Pablumese
« previous post | next post »
Knowing how much I like to invent terms for things that have no name ("topolect", "character amnesia", etc.), and needing a word for the parlance produced by ChatGPT-4 and kindred AI chatbots, Conal Boyce asked me to coin a term for it. I instantly obliged him by coming up with "pablumese" to designate the sort of language that is unremittingly neutral and takes no stance on any subject or topic it addresses.
Conal liked my invention and responded:
Here's one of the problems with ChatGPT and its brethren: Not only does it spew what Victor calls 'pablumese' but for technical questions it then mixes its pablumese with quantitative nonsense, creating a truly creepy kind of output.
I was curious to see how it would handle the question of how many copper atoms fit into the cross-section of a typical copper wire. It responded in a way that made it sound very knowledgeable, breaking everything down into tiny (sometimes condescending) steps, and yet, at the very end of its perfect logic, it botched its answer, because it was unable to do a conversion between millimeters and picometers correctly.
But here's the kicker: What makes this stuff maximally odious is that the creeps who design it will succeed in taking over the world anyway, because this week "version 4 is astonishingly better than the beta ChatGPT!!!" and version 5 next week will be astonishingly better than…. etc. etc. until they've improved it enough that it really will threaten the jobs of 3/4 of the human race. It must be an absolutely sickening time to be a young person, trying to plan one's career.
For your amusement, below I've copy/pasted one of the ChatGPT responses to my question. In this one, the numbers look correct until we get to this division step:
(0.7854 mm^2) / (51,393.3 pm^2)
ChatGPT is unable to do the conversion properly from millimeters to picometers. The result of the division should be 1.52×10^13, but it claims the answer to be just 15,291 (which many of us humans would suspect immediately as an unrealistic estimate of how many copper atoms fit into a cross-section of a copper wire). Let me know if you think I have it wrong, but I've tried checking this from several different angles.
To solve this circle-packing problem, we need to determine the maximum number of circles of diameter 256 pm that can fit inside a circle of diameter 1 mm.
The key to solving this problem is to realize that we can use the formula for the area of a circle to calculate how much space each individual circle takes up. The formula for the area of a circle is:
A = πr^2
where A is the area of the circle and r is the radius.
In our case, the diameter of each circle is 256 pm, so the radius is:
r = (256 pm) / 2 = 128 pm
Using this radius, we can calculate the area of each circle:
A = π(128 pm)^2 = 51,393.3 pm^2
Now, we need to determine how many of these circles can fit inside a circle of diameter 1 mm. The area of a circle with diameter 1 mm is:
A = π(0.5 mm)^2 = 0.7854 mm^2
To determine the maximum number of circles that can fit inside this larger circle, we need to divide the area of the larger circle by the area of each individual circle:
Number of circles = (Area of larger circle) / (Area of each individual circle)
Number of circles = (0.7854 mm^2) / (51,393.3 pm^2)
Number of circles = 15,291.5
Again, since we can't have a fraction of a circle, we need to round down to the nearest whole number. Therefore, we can fit 15,291 circles of diameter 256 pm inside a circle of diameter 1 mm.
I'm not sure exactly what Conal has proven here about AI chatbots, but, in general, I find them informative, innocuous, and bland.
Selected readings
- "The mind of artificial intelligence" (3/22/23)
- "Style? Stance? What?" (10/27/18)
Taylor, Philip said,
March 23, 2023 @ 2:45 am
It is very early in the morning (for me), so I may well be talking out of my fundament, but I cannot see how the following statement stands up to scrutiny —
That suggests that the individual circles can be packed together with no lost space, which is clearly untrue; there is some loss of space, and in reality one can fit only approximately 0.9 times the number of circles that the above formula yields.
liuyao said,
March 23, 2023 @ 4:02 am
Asking math or numerical problems sounds to me not a very good use of ChatGPT, at least for now. Unless you need to restore your self-worth.
What's your take that these AI models and "services" have rather unimaginative names (ChatGPT, Bard, LLaMA), while the Chinese ones that have come out are quite different in style: Wenxinyiyan 文心一言, Wudao 悟道? (Not to compare their capabilities.)
DeepMind has their own naming style: AlphaGo, AlphaFold, etc. It seems that the generative models do get better names: DALL·E (a play on Dalí and the movie Wall-e), Midjourney (don't know where that comes from). Maybe it's just me tired of seeing acronyms.
Jerry Packard said,
March 23, 2023 @ 5:40 am
C.f., pabulum.
Ross Presser said,
March 23, 2023 @ 8:27 am
I think that the so-called startling results that supporters sometimes give are being generated when the topic is something that humans have written a lot about. When the topic is rarer or obscure, nonsense tends to result. I asked ChatGPT what it could tell me about Mark-Jason Dominus, a well-known person in the world of Perl. It got one thing right — the name of his most popular book — and then told me three different stories about which college he attended for his bachelor's degree, all wrong. (Dominus was extremely amused when I forwarded him the conversation.)
Olaf Zimmermann said,
March 23, 2023 @ 10:54 am
I know there's no accounting for taste, but to my bi-atlantic ears "pablumese" scans better than "pabulumese".
see also "https://www.merriam-webster.com/dictionary/pablum"
cf German "Einheitsbrei" – first occurrence I can find is from 1930, hence priority over a certain baby food – rhymes with "Eins, zwei, drei" …
@Jerry Packard
What's your take on this?
Jerry Packard said,
March 23, 2023 @ 12:16 pm
@Olaf
I agree with you. ‘pabulumese’ is a mouthful. My reason for writing ‘pabulum’ was just to bring it to the attention of LL readers, as I was surprised when I became aware of that form’s existence a couple of months ago. Compare: ‘poinsetta’/poinsettia, aluminum/‘aluminium’.
Haamu said,
March 23, 2023 @ 2:05 pm
I'd like to understand the basis for this insult. I feel it's unwarranted. As a result, I have to filter the utterance rather than taking it at face value.
This doesn't surprise me. We've all spent our entire lives filtering the utterances of NIs around us, whether for factuality, or hidden motives, or lack of reasoning or math skills, or whatever. The idea that we should now have to do this with the utterances of AIs should be expected. AIs have been trained by us on us, and so they reflect us back to us.
So, I don't share the level of alarm and condemnation that I'm seeing in some others, even while I agree it's appropriate to point out where these models are falling short. Conal has shown that AIs need more awareness of their own limitations, and we are right to require AI creators to address these shortcomings, to give AIs more awareness of AI/NI limitations so they reach for an alternative algorithm when needed in much the same way that an NI should know when to reach for an encyclopedia or a calculator.
But that's only the half of it. We also need to reprogram ourselves. AIs have learned how to simulate one of our markers of trust: the ability to compose text in a human style. We will need to develop other markers and methods, but that feels to me like just an extension of the same struggle we've been involved in for the last 10,000 years.
Haamu said,
March 23, 2023 @ 2:18 pm
Edit — I intended to write:
… an NI should know when to reach for an encyclopedia or a calculator.
bks said,
March 23, 2023 @ 3:40 pm
Which jobs will be obviated first by ChatGPT-N (for large N)? For 30 years we've been told that robots will replace fast-food servers, but instead all we have is Roomba (still the only somewhat-useful consumer robot after 20 years) and highly edited, misleading, short videos of robots doing backflips. Writing essays for high-school homework doesn't seem all that highfalutin, nor job displacin'.
Jarek Weckwerth said,
March 23, 2023 @ 4:57 pm
@bks: Making autonomous hardware roaming the real world interacting with real life objects is difficult and expensive. It seems that many white-collar jobs will go before those "menial" on-the-ground jobs. Just the other day, I found a leaflet in my mail advertising a local business replacing house doors. And I thought, "these guys are likely to keep their jobs while mine is gone". Because, for example, in language tech, there's far less demand for human phoneticians than 10 years ago.
@Haamu: Thank you! It was refreshing to read such a measured comment among all the AI-bashing.
Anthony said,
March 23, 2023 @ 10:26 pm
Asking an AI to do a task more suited to Mathematica isn't a terrible idea. I wonder about its ability in another specialized area: music. I imagine it can't understand musical notation nor analyze written music. Asking what key Richard Strauss's Metamorphosen for 23 solo strings is in would be interesting. Strauss thought of it as being in C Minor, but it wasn't published with any key in the title and there is no key signature, implying C Major but with many accidentals pointing to C Minor. There is some but not much published material about this question.
NSBK said,
March 24, 2023 @ 1:43 am
I tend to take issue with sweeping, catastrophizing statements like "But here's the kicker: What makes this stuff maximally odious is that the creeps who design it will succeed in taking over the world anyway, because [reasons], etc. etc. until they've improved it enough that it really will threaten the jobs of 3/4 of the human race".
ChatGPT and its ilk are language models, i.e. the domain of their abilities is (natural) language processing [0]. You talk to it and it talks back — well, you write to it and it writes back.
Let's take a quick look at what the "jobs of the human race" are. By this I mean I'll do the American thing and take data for the USA and ignore the rest of the world.
In 2021, jobs in the USA [1] could be broken down into:
92.8% : non-agriculture wage and salary
1.4% : agriculture/forestry/fishing/hunting
5.8% : non-agriculture self-employed
Let's go out on a limb and say language models won't be taking agriculture jobs and won't be self-employed. The remaining 92.8% can be broken down into:
12 8% : goods producing, excluding agriculture
80.0% : service-providing, excluding special industries
I'll venture as well that language models won't be producing goods — that category is defined as mining/construction/manufacturing.
The remaining 80.0% labelled "service-providing" breaks down into:
0.3% : Utilities
3.6% : Wholesale trade
9.7% : Retail trade
3.9% : Transportation and warehousing
1.8% : Information
5.6% : Financial activities
13.4% : Professional and business services
2.3% : Educational services
12.7% : Health care and social assistance
8.9% : Leisure and hospitality
3.9% : Other services
1.8% : Federal government
12.1% : State and local government
I'd personally outright exclude utilities, trade, transportation/warehousing, healthcare, and leisure/hospitality. Which would leave us with sectors where some portion of jobs could conceivably be fulfilled using "just" talking and writing:
1.8% : Information
5.6% : Financial activities
13.4% : Professional and business services
2.3% : Educational services
3.9% : Other services
1.8% : Federal government
12.1% : State and local government
and that totals 40.9% as an absolute extreme-case maximum. During the time that it takes for language models to get good enough to do these kinds of jobs, though, what other new sectors and categories of jobs will be created?
Sources
[0] https://www.ie.edu/insights/articles/unpacking-chatgpt-the-pros-and-cons-of-ais-hottest-language-model/
[1] https://www.bls.gov/emp/tables/employment-by-major-industry-sector.htm
Jerry Packard said,
March 24, 2023 @ 7:54 am
I guess asking it to do something in MATLAB would be a chore as well.
Taylor, Philip said,
March 24, 2023 @ 8:19 am
Haamu said :
Although I do not believe that the creators of ChatGPT et al. are creeps per se, I do believe that it and its ilk are amongst the most dangerous tools ever to be released to humanity, and the sooner an enforced code of ethics for such tools is introduced the better.
At the moment, ChatGPT (etc) spouts seemingly authoritative prose, which many will undoubtedly (and unfortunately) take at face value. Such prose should, I believe, be both introduced and closed by a statutorily-required disclaimer along the lines of :
Conal Boyce said,
March 24, 2023 @ 10:11 am
For those who might be interested in the details of the session I had with ChatGPT, its response begins with the words "To solve this circle-packing problem…" That starting point is not clear in the formatting above.
As for the quantitative error, it's interesting to note that there already exist tools in cyberspace for converting between square mm and square pm — tools that ChatGPT clearly "doesn't know about." I tried one such tool and it agreed with my side-calculation that 0.7854mm^2 converts to 7.85×10^17pm^2 — which in turn means that the 'answer' to my query should have been (7.85×10^17pm^2 / 51,393pm^2 = ) 1.52×10^13, not 15,291.
Why do I find that error so noteworthy? By the time the reader arrives at "15,291" s/he will have read a very long chat full of friendly, reasonable-sounding verbiage, in a style that gives the overwhelming impression of an entity in full control of the situation — an entity addressing a question that is trivial for it though perhaps difficult for the user with a smaller brain. One gets LULLED by the tone. And because of that lulling effect, the very LAST thing one expects is that the AI will conclude its ramble with a number that is not just wrong but disastrously wrong. (Side-issue: Is this something the tool "learned from us" and that we humans should therefore blame ourselves for? Nonsense.)
Victor Mair said,
March 24, 2023 @ 10:57 am
AI chatbots (e.g., ChatGPT-4) exist. They are the product of an enormous amount of human ingenuity, labor, and resources. One is tempted to believe that they are good for something. Properly harnessed and utilized, they may indeed fulfill productive, useful purposes. Left to run amok, they may lead to disastrous results.
The same is true of nuclear power and countless other human creations.
Conal Boyce said,
March 24, 2023 @ 11:17 am
@Taylor, Philip
I agree totally. Didn't want to complicate the discussion with that. The algorithm itself is quite simple-minded and erroneous from the git-go.
Nathan said,
March 24, 2023 @ 11:47 am
The ability to confidently produce bullshit has never been rare. Anyone who is vulnerable to this new artificial version was always vulnerable to the natural one.
Just like the fake news problem, what's scary about this is the condition it reveals humanity has been in all along: a large percentage of people don't know how to read skeptically.
Victor Mair said,
March 24, 2023 @ 3:37 pm
What Nathan says about the production (and consumption) of BS makes a lot of sense. When you read something, caveat emptor.
Grading student papers, I'm more concerned about their ideas than their rhetoric. For me, it's substance over style. That's why I can be fair to my many students whose native language is not English.
Haamu said,
March 24, 2023 @ 5:09 pm
Nathan is 100% correct. This problem has always been with us.
It feels suddenly worse because we're at a turning point. We've been able to establish certain modes of communication as inherently trustworthy, but as we get better at simulating reality artificially, that's no longer the case. No one should assume anymore that a photograph or video is incontrovertible evidence of what is depicted. Nor should we assume that television news is actually telling us the truth, nor that an email that appears to come from our bank actually came from there.
I don't think this is a crisis. We need to become more discriminating consumers of information, and we need to learn to take less false comfort in subjective certainty and learn to navigate life with a risk-minded, probability-minded approach that is functional even in the face of uncertainty. Above all, we need to shore up the institutions and social compacts that will engender trust, like the justice system, the educational system, journalism, and the like. Unfortunately, we generally seem more interested lately in tearing all of those things down, right when we really need them.
Haamu said,
March 24, 2023 @ 5:16 pm
For those who are concerned about LLMs' ability to do math, as well as for those who are curious about their ability to understand or compose music (see Anthony above), it's worth checking out the new paper from Microsoft Research that's been in the news the last couple of days:
"Sparks of Artificial General Intelligence: Early experiments with GPT-4"
The authors compare the performance of ChatGPT with GPT-4 in a variety of areas — there are discussions not only of text generation, but also mathematical abilities, music, image generation, logical reasoning, and more — and the level of advance is remarkable. If the curve continues, it seems as though a lot of the issues we're concerned about here are going to get addressed in fairly short order.
Jonathan Smith said,
March 24, 2023 @ 8:23 pm
thanks for the link — quite a title
I don't see why "Artificial General Intelligence" needs to be a concern at all for these developers. I suppose GPT might go further to demonstrating that "deep learning" type architectures are sufficient proxies for biological brains… but re "AGI," something's missing, what is it? what is it? to quote Shel Silverstein
from the abstract: "In our exploration of GPT-4 […] we discuss the challenges ahead for advancing towards deeper and more comprehensive versions of AGI, including the possible need for pursuing a new paradigm that moves beyond next-word prediction."
that is to say "AGI" MAY but then again MAY NOT entail anything more than next-word prediction on text :/ :/ :/ :/
but let it not be said that no progress has been made: at long last, ~80 years post-ENIAC, the newest electronic brains can finally make basic computational errors :D
davep said,
March 24, 2023 @ 9:05 pm
Nathan above mentioned this….
These seem to be designed for bullshit. They very-confidently say stuff that might or might not be true.
They also appear to lack the ability to say “I don’t know”. And just bullshit instead.
So many of these sound like a high school student trying (unsuccessfully) to hide that they didn’t read the book while trying to reach a word-count requirement.
unekdoud said,
March 25, 2023 @ 2:24 pm
On the math side: I was able to get the value of 256pm only after looking up that a face-centered cubic sphere packing is ~74% efficient (minus a negligible amount for copper being polycrystalline).
So it seems that ChatGPT forgot that close-packing efficiencies are a thing, moments after using it in calculation.
(You can reuse the 74% value for a typical planar cross-section, as opposed to 90% for the optimal planar cross-section mentioned. But I believe a physical copper wire sliced in half would have much more atoms on the interface than that.)
That last step of rounding down is just hilarious. I can imagine xkcd doing it as a joke, but without the explanation.
TM said,
March 28, 2023 @ 11:47 pm
For anyone saying "it's the same old caveat emptor" about reading the garbage these things spew out, there are already companies out there spruiking integrations with this tech, including at least one company selling an "AI hiring tool", sadly intended to be used when hiring humans (not AIs, sadly): https://www.theguardian.com/technology/2023/mar/27/robot-recruiters-can-bias-be-banished-from-ai-recruitment-hiring-artificial-intelligence
I expect that it's not that much more sophisticated than a keywords analyser in an unholy alliance with some quasi-MBTI questionnaire, but it is being marketed as something that'd make real life decisions, at least during the initial application process.
Microsoft are heavily promoting "upcoming integrations" in their enterprise 365 product suites. I'm sure that'll justify the $10B they dropped on it immediately prior to firing 1000s of staff worldwide.
It's all very well if this stuff is used as fancy chatbots or fuzzy data sorting at the back end, but it won't be limited to that. I've already seen highly-ranked (by google) search results for technical subject matter, obviously constructed by one of these things, full of verbose rubbish that looks plausible enough, and attributed to a person's byline. Maybe the byline is merely due to some poor "content creator" taking a shortcut, but with the high SEO, it seems more likely that it's strategically being passed off by the publisher as authored content.