I wonder what a VC AI podcast thinks of AI?

This podcast episode Who’s Coding Now? AI and the Future of Software Development by AI + a16z (Apple Podcasts) was recommended to me by a colleague. These are my notes that I took (and later cleaned up) from listening to this single episode.

Send a check or money-order to…

Near the beginning, one of the hosts says,

AI Bubble

“There was a good blog debate about whether we’re overinvested in AI. I think the number was $200B annual investment. And I think the question was how we would recuperate it?

“Well, here we have a way to recuperate $3T, which makes the $200B look like peanuts.”

Sure, sure … except that people have to invest $200B first and the guy is saying that a $3T market will magically appear. it’s not even close to that after three years of the biggest brains in the world working intensely and ceaselessly on it. There is no evidence for that market yet, but everybody’s saying that it’s definitely coming. This is called an ⁠echo chamber (Wikipedia) and it’s the perfect place to brew up market bubbles. The nice thing for these people—A16Z—is that, even if the $3T never shows up, they’ll still have gotten the $200B.

Programming is hard, bro

A little bit later, they’re discussing how they use the tools but they don’t talk about which problems they’re solving. One person said that they start with specs, which is great (seriously). The others talk about how “no-one can remember all of the CSS classes like margin or padding…”, which makes my eye twitch. It’s like hearing your car mechanic say, right before they’re leaning in to fix your car, “no-one knows what all these wires are for…”

The host sounds like it’s an AI reading pre-canned text. I don’t think that it’s a person in the conversation. It basically throws up straw-man, leading questions, like,

Haters gonna hate, yo

“Is there some way to get the neckbeards engaged?”

Ah, yes, if people don’t jump on board with your scam—or they threaten to try to dissuade people from getting suckered themselves—then disparage those critics as nerds, training your minions to be unquestioning monkeys who don’t want to be called names. Don’t you want to be a cool-kid, AI-tool user making tons of money? Or would you rather be a neckbeard/hater/loser who’s going to lose his job to the cool kids?

If it’s such an obviously good thing, then why do you have to try so hard to sell it? Is it because you’re selling a solution to a problem that people don’t know they have? Is the problem that they don’t have a problem that your tool can solve? Or that they don’t recognize they that have a problem? Why can’t the tool’s performance speak for itself? Why does it need so much hype?

You’re not using it right

A bit later, the lady Yoko Li says,

“Given enough context and given enough tools…”

The problem, as far as Yoko is concerned, is that people aren’t able to use the tools enough yet, otherwise they’d be even better at helping you! And maybe you need to spend $200/month to get it working…and if it still doesn’t work, then it’s your fault.

They very lightly discuss context-poisoning and how the models will cheerfully offer wrong answers rather than admit when they don’t know something. They don’t offer any advice about what to do about it (e.g., resetting context in order to resolve poisoning, but that’s a “nuke it from orbit” solution that may throw out the baby with the bathwater). One of the guys says that LLMs are really good at more-complex tasks, which I think he misspoke, but I can’t be sure.

They admit that “models are not really creative…” and then say that if you’re doing something new, then it won’t help at all. I think that’s actually wrong! They can still be used as code-completion, even if it would be useless to try to have the LLM design the whole thing (which kind of works for tasks that have been done a million times before).

History never happened

One problem I have with these kinds of podcasts is that they sometimes feel so outside of history and prior work. The people seem to be considering problems of how we learn, how we create, and other questions of philosophy for the first time, which makes their analysis pretty superficial—because they’re retreading territory that many others have already covered, sometimes for centuries, if not millennia. I find myself thinking, yeah, that’s Kant, yup, there’s Hobbes; oooh, there’s Confuscius!

Yesterday was years ago

I love how Yoko Li says “I talked to a classic vibe-coder the other day…” when the term vibe-coding was introduced just 3.5 months ago. In the Silicon-Valley/VC world, one quarter is old and classic. Remember that that’s their context. Next up, she talks about the same Blender MCP example that I’d already heard about from one colleague and in a video that another colleague had sent to me.

You can’t control chaos

The more-technical host says something that we’re supposed to think sounds smart,

“A temperature-zero model is technically deterministic. The problem is that a miniscule change in the context will introduce a change in the output. … it’s chaotic…”

But for the end-user, it doesn’t really matter why the result seems chaotic, it just is. This observation is more of interest to those building tools on top of these LLMs, as it might give a hint as to how to improve reproducibility, which is paramount to establishing these tools as part of more workflows.

A narrow waist is an API

TIL I learned the term narrow waist, which is a concept, interface, or protocol that solves an interoperability problem (e.g., file-encodings, POSIX, IP, JSON, HTTP), which allows software to address or solve N variations of a problem with a single solution. They discuss whether the “prompt language” might be such a narrow waist. I don’t think we’re anywhere close to deciding that. It is far too vaguely defined and it’s utterly unclear whether the current paradigm will even survive in anything like its current form.

No-one knows how to make money with this yet

Remember, everyone: OpenAI is simultaneously the most successful AI company and the most unprofitable company of any kind in history. Don’t get too comfy using a tool that no-one has figured out how to provide in anything approaching an economically feasible way.

Try harder; be better

Overall, it was a much better discussion than I’d expected when I saw that it was an A16Z podcast. They weren’t very clear on which companies and which business models would benefit from writing software in this way, or when they should jump on board, and with which tools. The implication is, as usual, everybody should be using all the things, and they should have started yesterday.

Their context seems to be that, if you haven’t figured out how to profit from using AI, then it’s not a problem with the technology, but because you’re not trying hard enough.

Consider carefully

A more balanced take would at least leave open the possibility that some businesses might not need AI, or at least that there’s no business case for using the current iterations of it.

Businesses really have to consider what level of investment—in training and monthly licenses—makes sense for them. A16Z benefits from a world that considers the services they’re investing in to be essential to every facet of life.