thanks, indirectly, to Del. So, why did I think of M100. I mean this is an amazing galaxy, about 55 million light-years away, it is ~166 kilo-light-years in diameter, so why a bit bigger than the Milky Way. That might suggest it's 2-3x more massive than the Milky Way, but who knows. It's depleted in hydrogen because of  interactions with the intracluster medium of Virgo, so perhaps it has fewer stars than you would expect. Still, how will we be able to say we know the world until we have a complete audit of M100? (Yes, it would probably take billions of tears to get there.)

Nearer to home is the galaxy I actually wanted to talk about, the Southern Pinwheel, M83, another absolute beauty.

By CTIO/NOIRLab/DOE/NSF/AURA Image processing: T.A. Rector (University of Alaska Anchorage/NSF NOIRLab), D. de Martin (NSF NOIRLab) & M. Zamani (NSF NOIRLab) - The Outskirts of the Southern Pinwheel Galaxy, CC BY 4.0, https://commons.wikimedia.org/w/index.php?curid=156389522

The AAB again:

Twelve million light-years away lies the galactic masterpiece Messier 83, also known as the Southern Pinwheel Galaxy. Its swirling spiral arms display a high rate of star formation and host six detected supernovae. This image was captured with the Department of Energy-fabricated Dark Energy Camera, mounted on the U.S. National Science Foundation Víctor M. Blanco 4-meter Telescope at Cerro Tololo Inter-American Observatory in Chile, a Program of NSF NOIRLab.

This is much closer galaxy, APoD states that it's only ~40,000 light-years across, but AAB gives ~116 kilo-light-years, so go figure. Either way, it's amazing to contemplate everything that has ever happened there. Are there civilizations there, intellects vast and cool and unsympathetic? Has anyone there ever studied and wondered about the Milky Way. It seems too awful to imagine they have not.

University of Life

If the population of the US is about 5x that of the UK and Harvard takes about 1500 undergraduates a year compared to 6000 at Oxford and Cambridge, it must, to first approximation, leaving aside issues like legacy admissions, be 20x more difficult to get into Harvard than Oxford or Cambridge. Of course, there are other Ivy League schools (Brown University, Columbia University, Cornell University, Dartmouth College, University of Pennsylvania, Princeton University, and Yale University) plus several other universities that might be considered comparable (Caltech, MIT, Stanford). A further complicating factor are liberal arts colleges that don't have much or anything in the way of  postgraduate programmes. There is no equivalent to this in the UK. A difficulty then for us is how to understand how the university entry system works here. My experience is with the UK in the 1980s. Clearly, lots of people who were good enough to go to Oxford or Cambridge didn't because they were put off by the entrance exam. As I recall acceptance rates were something like 50% in the 1980s, but lower these days. In the US, there's no limit on the number of universities you can apply to and you can apply to lots of Ivies if desired. There seems to be a great deal of focus on extracurricular activities whereas here it's just academics.   

A 2001 list of Public "Ivies"

Northeastern

• Pennsylvania State University (University Park)
• Rutgers University (New Brunswick, New Jersey)
• State University of New York at Binghamton
• University of Connecticut (Storrs)

Mid-Atlantic

• College of William & Mary (Williamsburg, Virginia)
• University of Delaware (Newark)
• University of Maryland, College Park
• University of Virginia (Charlottesville)

Western

• University of Arizona (Tucson)
• University of California, Berkeley
• University of California, Davis
• University of California, Irvine
• University of California, Los Angeles
• University of California, San Diego
• University of California, Santa Barbara
• University of Colorado Boulder
• University of Washington (Seattle)

Great Lakes & Midwest

• Indiana University Bloomington
• Miami University (Oxford, Ohio)
• Michigan State University (East Lansing)
• Ohio State University (Columbus)
• University of Illinois Urbana-Champaign
• University of Iowa (Iowa City)
• University of Michigan (Ann Arbor)
• University of Minnesota, Twin Cities
• University of Wisconsin–Madison

Southern

• University of Florida (Gainesville)
• University of Georgia (Athens)
• University of North Carolina at Chapel Hill
• University of Texas at Austin

which is a lot of universities and even more campuses, although it sort of feels about right. It we consider the University of Washington

For the Class of 2025 (enrolled fall 2021), UW received 48,840 applications and accepted 26,121 (53.5%). Of those accepted, 7,252 enrolled, a yield rate (the percentage of accepted students who choose to attend the university) of 27.8%.

This doesn't seem very selective to me, but perhaps I am missing something. The undergraduate class is about 4x that of Harvard and I suspect that it's harder to get in the CS programme these days. But it's very hard to know what it would be like to be an undergraduate at UW, now or in the 1980s, or whether if one were living in Seattle as a high schooler, one would want to go there. My dad was an aircraft engineer. Seattle is Jet City, although the Americans he worked with on the Phantom where from MacDonnell-Douglas out of St Louis. But if one were in Seattle or St Louis in the 1980s, where might one have chosen or been able to go?    

 

 

The Thousandth and One Philosophy

François Chollet has announced his AGI startup:

Today, the acceleration of scientific progress hinges on one factor: AI capable of independent invention and discovery. This capacity is the gateway to advancements beyond our wildest imagination.

Which sounds a lot like Amodei or Hassabis. Which is probably no bad thing. Chollet has a reputation as if not an AGI sceptic, an AGI realist. But if he things there's an approach worth trying that's definitely interesting. That's quite different thing from writing a wrapper library for TensorFlow. His company is called Ndea and you can apply to be part of its journey without CV or covering letter. Well, I guess they would ask you for a resume at some point. Sadly, I suspect my (joint) patent and 1997 PhD thesis aren't going to cut the mustard. But it does incentivise me to upload my PhD thesis to my GitHub.

I have for many years said that we should let a thousand philosophy's bloom. Chollet's idea is guided program synthesis:

We believe program synthesis holds the key to unlocking AGI. Instead of interpolating between data points in a continuous embedding space, program synthesis searches for discrete programs, or models, that perfectly explain observed data. This allows it to achieve much greater generalization power with extreme data-efficiency, requiring only a few examples to learn.

Apparently this is something all the big labs are working on now, but it's still an immature field; Chollet suggests it's comparable to Deep Learning in 2012. There's an AAB article on program synthesis, but it doesn't look terribly helpful. 

On Bluesky, Chollet states

I've been talking about some of these ideas -- merging "System 1" deep learning with "System 2" program search, on-the-fly knowledge recombination in the face of new tasks, lifelong learning, autonomous abstraction generation -- since 2017. 

So, we are looking at combining

• "System 1"
  • Deep Learning
•  'System 2"
  • Program search
  • On-the-fly knowledge recombination in the face of new tasks
  • Lifelong Learning
  • Autonomous abstraction generation

This at least gives me a goal to understand at least something program search and something to try and do with the Nvidia Jetson, if it arrives, or the DIGITS, if I can get one. Combining a reasoning model with a program synthesis system sounds like a fun project. Are there any actual resources yet though on program synthesis. There surely must at least be survey papers mustn't there? Perhaps this is royal route to PhD-level AI skills... 

Missing In Action

I had a massage last evening, which was why I didn't emit a blog. I am not sure how much better I feel for it, but I would probably feel worse if I hadn't had it. Which is like Dry January. I am definitely sleeping better, but it's hard to say I feel better. But I do know that I would feel worse if I were still drinking (this month; all bets off next, natch).

But here's the thing. How do you feel good? Not in the sense of not bad, but in terms of actually good? It's the same as enjoying something. There's lots of things I like, but how does one enjoy something as an unmediated experience? How does one find something that is so engaging that you would rather be doing that than something else? Some of it is concentrating on stuff. It's harder these days with the internet, but it's not like we didn't used to have books and magazines to distract when we were watching the telly. I can remember Civilization being a way of time-travelling into the future in 1992, but I have never been able to get FreeCiv to work and I probably don't have the patience to play it these days even if I could. Is there anything, at 56, for which the thing is more engaging than the idea of the thing? This links to the idea of whether you enjoy reading or having read. If you could download a book instantly into your brain and assimilate it somehow, would that be as good as actually spending hours reading the book. I feel there's at least a short story in that. 

TOPS, POPS and EOPS

Apparently my Nvidia Jetson is still on back order, at least I got another email about it, so that is good. Although who knows when it might arrive. As I have said, it supposedly delivers up to 67 TOPS. I am guessing those are INT8 ops or possibly FP4. The DIGITS, which I want, of course, has up to 1 petaflop at FP4. Or 1 POP. Which vaguely assumed me. I mean, one casually thinks, 67 TOPs, 67 billion operations per second. That's a lot! Then you realise, no, it's 67 trillion operations per second. The DIGITS is 1 quadrillion operations per second. At that point words have almost lost their meaning. Better, perhaps, to write 10¹⁵ operations per second. I mean, flipping heck.

https://commons.wikimedia.org/wiki/File:Supercomputers-history.svg#/media/File:Supercomputers-history.svg   

Moore's aw (broadly construed) is a remarkable thing as we can see here. In 1995, the most powerful computer had a performance of about 100 Gigaflops. The 500th fastest, which would have been a beast, 1 GFLOP. That's 67,000 times slower than the Jetson and a million times slower than the DIGITS. The fastest computer in the world currently is El Capitan, manages ~2 EOPS, so perhaps 20 million times faster than the fastest machine in 1995. That's 24 Moore's law generations, so the doubling period here is just over a year, although the 500th fastest and cumulative compute metrics indicate the period has lengthened over the years. It's still remarkable and I am hoping that we shall that trend reversed over the next few years. It is, as I have said, one of the few things that keeps me going. 

Fun in the Sun

In our post-political area, what can we do have fun? And by fun, I mean genuine fun. Not pretend fun or things that are worthy, but actual fun? It's a question I have struggled with for many years. It's not helpful to ask a foundational model. They either propose long term things such as taking a glass-blowing class or things I have no interest in like home mocktail competitions. Of course, there are lots of things that one can do on one's own, such as blogging or wilfing or writing or reading or doing a course or thinking deep thoughts, but I am not sure how much fun they are, and they certainly aren't things that you can do with someone else. Take reading, the last time I got lost in a book was, I think, Posy Simmonds's Gemma Bovary and that must be nearly a quarter of a century ago. That leaves games. Perhaps that's a possibility. I like the idea of games and I suppose there are whole classes of modern games for two people. (A problem is that I like the idea of a lot of things, but the idea of something is almost always better than the thing itself.) Of course, lots of classic games are for two people, chess or go, but I don't really see those as M's thing. We are going to give Avalon's Hill's Baseball Strategy a go finally after 37 years,so we shall see, but it would be good to have some others that worked for both of us. I like games, but, more particularly, I like the idea of games. This perhaps links me to Banksie and Empire. Indoor climbing seems improbable. Golf simulators? Perhaps.

Obviously, a lot of professionals (whom I am seeking currently) would argue that virtue is its own reward. Sure, nothing is really fun in the way that it was when you were a kid, but that's OK for reasons. It's not the games of gin rummy themselves, but that playing them is a way of keeping the shuttlecock in the air. The same with cooking a meal. It's just a meal, but it's all part of the life well lived (Beyond Beef mince and vegetable tagine with couscous tonight). The world is absolutely stuffed to the gunnels with interesting things, but it is difficult to maintain one's equilibrium at all times especially in an epoch of perhaps unprecedented idiocy. But one must try. There is no alternative. Let's go on. Perhaps the light in the distance is the dawn.

The World Builder's Handbook

I am currently read ing the Mongoose Traveller The World Builder's Handbook. How was it when I was in m my mid-teens that I could read everything? Well, there was no internet to wilf on, but even so. Anyway, this certainly goes a lot further than Book 6 Scouts. I mean we have pretty everything now, not just main sequence stars, but sub-giants and giants, white dwarfs, pulsars, neutron stars, black holes, even protostars and brown dwarfs! You can even have exotic stars if you like. This is sf after all! So, that's a pretty comprehensive menagerie of objects. But why not? One of the odd things about the stars in the sky is that almost all of them are much brighter than the Sun. We can't see Alpha Centauri in the Northern Hemisphere, but it's almost identical to the Sun and is the third brightest star in the sky. We can't even see the closest star, Proxima Centauri, with the naked eye or any other red dwarf, by far the most common type of star in the universe! So, every other star we can see, to first approximation, is either more massive than the Sun, like Sirius or an evolved star or both, like Betelgeuse.

So, though we might tend to think in terms of planetary systems like the Solar System, most are very different. But, in Traveller, this doesn't matter. Consider a planet in the habitable zone of a red dwarf. It's probably tidally locked (but then again Mercury isn't) and might experience severe volcanism because of tidal-heating of the interior. Worse, the massive stellar flares that red dwarfs are prone to probably blasted off any atmosphere billions if years ago. But a space-dwelling civilisation is quite happy to live on a vacuum planet, especially if it has near Earth-level gravity.

AS the Sun ages its luminosity increases. That's not necessarily good for the Earth, but Mars will get balmier, especially when the Sun leaves the main sequence. Of course, it might not have much atmosphere left billions of years in the future, but ice worlds will melt. The water vapour might just escape into space, but perhaps there's a period when Europa, Ganymede and Callisto might be water worlds. It doesn't even need to be for long on a planetological scale. AS long as conditions are stable enough for a few hundred years, you can seastead the outer Galileans for a while. Although nothing is dissolving in the lakes of Titan right now, when the ice melts all bets are off. And a "habitable" Titan will probably exist for millions of years with a thick atmosphere and water lakes. You might not get life evolving in that time, but it's fine to colonise.

Brown dwarfs are perfect for refuelling in the Traveller universe plyus why not have aerostats in the atmosphere. Indeed, Traveller has gravitics, so floating platforms are very much going to be a thing as are settlements at the bottom of hundreds of kilometres-deep oceans. Protostars might seem unpromising, but then why not live in a habitat if there are no suitable planetary surfaces yet.  Neutron stars and white dwarfs can host some pretty exotic objects that might not just of purely scientific interest to a starfaring civilisation, such as carbon stars.  

By NASA/JPL-CalTech Artist:Joby Harris - https://exoplanets.nasa.gov/multimedia/exoplanet-travel-bureau/?page=0&per_page=25&order=pub_date+desc%2C+id+desc&search=&category=122%3A174, Public Domain, https://commons.wikimedia.org/w/index.php?curid=83600821

There is that feeling I have discussed before that we missed the New Space Opera moment, especially given that it was a movement largely led by British writers. Doubtless, Al and Steve and Pete and Paul and CHarlie and all the other writers have anticipated some or many of the ideas I am outlining here, but it's not as though we are going to run out of interesting and different systems to think about and explore. G. Harry Stine wrote an Analog article called "Something's Strange in the System". This was my second Analog and the one that included Greg Bear's novelette, "Blood Music". So, I wonder if Thomas Pynchon had this issue. He might have read Stine's article. The point is that the Solar System is stranger now than in the 1980s (we will talk about this later), but every system is going to be, in its way, equally quirky. It would be fabulous to live at at an epoch at which we had audits of thousands or billions of systems. At the very least, there are still stories in that and worlds to be built.

Tap Dancing Towards the Abyss

Sometimes it's hard to keep the shuttlecock in the air given the sheer levels of willed fuckwittery we are experiencing right now. And things haven't even got started fully yet. So, the question is, what can we do that's genuinely engaging to distract us from the world crisis? Like what could I do now, right this minute, on a Friday evening? That's not watching telly. Or wilfing on the internet. Or exercise. I have already done an hour on the elliptical. And it's Dry January, so I can't drink. And I feel heavy and slow. It's hard because most of the things that one might propose are large-scale, long-term projects that are of no help now or, indeed, over the weekend or the next few days and weeks. Like having a massage, but I don't really even particularly like massages that much. There's a lot of worth stuff that could be done, but what's actually fun?    

Julia!

Via GeekWire's list of 21 startups selected for 12th WTIA Founder Cohort, I came across an interesting company explicitly describing itself as AGI company. Well, that sounds very exciting and I looked to see if they had any jobs going. They do, the catch being that they want a Julia Architect. Julia? I know that the Ju in Jupyter Notebook stands for Julia. Looking at the most popular programming languages in 2025, I find the TIOBE Index.This puts Julia at 31. Fortran is 10 up from 12 in 2024 (so I guess these are really 2024 figures). So, in terms of scientific programming one might be better with Fortran or Python one might have thought. The world is a very big place and, I suppose, even with a rating of 0.49% perhaps there are a lot of Julia programmers out there. It's perhaps not surprising that Udacity has nothing on Julia and Coursera just one course from the University of Cape Town in Scientific Julia. I'd be very happy to learn it , so I am going to take the American approach and apply anyway. I have nothing to lose.   

Programming Language	2025	2020	2015	2010	2005	2000	1995	1990	1985
Python	1	3	7	7	7	25	23	-	-
C++	2	4	4	3	3	2	1	2	13
C	3	2	1	2	1	1	2	1	1
Java	4	1	2	1	2	3	-	-	-
C#	5	5	5	6	9	9	-	-	-
JavaScript	6	7	8	9	10	7	-	-	-
Go	7	17	37	184	-	-	-	-	-
Visual Basic	8	19	234	-	-	-	-	-	-
SQL	9	9	-	-	100	-	-	-	-
Fortran	10	29	31	25	15	18	5	9	11
Ada	25	35	30	27	16	17	7	4	3
Lisp	29	31	18	17	13	16	6	3	2
Objective-C	35	10	3	24	40	-	-	-	-
(Visual) Basic	-	-	54	5	4	4	3	5	4

Great to see that Fortran has been making a concerted fightback over the last few years. The Original and the Best.

The languages in positions 51-100 make interesting reading. In alphabetical order:

• ABC, ActionScript, Algol, Apex, APL, CFML, CHILL, Clipper, CLIPS, Clojure, Curl, Elm, Erlang, F#, Forth, Groovy, Hack, Icon, Inform, Io, J#, JScript, LabVIEW, Ladder Logic, Modula-2, Mojo, MQL5, NATURAL, Nim, OCaml, Occam, OpenCL, OpenEdge ABL, PL/I, PostScript, Q, Ring, S, Scheme, Simulink, Smalltalk, SPARK, SPSS, Stata, SystemVerilog, Tcl, Vala/Genie, VHDL, Wolfram, Zig

I think I have used four of those (Curl, Erlang, PostScript and Wolfram) at some level though not necessarily for real programming, but certainly in the case of Curl for real work. I have heard of, for certain values of heard of, probably 15 or 20 others. I guess a lot of this for the maintenance of legacy systems. Imagine doing Algol in 2025!