Note #24435

According to news media: today, a billionaire is being taken into space, where the crew will open the rocket door and put him outside.

The Polaris Dawn crew (billionaire Jared Isaacman, accompanied by a retired US Air Force pilot and two SpaceX employees) preparing to launch.

This seems like a great start, and I hope that the remaining 1,780+ billionaires left here on earth will be following suit soon.

×

Pickled onion crisps

Why don’t I like pickled onion crisps?

I like pickled onions. And I imagine that the flavourings used in pickled onion crisps are basically the onion flavouring from cheese & onion and the vinegar flavouring from salt & vinegar, both of which are varieties I like.

AI posts on social media are the chicken nuggets of human interaction

This is a repost promoting content originally published elsewhere. See more things Dan's reposted.

Perhaps inspired by my resharing of Thomas‘s thoughts about the biggest problem in AI (tl;dr: he thinks it’s nomenclature; I agree that’s a problem but I don’t know if it’s the biggest issue), Ruth posted some thoughts to LinkedIn that I think are quite well-put:

I was going to write about something else but since LinkedIn suggested I should get AI to do it for me, here’s where I currently stand on GenAI.

As a person working in computing, I view it as a tool that is being treated as a silver bullet and is probably self-limiting in its current form. By design, it produces average code. Most companies prior to having access to cheap average code would have said they wanted good code. Since the average code produced by the tools is being fed back into those tools, mathematically this can’t lead anywhere good in terms of quality.

However, as a manager in tech I’m really alarmed by it. If we have tools to write code that is ok but needs a lot of double checking, we might be tempted to stop hiring people at that level. There already aren’t enough jobs for entry level programmers to feed the talent pipeline, and this is likely to make it worse. I’m not sure where the next generation of great programmers are supposed to come from if we move to an ecosystem where the junior roles are replaced by Copilot.

I think there’s a lot of potential for targeted tools to speed up productivity. I just don’t think GenAI is where they should come from.

This is an excellent explanation of no fewer than four of the big problems with “AI” as we’re seeing it marketed today:

  1. It produces mediocre output, (more on that below!)
  2. It’s a snake that eats its own tail,
  3. It’s treated as a silver bullet, and
  4. By pricing out certain types of low-tier knowledge work, it damages the pipeline for training higher-tiers of those knowledge workers (e.g. if we outsource all first-level tech support to chatbots, where will the next generation of third-level tech support come from, if they can’t work their way up the ranks, learning as they go?)

Let’s stop and take a deeper look at the “mediocre output” claim. Ruth’s right, but if you don’t already understand why generative AI does this, it’s worth a little bit of consideration about the reason for it… and the consequences of it:

Mathematically-speaking, that’s exactly what you would expect for something that is literally statistically averaging content, but that still comes as a surprise to people.

Bear in mind, of course, that there are plenty of topics in which the average person is less-knowledgable than the average of the content that was made available to the model. For example, I know next to noting about fertiliser application in large-scale agriculture. ChatGPT has doubtless ingested a lot of literature about it, and if I ask it what fertiliser I should use for a field of black beans in silty soil in the UK, it delivers me a confident-sounding answer:

ChatGPT screenshot. I ask 'I'm planting a field of black beans in silty soil in the UK in Spring. What fertiliser should I use to maximise my yield?' and it responds with ~560 words suggesting 30-40 kg/ha of phosphorus (P) and 60-70 kg/ha of potassium (K) at planting, among other things.
Who knows if this answer is right, of course! If the answer mattered to me – because I was about to drill my field – I’d have to do my own research to check, by which point I might as well have just done the research in the first place. If all I cared about was a quick sense-check to an answer I already knew, and it didn’t matter too much, this might be okay output. (It’s pretty verbose and repeats itself a lot, like it’s learned how to talk from YouTube tutorials: I’m surprised it didn’t finish by exhorting me to like and subscribe!)

When LLMs produce exceptional output (I use the term exceptional in the sense of unusual and not-average, not to mean “good”), it appears more-creative and interesting but is even more-likely to be riddled with fanciful hallucinations.

There’s a fine line in getting the creativity dial set just right, and even when you do there’s no guarantee of accuracy, but the way in which many chatbots are told to talk makes them sound authoritative on basically every subject. When you know it’s lying, that’s easy. But people don’t always use LLMs for subjects they’re knowledgeable about!

ChatGPT defined several words - 1. Quantifiable: Something that can be measured or expressed in numerical terms. 2. No cap: A slang term meaning "no lie" or "I'm being truthful." 3. Erinaceiophobia: An irrational fear of hedgehogs. 4. Undercontrastised: A medical term referring to an image, usually from a scan, that lacks sufficient contrast for clear diagnosis. (I made this word up, but ChatGPT defined it anyway!). 5. Ethology: The scientific study of animal behavior, particularly in natural environments.
I asked ChatGPT to define five words for me. Two (“quantifiable” and “ethology”) are real words that somebody might have trouble with. One (“no cap”) is a slang term. One (“erinaceiophobia” is a logically-sound construction from the Latin name for the biological family that hedgehogs belong to and the Greek suffix that’s applied to irrational fears). ChatGPT came up with perfectly reasonable definitions of all of these. But it also confidently defined “undercontrastised”, a word I made up and which I can’t find used anywhere at all!

In my example above, a more-useful robot would have stated that it didn’t know the answer to the question rather than, y’know, lying. But the nature of the statistical models used by LLMs means that they can’t know what they don’t know: they don’t have a “known unknowns” space.

Regarding the “damages the training pipeline”: I’m undecided on whether or not I agree with Ruth. She might be on to something there, but I’m not sure. Needs more thought before I commit to an opinion on that one.

Ruth followed-up to say:

Oh, and an addendum to this – as a human, I find the proliferation of AI tools in spaces that are all about creating connections with other humans deeply concerning. I saw a lot of job applications through Otta at my previous role, and they were all kind of the same – I had no sense of the person behind the averaged out CV I was looking at. We already have a huge problem with people presenting inauthentic versions of themselves on social media which makes it harder to have genuine interactions, smoothing off the rough edges of real people to get something glossy and processed is only going to make this worse.

AI posts on social media are the chicken nuggets of human interaction and I’d rather have something real every time.

Emphasis mine… because that’s a fantastic metaphor. Content generated where a generative AI is trying to “look human” are so-often bland, flat, and unexciting: a mass-produced most-basic form of social sustenance. So yeah: chicken nuggets.

Photo of chicken nuggets with "AI" written on each of them.
Ironically, I might’ve gotten a better picture here if I’d asked AI to draw this for me, because I couldn’t find any really unappetising-looking McDonalds-grade chicken nuggets on the stock photography site I used.
× × ×

Note #24429

This post is part of 🐶 Bleptember, a month-long celebration of our dog's inability to keep her tongue inside her mouth.

How is it the Tenth of Bleptember already? This young lady has so-far put off her morning nap and is instead intently watching me to see what I do next with my workday. Maybe it’ll involve dog treats! (Spoiler: it probably won’t. But you never know…)

Exaggerated soft-focus photo of a sitting French Bulldog looking up from her basket, her tongue sticking out.

×

The biggest problem with AI…

This is a repost promoting content originally published elsewhere. See more things Dan's reposted.

The biggest problem with “AI” is probably that it’s used as label for two completely different things:

1. Specialized neural networks trained to do highly specific tasks (e.g. cancer screening) which often work reasonably well as a tool to support human experts

2. Generative AI which thoroughly produces the most mid bullshit

It doesn’t help that neither are intelligent in any way, they’re both statistical pattern matching.

Fundamentally, Thomas seems to be arguing that the biggest problem with AI is how it is marketed, or things-that-are-called-AI are marketed as AI. Also that LLMs, by producing s statistical average of their input data, produce output that’s pretty-average (which is, of course, statistically that you’d expect)1.

I’m not sure he’s right: the energy footprint and the copyright issues of generative AI might be the biggest problems. But maybe.

Footnotes

1 That’s not entirely true, of course: sometimes they produce output that’s wild and random, but we describe those as “hallucinations” and for many purposes they’re even worse. At least “mid bullshit” can be useful if you’re specifically looking to summarise existing content (and don’t mind fact-checking it later if it’s important): y’know, the thing people use Wikipedia for.