Category Archives: AI

AI’s Proliferation Requires a new “Proof of Work”

The proliferation of AI demands a new “proof of work.” Personalized, intelligent spam is still spam. Personalized, intelligent unwanted sales calls are still…unwanted. Humans “just checking in” can now be superpowered and never drop a ball again…which means all communication is going to be so crowded as to be unusable.

Bitcoin’s origins date back to a system called Hashcash, proposed by Adam Back in 1997 (“Hashcash was originally proposed as a mechanism to throttle systematic abuse of un-metered internet resources such as email, and anonymous remailers in May 1997”).

Postal mail requires a postage stamp, and that small cost prevented abuse. Want to send out a billion letters? That’s going to cost you a few hundred million dollars. That explains why you don’t get 1000 pieces of physical junk mail every day.

But email? Virtually free. Hence subject to abuse.

Hashcash would force the *sender* to do a certain amount of [then!] CPU work, which the recipient could instantly verify. An intentional asymmetry. 20 seconds to send, .001 seconds to verify.

It never took off because Bayesian etc spam filtering got better, things like CAN-SPAM were passed, etc. But I implemented Hashcash back in the day, and thought it was the right solution since it used the laws of economics to control the problem. Increase cost, decrease supply. Ensure it’s not worth the cost unless enough economic value is created.

Fast forward to 2026. AI-powered email, phone calls, text messages, and all other forms of communication are about to explode. And given AI’s “computer use” wizardry, everyone can just have AI use existing systems to pump out more, more, more…and look indistinguishable from humans.

The Turing test has been rendered essentially obsolete, so we don’t need a better Captcha. We need an economic solution.

Bitcoin took proof of work and turned it into a currency / a store of value. One option is to simply “charge” per receipt/connection, to create an economic constraint. Another is to force/throttle based on proof of work in a way that hopefully is brute-force GPU resistant — which is the exact same thing as “charging,” but without a currency.

But we are quickly headed towards a communications catastrophe, and rather than forcing agents to get “smarter” and sneak past more filters (a never-ending virus v anti-virus battle), there’s a real opportunity to create a proof of work standard and use an economic solution.

Digital Payments are Going to REALLY Grow

The payments market is going to massively expand over the next decade because:

1. ANYONE can now build anything digital — AI code creation means exponentially more digital SKUs that can be created and, of course, paid for. We are in the very early innings here. The gating item is just human creativity. It’s not just software. Can you whistle or come up with a tune? Then you can compose music (no need to learn to read music or know music theory). Can you think of an idea for a movie? You can just…create one. Etc.

Combined with:

2. Almost anything that was “payroll” (paying PEOPLE) can now be “payments” (paying for THINGS). For example: “Hiring an assistant” or “hiring a paralegal” (both payroll) -> paying for a SKU.

We don’t think of ADP or Paychex as payments companies because they aren’t; they are payroll companies. Paying people != paying things.

But more tasks/outputs that were once only available through “paying for people” now become available for purchase on a credit or debit card. This is already starting to happen and accelerate.

And of course, this is not zero sum! Much of this is “everything to the right” of the supply-demand equilibrium point, where there’s conceptually high quantity demanded at a very low price where there’s heretofore no (human) labor supplied. Lots of people will want to purchase a SKU who were unable to hire a person historically.

Redoing Education: Lysenkoism and Preference Falsification

Somebody needs to build a parallel education stack, from top to bottom. The current one is just too broken.

There’s an incredible new input to education (learn anything for free with AI!) and a very different world in terms of needed skills (and economic realities) upon graduation. Against a legacy cost curve that seemingly is inflating to infinity and is immune to productivity gains.

Here’s a math course at Exeter and the introductory, mandatory Biology class at Andover. Andover and Exeter are two of the “best” high schools in America, something that other schools try their best to emulate. The kids that go there want to go to a “good” college so must pretend to enjoy / agree with these courses in order to get the grades and recommendations to get into the “good” colleges. Got a C in Biology? Take a stand against the English teacher who generates grades based on vibes? No good school for you!

The “good” colleges thus are increasingly filled either with people who have excelled at preference falsification and politics, or who genuinely believe in Lysenkoism.

It’s also why college grade inflation is not surprising. Force everyone to take hardcore quantum mechanics and grades will deflate. But the students who have perfected themselves through superficial perfectionism have a real skill (superficial perfectionism!) and will continue that throughout college and life. It just stops accruing benefits outside of academia.

Software Clone Wars of 2004, meet AI Cloning of 2026

History doesn’t repeat, but it rhymes.

Before SaaS, and before freemium, there was “shareware” — try before you buy software. This was a concept dating back to the 1980s, where software would be freely distributed on floppy discs attached to PC magazines…dozens of products on one floppy! Written by hobbyists and even upstart companies.

id Software of Doom fame started out like this, as did McAfee. As did I!

As things like BBSs, AOL, Compuserve, and eventually the Internet grew in the 1990s, one of the main use cases was downloading shareware.

And it eventually started becoming a big business. The biggest download site was the appropriately named Download.com, owned by CNET.

Around the same time, more people in more countries got access to the internet. And this little site called Elance (now Upwork!) survived the dotcom bust and ended up being a leading outsourcing site for everything from translation to, you guessed it, software engineering.

So now there was a huge opportunity. You pick the number one or even number twenty product on Download.com that’s printing money. You go to Elance. You get dozens of predominantly Indian and Eastern European outsourcing shops to compete / bid on “cloning” it.

I had a pop-up blocker (how I met @jonoringer), a couple of security products, and a bunch of utilities like a cool macro tool, an email tracker, etc.

But now I could hire somebody for $500 and have them replicate anything on the top download site on the Internet! It was incredible.

But it wasn’t. There is such much complexity under the hood that you never see merely by using the product. You see it when designing the product, when receiving hundreds of customer complaints, when realizing how much you could improve your conversion funnel, etc.

You can replicate something “skin deep” but miss most vital organs. Who knew you needed a Pancreas or two kidneys?

Elance fundamentally changed the shareware business. Anyone with agency could now hire somebody to clone a product or build a product.

But here’s what I noticed:

-cloning almost never worked, because there was too much “dark matter” in these products to be understood or seen when the goal is just rote replication

-coming up with a NEW idea — much better path, since you have to conceive of all of the myriad corner cases. No free ride to rest on. -technical people still reigned supreme, since they could edit the resulting code from the outsourced shops

-distribution > product. Now that it was so easy to build (or hire to build!), the advantage went to those with a real knack for acquiring customers. And it couldn’t just be “I uploaded it to the file library” like it used to be in the good old days of the 90s

Now replace Elance with Claude or Cursor, and repeat this exercise Distribution will rule supreme. Original thought and insight will rule supreme. “Cloning” things at a shallow depth is a fool’s errand.

Good luck.

Not All SaaS is the Same

There are effectively three kinds of SaaS and it seems the (public) markets can’t tell the difference between the three with the coming AI wave.

Group 1: Software utility is not tied to heads, or if tied to heads not based on those heads delivering an outcome WITH the software. Companies can’t cut back on Workday seats because of AI! Quickbooks is used in small businesses. These systems of record will add AI features which will be accretive to revenue — think background checks for Workday, collections for QuickBooks, etc.

Group 2: AI potentially lowers # of users of the product but potentially introduces more usage? If you need fewer graphics designers you might need less Adobe licenses, but it’s possible you need more? Or the expanded output and productivity gains of AI increases usage?

Group 3: Software utility and pricing are DIRECTLY based on heads using software, where AI directly erases heads for the vertical. Zendesk falls squarely in this category. Theoretically CRM could, too. Without a pivot to outcome based pricing, these guys are in trouble.

But there’s a huge difference between the three. The best companies often have hostages, not customers — and they will maintain pricing irrespective of AI usage.

There’s another thread of “companies will vibe code their own software” but unlikely for critical systems of record where renting is cheaper than owning (hence the shift to SaaS from On Prem starting 20 years ago!)

More here on the value of filing cabinets:

https://a16z.com/ai-turns-capital-to-labor

Cost > Value as a Model

Originally posted as a Twitter thread on September 30, 2025


Cost > Value

You might not realize it, but you constantly face decisions where cost is more than value — and you do not proceed.

I recently stayed at a hotel where I left my (stinky) gym clothes in the room by mistake. $50 of clothes. Look at the cost to ship them back:

There’s just no scenario where this action makes sense. The cost is more than the value and because it’s overseas shipping, there’s likely no hidden margin that would unlock this. Cost is just always going to be more than value.

Of course I have to make this about AI 🙂

What is SO exciting about many AI companies is they are flipping cost and value. Can every business have a multilingual omniscient receptionist? Sure, but that would have been too much cost for too little value.
Now it works!

Which is why the “AI will take all the jobs” narrative rings somewhat hollow based on what I’m seeing. Where AI Applications are really working is where suddenly value > cost. It’s replacing inertia — a world where cost > value for a seemingly infinite group of things.

AI Commerce

Originally posted as a Twitter thread on August 29, 2025


How will AI commerce start?

Sometimes the best way to predict the future is to observe the present.

Millions of people who value their money more than their time already hunt or wait for deals.

Including me:) I had an alert for Spindrift, it cleared my price, and I bought.

All that is missing from this example is the “do it for me” — as in “BUY it for me when this item clears a specific price”

There’s no question that “buy this SKU for me at the lowest price” (already an observed behavior for those who value money > time) will migrate to AI

More in this piece from @venturetwins and me:

https://a16z.com/ai-x-commerce/

Help Wanted? Hire my software!

Originally posted as a Twitter thread on October 08, 2024


Help Wanted? Hire my software!

Airbnb famously grew by “[growth] hacking” Craigslist — automatically responding to listings to encourage a cross-listing to Airbnb:

https://hackernoon.com/how-airbnb-hacked-craigslist-for-viral-growth-24l35eg

As software “becomes” labor (https://a16z.com/ai-turns-capital-to-labor/), we will see similar hacks to sell software — responding to job postings with software products that do a meaningful percentage of the job’s responsibilities. “I know you’re looking for a ‘Bachelor’s Degree preferred with 5 years of experience X’, but our AI-powered software product can do everything you need, is fully trained, and only costs $20,000/year…and can start tomorrow.”

Try going through Craigslist or Monster for job listings — and notice for each job how many of the responsibilities could be done with AI *in its current form* or where AI will be very soon, particularly with more integrations…here’s a local optometrist:

One of my favorite restaurants where I live uses *zero* system for reservations. You have to call, they’re closed in the morning, and the number is always busy during peak dining times. I asked the owner — why not use an online reservation system? They’re so inexpensive now!

His answer: “We are *always* at capacity. We have never had a day where our tables are not full. This would just be an additional expense, and the employee who answers the phone is needed to greet/seat diners anyway — so the reservation taking has no extra cost.”

I asked him: “What if she retires or quits?”
“Well, then I’d hire somebody else”
“What if a software products could do everything she did, for half the cost, spoke 100 languages, never called in sick, and was never late. Would you hire that ‘product’?”
“That sounds crazy, but I guess?”

We’ve already started to see some of this. AI-powered products focused on specialized “complex workflows” are taking on tasks like the ones seen in the prior Optometrist advertisement. And some technology-laggards are adopting these tools first because (like with my restaurant example) they’ve just avoided tech because it’s felt like an extraneous expense…

New Essay: The Transmutation of Capital into Labor

Originally posted as a Twitter thread on August 22, 2024


New Essay: The Transmutation of Capital into Labor

https://a16z.com/ai-turns-capital-to-labor/

The first era of software took analog files, digitized them, and made them accessible with a specialized interface. Think PeopleSoft for HR files, Quickbooks for ledgers, Epic/Cerner for health…

This has played out for 50+ years as more industries have moved to running on software, not files. Cloud lowered the adoption barrier. Adding financial services to cloud made more markets “big enough” for specialized companies (e.g., Toast, ServiceTitan) to exist.

But the same humans that acted on the analog files now act on the digital files! And sometimes it’s impossible to align hiring and training (of those humans) with business needs.

This is what’s exciting about AI. It’s not filling software budget. It’s filling “labor” budget.

Wages in the US alone are $10T+ per year. The worldwide software market is a few hundred billion dollars.

The original “digital filing cabinet” winners have a tremendous amount of potential to add AI, but also have a daunting task of shifting from “per seat” pricing to “per outcome” pricing. Zendesk monetizes per seat. What if a business needs 95% fewer seats because of AI?

Some of the biggest startup outcomes will likely be “net new” industries where a business runs on nothing but Excel…because the software budget was small, the human budget large, and the ability to hire humans was so hard…think compliance officers at a bank.

There’s a saying in economics: “the cure for high prices, is high prices.” As the price goes up, more widgets get manufactured, which increases supply, which lowers the price.
But when it comes to humans and wages, there’s too much latency because of training, licensing, etc.

AI will largely augment employment, and fix many of the “market failures” present with highly skilled yet episodic labor. Imagine: I need your skill for 3 days a year (peak demand), but you need to go to school for 3 years to earn it.

Outcome Based Pricing

Originally posted as a Twitter thread on June 21, 2024


Can’t wait to see the first “incumbent” (in a large software field…like support, CRM, HR, etc) switch from “per-seat” pricing to **per-outcome** pricing.

I’m writing an essay on this now, but consider Zendesk at $115/seat per month…or ~$1.4M/year for 1000 agents:

Let’s say an agent is paid all-in $75,000/year and answers 2000 tickets per year.

This makes the human cost of a ticket $37.50, and the software cost $.69.

The human cost obviously massively outstrips the software cost…and unlike software licenses, it can take months to “install” (find, hire, train) a human to occupy that seat. And in many areas there is simply a dearth of qualified humans given licensing latency

In other words, you can’t simply lift wages and produce more workers…if it’s a role that requires licensing or sufficient training (think mortgage brokers, nurses, etc)

Not to mention the fact that it’s hard (and cruel!) to “flex” humans. Southwest Airlines can’t hire tons of humans when bad weather threatens to cancel flights and then fire tons of humans when weather is clear. But software is perfect for this

So: given the rate of improvement in AI for asynchronous support — what will it take for Zendesk to switch from (in the prior example) $115 per seat per month to, say, $10 per successful ticket answered BY Zendesk? Still much cheaper, more flexible, instant provisioning

It’s obviously going to happen, but how should they price this — it’s the ultimate example of value-based pricing? How to have this interact with existing “seats”? How to have teams not feel threatened by their new AI colleagues filling “seats”?

Whole industries will change, and new ones will be created now that software can produce the outcome vs simply be the tool.

Salesforce charges per-seat pricing for salespeople…why not charge per sale?
Maybe Workday can charge for HR “resolutions”
Etc