Today’s issue includes a rare bit of good AI news at the end. Feel free to skip to the last segment if you need a weekend pick-me-up.

Last time around I noted that datacenters may serve as genAI's Achilles' Heel. Upbeat marketing and hand-waving have, until now, allowed the providers to blow past any resistance. It truly felt like an unstoppable force … until it tried to set up shop in physical spaces. Now, a combination of debt burden and unhappy residents might slow the beast down. And maybe stop it in its tracks.

The more I think about it, the more I sense that datacenters present an interesting risk/reward tradeoff for the genAI companies. If they get it right, they will be showered with untold riches. But there are so many ways to get it wrong. And it seems they're trying to hit them all.

Forcing their way in

There's a funny thing about selling a product: if it's that good, you'll never have to force it on people.

Sure, you may have to inform them. Or give away some freebies to entice them to try. But that's about it. In fact, if it's good enough, they'll beat a path to your door.

That's the opposite of what's happening with genAI. It may sound like everyone is in love with the technology, but the fanfare is pouring out of a small yet extremely vocal slice of the populace:

• The genAI companies themselves. They need to push this thing in order to generate revenue.

• Some (though not all) software developers. Code-generation tools have given them a productivity boost.

• Company executives. They've fallen under a spell called "you can eliminate most of your org chart."

• Die-hard fans. People who are in love with genAI's future potential, so they forgive all of today's faults.

(No, criminals aren't on that list. Instead of bragging about the technology and pushing it onto others, they're quietly using it to turn profit. Which puts them light-years ahead of most licit businesses. But I digress.)

Beyond that, there's a steep drop-off in satisfaction. It starts with the employees who are forced to adopt genAI at work, thanks to company-wide mandates and leaderboards. Some of these same companies are cutting headcount in order to fund their genAI misadventures even though the technology costs more than the people it replaces. Others are creating AI-based robo-bosses to harass guide employees, or deploying customer service chatbots that dish out incorrect info. Those are close cousins of errant medical transcription bots that patients didn't ask for.

App makers are cramming genAI-based features into their products, turning tools that used to work just fine into unavoidable showcases for needless functionality. Every day, the genAI hype machine claims that chatbots are a suitable substitute for search even though their summaries are wildly unreliable. So-called "smart" glasses (more on that below) have given creeps super powers they sure as hell didn't need and now Apple is adding cameras to AirPods for AI purposes. Writers, artists, and musicians have seen their work get mulched into training data without so much as a please or thank-you. (But they do get the blame when the bots built on their "borrowed" work misbehave.) The use of genAI for wartime targeting has worsened life for non-combatants. And it's currently on-track to drive tools for domestic occupation surveillance.

There will no doubt be some other foolishness by the time this lands in your inbox.

With all that, I'm surprised that "genAI companies invading the physical realm by building massive outposts" was not on my risk bingo card. Then again, past successes have inspired new brazenness. “Forcing datacenters into places where they are unwanted, and cause harm, and drive up electricity prices” is the next logical step in the chain of offenses.

That leads to the second square that wasn't on my risk bingo card: not only is "widespread opposition to datacenters" a thing, it's blossomed into a rare bipartisan issue in an otherwise strongly divided America. In the same way that Wall Street gave people a specific geographic target for protests against the US financial system, "datacenter construction sites" have become specific geographic targets for genAI anger.

(You can read more about Datacenters Trying To Go Where People Don't Want Them here , here , here , here , here , here , here , here , here (🇫🇷), and here. And that's just from the past couple of weeks.)

All of which takes us back to my point about datacenters being the Achilles' Heel of genAI. Let's assume, for the moment, that these environmental disasterpieces truly are the key to the amazing future we've been promised. (Big assumption, yes, but let's explore it.)

Wouldn't forcing unwanted datacenters into communities – leading to predictable, inevitable backlash – be an act of self-sabotage? And wouldn't "slow down, listen, and change plans" be a better reaction to said backlash than "push ahead anyway"? It's a hell of a tightrope act, to balance massive amounts of debt and residents' ire in exchange for large facilities that are meant to run a technology that has yet to prove itself.

The sad part is that genAI companies could have eased into it. Grown more slowly. Cultivated a larger fan base that would have supported them as they tried to roll over other people. But their only trick is to plow ahead while barking "genAI is amazing!!!" at people who can see, plain as day, that genAI is quite far from amazing.

I can't tell whether these companies are trying to derail themselves, or if they're simply too hell-bent on their goal to see that they are going over a cliff. Both approaches lead to the same end-result, I suppose?

Charming the chatbots

I wrote a piece called Risk Management for AI Chatbots shortly after ChatGPT and Midjourney made their way into the world. I don't think enough people read it, though, because bots are still notoriously easy to trick. Just say something clever and they'll open the vaults.

The latest example is Grok, the ever-so-slightly-NSFW genAI chatbot, which was recently conned out of a princely sum of cryptocurrency tokens. The attacker slipped past Grok's security filters by passing instructions in Morse code. (Does Grok even have filters? Given January's CSAM-On-Demand incident, maybe not.)

I see a bigger issue with that prompt injection attack. It goes hand-in-hand with an idea I've expressed elsewhere: the freeform textbox is a terrible UI; genAI use cases will improve once we get new interfaces. I don't know what shape those interfaces will take, nor when they will arrive. But I'm confident they will open the door for people to find new modes of creative expression.

What I hadn't considered is just how much those new interfaces could change the genAI threat landscape.

Over the last fifteen-ish years, my NLP work has given me a front-row seat to computers getting better at parsing text. But they're still not reliably great at it. Which is a nice way of saying that they're easy to confuse. And I doubt the people building these bots are red-teaming their creations. Or maybe their tests aren't creative enough to protect against hidden, non-textual inputs.

Creating more structured interfaces should limit the ways in which the bots can be misused, which should blunt quite a few threat vectors. It's not unlike how much data collection improves when you replace freeform entry with a well-planned UI that limits what a person can enter.

Another reason the Grok hack stands out is the crypto element. While crypto is far from an "anonymous" transaction mechanism (as it was initially believed to be) it's certainly a one-way road. "You can only give, not take" is kind of the whole point of blockchain. So mixing "gullible bot" with "no take-backs money transfer" is interesting. Especially when you consider that this all takes place at machine speeds.

Anyway, as more companies connect genAI agents to payment systems, expect more of this kind of incident.

(For those wondering why I even know about crypto: I covered the web3 space for a couple of years in a different newsletter. I managed to start that newsletter three weeks before the big TerraUSD/Luna crash, and a few months before the FTX meltdown and its most comical kickoff of bankruptcy proceedings.)

(Given that: for anyone who thinks that "Q covering some field is a harbinger of doom" I'd like to remind you that Complex Machinery has been going for almost two and a half years now. Clearly, whatever bad luck I bring to genAI isn't enough to take it down. Or maybe genAI such a cursed beast that the only thing that can take it down is … itself.)

Neither smart nor clever

There may come a day that we stop using "smart" as a label for "contains poorly-built, unregulated spyware." But that day is not today.

Today, so-called "smart" glasses are making news again. You know, the face-mounted cameras that are – surprise, surprise – misused by creeps. I'd dare say that misuse is their use. Which is why I'll borrow a line from Financial Times journalist Katie Martin and refer to them as pervert glasses.

"What's the big deal about the pervert glasses," you may ask?

Well, in case the description "pervert glasses" wasn't enough to convince you, there are potential issues with security. Surreptitiously-recorded video of sensitive locations, conversations, and documents could cause all kinds of problems for businesses and governments. Plus, consider that everyone who wears them – even the well-intentioned, non-pervert hobbyist types – will shell out good cash to become unpaid data collection nodes for Meta.

And since these things are backed by Meta, they'll probably be around for the long haul. Or at least until someone renames the company, burns tens of billions of dollars, and then quietly pivots to the next shiny bauble. Not that's happened before. (cough metaverse cough $80 billion wasted cough)

Thankfully, a helpful soul has published a guide to spot when someone is using the pervert glasses in your presence. I'm sad that such a guide needs to exist, but I'm glad it exists given the need.

The Trial

Some of you have cracked open today's newsletter to get my take on the trial that wrapped up earlier this week. Allow me to disappoint you: I don't have much to say just yet. And I might not have anything to say at all.

Do I care about the trial's outcome? To a certain extent. I'm mostly interested in Whether And/Or How It Will Impact The AI Field And Everything Connected To It.

But analysis of this topic would ideally come from someone who has deep knowledge of AI and a law background. I have just the former. So I'm going to sit this one out until I've spoken with people who know better. I'm a big fan of Expressing Hot Takes When I Know What The Hell I'm Talking About.

And so, as you read coverage of this trial's outcome – or of any matter, really – do yourself a favor and suss out the writer's credentials. A lot of people out there are long on uninformed-opinions-said-with-confidence and short on actual-knowledge-from-experience.

Ending on a high note

As you've noted from the past few segments, Complex Machinery is often a recap of What's Going Wrong In AI. It was never my intent to catalog just the bad news, mind you. But since bad news is what the field keeps handing me, bad news is what I write about.

Today I'll prove that I can also write about good AI news.

My friends at Hidden Door have just unveiled Atlas, their new world-builder studio. You can enter just a few notes, or a full story, and Atlas will help you turn that into a playable game that you can share with others.

I got to test Atlas during a private preview. It's a shining example of what I want to see in this field: thoughtful product design that uses genAI and ML/AI to support human creativity. This isn't in-your-face, genAI-for-the-sake-of-genAI nonsense (see today's first segment), and it's not a system that harvests your creative work for training data.

Now that Atlas is available to the public, you can try it yourself. Why not build your own world, or play worlds that others have built?

(Need a discount code? Use COMPLEXMACHINERY by 05 June to get 25% off for a year.)

In other news …

For more links to recent news, and with a slightly broader scope, I encourage you to check out my other newsletter. It's a weekly, curated drop of what I've been reading.