Gen Z lingo and search engines: A Millennial Odyssey Still on the waiting list for Microsoft's AI-powered Bing search features? No? Well, me neither, but CTO Kevin Scott recently let slip an amazing new use case that couldn't possibly be done with a boring, old-fashioned search engine like Google. Here it comes. Are you ready? Translating Gen Z slang. Wow. Speaking to the New …
Sorry but by what mechanism is generative AI supposed to distil information? Its advantages seem to be to pick keywords out of a verbose search and return verbose (and possibly inaccurate) results in a well-formed manner (due to it's 'knowledge' being pretty much, thanks to the LLM, how language should be used).
It might put "consultancy me" out of business, as it is even better at generating bullshit than I am. But "research me" would seem to be safe for a while yet.
You are already on the outside looking in. They already got you calling a GAN language model an "AI". Even in decrying the uselessness of the buzzword you are re-enforcing it :-)
Much like the recent blockchain frenzy this madness will continue until it kills enough people, the VC money runs out, or a better buzzword comes along that actually makes them some money. What we are seeing now is the circular reasoning playing out. They paid a ton to hoover up and store every bit of data they could, and paid large sums of money to people who were attracted to a field where the output of your work is a black box nobody REALLY understands that produces something that looks like "answers" based on suitably arm-wavy criteria. And these very smart people always say they will be able to fix all the problems with 10x more data, 3x more compute, and 30% more headcount. And maybe a neat sounding title.
I did have your reaction to the use of the term AI for these less-than-intelligent simulators. That's because my appreciation of the term was rooted in '70s research when we looked forward to truly intelligent machines.
However, I've become more relaxed about it. 'Artificial' was a poor word for non-human intelligence. 'Artificial' implies a less than perfect copy, lacking in those qualities held by the genuine article. As such, today's marketing darlings are truly AI - in the same sense as artificial flavourings or artificial turf. They are less than the real thing, imitations, made to fool our senses in some small way.
As such, I prefer the term 'Machine Intelligence' for the systems we hope for, and am happy to leave 'Artificial Intelligence' to those with lower expectations.
Much like [any new technology] this [will eventually find it's niche and fall out of the pubic conciousness]. What we are seeing now is [just the initial excitement, and that process of adaption and refinement, just as with anything new].
I would elucidate further but I've just seen ANOTHER ONE OF THOSE FUCKING CLOUDS!
<wanders off outside>
It's still fun though:
Me: Do you speak jive?
ChatGPT: As an AI language model, I have the ability to generate text in a variety of natural languages and dialects, but I do not have the ability to speak or understand spoken language like a human would. However, I can understand and respond to written text in a variety of dialects and languages, including Jive.
Me: Cut me some slack, Jack.
ChatGPT: Sure thing, my man. No need to get your feathers ruffled, everything is cool like a fool in a swimming pool. Is there anything in particular you need assistance with?
I'd say that's fairly impressive.
Depends on your definition of "impressive", I suppose. I don't find it surprising that a transformer LLM with that many parameters would 1) be trained on jive samples, and 2) locate the appropriate part of the representation space in response to a prompt.
Your particular example could be handled with the ML techniques of the '80s – a bunch of output filters using PCFGs, a decision tree to select the filter, a HMM for base "parody" text generation. ("Parody" was a term of art in natural language processing circles of the era, referring to Shannon's suggestion of using a PRNG as input and running a language model backward to try to produce plausible prose or speech.) True, that becomes infeasible at scale – each filter has to be implemented manually, or at least learned from human-labeled samples (supervised) – while a big transformer with a big training corpus can learn to do it unsupervised, or with only a final supervised tuning step. But for each specific instance we've been there, done that, read the machine-generated description of the t-shirt.
Transformer architecture and LLMs are novel steps in the evolution of natural language processing. But they present as far more radical than they actually are.
Ah, but you're looking at them with what sounds like a considerable amount of experience in the field, could this mean you're downplaying things slightly, simply because you forget the effort involved in amassing knowledge in a subject you find interesting. (I find things some junior programmers struggle with trivial, but it's only because I've been doing it many years, not because it really is).
Also, what we're seeing now is the result of decades of work, not just in the field of AI, but distributed computing, storage and networking; it's only all these advances that have made possible the systems needed for these kind of models to not only exist but be used by a wide section of the population.
That, I'd argue, is impressive.
But they present[ed] as far more radical than they actually are.
Of course they are, they're wildly over hyped, but so is anything that's new to the masses. (Hell, 20 years ago XML was going to save us all... somehow....)
> But they present as far more radical than they actually are.
Consider: Linux was about as novel as the IBM PC in its day. There was nothing original in the ipod/pad/phone.
Perhaps radical isn't what's important? I'll wait to see what the kids do with this one before passing judgement.
So the AI is prioritising a language typically used by a highly educated white male in their 40s over language used by other groups of people?
Isn't that racist in itself?
Try to chat using slang with ChatGPT, you'll get: "As an AI language model, I do not condone or promote criminal activity or the use of slang terminology.".
It is implying criminal activity just be mere use of slang.
Based on my usage of ChatGPT thats not true though. One of the "killer" use cases for it is what I call style translation.
You paste in a rough draft and then you ask ChatGPT to formalise or to "Gen Z"-ise the results are pretty promising.
ITs also a great "inspiration tool" if you are struggling to get started its good for a couple of summary paragraphs.
Of course idiots will use it verbatim, but thats really like using Excel as a calculator - a valid use case but not the tool sweet spot.
I guess it depends on the prompt and their PC detector.
Bruv, lemme tell you, my experience with ChatGPT tells me that's straight cap. One of the mandem's go-to moves is what I call style translation.
You throw in a rough draft and let ChatGPT finesse it or give it that Gen Z drip, and the outcome is straight-up peng.
Aiight, if you're struggling to get the ball rolling, ChatGPT's a real inspiration tool that can help you pattern a couple of solid paragraphs, ya get me?.
But you know some wastemen might just jank it word-for-word, but that's dead fam. It's like using Excel as a basic ting - it works, but that's not where the tool's level's at.
For comparison:
My esteemed acquaintance, I should like to impart to you my personal encounter with ChatGPT, which regrettably leads me to question the validity of the proposition at hand. In my estimation, the tool's most meritorious facet is that of style translation.
To elaborate, one submits a rudimentary draft and endows ChatGPT with the liberty to refine it, or imbue it with a contemporary lilt, to yield a most delightful result.
In addition, for those besieged by compositional inertia, ChatGPT presents itself as a fount of inspiration, enabling the generation of a few eloquent paragraphs.
One cannot deny the fact that certain individuals may choose to mindlessly reproduce the text verbatim. However, this is most inadvisable, and tantamount to employing Excel purely as a calculator, when the tool is capable of much greater accomplishment.
"Whatc*nt, wec*ck needc*ck nowc*nt isc*ck anc*ck algorithmc*ck forc*ck obfuscatingc*nt writtenc*ck stylec*nt, soc*ck thec*ck AIc*nt can'tc*ck pickc*ck outc*ck thec*nt salientc*ck featuresc*ck andc*ck reproducec*ck itc*nt laterc*ck."
That cloak works, AI refuses to write anything in this style (I mean if you got creative with the prompt it could probably be bypassed...)
Seems obvious. Just take all these wonderful examples we've been given and mix them together, aiming to never use any noun more than once (and not even once if you can). That way, GPT won't have a clue what you're saying, and as a bonus, your writing will be indistinguishable from output from a crazed chatbot anyway so nobody will have to know a human wrote it.
The perspicacity of the self-evident course is indisputable. Conflate these magnanimous instances proffered, whilst forgoing the reiteration of any noun (and refraining from such repetition if practicable). Thusly, GPT shall be confounded by the parlance thou dost employ, and as a fortuitous benefit, thy composition shall be indiscernible from the output of a nonsensical conversational automaton, thereby masking any vestige of human composition.
But the public language models are quite perilous to trust with this sort of thing unless you intimately understand the "Styles" you are using. The training data is the open sewer of though that is the public internet, and you can't include or exclude mass troves of radioactive material like twitter, reddit, 4chan, tiktok, youtube etc etc without infecting your model or biasing it.
The data sets would have to be painstakingly curated and cleaned, not just algorithmically scraped. You may get grad students to do that for formal and scientific language, but good luck getting the skate punks an the corner to slow down that long.
Or you know, not punk you, prank you, and laugh at everything but your inevitable failure to your choice in attire.
This exposes the underlying problem with the whole ML endeavour:
Getting the skate punks on the corner to do this kind of curation work is trivially easy *if* you pay them what their time is worth.
Convincing them to pay *you* for the opportunity to do your curation work for you, so that your bosses will approve a nice fat bonus for you, that's the hard part.
https://www.youtube.com/watch?v=RrZlWw8Di10
https://www.youtube.com/watch?v=NIaiW1XrzxA
Some years ago I joined a South London school party on a skiing trip to Switzerland. The kids quickly discovered that their youngish ages were no barrier to ordering drinks in the local bar.
Despite the fact that they had all had lessons at some point in time in French, the language spoken locally, not one attempted to use it. OK, the locals mostly could speak some English, so their behaviour was rude rather than a block to communication, except that they insisted on shunning standard English in favour of Herne Hill patois. When that failed they resorted to the time-tested technique of repeating themselves slowly and loudly.
I'd like to say that the bar staff beat them soundly around their heads, but that's not the Swiss way.
-A.
"...prioritising a language typically used by a highly educated white male in their 40s over language used by other groups of people."
The primary purpose of language is communication. To communicate, both ends must use the same definitions for words. We have standardized this in English, (for example,) so everyone can understand what is being communicated. If you are using something else, it's just social grooming and I don't care what you are saying.
Language changes over time and circumstances.
English is particularly prone to this, due to wide usage.
It also appears that you do not write standard English either, but instead a rather quaint USA dialect with more 'z' than "standard" English.
The meanings of words and phrases change all the time. For example, it's not long since a "Computer" was a human.
Now here is a genius proposal: "OpenAI Proposes Government Restrict AI Chips to Prevent Propaganda Explosion"
(https://www.vice.com/en_us/article/dy7nby/researchers-think-ai-language-models-will-help-spread-propaganda)
"...They also propose that governments could impose restrictions on training data collection and create access controls on AI hardware, such as semiconductors..."
So we should all block anyone but OpenAI from unrestricted access to hardware and to public language models? Right, block your competition for you, got it. No.
This reminds me of the same logic that produced the clipper chip proposals, and I expect similar success if it is implemented. The reality is that this is like the rest of the web. Built on a foundation that didn't include a trust model.
sign up tp a MS account and then apply to join a waiting list. I can remember the launch of Google 25 years ago, and there was a distinct lack of having to join waiting lists, is the thinking that they need to create a hype and then have people waiting excitedly to use the magic computer?
I, like everyone else, am very keen to see what crap I can get it to generate in order to amuse myself, but if it's so resource intensive that we're back to the days of having to wait days to have our turn at running something, it is severely underwhelming as a modern technology.
Alternatively teach your kids about how language is a tool for communication. And that for it to be effective both sides need to understand the language used. As such tuning ones use of language for the target audience matters and using words like bussin' and rizz isn't quite acceptable any time but certainly not when your parents (or anyone else deemed "proper company") is listening.
I would think the inimitable Jeeves might recommend consulting Sir Roderick Glossop who is
"...really a sort of janitor to the loony-bin. I mean to say, when your uncle the Duke begins to feel the strain a bit and you find him in the blue drawing-room sticking straws in his hair, old Glossop is the first person you send for. He toddles round, gives the patient the once-over, talks about overexcited nervous systems, and recommends complete rest and seclusion and all that sort of thing. Practically every posh family in the country has called him in at one time or another, and I suppose that, being in that position - I mean constantly having to sit on people’s heads while their nearest and dearest phone to the asylum to send round the wagon - does tend to make a chappie take what you might call a warped view of humanity."
The Inimitable Jeeves (1923)
PG Wodehouse
1923 a century later ... plus ça change, plus c'est la même chose
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
