if you built your saas business before ai, your entire business was designed around one assumption which was that software has ~zero marginal cost.
ai blows that up. every ai action costs money. incumbents now have to create more expensive tiers or introduce usage based pricing. both force customers to make a new purchasing decisions while revenue doesn’t automatically increase just cuz the product became more expensive to operate. so your cost per seat rises, your gross margins compress, & adoption remains uncertain.
meanwhile, the model labs are subsidizing ai usage like crazy, so they can undercut you. ai native startups are burning venture capital to acquire users, so they can undercut you too.
(lack of capitalization in original)
I really don’t know what to think of AI for coding. I’m no longer a coder, haven’t been for almost thirty years now, so I don’t have enough firsthand knowledge, though I did have an AI generate basic sorting and random number generation code just to get my hands a little wet. Among devs I trust, opinion is split. Some love it, some hate it, some are on the margins. Here’s one in the hate category:
Are companies actually seeing massive productivity gains from their AI adoption?…
… All of the AI projects we have observed as a team are failing. Every single one – we have seen 0% success in a year and a half, not only amongst projects we have been asked to participate in2, but even within projects that we have observed in passing while doing totally unrelated work. Even if you grant that AI tooling accelerates specific workloads, the method and scale of the current investments is senseless. Frequently the failure is not related to AI itself, but rather that companies are terminally bad at running software projects effectively, and as I have remarked previously, AI projects are subject to all the failure modes of normal projects plus you can get everything right and then still fail because of the method’s novelty. Very few companies are so good at shipping software that they can afford the extra risk profile…
…
our team has quickly learned while on an engagement not to ask anything about ongoing AI projects in any context – by the time that project has started, it is too late for the management team, and intervention is not possible until a crisis point is inevitably reached. There is no conceivable positive outcome. The failure rate is so high that even basic inquiry leaves us in an untenable position. Any coherent question about how it’s going, what the goal is, who is using it, constitutes an inadvertent attack on the chain of command responsible for the work because there are no good answers to anything. Even in rare cases where my interlocutor has stated that things are going well (usually while the project is still mid-flight and failure has not had a chance to manifest), it is generally obvious that they are doomed, but at least in these cases I can simply agree and then go home to scream into a pillow for six hours straight3.
All of this is to say that I am very confident that almost every report at a company about “massive AI productivity gains” is untrue as a matter of brute fact. Even if some companies are seeing clear gains, this is the exception, not the norm.
It’s a long post and the entire thing is worth reading. This is a high end engineering consultant who gets a view on a lot of companies.
My feeling/guess is that AI is not mature, that the use cases exist, but that it introduces significant risks of new failure modes because AIs aren’t reliable and aren’t actually intelligent: they don’t know what they’re doing. If you don’t do the work yourself or audit it to the level where you might as well have done it yourself, it seems easy for errors to creep in, for hard to maintain or understand code to be created, and for new failure modes to exist.
If I were an executive in most businesses I wouldn’t be using AI for much, if anything yet, though I’d let a few people have an account and test it, and would do so myself. It’s moving fast, it clearly adds risks and mistakes that we don’t know how to mitigate yet, it’s expensive and that’s with massive subsidies, so becoming dependent on it when prices are likely to increase significantly is unwise.
There are going to be exceptions: if you’re Google, not sure you have much choice. But if you manufacture widgets or sell hamburgers or build homes or are a lawyer (who can’t afford to have the AI hallucinate cites) I’d give it a pass for now, or use it very warily in exploration mode. I use it for research myself, especially as search engines get worse and worse, but I also check the sources it uses.
And, as we’ve discussed before, non open models have huge risks, since prices can easily be raised, the government can intervene to cripple the model or deny to other nations, and the company itself can make changes to the model which make it less useful to you. So even if using AI I’d be primarily focusing on Open models (which means mostly Chinese.)
But my best guess is that this is a real tech, with real uses, which is not yet near to mature and which is being deployed before people understand what it’s good at or how to use it safely and cost effectively, or, more importantly, when NOT to use it. I’d also guess that at least so far, it’s not the second coming, the next great thing, in the way that its evangelists preach.
So far it seems to increase failure rates on real projects and raise costs significantly at the same time. Unless I’m running a business where I must be in it, I think I’d take it slow.
There are very few businesses where “the way we did it in 2022” will expose you to significant risk and costs. If you’re in one of those businesses, chill a little and observe. Let other people pay the price and make the mistakes. If you need to, because it’s the current “everyone must do” just lie and say you’re into it, while doing the minimum.
This is certainly something I could be wrong about. I’m confident that Chinese AI will win, for example, I’m less certain about how useful AI will turn out to be. I do think that I’m almost certainly right that most companies adopting it in a big way, NOW, are making a mistake. Wait. See how it turns out. Let other people figure out what it’s good for, how to manage it, how to mitigate the risks and how to drive costs down.
And remember, even if it’s the true next big thing, early rushers often get badly burned, as with the dot-com bubble. Don’t buy hype, make sure what you’re investing in is real and you understand what it’s good for and what it sucks at.
What I write here is for the benefit of everyone, but alas, I live in capitalism and I, and the site, take money to keep running. If you value the writing here and can, please subscribe or donate.
Feral Finster
Is it not written that the early bird may get the worm, but the second mouse gets the cheese?
Purple Library Guy
I think Google are foolish to be going into AI the way they are. So the value proposition for Google is, when anyone does a search, they just look at the “AI summary” instead of following a link to some other website. As a result, in theory Google can charge more for ads right on the search page and thus make big bank.
But there’s a basic problem here: Google owns most of the ad space on the web. All the ads in all those places people are no longer going, become worth LESS. So it’s mostly a wash–you gain revenue on Google’s own pages, you lose revenues on everywhere else. Meanwhile, each and every AI summary costs money. And people’s rate of genuinely buying something as a result of a Google search is surely not that high, so you’re running a lot of compute for every time someone spends actual money, and for it to be worth people buying the advertising they have to be making more (net) from people buying things than they spend on the ads. I don’t see how Google can be making net money on this. So that makes it tactically stupid.
And that’s before Google influencing people to never go to the actual web any more starts causing the web to die, meaning there’s less web around to buy ads or provide information to scrape for “AI summary”. That makes it strategically stupid. It’s at the level of “This goose lays golden eggs! If we kill it, we should be able to loot a couple of un-laid eggs from its corpse!”
mago
In what universe is ai beneficial to human and environmental wellbeing? Just asking. Not expecting an answer, especially from Hal.
spud
there is nothing intelligent about AI. its simply a search engine that tries to assemble what you want, from other peoples work.
your correct Ian, keen say the same thing, the chinese will get it right, and american capitalism will once again, be exposed for the scams.
https://www.youtube.com/watch?v=GjBJLdQ7uRI
YOU WILL PAY For Inevitable AI Bailout
more,
https://www.youtube.com/watch?v=WYfIioClpjo&t=136s
China SHUTS DOWN Billionaire Wealth Funnel
Krystal and Saagar discuss China cracking down on financial billionaires.
both sides do it
Great post.
I’m a software / data (not AI) consultant at a very small shop (me and a buddy) who’s been doing it for a long time.
Everything here rings true.
Here’s a slightly different way of re-phrasing and expanding Ian’s points, would love to know how this lands with folks:
At an abstract level, the AI that generates value in a business context is not generating words (email, powerpoint slides, etc.). This operates at the “bullshit jobs” level, where the only mark of effectiveness is how well it complies with arbitrary social convention. Who gives a shit. Effectiveness there is not going to generate any economic value because those tasks are not used to generate economic value, they are used as social signaling for how well someone complies to a given culture in a given role.
Ignoring graphics / audio / etc. generation for a moment, the AI that matters in a business context is two things:
– extremely cheap custom software on demand
– computed mapping, where costs vary widely by task and amount of computation to be effective
That last bit doesn’t sound like a big deal but it’s enormous.
Taking arbitrary inputs — chunks of text, images, sounds — and effectively mapping them into arbitrary buckets based on context — pertinent / not pertinent, effective / not effective — is a non-negligible portion of cognitive work every day.
A machine that can do that is almost like a steam engine of cognitive work.
AI is that machine.
BUT.
It can only effectively do those tasks — both the quick custom software generation and the arbitrary mapping — when its been properly set up, given the proper inputs, and the outputs are put through an effective and iterative QA process.
Because of this, AI functions as essentially an acid test.
AI is an acid that burns through institutional cruft to expose the effectiveness of workflows and, more subtly, incentive structures within / between departments.
If workflows are solid and incentive structures are aligned, AI tools can reliably be set up correctly, given proper inputs, and put through an effective iterative QA cycle to provide value at effective cost *for almost any cognitive job*.
Almost no firms have solid workflows and aligned incentive structures.
Executives want a shiny thing built they can point to having championed regardless of how effective it is. Managers are out of touch with what the employees they manage are actually doing. Employees aren’t properly incentivized or organized to either figure out problems or communicate solutions on their own.
And even if all those happen to not hold: workflows still need to be intelligently mapped out for an AI to help improve them. A bad process is not going to be helped by jamming AI into it.
Any of those failure states above are a recipe for “we spent millions buying everyone a Copilot subscription, have at it” and six months later the only thing it’s being used for is to generate pictures of cats in barely amusing outfits.
Or, software teams going over budget by millions of dollars because the steam engine for cognitive work is being used to power the same process that previously just went around in circles. Only now the traversal of the circles is faster.
A business that is a business and knows what it’s doing can adopt this tool widely and to great effect.
A business that doesn’t know what it’s doing, or a business that is a shell for rent-seeking execs paying game-playing managers administering disempowered employees running a gate-keeping political economic function, is going to waste a fortune on this stuff.
But it’d be wasted anyway.
vmsmith
There will always, by definition, be early adopters. I mean, if no one is willing to take the first step, then it never happens. And, you know, that’s how problems get identified and things improve.
Having said that, I think a lot of organizations are wrapped up in FOMO and going about AI adoption in really bad ways. The thing I’m paying particular attention to is the librarian community. Their position is that so many enterprise projects are failing because the enterprises are just throwing AI into the mix without having first developed a good semantic layer to make sure that the AI apps and agents in the various parts of the enterprise all share the same meanings. It’s a fascinating topic, and I can see librarians and MLIS programs getting a real boost as enterprises start racing to develop their controlled vocabularies and taxonomies and thesauri and ontologies.
And I, too, am reminded of the dot com bubble. I have a small little enterprise that rests on AI, and one of our three maxims is, “Turtle beats hare.” I tell everyone just that: let the hares rush by now in their FOMO-induced intoxication, while we plod along and focus on the fundamentals that will matter in the long run. I am happy to let AI1.0 rush by while preparing to take full advantage of AI 2.0.
All that aside, I’ve asked this before, here and elsewhere, and never really gotten an answer. When people say that China will win the AI race, what’s the race, and what does it mean to win it? What’s the finish line that someone or some country is going to cross first?
TM
There are some very useful cases for machine learning; voice recognition and transcription services have gotten very good, no longer have to split focus during meetings to get good, solid notes. Good at sorting through a large volume of logs or command history to surface unusual patterns. Boilerplate code happens instantly, useful for little glue scripts and toy examples in languages I’m not familiar with, as long as I have the logic and priors right. In my previous industry of VFX there have been some really useful tools released– better timewarp, better denoise, the ability to upscale low-resolution frames, easier roto and matte generation. Nice handy little tools, great!
But the level of investment and hype is nowhere near commensurate to the actual outcomes provided by these tools. Generative AI is being treated like it is a quick hop to general AI, when there is no evidence that this is possible. Because of this, Business Idiots (thanks Ed Zitron) are throwing everything they have at the assumption that Fully Automated Luxury Technofeudalism is just around the corner. There are serious externalities that are being completely ignored that will cause a great deal of suffering, at an ecological, social and economic level.
Carborundum
More than anything, I think this is “Maybe early AI managers are stupid?”. The technology looks to me to be pretty promising / effective when used appropriately, but “appropriately” is doing a lot of work in this statement.
I don’t think this is as much the emergence of entirely new modes of engineering failure as it is what happens culturally when the pendulum swings back. I get why managers are latching onto this tech with religious zeal – they’ve been increasingly dependent on ever-larger armies of well compensated technical specialists who do things they don’t entirely understand and are prone to being kind of assholes when shit goes sideways. (I’ve worked with a very non-zero number of devs who would say it wasn’t their fault they picked their noses until their foreheads caved in if it wasn’t clearly set out in the specifications to stop once they started experiencing pain and producing bloody grey matter.) If I were in their shoes, the notion that I could plain language query to generate production code would seem frickin’ awesome – sign me up with a cherry on top.
Personally, the use of AI has changed my work quite a bit since the release of the first capable LLMs. Not so much in the fundamentals as in making the development of new capabilities much, much more viable. It has not particularly resulted in greater productivity in the sense of more absolute product out the door, but it has very significantly changed the nature of both the product and the production processes.
Tangentially, I suspect that what we’re going to see in situations like Canva’s in the linked tweet is that they’ll offer versions of the product with the capability to use the user’s own AI, either local or remote (i.e., they’ll download the cost driver of the AI features to the user).
Olivier
Google has to be active in AI to protect its ad crown jewel. Meanwhile its cloud business is finally picking up, with growth rates exceeding those of AWS and Azure, and that’s mostly due to AI (its custom TPU chips are proving very popular). Thus it is also in the for-rent AI platform business, which is exactly the right strategy given its historic strength as an hyperscaler and vast data center acreage.
Bullweather
@both sides do it
I’m in a job that many (fairly) have called out as being a “bullshit job”.
You’re are correct that the value for this is in the cheap, custom software. I have the AI make small python scripts, or little apps that take away like 75-80% of the friction of my varying monthly/quarterly tasks so I can focus on the meat and potatoes.
Personally I haven’t had the new workflows in place for long enough to determine how much of a time saver they are, but the (very small; i.e. about 2 weeks worth of work) investment to set up the new workflows/apps will be worth it as a sample one way or the other.
That said, this is entirely local on my machine, and not being shipped as a revenue generating product exposed to hundreds, thousands, or millions of customers. That seems irresponsible.
Purple Library Guy
One thing to keep in mind is that the progression from a technology that has early adopters to a technology that is in widespread, long term use is not inevitable. Take for instance hydrofoils–remember hydrofoils? They let boats of decent size go REALLY FAST because they mostly eliminated the drag of having a hull ploughing through the water. Brilliant idea, wave of the future, saw a significant amount of adoption. They were being used for smaller ferries to make for quick trips, and so on and so forth.
See any lately? No, because there’s hardly any left. Turns out there were significant disadvantages. They needed more fuel, they needed more maintenance, they were picky about what kind of water they would work on, which in turn meant they could not handle bad weather, which meant schedules went to hell if there was any. In the end the costs outweighed the benefits and what filled the role of “somewhat faster boat” ended up being the catamaran (something similar happened to hovercraft a few decades earlier). Just because a technology is new, does not mean it is inevitable.
Forecasting Intelligence
Super interesting.
I’ve been using AI more and more personally and at work. Overall conclusion, you do need to be on top of it but the prompts make a big difference. I get AI to write the prompt for more complex project which helps a lot.
It still needs auditing and curating. AI is not a click a button and the final version output pops out.
different clue
. . . ” There are serious externalities that are being completely ignored that will cause a great deal of suffering, at an ecological, social and economic level. ” . . .
Perhaps the BizId FALTs ( Bizniss Idiot Fully Automated Luxury Technofeudalists) are not ignoring these serious externalities. Perhaps they are fostering them on purpose and reveling in them, secure in their wisdom that ” the externalities shall be visited upon the peasants and peons”. Perhaps they are hoping that the externalities they deliberately foster and amplify on purpose will kill many millions of peasants and peons . . . in line with the Long Jackpot Prime Directive.
Brian M
I think Purple Library Guy is basically right. Google is eating its seed corn.
The Google Search and Advertising model is basically a deal. Businesses pay Google to drive traffic to their websites and Google promotes their links accordingly. But what Google is doing now is undercutting that. The Google AI summary provides direct answers without prominently promoting the associated links and companies are seeing huge drops in their traffic. At some point, businesses are going to ask themselves why they are paying Google to NOT drive traffic. I mean, this is not a good deal. Companies (especially news and information sources) will likely start restricting or blocking Google access to their websites as opposed to actively encouraging and paying Google to reference them. As this happens, Search becomes less useful because it has less access to relevant data and AI becomes less useful because it relies on a more limited Search. AI can’t replace Search revenue. End users won’t pay for the summaries that are jammed down their throats and are increasingly useless. Companies need traffic and the AI summary model decreases traffic so they won’t pay either. This seems like a business model designed to crush their cash cow business.