On The End of A Con

CLVI

[The Antithesis of the Bubble; PLG - Week One; The Code Base of Theseus; It's Still Iterative; Not To Burst Your Bubble...; Learned Helplessness; Fear & Loathing]

Thesis: Building good things is hard; when the bubble bursts, don’t mistake the fallout for evidence that technology is bad.

[The Antithesis of the Bubble]

This post is a bit disjoint because I first follow up with some exciting but very early results on WhiteWhale's swap to Product Led Growth (PLG) from last week, and then I dive back into the AI bubble.

Still, the two are connected in a few ways.

In one sense, we are using the leverage baked into a PLG motion to bet against the AI Bubble.

More than that, though, I think the process we've been using to iterate on WhiteWhale is a really good contrast for the way that the loudest people in the AI bubble are implying or outright saying that things get built.

This is my biggest gripe with the AI bubble: how thoroughly AI Labs & the media & generally AI boosters seem to have convinced everyone that LLMs are obsolescing people & skills.

No matter what they've told you about the death of software engineering or the end of this skill or that, building cool things is still really, really hard & takes focused time and effort.

Dario & Altman are finally ostensibly rolling back on their fear mongering everyone-is-losing-their-jobs-claims, but this is only after or in conjunction with companies like Uber admitting they are struggling to find ROI on the ai spend.

I'll explain both why it's too little, too late, and why their should be no quarter given for the lies and fantasies spewed over the last few years about what AI maybe will one day possibly do.

I'll also clarifying how this stance of mine is coherent and actually amplified by another, seemingly contradictory claim: I adore technological progress and think AI is a core technology we should continue exploring.

[PLG - Week One]

Product Led Growth (PLG) is going unreasonably well for WhiteWhale.

In the first 8 days of having a buy now button on our website, we had two customers buy without speaking to us, plus a third purchaser who I demo'd a year ago and dm'd when I saw him in the flow.

I say this is unreasonable because the volume of people who have clicked on our flow is hardly significant.

I hesitate to share this because I generally don’t like sharing numbers, but what the hell: in 8 days of launching, we added $900 MRR from the PLG motion with 40 people going through on onboarding flow & giving no free trial.

For a VC backed company, this is small potatoes; but for two guys with the goal of hitting $1M ARR without hiring anyone or raising any money, this is a lot better of a start than I thought we’d get with PLG.

Regardless, this first week could literally just be a fluke, and there are other levers we need to pull to make it work in the long run. I’m super grateful for the start, though.

[The Code Base of Theseus]

I speak about iterative engineering a lot (eg, here & here). Meaning, you start with one solution and then, overtime, you work your way towards a 'more optimal' solution

Take the onboarding flow we just wrote. After I had the simplest first draft I could ship, I've already gone through it nearly 70 times to fix bugs and iterate on the experience. Perhaps 50 of those were before we put the button on the website; and Jack went through it, too.

Still, we both missed things that became obvious once we started watching user sessions*.

Copying the RB2B flow, we have our users add our bot to their Slack in the onboarding flow itself. As part of this, they need to pick which slack channel they want to add it to.

When I was testing this, I was doing it with a dummy slack I made for these purposes. That means, there were only 3 channels in this slack, so I could find the right one from a list very quickly.

But, when I was watching recordings of real users, I quickly realized that most people have a TON of channels in their real slacks (we do, too!).

So, in practice, I realized we were missing two bits of functionality on this page that neither myself nor an LLM picked up:

1. Freezing the refresh button to pull in new channels they might make at the top of the element; otherwise they could scroll down, make a new channel back in slack, and because of the hidden refresh button, not realize they needed to click refresh to find the channel

2. Adding a search bar for the slack channels, frozen at the top next to the refresh button. Otherwise, you might have to scroll through 50 channels to find the right one!

Maybe if you're actually a good front end engineer (truly, I'm a novice), you'd just 'know' you should do these things when designing. But, for the rest of us, just shipping and watching is a super power. Users will pick up things you miss.

The same about iterative design goes on the backend: we had some performance issues this week because the backend was putting too much load onto the database.

So, I played the game of finding the most resource hungry query across all sessions and making changes until it was not the most resource hungry query.

I did this across 3 queries and severely reduced the load on the database by an order of magnitude from peak to new average, and by a factor of 3 or 4 from old average to new average.

Again, maybe a better engineer would have done this 'right' the first time around, but one of these changes involved creating another table. In other words, it was a heavier fix that I was able to make more clearly with a lot of information I gathered from the time I originally wrote it. I do think in this case, even a more competent engineer would not have seen this right away; even if they did, they wouldn’t know if it was the ‘right’ call.

Codebases are like the Ship of Theseus: you replace one part at a time until it's no longer what it once was. Unlike the Ship of Theseus, each replacement is not just maintenance, but also often an improvement.

_{*Being able to watch the real playback of a user session is insanely valuable, by the way. This is not a paid ad, I'm literally still on their free tier, but I am in love with Posthog - it took literally 30 seconds to set up. Perhaps I'm exaggerating, but it can't have been more than 5 minutes.}

[It's Still Iterative]

I share all of this because, for one, the engineering process is deeply interesting to me.

The second reason, though, is to remind you that the engineering process really is a process.

You discover more information over time as people use your contraption, and as it interacts with the world.

An LLM cannot 'see the future.' Perhaps it can come up with a better solution on average than a novice in some fields, and it can often do this faster, too.

This doesn't inherently mean it can get you to the end state faster, though. There is still a process that takes time and iteration.

And if you let the LLM produce a bunch of code you can't understand, it can actually slow you down and create structures and patterns that simply don't make sense or are about 490,000 lines too long but now have to be maintained by humans, who can't or simply don't want to do this now overwhelming jobs without LLMs!

And to be clear, for my front end, I heavily use LLMs for the CSS / HTML / JS. It's not like the LLM magically caught the slack channel issue above, nor would I expect it to, just as I wouldn't even expect myself to.

The lie that software engineering is dead or LLMs can do it all is simply that: a lie.

Sure, it can spit out some code that kind of works, but so can an underpaid intern. It's not done yet, just like I don't think a competent developer who care about their output would ever think their first pass was done, ever.

'Done' comes in a gradient and you get a lot closer if you go through it 50 times before you launch it, with an obsessive attention to detail. Even then, you will miss things.

[Not To Burst Your Bubble...]

_{Disclosure: In May, I put my money where my mouth is and took out a position against Oracle. I made this trade because I think putting more skin in the game on something I talk about so much aligns my incentives more than not trading does; I lose money if I'm wrong, not just credibility. Unfortunately, this also means I can lose money if I'm right but place a sloppy trade; I'll likely double down on it next week to fix that.}

Ed Zitron wrote this week that he thinks we're near the end of the AI Bubble. I do, too.

Some of the executive driving AI spend are starting to realize they're more or less burning piles of money on dubious to 0 roi.

Anthropic & OpenAI have raised so much money, it's not obvious where any more capital can come from barring the exit liquidity provided by public markets.

Elon is actually a role model of mine in number of ways, and I really like the fact that the richest guy right now has been the driving force behind building real technology that has changed the world. But man, I'm not a fan of SpaceX's impending rushed inclusion in the NASDAQ-100 at all.* But hey, I guess I'd rather have people throwing money at a company that I believe will actually have a long term positive impact on the world instead Anthropic & OpenAI.

After all, Antrhopic & OpenAI, are deeply unprofitable & fundamentally flawed businesses kept afloat by massive quarterly capital injections with no path to profitability. And no, some leaked deck from a vc that vaguely says ‘they'll have as much revenue as Microsoft' is NOT a path to profitability. It's just more marketing.

These companies make me really, really mad, because they've spent the last 3 or 4 years fear mongering about how people would lose their jobs and driving this obscene wave of doubt around the value of learning real skills.

I don't care that Sam Altman and Dario are rolling back on this nonsense now, that doesn't make up for the last 3 or 4 years.

I can't tell you how many times I've spoken to people who were genuinely concerned they weren't needed anymore and their skills were being obscolesced and it wasn't worth it to learn things vaguely because AI will replace them and do it all and that all that's left for humans is relegation to some ambiguous and ill defined skill like "taste".

_{*In short, some rules were changed that allowed SpaceX to be included in a major index very rapidly upon listing. This means that the stock price gets additional support by purchases made by large etfs and index funds. If you have a 'vanilla' financial advisor or retirement fund, there is an incredibly high chance that you will soon be an owner of some shares of SpaceX without realizing it.}

[Learned Helplessness]

Part of the reason I focused so much on the iterative engineering process above was to contrast it with some of the things I've seen in the past few moons among the kind of person who was sold the narrative that if they downloaded Claude Code salvation would be theirs.

I had one session where I watched a person attempt to use Claude Code to install and run a git repo for about 30 minutes. For context, there were 15 minutes before this where the person was not actually signed into Claude Code and therefore couldn't figure out why nothing was working.

The most painful thing about this whole endeavor was not a non technical person trying to do something technical; that is actually a beautiful thing, in my opinion - everyone starts as a beginner in everything.

What was painful was a sort of learned helplessness I observed in this person. There was no attempt to conceptualize what he was doing start to finish, no planning. There didn't seem to be a conception of where he was in the process, either. When he had to restart Claude Code and it came back and said "I don't see the thing you asked for," he didn't even try to give it a history of what it had just done; I found myself dictating one sentence prompts as simple as "You just said you installed this thing last session, where is it?"

LLMs have offered to teach helplessness en masse and so many people got in line to learn it.

"Oh, you don't need to be spending your time on that, here, let me do it for you!"

My brother in Christ, sometimes you do need to spend your time on that. It's the only way you'll learn and get better.

Fumbling around in a GUI disguised as a terminal will not solve your problems or make your life better.*

I've said it before and I'll say it again: LLMs really can be useful! I use them to help with coding often, particularly my front end. But, as you may be able to guess with the process described above, even using them to help, it took me well over 100 hours to rewrite our front end from scratch and get it to a place that I felt was shippable. Since we started rolling it out in mid May, I wouldn't be surprised if I'm approaching another 100 hours spent on it!

Building cool things is hard and takes effort!

_{*I still can't get over the fact that a 'terminal' application was written in fucking react & has taken over a year to maybe kind of solve a screen flickering issue by introducing a ‘no flicker mode’ that sometimes works after failing with an 80% fix that made other things worse. To put this in context for non technical people, this might be like someone whose never mowed the lawn or trimmed a hedge trying to trim a hedge with a lawn mower, cutting off their foot, and then tweeting, 'Hey, I'm still 50% able to stand and the hedge looks a little better, so this is a win!’ only to later introduce ‘no amputation mode’ as if it were a fucking feature.}

If you enjoyed this post, feel free to subscribe, or don’t. I’ll be here each weekend regardless.

[Fear & Loathing]

I've written about the AI bubble a lot, and these last few weeks I felt like I didn't need to write about it as much anymore because we really are likely near the end.

While I understand that a lot of people hate AI now, I hope it's for the right reason.

I really don't think you should hate OpenAI and Anthropic because of some vague fear that they'll steal your jobs and drive unemployment. Unfortunately, this was part of the con. There are plenty of other reasons the layoffs happened, including overhiring and an economic downturn.

I am actually a huge proponent of continuing to build and grow towards Artificial Intelligence in it's myriad forms, probably even cybernetics.

What I am upset about is that the market was sold a lie about what specifically LLMs could possibly one day maybe do and they bought it like it was the truth today. Something like a trillion dollars was dumped into what I believe to frankly be a dead end AI research path when there are a million other things we can do to further understand & augment intelligence & build useful things.

Expensive GPUs are tirelessly laboring away in data centers to write emails no one will read and summarize emails no one paid attention too and probably didn't need to anyways.

People were told AI was scary and dangerous. It gave governments another fucking thing to waste billions of dollars on talking about maybe regulating. Think about how many meetings happened in Washington DC discussing what needed to be done about a fucking token prediction engine, how many c suite or board meetings started with someone so far removed from the actual fundamental business that they thought the best use of everyone's time was to make the whole fucking meeting about the "AI Strategy".

Obviously, OpenAI & Antrhopic are not the only ones to blame for this - if our society didn't have such a strong disposition for bullshit, no one would've bought these narratives to begin with.

Again, technological progress is fucking amazing.

It put us on the moon, let us harness the power of stars, lets me send this blog at the click of a button to hundreds of readers, allows my startup to process terabytes of data to find valuable information at a cadence and accuracy that you'd need hundreds of human researchers for, lets us grow food in our caves of steel, transplant hearts, delays or stops the onset of crippling diseases, and create aluminum beasts that take us from point a to point b with no human driver.

But Jesus Christ, why'd we have to waste a trillion dollars on fear mongering campaigns and fucking summarizing emails & writing a bunch of slop code that no one will use?

I fear that the fallout of this bubble will be bad for science and progress. I worry that the narrative will not be, 'decision makers acted like children who got into a strangers van because he offered them candy, shame on them for being so willfully blind to reality and such terrible capital allocators.' I worry that the narrative will land closer to fear and loathing towards wealth and technology.

These really are two categorically different things; the former points us towards some potential incremental improvements and changes while the latter leads us away from the stuff that drive progress.

I cannot stress enough: technology, AI, and investment in the future are not bad.

What's bad is selling & buying a narrative that this one piece of technology with no provable ROI is this dark and dangerous power that can concurrently jepoardizes and can save the world. It’s a con and a power grab.

Please, remember, always: there are no silver bullets.

Live Deeply,