> Although I was on an AI team, I often pushed back against applying AI unless we had a really compelling reason. What is the user problem we are trying to solve? Do we really need an LLM or could a few if statements suffice? Are we sure that natural language is the appropriate interface for this?
This practical approach to AI feels refreshing in a field drowning in buzzwords. I’ve built tools where simple regression models outperformed neural networks, and convincing teams was an uphill battle. It's hard to not get pushback from teams for not going all-in on AI when it seems decisions and budgets are hype-driven.
I have pushed back in a similar way many times (wrt LLMs), the response I typically get is some combination of:
- A custom/heuristic driven approach would perform better but will take much longer to build so would be lower ROI.
- There is a strategic angle to using AI here (building competency). We aren't sure what new use cases will open up in the medium term and we need to be fluent with building AI products.
- There is a perceptual/marketing angle to using AI here. We need to convince the market/investors we are on the bleeding edge (hype).
3 is widely mocked but is a completely rational allocation of resources when you need to compete in a market for funding.
Prolog is AI, so whenever I see such a problem, I use miniKanren, implementing Relational/Logic programming in a light-weight way. Bleeding edge AI it is.
> will take much longer to build so would be lower ROI
This one is funny because my experience has been that ekeing out the issues in this sort of thing is enormously complicated and unreliable and takes an inordinate amount of time. Often the 'bugs' aren't trivially fixable. One we had was the LLM formatting URIs given in the prompt wrongly meaning they're no longer valid. Most of the time it works fine, but sometimes it doesn't, and it's not reproducible easily.
It's true, it can be maddening (impossible?) to chase down all the edge-case failures LLMs produce. But outside of life/death applications with extreme accuracy requirements (eg: medical diagnostics) the attitude I've seen is: who cares? A lot of users "get" AI now and don't really expect it to be 100% reliable. They're satisfied with a 95% solution, especially if it was deployed quickly and produces something they can iterate on for the last 5%.
- Assume LLMs will be more intelligent and cheaper, and the cost of switching to a new LLM model is non-existent. How does improving the custom/heuristic compare in that future?
That's kind of what I was getting at in point 2, about "new use cases" opening up, but yeah you stated it more directly. It's hard to argue with. With a heuristic driven approach we know we will need expertise, dev hours, etc to improve the feature. With LLMs, well, some lab out there is basically doing all the hard work for us, all we need to do is sit back and wait for a year or two and then change one line of code, model="gpt-4o" to model="gpt-5o" or whatever.
for 1/2, surprised to hear this because debugging models is usually a total black box and practically impossible. for 2, it's a similar problem where getting performance and accuracy using the same model over and over again on different problem sets can be challenging. not an AI expert or anything this has been my experience on the product side.
Responded to the same sentiment elsewhere but my general sense is that for many use cases users simply do not care about high 9s accuracy/consistency. A 95% solution using AI is "good enough" if you can ship it quickly and give them the tools to iterate on that last 5%.
95% solution might work for small startup X or small biz y but at large company scale 5% is a huge deviation to correct on. Maybe just depends on the client and how touchy they are. At my company, we measure metrics in bps and moving something 50 bps is a huge win. 500 bps would be unheard of.
IMO it's less about the size of the company and moreso the nature of the integration. Users are more forgiving of 95% accuracy when it's used to enhance/complement an existing (manual?) workflow than when it's used to wholesale replace it. The comparison would be building an AI tool to make data entry easier/faster for a human employee (making them say, 2x as productive even at 95%) versus an AI tool that bills itself as a full replacement for hiring a data entry function at all (requiring human or superhuman accuracy, edge case handling, maddening LLM debugging, etc).
In the long run the latter is of course more valuable and has a larger market, so it's understandable large corps would try to "shoot for the moon" and unlock that value, but for now the former is far far more practical. It's just a more natural way for the tech to get integrated and come to market, in most large corp settings per-head productivity is already a measurable and well understood metric. "Hands off" LLM workflows are totally new and are a much less certain value proposition, there will be some hesitation at adoption until solutions are proven and mature.
I have a friend who was tasked by a think tank to build an AI model for which they'd managed to get an enormous amount of funding. My friend pointed out that it would be much more effective if they used a regression model that could be built in a couple of days by a junior developer. Eventually he was fired from the project for sticking to his guns and refusing to pad his time while building something worse just so they could say they'd used "AI".
For years I worked at a company doing manufacturing automation with a generative CAD component. People would constantly throw out, "use ML" "use AI", everything was just regressions and open cascade... and those people never understood that it worked just fine without buzzwords.
This is a perfect lesson in why strong communication skills are important in engineering organizations. It's leaderships responsibility to learn from engineering what is technically feasible but its also engineering's responsibility to communicate well enough to convince their organization on the right path forward.
I've found having a cascade of things helps. Trying to split things into "decision at this later, or pass on to the next" with increasingly complicated models/approaches.
Start with ifs, then svms then something else for example.
This has some technical benefits, like speed, and gives you a place to put important hard coded fixes for where a better model makes a small but key mistake. But the bigger benefit imo is getting something to solve the bulk of the problem quicker, and a organisationally it means not saying no to an approach - just where it fits and at what level of improvement it's worth it.
It's important to consider if existing tech can do something as well if not better.
LLMs can have have great application where existing tech can't reach.
Too often, seeing LLMs doing something that's done better already by an existing tech or something it's not designed for seems to miss the impact being sought.
I have had a somewhat similar journey but 14 years instead of 25 and I always wonder how it would be different today.
We were lucky enough to grow up with the industry and progressively learn more complexity. The kids out of school today are faced with decades worth of complexity on day one on the job.
Physicists have centuries to catch up on just to get started. I think they will survive. The main issue today is more the saturation of useless information in my opinion. There’s little time for your own thoughts as too much time is spent sorting the thoughts others want you to think.
This is true for every field. Everyone has had to step into a field that was built upon the hard-won experience of others and had to get up to speed, and the easiest way to do so is to recognize that fact and take advantage of the wisdom of those who came before.
Your comment made me curious so I looked at his posts and he has a one about leaving academia because he wasn't happy in 2022, and a more recent one about rejoining it some months ago.
I didn't see the more recent post, so thanks for the link.
I should say that I'm still in grad school (nearing the end), so the decision hasn't been made yet. The direction I'm thinking is away from academia.
I love the academic environment, access to university resources and close proximity to lots of domain experts. However my experience as of late has been pretty isolating, as my group is almost fully remote despite nearly everyone living in the same town making motivation difficult some times. I also sometimes miss exercising my practical engineering skills, as my current work is entirely analytical/simulation. Overall its been less rewarding than I had hoped.
It makes such a refreshing change to have a web page not cluttered with adverts and popups. Just nice, clean, well-spaced-out text and simple organisation.
For a lot of people, AI is a fun journey where they create things that are amazing. And I agree, the results are quite amazing. But it's also a sad thing that the world works this way because scientists like this never think of the larger social consequences of their work. They are insulated and elevated to lofty social positions while their creations fundamentally alter the social fabric. AI is one of those things that is quite dangerous and the fact that large corporations attract people by glorifying their intellect is a recipe for disaster.
I'm not sure what this means exactly, because AI is a wide field that covers so much and angers a great many people for many different reasons.
But IMO it's pointless to hope for something else. AI at its core turns out to be pretty simple. No matter what the best intentioned scientist did, somebody else would think differently.
For example, Stable Diffusion was originally released with a filter that refused to generate porn. There's your scientist thinking of social consequences. But does anyone even still remember that was a thing? Because I'm pretty sure every SD UI in existence at this point has it disabled by default.
> For example, Stable Diffusion was originally released with a filter that refused to generate porn. There's your scientist thinking of social consequences
We would have to know their internal conversations to know whether that filter was being driven by scientific concern over social consequences or any number of business goals/concerns. We can't assume the reasoning behind it when we only have the end result.
That doesn't matter for the point I'm making: which is that this attempt (and any other) are trivially nullified by those that come next. The SD devs couldn't have created a state of affairs in which AI never ever would generate porn.
And transformer architecture is too low level to care about things like that, there was no way for the people who made the guts of the modern AI systems to make it so that they only can ever make cute fluffy kittens or give accurate advice.
So what I'm saying is that there's no timeline in which socially conscious scientists would have succeeded in ensuring the the current gen AI landscape with its porn, deepfakes and propaganda didn't come to exist.
> No matter what the best intentioned scientist did, somebody else would think differently.
This is exactly an argument that supports technological determinism. We simply can't decide -- we have no ability for oversight to stop technology from evolving. That's precisely why I think AI is so dangerous.
IMO, the dangers of AI are mostly overrated. AI is just a new fancy way of generating pictures and text. It does those things better in some regards, but the danger is the same we already had.
We're only just starting to get to the point that AI, if unconstrained, is capable enough to be dangerous. The danger is getting to the point where the not-so-bright malevolent actor can tap into AI to get detailed instructions to do something highly destructive, or have it do it on their behalf (e.g. hack into some system), that they wouldn't previously have been able to figure out just by Googling for information and trying to piece it together themself.
Of course not all malevolent actors are dimwits, but there are also many things that even a highly intelligent individual couldn't do on their own, such as a Stuxnet level attack, that AI will eventually (how soon?) be able to facilitate.
This is actually where the AI concern arguments seem to get misunderstood in my opinion.
I've never heard anyone raise serious concerns over fancier ML and generative algorithms - maybe concerns over job loss but I don't think that's what you had in mind (correct me if I'm wrong).
The more serious concerns I hear are related to actual artificial intelligence, something much smarter than humans acting on a time scale drastically different than humans.
I'm not vouching for those concerns here, but I would say its more fair to keep them in the context of AI rather than ML, LLMs, and generative tools.
It's also a way of mechanizing even further large amounts of human labor and reducing the importance of art. I guess it depends on what you value: for you, apparently a world with AI is not so bad. For me, it's disgusting.
I honestly don't see it fundamentally different from most other code. I generated images and music (PLAY instruction) with GWBASIC back when I was a teenager. I generated text with Perl.
This is just the continuation of the same old, just a bit fancier.
I don't think it is. One could say that getting hit by a car is the continuation of getting hit by a person, but one is much more powerful than another. AI allows mass creation of much more complicated works at a speed much greater than before. the PLAY instruction might create some music, but it won't be the sort of music that can compete with human-made music. AI music is very close to it.
Speed is important, strength is important. There is no obvious qualitative difference, but qualitative differences emerge due to a massive increase in complexity, just like consciousness emerges in us but (probably) not in a bacteria due to the massive difference in complexity, even though we are just a scaling of the former.
Your text generation with Perl wouldn't be able to write an article, but ChatGPT can, and the magnitude difference is precisely what we cannot handle, just like I can't be hit by a speeding car at 100km/h and survive but I'd probably walk away from being hit at 2km/h (and once this actually happened to me, without injury). Would you say there's not much difference between the two?
This practical approach to AI feels refreshing in a field drowning in buzzwords. I’ve built tools where simple regression models outperformed neural networks, and convincing teams was an uphill battle. It's hard to not get pushback from teams for not going all-in on AI when it seems decisions and budgets are hype-driven.
- A custom/heuristic driven approach would perform better but will take much longer to build so would be lower ROI.
- There is a strategic angle to using AI here (building competency). We aren't sure what new use cases will open up in the medium term and we need to be fluent with building AI products.
- There is a perceptual/marketing angle to using AI here. We need to convince the market/investors we are on the bleeding edge (hype).
3 is widely mocked but is a completely rational allocation of resources when you need to compete in a market for funding.
This one is funny because my experience has been that ekeing out the issues in this sort of thing is enormously complicated and unreliable and takes an inordinate amount of time. Often the 'bugs' aren't trivially fixable. One we had was the LLM formatting URIs given in the prompt wrongly meaning they're no longer valid. Most of the time it works fine, but sometimes it doesn't, and it's not reproducible easily.
- Assume LLMs will be more intelligent and cheaper, and the cost of switching to a new LLM model is non-existent. How does improving the custom/heuristic compare in that future?
In the long run the latter is of course more valuable and has a larger market, so it's understandable large corps would try to "shoot for the moon" and unlock that value, but for now the former is far far more practical. It's just a more natural way for the tech to get integrated and come to market, in most large corp settings per-head productivity is already a measurable and well understood metric. "Hands off" LLM workflows are totally new and are a much less certain value proposition, there will be some hesitation at adoption until solutions are proven and mature.
Start with ifs, then svms then something else for example.
This has some technical benefits, like speed, and gives you a place to put important hard coded fixes for where a better model makes a small but key mistake. But the bigger benefit imo is getting something to solve the bulk of the problem quicker, and a organisationally it means not saying no to an approach - just where it fits and at what level of improvement it's worth it.
LLMs can have have great application where existing tech can't reach.
Too often, seeing LLMs doing something that's done better already by an existing tech or something it's not designed for seems to miss the impact being sought.
We were lucky enough to grow up with the industry and progressively learn more complexity. The kids out of school today are faced with decades worth of complexity on day one on the job.
Your comment made me curious so I looked at his posts and he has a one about leaving academia because he wasn't happy in 2022, and a more recent one about rejoining it some months ago.
https://austinhenley.com/blog/leavingacademia.html
https://austinhenley.com/blog/rejoiningacademia.html
I should say that I'm still in grad school (nearing the end), so the decision hasn't been made yet. The direction I'm thinking is away from academia.
I love the academic environment, access to university resources and close proximity to lots of domain experts. However my experience as of late has been pretty isolating, as my group is almost fully remote despite nearly everyone living in the same town making motivation difficult some times. I also sometimes miss exercising my practical engineering skills, as my current work is entirely analytical/simulation. Overall its been less rewarding than I had hoped.
But IMO it's pointless to hope for something else. AI at its core turns out to be pretty simple. No matter what the best intentioned scientist did, somebody else would think differently.
For example, Stable Diffusion was originally released with a filter that refused to generate porn. There's your scientist thinking of social consequences. But does anyone even still remember that was a thing? Because I'm pretty sure every SD UI in existence at this point has it disabled by default.
We would have to know their internal conversations to know whether that filter was being driven by scientific concern over social consequences or any number of business goals/concerns. We can't assume the reasoning behind it when we only have the end result.
And transformer architecture is too low level to care about things like that, there was no way for the people who made the guts of the modern AI systems to make it so that they only can ever make cute fluffy kittens or give accurate advice.
So what I'm saying is that there's no timeline in which socially conscious scientists would have succeeded in ensuring the the current gen AI landscape with its porn, deepfakes and propaganda didn't come to exist.
This is exactly an argument that supports technological determinism. We simply can't decide -- we have no ability for oversight to stop technology from evolving. That's precisely why I think AI is so dangerous.
Of course not all malevolent actors are dimwits, but there are also many things that even a highly intelligent individual couldn't do on their own, such as a Stuxnet level attack, that AI will eventually (how soon?) be able to facilitate.
I've never heard anyone raise serious concerns over fancier ML and generative algorithms - maybe concerns over job loss but I don't think that's what you had in mind (correct me if I'm wrong).
The more serious concerns I hear are related to actual artificial intelligence, something much smarter than humans acting on a time scale drastically different than humans.
I'm not vouching for those concerns here, but I would say its more fair to keep them in the context of AI rather than ML, LLMs, and generative tools.
This is just the continuation of the same old, just a bit fancier.
Speed is important, strength is important. There is no obvious qualitative difference, but qualitative differences emerge due to a massive increase in complexity, just like consciousness emerges in us but (probably) not in a bacteria due to the massive difference in complexity, even though we are just a scaling of the former.
Your text generation with Perl wouldn't be able to write an article, but ChatGPT can, and the magnitude difference is precisely what we cannot handle, just like I can't be hit by a speeding car at 100km/h and survive but I'd probably walk away from being hit at 2km/h (and once this actually happened to me, without injury). Would you say there's not much difference between the two?