I mean, I’ve been aware of at least reasonably good voice-to-text tools for… long enough ago that it’s a little fuzzy, but I think I remember my dad starting to use one in the late 90s or early 2000s to dictate notes, rather than recording them and having a secretary do it. So, close to 30 years? I ended up marrying a doctor as well and I should probably ask her what they do today, but as someone whos spent the last 20 years in an office environment, I can think of less than a handful of people who use voice-to-text tools as part of their work. Certainly, on a trading desk, it would be a clusterfuck.
But, ultimate point being, voice-to-text is a 30-year-old technology, and it’s STILL not widely adapted in the business world. That’s worth thinking about.
I think there’s an element of “you people were so concerned whether it COULD be done, that you never stopped to think if it SHOULD” here. Something like voice-to-text just isn’t going to fly in an office environment where you have a team working in an open layout, and not in closed-off private offices. And it almost definitely wouldn’t work in any situation where it would need to analyze voice commands including confidential client information - names, addresses, socials, etc, any information that we’re legally obligated to keep secure and a voice-analysis software suite would represent a potential vulnerability. This is also true of AI - our policy is anything involving the potential use of AI tools needs to be precleared by our compliance department for a number of reasons, but confidentiality and data security are two of the biggest. Copyright is also a growing concern in the AI space, both the inputs used to “train” a LLM, as well as the output it then produces.
I think there’s this attitude in AI enthusiasts that we’re going to see a productivity revolution now, overnight. And I think that’s wildly optimistic. There are too many practical issues to work through first that I think are being badly underestimated.
Right, but somebody could write it down on a piece of paper and the computer’s camera would know what it is. I suspect that young people won’t know how to type in a decade.
I’d bet heavily against that. Any number of reasons, but including:
*i can probably type faster than I can write, and I’m NOT an especially fast typer - extremely fast for a “2 fingers and 1 thumb” typer, but not properly trained. I can almost certainly type faster than I can hand-write-and-scan, though. I could see that being a little more common for scanning notes into a non-image electronic format, but I think that’s ultimately a stopgap because…
*the general arc of technology has been moving away from pen-on-paper for a long time now. Boomers love to bitch about how millennials no longer remember how to write in cursive. Is AI going to change that? As it stands it’s becoming far more common to see people coming into meetings with laptops and taking notes on them, than on paper.
*all the data privacy issues still stand with using a third party provider to analyze handwriting and turn it into digital text, that may include confidential client data. Not a dealbreaker… but it will slow progress, at a minimum.
We’re at a point in the AI cycle where we’re all really hung up on the sci-fi like potential for what AI COULD do. But, we’re not really thinking much about where it would actually add value. And I think this is a pretty clear example of where it doens’t meet the latter test - why would I write out an email by hand and scan it, when I already can type it at least as fast, and in a fully AI-integrated world with adequate controls to maintain the privacy of client data I could speak or type into a dialogue box, “please write an email to Bob about account number 67821 of three to five paragraphs politely but with a slight hint of a threat telling him that we’re prepared to negotiate down 3.5% but at the 8% he’s asking for he can go elsewhere, reference the data in fee schedule #16, as well as preliminary 3Q revenue, and make some sort of a reference to his son’s t-ball game win last night, look up the name because I don’t remember it” and then edit whatever the computer spits out? I think that’s more aligned with where business has been going and where AI could potentially be used as a way to bridge multiple systems - CRM, billing, accounting, etc. It’s just I think we’re a long time from being there, and I’d still expect a lot of human attention being needed to the output from that sort of a hypothetical situation before you could hit send.
But, I’d bet you a beer that in 2034 kids still know how to type. Easy. If nothing else, in an industry where billions of dollars are at stake, that’s a LOT of liability to leave in the hands of a computer system being able to recognize letters in handwriting.
Fleshing out the above a little more fully… I think there’s only so much detail I cann go into here for obvious reasons, but where I see potential for AI in my corner of finance would be things like this:
*“boilerplate” content generation, give an LLM a series of objectives and have it spit out a rough draft for a letter or white paper
*first-pass research tool - load a 10K or 10Q from a company and have the tool scan it for anything related to a number of indicated themes. Basically, a smarter search or “red-line” tool.
*potentially allowing more sophisticated solutions for things like trade allocation tools - answering questions like, “we have 10mm bonds to place and 467 accounts could use a combined 32.5mm. Which accounts should get bonds, based on these criteria?” We alreay have tools that do this - I’d think AI could potentially offer the ability to make them, if not better, then easier to create.
Here though I think - aside for spitting out boilerplate - we’re mostly looking at incrementally more efficient versions of existing processes. So far, as far as actual investment management, most of what we’ve seen is “AI-washing,” firms claiming their investment process uses AI when the actual claims/effect is fairly dubious, and the SEC is already issuing findings against firms for this. It hasn’t really changed the landscape, and most of the applications I’ve seen really are not THAT impressive, compared to what a reasonably intelligent and experienced human could do.
Sorry, I meant to say that I expect speech to be the primary interface, but if there is (say) a Social Security number, perhaps they might write that down if they don’t want others to hear, but only that. The cameras will always be on, so there would be no explicit scanning.
I’m going to guess you don’t have experience with software development on a global project involving multple appication layers, technologies, discrete codebases and teams, because you don’t seem to have a clear how complex this stuff can be.
I cannot see that: even if the models were a few decades ahead of where they are now, they’ll be far more expensive and high-maintenence than people. You’ll have to train them everytime requirements change (quarterly at best, more likely on a weekly/monthly basis), and there’s the matter of data storage: to bypass data liability concerns (esp in the EU) they’re going to have to adopt on-prem storage, which is massively expensive. Data is like toxic waste in that you don’t want to store it near you unless you have no other choice. Why pay to store the models for something like the “perfect” code review skills (in addtion to all the other skills they’ll need) when storage in your human devs do it basically for free?
I’ve used my experience in a massive Fortune 500 company as an example of the problem with AI and scale, but it presents unique problems at other levels of shops as well. The insane pace of startups means a machine trying to keep up with any human insane or driven to work in that environment is not viable.
At the end of the day, ilke other tech crazes such as cryptocurrency, AI is largely a bill of goods for all but the most straightforward tasks. It’s an automation tool, and like other automation tools it either currently does or will be a compelling alternative to humans doing things too boring or granular for them to excel at (like the aforementioned assistance to radiologists in interpreting film results), but will always fall behind human intuition and flexibility for more complex or creative workflows, to say nothing of the fact the people can address unprecendented scenarios in the workplace without needing a prompt of some kind, which is a near daily occurence in high-thoroughput environments.
For anyone who didn’t read the Cory Doctorow article posted by @mscaveney, Doctorow makes a really crucial point:
In high stakes fields (e.g. radiology), AI still makes egregious mistakes and needs supervision from human experts to provide correction when it does. The human labor still needs to be done, but the AI provides something for the expert to “compare notes” with. AI as it stands now can potentially help the expert discover paths for investigation the human expert may have otherwise missed, but can’t safely replace the human expert. Doctorow argues the continuing need for expert supervision means in that in high stakes fields, AI should not be viewed as a way to cut costs by reducing human labor, but as a way to improve performance while paying a new cost in addition to the human labor cost. Doctorow argues a major problem with the current AI goldrush is that it’s being pitched as reducing labor costs, but Doctorow argues that’s not true for the most high stakes (and most profitable) use cases.
I cannot even “see” the frontier model of ten months in the future, not to mention ten years in the future. It could be that progress will stall and programmers will still be needed, but if the sustained rate of change is anything like we’ve seen in the last few years, programmers will surely be unnecessary (except perhaps for the most elite and gifted).
The more complex/distributed the project, the better suited it will be for machines to read it and then just fix or rewrite it as they see fit, where this will likely all be done in hours or days. Humans will not have a chance, I fear.
That is certainly right in 2024. But it’s not clear that this will be the case in even a few years from now. I’m scared of the upcoming frontier models over the coming years.
Interesting that you haven’t offered up what your experience is with technology: I suspect you’ve bought into the kool-aid of what is basically a very complicated if/else loop, but don’t have any actual experience in professional software development,and aren’t really interested in hearing other perspectives apart from what confimation bias has cooked in for you here. So, have fun being afraid, I guess???
Well… it’s hard to say. While I live in the Silicon Valley and have worked for several FAANG companies and done AI-related activities over the last few years, I don’t think that my experience really qualifies me to predict what the LLMs will be doing, particularly as they always catch me by surprise. (I also suspect that new and better architectures will emerge.) So, overall, I generally disqualify myself from making predictions. Now, that said, if the frontier model progress continues (the current frontier models seem to be a bit under 2T parameters?), things will get really scary for programmers almost immediately! So, will this progress continue? That’s the question… so eventually (I’m not sure when) there will be the first artificial general intelligence (and if it’s cheap enough to run, watch out), and perhaps enough of those will make the first artificial super-intelligence, and who knows what the hell that will do. So am I scared? Yah, I’d say so! But is this inevitable? I’d say yes, but the question is time… how many years? I have no idea.
I have done some different things from embedded systems to desktop apps but most of my career and what I’m doing now is e-commerce web apps. I work for a small company with a niche integrating a ticketing system (Tessitura, a competitor of Ticket Master) into the sites and systems of performing arts orgs.
Yeah I didn’t spend tons of time in that space lol! Just when I was getting comfortable I got moved to another project. Web is where I am most comfortable and I do need to spend some time mitigating/future proofing…or new career searching
I am good on that front. Everywhere I’ve worked I’ve always been a key player and even though my company is tiny I am the top developer there. Still, we can never rest on our laurels. It’s always been a challenging space to stay relevant in.
I’ll second this, and say that this has been my experience in finance. And, a thought experiment I’ve suggested to a lot of other folks elsewhere about AI, is more likely than not you’re an expert with fairly cutting-edge knowledge of at least one specialized field. Go out and ask AI to solve a problem or make a plan you could use in that field, and dispassionately look at the output.
I won’t even go finance - I’m a pretty serious cyclist, and while I’m not myself a professional coach I self-coach and have done a fair amount of research into building training plans, maybe not enough to call myself an expert (the devil is in the details) but certainly enough to evaluate something at a “gut check” level, - “is this plausibly a right answer?”
Asking generative AI to build a training plan to raise my functional threshold power, FTP, was sort of an eye-opening experience in this respect. It absolutely generated a multi-week training plan, and at least did show some progression within workouts. An untrained cyclist who started on that plan would likely be fitter at the end than they were now.
But, a trained cyclist, looking to raise FTP (basically, the power threshold at which you exceed your aerobic capacity and start to accumulate fatigue quickly rather than slowly), it wouldn’t have worked. right off the bat, it called for a lot of time in the gym lifting weights and while they were at least all leg exercises, that’s the kind of stuff you work when you’re trying to build your maximal power - training your sprint, not your aerobic base. It was just a general all-around “do this if you’re on a bike” plan that wasn’t very likely at all to produce the sort of specific training adaptations you need to increase your aerobic threshold, and I probably could have myself come up with at least three or four different approaches that were likely to work better.
I think there’s an easy way to explain this, and a more nuanced way to explain this.
The easy way - generative AI is still not THAT smart. you feed a ton of data in, and it doesn’t necessarily “understand” that data so much as get really good at returning the data you need if you aska question. For relatively simple or mainstream questions, this is enough - one of my favorite LLM AI examples I’ve seen yet was my wife asked it to write her a christmas card in the voice of a millenial, and it came back with something like “Seasons yeetings, fam!” which I think is fucking hilarious. But, as the complexity of the question increases, it becomes a lot harder to produce truly useful responses. And this is where it really starts to fall down, and can’t measure up to an expert in an area of specializaiton.
And the more sophisticated critique - it’s important to never lose sight of the fact that garbage in is garbage out. ChatGPT is trained on a fixed “slice” of the internet, so it can’t start creating feedback issues where its own output starts to become input. This works for things that don’t change very much. But, for all the obvious doping jokes this will probably entail, our understanding of how training can acheive specific goals has actually evolved quite a lot even in the last five or six years, and training protocols today have changed significantly from the 2010s, with pretty staggering results in the peloton (Jonas Vingegaard’s Stage 16 time trial, which I had the good fortune to see in the flesh last year, had him putting down a reported average 7.4 watts per kilogram for about 32 minutes, which I’m a well-trained amateur struggling to break through 4, and the average untrained male is probably 1.5-2w/kg, so those are otherworldly numbers). Basically, ChatGPT isn’t going to be able to produce a training plan that, even if I was still in my early 20s, and my whole job was training to ride as fast as possible, that could take me from 4w/kg to 7w/kg, because all the data it has to pull from is from an era we now consider outdated (in 2009 I wouldn’t have questioned the plan at all), and knowledge has moved on. Beyond that, by the time you get to this level, you’re kind of exceeding the limits of current knowledge, and figuring it out on real time - Vingegaard’s coaches are responding to incoming data as much as they are taking a plan and adjusting on the fly based on how he’s responding to it.
But, I’m probably getting too far into the weeds here - the main point is, if you take something that you have a high level of expertise in, and ask AI to do it, you’ll be able to assess for yourself how well it can produce expert-quality work. and, at least today, it can’t. I think Frylock’s example is far more realistic - it can do a lot of the heavy lifting for an expert, and then allow that expert to provide the truly differentiated cutting edge knowledge, improving their productivity (and in turn value) rather than eliminating them.
One of the other big issues with using public discourse as training data is there are so many factual errors in it. In particular, if a misconception is popular enough, ChatGPT will treat it as a fact. It is a literal manifestation of the concept “tyranny of the ignorant majority”. It is the very antithesis of expert knowledge. Look at the way expert answers get downvoted on places like reddit. This system of “truth measured through popularity” is a profound problem for AI training to overcome. The current workaround is for organizations to train models themselves on data that they trust, but that’s a far cry from the more ambitious claims of what AI can become. And on a broader scale, trying to train a general AI in a way that gives more weight to the opinions of experts is a more difficult problem than it sounds on the surface: resolving disagreements among experts, measuring the reputations of experts, dealing with fields where the recognition of expertise is not widely formalized.
Indeed, it doesn’t even know the difference between true and false, it is autoregressive and simply picks the next most likely token given the tokens in its context window. So in many ways it is astonishing that they work as well as they do.
Oh, that is an outstanding point I had totally forgotten about. Wasn’t there some AI generated Seinfeld stream that got shut down because it got transphobic real quick? Still, it’s a really salient point that “more information !== good information”. And I don’t think you can train nuance, which is something humans will excel at over machines
