Risk management lets you explore possible future events, so you can surface greater upsides and steer clear of downsides. It's about the closest you'll get to a corporate crystal ball.

Not everyone wants that future vision, though. Some of your favorite tokenmaxxing companies insist on learning key lessons the hard way: first-hand and in hindsight.

A rare and clear connection

AI-eager companies are feeling the pain from their LLM usage. Some are slowing their token consumption while others may reduce developer headcount in order to keep feeding their genAI addiction.

At least one company is staring dumbfounded at the latest bill, not sure what to do. Their tokenmaxxing efforts have burned a $500 million hole in the balance sheet.Yes, that's million, with an M. Enough millions to make half a billion, with a B. In just one month.

(They've yet to come forward, so in this post-truth society it's reasonable to question the story's veracity. But for the sake of this newsletter I'll proceed as though it is. Mostly because I want to imagine the uncomfortable Slack thread between the CEO, CTO, and CFO as they try to figure out what happened. In an ideal world the culprit would be people using token-wasting tools to comply with their employer's AI mandate.)

You could write this off as a simple tale of Apply Token Usage Cap, Get On With My Day. You could. But there's so much more here.

Large incidents typically stem from unfortunate collisions of several smaller problems. Those problems don't add up in linear fashion; they compound. So it logically follows that finding and addressing smaller issues should prevent large, painful, costly experiences.

Yet one interesting – and often frustrating – aspect of risk management is that it's hard to see the precise shape of an incident in advance; you only get the specifics in retrospect. Since you can't say exactly what will happen, you can't really demonstrate how much money you'd protect through appropriate planning and preventative action. (The technical terms here would be "loss amount" and "non-event.") Which makes it easier for company leadership to pass off risk management as worried hand-waving, instead of the strategic asset that it is.

So this (alleged) loss of $500M on a runaway token spending spree?

This is just … delicious.

The risk managers, CFOs, and other sensibly-minded execs are grinning ear to ear as they read this. It's not that they take pleasure in someone else's misfortunes; it's that they rarely get concrete examples of transferable lessons. Company leadership usually claims that the other, affected party is too dissimilar to treat their mishap as a learning experience.

This half-billion-dollar crater has changed the game. It's given every risk manager a rebuttal to their CEO's we-must-do-genAI-everywhere argument. They can show a clear, direct connection between tokenmaxxing-without-controls and oh-fuck-what-just-happened. All because someone learned the hard way that "AI at all costs" really means all costs.

To you risk managers and CFOs: I know you're gearing up for your victory lap. I know how much you want to storm into the next leadership meeting and raise hell about instituting token caps and other risk controls around AI usage. May I suggest a softer approach? (Note: not professional advice.) You're welcome to borrow my analogy to open the discussion:

Investment banks track each trader's gains and losses, and limit the amount of money they can place in a single sector or strategy. The former is a simple measure of performance while the latter caps the bank's losses should that investment go awry.

As they stare at you, confused, you can make your move:

Everyone here who gets a Claude (or Gemini, or ChatGPT) license is our company's equivalent of a trader. They're putting tokens out into the proverbial market in search of business improvements. Sometimes they win and sometimes they lose. But if we don't keep track of their token spend, and set usage limits, they may as well be at the tables in Vegas with the rent money.

If the other execs dig their heels in, then you can bring up the story of the lost half-billion dollars. Along with the other tokenmaxxing-gone-awry articles I linked to at the start of this segment. Maybe drop a comment about how well the lack of risk controls turned out for the 2007 mortgage market.

Now if you'll excuse me, I need to add "risk controls may save you up to $500M per month" to my marketing materials.

Not really selling us on the idea

Podcasts really bring out the best in genAI fans.

First up was an interview with Google CEO Sundar Pichai. When asked about genAI's low adoption rate and people's distrust of the technology, he said:

“I’ve always viewed [genAI] as the most profound technology humanity will ever work on. It’s progressing at an extraordinary pace. Humans aren’t evolved to process that much change.

Hmm.

That statement pairs well with how a16z cofounder Marc Andreessen explained genAI's value proposition. I don't usually include excerpts this long, but I want to make it clear that I'm not cherry-picking a single sentence without its supporting context:

“Yes — oh, sell it, I mean, look, so it, it is, alright — I mean, alright I’m gonna give you the deepest of all pitches, I’m gonna give you the, the — okay,” Andreessen stammered right out of the gate. “So, uh, Isaac Newton spent 20 years looking for this key to what he called ‘alchemy.’ Uhm, and the idea of alchemy was to transmute something that was very common into something that was very rare.”

Andreessen goes on like this for the next minute, trying explain that Newton wanted to turn lead into gold, seemingly trying to draw a parallel to the tech industry’s drive to turn sand — silicon — into thought.

“In any event, you may know that he never — we have never figured out how to do that,” Andreessen continued. “And gold is still rare and valuable, so, imagine a form of alchemy that turns sand into thought. Pause on that for a moment.”

(This wasn't some random hallway-ambush of an interview, either. Andreessen was a guest on a podcast, so it would stand to reason that he 1/ knew the host would ask about his genAI investments and 2/ that he should come prepared with some talking points, especially since 3/ this question came after Andreessen himself complained that genAI providers weren't showing people why the technology was so great. As such, 4/ I figure his media relations team could use a drink.)

Apparently, between Andreessen and Pichai, the rest of us are too slow to appreciate something that doesn't do anything?

Cool. Cool.

These statements may sound like a tacit admission that they – and perhaps their entire field – don't care about genAI's value prop so long as it brings them money. And maybe that's true? But there's something else here.

To set the stage, let me give you a quick story from earlier in my career.

The company in question – let's call it Redacted, Inc – had just acquired a massive piece of computer hardware. I could hear the software engineering team's elated reaction a couple of rows over. "It has sixty-four CPUs! That is so cool!!"

Me being me, I went straight to weighing the benefits of having that much horsepower in a single machine against the challenges of updating code and workflows to make use of that extra compute capacity. So I foolishly asked the engineering team what sorts of workloads Redacted, Inc had that might take advantage of that many processors.

Their blank stares told me a lot about myself:

1. "I'm not as much of a 'tech guy' as I'd thought."
2. "That was clearly not the right question to ask of the revelers."
3. "This is why I don't get invited to parties."

(OK, I do get invited to parties. Which is great! But that tells you a lot about the hosts, and what they think of their other guests. Perhaps my role is that of a Vegas-style cooler?)

Once they found words again, the developers repeatedly tried to explain what was so great about this massive machine. But all they could say was "sixty-four CPUs." No use cases; just hardware stats. It was very much a Spinal Tap "these go to eleven" moment.

That brings us back to those statements from Andreessen and Pichai.

The "all tooling, no use cases" mindset I witnessed in the Redacted, Inc engineering team is now leading the genAI charge. Instead of bragging about CPU count, the peddlers simply repeat "generative AI!!!" without explaining why the hell we should care. And their fan base stares at you, dumbfounded, for not picking up on why this is a Big Deal.

The key conceit of the genAI boosters is that they throw away the Sales 101 rulebook – they don't try to get into the buyer's head to see how their widget meets that person's needs. They just whip you into a frenzy, and you will scramble to find the use cases yourself.

Their behavior reeks of that special kind of peer pressure, of the loud people who are insecure and are hoping you don't notice. It makes me wonder how much of that insecurity stems from the stress of backing a technology that's losing races against both public relations and reality, and how much is just the standard Naive Little Tech Boy attitude.

Whatever the case, my biggest concern with the genAI-with-no-use-cases push is that it leaves buyers to stumble around, groping in the dark for a reason to use the thing that is supposedly a lifeline for their business. That rush to push genAI anywhere, everywhere, guarantees it will be used where it is a bad fit. Which will leave the rest of us to deal with the fallout while the genAI fanboys count their money.

White plastic fences

This Washington Post article on fences has nothing to do with genAI. And yet, it has a lot to do with genAI.

The gist is that homeowners are moving away from the old "white picket fences" trope in favor of walls made of vinyl slabs. The slabs win for privacy, sure – unlike the fences, they offer no slits for a voyeur's gaze – but the real appeal is that they're cheaper and lower-maintenance than their wooden counterparts. This makes work less appealing for traditional fence-builders, who prefer to craft out of natural materials.

It's hard to not see parallels to the way genAI is upending software development. I used to work as a developer way back when, and a good portion of my professional network still writes code for a living. Conversations with them have shown me the range of disruptions, enhancements, and emotions brought about by code-generation tools.

For some, the bots have super-charged their work. Others have missed opportunities as would-be clients replace them with LLMs. And then there are those who spend their days cleaning up after what the bots have produced. Along the way, prospective employers and clients are asking developers whether they're using genAI in their work – and they expect the answer to be an enthusiastic "yes."

The code coming out of the bots may as well be vinyl panels: prefab, machine-built, with a little less humanity for builders and customers alike. But good enough to meet the overall task at hand while also being easier to maintain. (Some developers say they'll avoid updating AI-generated code by simply generating it again. Hopefully the LLMs will keep up with the newest code libraries.)

The ability to generate code on the cheap creates a new audience for custom software. Smaller companies that couldn't afford to hire a professional developer can now throw their wish list into an LLM. Will the result be as good as something built and maintained by a professional? Probably not. Will some of these generated apps exhibit critical bugs that put these companies and their clients at risk? In some cases, sure. But do the business owners now have access to new power? They sure do.

We've seen this before. The tech sector has given us sorta-chauffeurs in rideshare apps, sorta-assistants in Siri and other voice-activated tools, and a host of other not-quite-the-real-thing-but-good-enoughs. I was introduced to this idea – offering people a cheaper variant of a service that was heretofore out of their reach – in Blue Ocean Strategy. (The authors gave it a fancy name, but for the life of me I can't find it now. Something about "class" and "the masses," I think? Please let me know if you have the book and can find the term. Print books are not so good for search…)

Software developers aren't the only professionals to get AI-based good-enough doppelgängers. People have already used genAI to create therapists, research assistants, accountants, management consultants, realtors … You name it. With enough time and curiosity, we'll see even more of these lower-cost – and, I might add, unlicensed and untrained – machines filling buyers' needs.

Recommended reading

For the next several weeks I'll point you to newsletters written by some people I've met on my travels. Some are data-related, others focus on risk and finance.

Consider this your summer reading list.

• Counting Stuff (Randy Au) – Weekly musings on all things data-related.
• The Gray Rhino Wrangler (Michele Wucker) – All things risk, from climate to policy to finance. From the author of The Gray Rhino and You Are What You Risk.
• Things I Think Are Awesome (Lynn Cherny) – AI, NLP, games, and so much more. Lands twice a month.

I'll add to this list over the course of the summer. Stay tuned …

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.