Friday, April 28, 2023

Large Language Models - evolutionary tree and selection algorithm 2023

The following two diagrams were taken from Mooler et al's GitHub LLM guide (April 2023). I'm posting them here as a snapshot of this historic moment. The root of the LLM evolutionary tree goes far back in time, almost five years ago. Click image for full res.

Algorithm for selecting an AI technology

LLM Evolutionary Tree

Friday, April 21, 2023

Prosumer digital image management has not progressed to our advantage

Every so often the software market fails. I've had this happen to me a few times. From 1997 to 2007 I used a variety of PalmOS devices for what we used to call "personal information management" (PIM) - including Contacts, Notes, Tasks and Calendaring. My many PalmPilot/Palm handheld stylus devices synchronized by cable connection with Palm desktop software.

PalmOS died around the time the first iPhone came out. That original iPhone was both revolutionary and crappy. Functionally it was a huge regression from PalmOS Calendaring and other PIM solutions, but it was immediately clear that the iPhone was the future (seriously, there were no honest skeptics). Palm had been ailing already, but at that moment it was utterly dead.

It took three years for the iPhone to develop useable solutions for the "PIM-4" that worked across devices (often using either Google or Microsoft Exchange). During that time I had no handheld solution; I returned to using a paper Franklin planner. Finally, in 2010 or so, I was able to transition to the iPhone and iOS.

The market failure of digital image (and video) management has lasted longer and there's no end in site. This means something.

Things were actually looking pretty good for image and video management in 2015. Apple had consumer (iPhoto) and prosumer/professional (Aperture) applications that (mostly) shared the same image library. Things were not perfect -- Aperture had had years of horrible bugs and performance issues, but in retrospect this was a golden age. SSDs were fixing the iPhoto/Aperture performance issues and there were several reasonably priced alternatives including Adobe Lightroom. We didn't know how well we had it.

And then 2015 was when Apple killed both Aperture and iPhoto. There was no replacement for Aperture; users were left stranded with limited ability to migrate to another platform. Photos replaced iPhoto, but in most ways it was a functional regression. There was only one Photos advantage -- it promised a cloud-centric approach to image management with some limited backup features. If your iPhone or laptop was lost or destroyed your Apple Cloud images were probably safe -- as long as you paid for storage or didn't get locked out of iCloud by a phone thief.

Several alternative prosumer image management solutions emerged. But they all had the same problem Aperture had -- they all had severe data lock. If the software were to be discontinued, as happens to most products, there would be no way to extract one's images, image edits, and image metadata (ratings, keywords, titles, descriptions, albums, and on and on). In addition, perhaps inspired by the power of this data lock, many vendors moved to a subscription model. Adobe Lightroom now costs $120 a year, if you don't pay your photo library is essentially dead. Adobe can, if they wish, double or triple that price and customers will simply have to pay up. (I don't know what happens to the image library when a subscriber dies.)

I hoped Apple Photos would mature and develop more advanced features, but it has essentially languished. Recently Apple introduced a "Shared Library" model that is complex to use and, in my experience, has weird bugs and permission problems. (Lesson to users - if you ask Apple for something be prepared to regret your request.)

Eight years after Aperture died there still is not a great prosumer photo management solution for macOS customers. All the options have Hotel California Syndrome -- you can check-in but you can never leave. Apple's only option, the most natural fit for a macOS users, is dreadful and may be deteriorating. Many choices are subscription based and it's very easy for vendors to raise costs.

It's not hard to create a new standards and file based photo management solution. The file system does much of the work. Adobe has an open specification for image metadata management (XMP). Image to album, project, folder relationships are simple row triples. We've known how integrate external image editors for decades [1].

It's not hard ... but it hasn't happened. No vendor has decided to disrupt the marketplace and no open source (really open data structure is what we care about) solution has emerged.

Why not?

My best guess is that the Cloud is the problem. We've only gradually learned how to build responsive synchronizing Cloud products and they are not intrinsically file based. Development is much more challenging and the data lock advantage is irresistible for incumbents.

In the absence of a decent solution vendors are starting to build around the Apple Photos framework. This week Power Photos has a migration and access project. CYME Peakto is some mixture of Photos extension and standalone management solution. Houdah Photos Workbench adds a minuscule number of missing features to I can sort of imagine who these products might work, but Photos is a terrible foundation on which to build.

It's easy to image ways Apple could help, but they've been butchering photo management for a long time. They appear to be broken. The more realistic hope is that it will become easier for open source and other vendors to implement a standards based Cloud solution that would allow library migration between cooperating vendors - either through direct Cloud-Cloud communication or (better) a file based interchange format (what's a TB or two between friends?). I would be happy to pay a $200/year subscription fee for that kind of data freedom solution.

I've spent 7-8 years sitting on Mojave preparing to migrate to Apple Photos. The more I use Apple Photos the less I like this idea.  At this point I expect to convert my beloved 2015 MacBook Air to a non-networked Aperture machine and purchase a new M2 machine for my other work. Since Ventura no longer supports importing Aperture Libraries. I'll be looking for other migration options over the next one to two years. Maybe some vendor will decide to disrupt the data-lock. In the meanwhile I'll test Power Photos migration by periodically migrating my Aperture library to Photos.

[1] For each image store original, the proprietary image editor non-destructive edit recipe, and the most recent edited version in a user-defined format (lossy or lossless). If the editor is or changed the edit recipe is useless, bu the edited version is good.

Monday, April 03, 2023

We need a new word for the historical singularity.

TLDR: The "technological singularity" was an important and useful term with a clear meaning. Then it became the "Rapture of the Nerds". We need a new term.


I first heard the word "singularity" in the context of black hole physics; it dates back at least to the early 20th century:

ChatGPT 4 2023: "At the singularity, the laws of physics as we know them, including space and time, break down, and our current understanding of the universe is insufficient to predict what happens within it."

Not much later, in the 1950s, the term was applied by von Neumann in a technological context (from a 1993 Vernor Vinge essay):

Stan Ulam paraphrased John von Neumann as saying: "One conversation centered on the ever-accelerating progress of technology and changes in the mode of human life, which gives the appearance of approaching some essential singularity in the history of the race beyond which human affairs, as we know them, could not continue."

Brad Delong used to write about this kind of non-AI historical singularity. My favorite description of what it would be like to a approach at technological singularity was Vinge's short story "Fast Times at Fairmount High". (This prescient story appears to be lost to time; he wrote a similar full length novel but I think the short story was better).

The core idea is there's a (virtuous?) recursive loop where technology improves technology with shorter and shorter cycle times. Many processes go exponential and even near term developments become unpredictable. One may assume social end economic structures train to keep pace. The historical singularity exponential curve was part of The Economist's y2K Millennium issue GDP per person historical graph:

In a January 1983 essay for Omni Magazine Vinge focused on a particular aspect of the the technological singularity arising from superhuman intelligence (aka "super intelligence"):

We will soon create intelligences greater than our own ... When this happens there will be a technological and social transition similar in some sense to "the knotted space-time at the center of a black hole" 

A decade later, in his 1993 essay later published in Whole Earth Review (non-Olds cannot imagine what Whole Earth Review was like), Vinge revised what he meant by "soon":

... Based on this trend, I believe that the creation of greater-than-human intelligence will occur during the next thirty years. (Charles Platt has pointed out that AI enthusiasts have been making claims like this for thirty years. Just so I'm not guilty of a relative-time ambiguity, let me be more specific: I'll be surprised if this event occurs before 2005 or after 2030.) ...

So by the year 2000 we had the concept of a historical technological singularity (eminently sensible) that had become focused on a specific kind of self-improving technology - the superhuman intelligence with an upper-case S Singularity (presumably AI). Those were useful concepts - "technological singularity" and "superintelligence" Singularity. 

In 1993 Vinge predicted the Singularity would happen before 2030, later experts like Scott Aaronson predicted after 2080. (Aaronson has since revised that prediction and works for OpenAI; Vinge's 2030 dates looks pretty good.)

After 2000 though the word Singularity went off the semantic rails. It came to be used for for a peculiar future state in which human minds were uploaded into simulation environments that were usually described as pleasant rather than hellish. This is, of course, antithetical to the original idea of unpredictability! This peculiar meaning of "The Singularity" came to be known as "The Rapture of the Nerds" based on the title of a book by Charles Stross and Cory Doctorow. More recently that vision underlies a weird cult called longtermism that seems to have infected some vulnerable minds.

The "Rapture" meaning of "Singularity" has effectively taken over. We no longer have a term for the original von Neumann concept of self-reinforcing technological change that makes even the near term future unpredictable. That perversion of meaning is why I use the tag "Whitewater world" rather than Singularity for my own blog posts (and "Skynet" rather than super intelligence).

So we need a new term. I don't think "whitewater world" is it.

Anyone have a new term? We probably have a few years in which it might still be useful.