April managed to slip into May so quickly I barely had time to blink, so some of the elements in this issue have been floating around since then, though updated for the current release.

Table of contents

• Conspicuous Non-consumption
• Current Reading: The Illusion of Thinking
• ChessGPT
• Into the Darkroom
• Newsletter Exclusives
• Multi-melting
• New to Dr Implausible's Bookshelf
• Recently on the podcast
• Past Story Updates

Published in the last week or so by the Machine Learning Research Lab at Apple, the Illusion of Thinking paper (linked here) challenges public assumptions about the reasoning ability of LLMs like ChatGPT, even as their capabilities expand with the addition of LRMs (Large Reasoning Models) to their abilities.

However, testing the capabilities of the LRMs (which granted, are very new) is largely uncharted territory, mostly focussed on solving math problems, which have returned some dissapointing results, and might not be indicative of "reasoning" per se. So the authors have created a method by which they can test them.  Gathering some of the major LRMs on the market (Claude 3.7 Sonnet, Deepseek-R1, and OpenAI's o3-mini), they ran them through a battery of tests.

But rather than math, they've compared the ability of the selected LLMs and LRMs to solve a number of problems, the kinds of word puzzles or things that most humans would have encountered at some point during their primary education. These include the classic tower of hanoi (stacking discs), checkers (or Hi-C), the river crossing (often presented as a wolf and sheep problem), and stacking blocks. This snip from the article illustrates them well enough.

The results are pretty interesting.  They saw the LRMs get stuck at various parts; often spending a lot of time and resources trying to push through an initial wrong assumption.  As the problems got more complex, the accuracy would nose-dive, leaving the model in a state of collapse.  The LRMs are supposed to break the complex problems down into smaller steps, which could also be performed by the non-reasoning LLM's that were also tested in comparison.

So, like the title goes: it's the Illusion of Thinking

However, like anything in the AI space, a field which is moving incredibly fast, this study is best understood as a snapshot in time.  Given the attention this paper is getting, both the model-makers and the model-testers will be adapting quickly, working towards this as a new baseline of problem to solve. They may look at addressing the problems directly - either through enrolling task-specific agents or improving the general reasoning. They will also likely look at dealing with the model collapse, and seeing what can be done to prevent the AIs from turning into a quivering heap during tough exam.

Which, to be fair, has happened to more than a few students over the years as well. Maybe they're closer to AGI than they think.

And continuing along the trend noted in the Illusion of Thinking paper, it was reported in the past week or so that a coder was able to pit ChatGPT versus Video Chess, released for the Atari 2600 VCS in 1979...

...and the Atari came out on top.

Queue the shock and horror from online, as the cutting edge of current consumer grade AI was laid low by 4K worth of memory and transistors from the last century.

But it isn't about ChatGPT getting beat by an Atari. My 3D Printer sucks at making toast too (it'll do it, kinda, but it takes forever and there's also a machine from the 70s that'll get it done faster).

There's no reason to expect that a text interpreter would inherently be good at chess, no matter how much data has been slurped to fuel its model.

It highlights the audiences assumption of use for GPTs and LLMs, but they're not there. Conceivably, if it had a chess module or could reach out to a different agent, that could act in this domain specific area, then it might be something, but it's not quite there yet. There is work being done in this area, to allow for the AIs to serve as controllers sending out commands to various agentive systems.  We'll look further into that in the coming weeks.

For now though, I got schooled by Video Chess back in the day too, so there's no shame there.  Depends on whether one learns from their mistakes.

Following a photography course taken in 2024, I decided to further my skills on the production side - in the darkroom.  Our local tech college has a fantastic darkroom facility, with people who know things (critically important), and it was worth following through.

While relatively inexpensive as a hobby, it was by no means "cheap". Even doing it on the low end things add up fast.  I used what's called a "simple use" camera - basically a disposable camera that has enough extras that you can reload it with a fresh roll. (I use one from lomography that I picked up on a trip last year, frex.)  The materials cost is about $13/roll Ilford HP5 400, and $140 per box (of 100 sheets) of print paper (again Ilford Multigrade RC Glossy), and I ended up going through quite a lot.

Not bad though: maybe $1.80 per 8x10 glossy? for a roll of 36?b  This rises or falls depending on how many test prints you have to do to get things just right.

Some of the materials like the developer and fixer were included in the cost of the course, which is nice, so the course fee basically becomes a facility fee for the 12 week or so duration.  And depending on how focused you are, you can get a lot done in 12 weeks.

So I think it makes a lot of cents. As I said, there were some seasoned professionals in there just for access to the facilities, and they were either showing or selling their work to interested clients.

But at my skill level (rank beginnner), I was making a lot of mistakes, but learning lots.  Because the darkroom side of photography is a matter of skill, where the art comes in the choices made in developing the picture.  Yes, the shot choice matters tons, but interesting things can be done in the dark.

In the background, I've been updating the indie version of the website, available at indie.implausi.blog, and making sure the past issues of the newsletter are available there.

We've still got a few months of blog posts to transcribe, but it's mostly there.  As detailed last year, this is partly a back-up, in case either WordPress or the current webhost takes a dive to the worse, and also an opportunity to engage with the web in a simpler and more knowable way.

In updating the newsletters, I realize a number of the exclusive articles don't really have a presence on the blog, and the I need to promote the new ones as well.  So we'll move exclusives to unique posts after a one month delay, and publish the full posts in the timeline as well.  I hope that isn't too much of a flood if you follow both the site and the newsletter, but I think it'll help us spread the word.

There's a lot more on the bookshelf, as the local Kinsmen book sale added another 50-odd titles to the stack. We'll dig into the results of that haul over the next issue or two.  One of those titles came recommended:

Book 1 of The Deptford Trilogy, along with 1972's The Manticore and 1975's World of Wonders, portrays an image of Canada that once was, in the early 20th century: more rural, more parochial, in ways that were echoed by Stuart McLean's Vinyl Cafe series and shows like Corner Gas and Schitt's Creek.  It's a unique window on that time, written in vivid way to bring the images to life in the reader's mind.  Enjoying it so far.

It's been a media-centric few months, with some amazing new works coming across the desk.  Here's a few of the other items that I've enjoyed recently.

Of course, the news keep on happening, and occasionally a story will come up that involves something we've addressed before, either here or on the podcast. Case in point:

AI Pharmaceutical discovery

In episode 38 of the podcast, we talked about how AI tools might be useful in the pursuit of science, by allowing for novel combinations of existing materials in ways that might not necessarily occur to human researchers.  Evidence of this can be seen in the above article on the use of AI tools in drug discovery.  This is rare, but promising, and I think we're going to see a lot more work done along these in the future.  For now, we'll add it to the Positive AI Archive, and keep looking for more.

We'll have a few new issues in the weeks ahead, recapping some of the podcast episodes that are being released as we speak, and showing the roadmap for the summer of 2025.

Up next: we'll be getting a little Rebellious in issue 10, looking at the season 2 of Star Wars: Andor and a visit to the Banksy exhibit in Vancouver.

Following that, we'll be getting Adaptive, with a look at our forthcoming coverage of the WYCU, the Weyland-Yutani Cinematic Universe over the course of June and July. We've talked about this a bit on the blog, and we'll be covering some related academic articles here too.

Stay tuned!