PG had an essay about this during the dotcom, when he worked at yahoo. Iirc...Yahoo's share price and other big successes in the space attracted investment into startups. Startups used that money to advertise on yahoo. Yahoo bought some of these the startups.
So... a lot of the revenue used to analyze companies for investment was actually a 2nd order side effect of these investments.
Here the risk is that we have Ai investments servicing Ai investments for other Ai investments.
Google buys Nvidia chips to sell anthropic compute. Anthropic sells coding assist to Ai companies (including Google and Nvidia). They buy anthropic services with investor money that is flowing because of all this hype.
Imo the general risk factor is trying to get ahead of actual worldly use.
The Ai optimists have a sense that Ai produces things that are valuable (like software) at massive scale...that is output.
But... even if true, it will take a lot of time, and lot of software for the Econony to discover this, go through the path dependencies and actually produce value.
The most valuable, known software has already afy been written. The stuff that you could do, but haven't yet is stuff that hasn't made the cut. Value isn't linear.
I can't continue the current model. The dev that gets AI is done in five hours, the ones that don't are thrashing for the next two weeks. I have to unleash the good AI dev. I have the Product team handing us markdown files now with an overview of the project and all the details and stories built into them. I'm literally transforming how a billion dollar company works right now because of this. I have Codex, Claude and GitHub Copilot enterprise accounts on top of Office 365. Everyone is being trained right now as most devs are behind, even.
The (imo) question isn't how you produce software, but what the value of this software is. Are you going to make make/better software such that customers pay more, or buy more? Are those customers getting value of this kind?
The answer may be yes. But... it's not an automatic yes.
Instead of programming think of accounting. Say you experience what you are experiencing, but as an accountant. 6 person team replaced by 2-3 hotshots.
So... Maybe you can sell more/better accounting for a higher price. But... potential is probably pretty limited. Over time, maybe business practices will adjust and find uses for this newly abundant capacity.
Maybe you lower prices. Maybe the two hotshot earn as much as the previous team.
If you are reducing team size, and that's the primary benefit... the fired employees need to find useful emplyment elsewhere in the economy for surplus value to be realized.
Mediating all this is the law of diminishing returns. At any given moment, new marginal resources have less productive value than the current allocation.
If you write a program in Python or JavaScript, you have a terrible mental model for how that code is actually executed in machine code. It's irrelevant though, you figure it out only when it's a problem.
Even if you don't have a great mental model, now you have AI to identify the problems and generate an explanation of the structure for you.
Outsourcing that to an AI SaaS might be ok I guess. Given past form there's going to be a rug-pull/bait-and-switch moment and dividends to start paying out.
For the past decade people have been clawing their eyes out over how sluggish their computers have become due to everything becoming a bloated Electron app. It's extremely relevant. Meanwhile, here you are seemingly trying to suggest that not only should everything be a bloated, inefficient mess, it should also be buggy and inscrutable, even moreso than it already is. The entire experience of using a computer is about to descend into a heretofore unimaginable nightmare, but hey, at least Jensen Huang got his bag.
Only time will tell which version of the future we end up with. It could be good or bad and we will have to see.
The performance - how that is executed on the machine is what you were referring to. “As if” is the key to optimization
That dev is productive with AI precisely _because_ they have a good mental model.
AI like other tools is a multiplier - it doesn’t make bad devs good, but it makes good devs significantly more productive.
Understanding the problem space is becoming more valuable. Strength in architecture of a solution is another skill that is becoming very valuable.
We are close to getting to a point where someone with overall general (and perhaps not very detailed) understanding of arch and design and a good understanding of the problem space and having a good taste in usability will be able to create awesome solutinos.
I can't wait to see these solutions being created by one or two person teams.
This doesn't tell me anything. Two devs who cared and didn't have a bunch of pointless meetings could already, and regularly did, scoop the big tech teams.
There were always 2 ways to complete a ticket. One that did what the stakeholder wanted, and one that does what the ticket says.
But devs that care about the product and what the stakeholders need are rare, and finding one of them was already a significant bottleneck on most projects.
AI might be an accelerator, but we've yet to see if it's optimizing the part that was actually the bottleneck yet.
A lot of stuff that doesn't make the cut is the the stuff that does have value. When you're lowering the bar, remember it's a noisy bar - so a lot more good stuff is going to come through as well.
.. and that entropy can be where all the ultimate value is. That said... considering the point at hand is the context, it's important to start with the diminishing marginal returns.
To give a simple example... Google and FB do not have "invest able software opportunities" at hand. They've been searching everywhere for nails for their "build software" hammer. They are well resourced and risk tolerant.
The diminishing returns curve for "more software" is steep.
Good stuff coming through often starts with $100m markets becoming $1bn markets. That's not even noise at the scale they're thinking about. Long term, sure. Plausibility range is as wide at it has maybe ever been.
But... systemic value is hard to make.
If you can burn through lower priority experiments quickly it's great!
They might be working on all of the super high level things they can think of, but there are always more A/B tests, more features, etc. that are just lower priority, and the chaos of scaling up the org to address them all is super linear whereas the return on going down the list is sub linear.
So you end up with an equilibrium. If the cost shifts, just like in econ 101, the output will change.
I personally make sure I really diversify, so that when I buy funds, I buy those with stocks of EU companies which pay dividends. AFAICT there are 0 European AI companies that pay dividends.
You have to go pretty far down the list of holdings (under "Holding details") to find any big bets on AI:
https://www.vanguardinvestor.co.uk/investments/vanguard-ftse...