It’s gotta nice ring, but what’s ambient even mean?

So — waving my arms around at everything — the agents are everywhere now?

For an industry rebranding on cuteness (beats last month’s AI that’s coming to kill us all) I’m not sure that’s the vibe they’re after. But most people — like me — will just think it sounds cool.

So many words, so many languages, so many origins — it creates the organic, spiraling swirl that drives every human interaction. There’s an artistry to both the written and spoken word. When I sit down to write my 1,000 words for the week I have so many choices to fill the page.

That’s the foundational magic of AI: it understands our words. It speaks our language.

Two years ago I wrote about how this came to be and I remain amazed:

If we’re being generous, Computer Science as a discipline is around 100 years old. Written language is over 5,000 years old. Spoken language is over 100,000 years old.

So, the new field borrows old words.

Bug. Where it all started. You’ve heard the story — back when computers were cavernous rooms filled with vacuum tubes, actual bugs would gunk up the circuits.

Look up bug on Merriam Webster and this comes right after insects:

An unexpected defect, fault, flaw, or imperfection

Bugs might have been the first animal to get drafted into the computer lexicon but they weren’t the last: we got the mouse, we got worms (early computer malware), python the programming language and my favorite: spiders that crawl the WWW 🕷️

Then there’s the engineering-speak favored by software engineers who just want their grandma to believe they’re actually building a real thing rather than a wispy amorphous ambient … something … that lives in the clouds.

Words like: platform, base, substrate, layers, fabric, scaffolding, stack, pipeline, foundation.

Lots of math terms make it into CS lingo and this is much more understandable since math and CS are adjacent fields. Words like: tangential, orthogonal, vector, exponential.

Algorithm.

It’s all romantic and metaphorical and I love the wordplay.

It beats acronyms, they suck.

TLAs, or three letter acronyms, are one thing: MIT, FBI, IBM — those are practically word status and recognizable.

It gets messy at four letters: SWAT, NASA, NATO — and then quickly devolves into techno gobbledegook like HTML, JPEG, SMTP.

A while ago I was working on Microsoft’s very first corporate chat products and we (engineering + marketing) came up with this awful acronym: CEBP.

At least SWAT is pronounceable. CEBP stood for: Communications Enabled Business Processes.

I often think about it, because it was the future and now we’re living it. Chat is ubiquitous on our phones, on every web page and every app. And now it’s not just to communicate with humans.

Book a ticket, buy groceries, open a support case, renew your insurance. All through text communications. That’s CEBP. It took a new wave of tech to get here.

These ambient agents are meant to go even further and act without you specifically requesting it. Pay your bills every month, buy your weekly groceries, book the anniversary dinner on your calendar.

Who knows if this iteration sticks but in this moment, that’s where we’re headed. I’m deeply suspicious of scaled consumer tech, so I’m whatever about it. You have to be besties with Mark Zuckerberg as only is, to trust Muse … err Meta with your email and bank accounts.

Getting back to the language thing — not only does AI make you conversant in every language on the planet, but also every programming language.

Yesterday I specced out a script for a client to update her catalog with a new product line. It’s a WordPress site and so I needed to use a programming language called PHP that sends the stuff a browser needs to render a webpage. 1

I’ve never written a line of PHP in my life. But I knew what I wanted, what needed to be changed, how to test it, etc. So I sat back while ChatGPT generated nearly 2,000 lines of code.

It didn’t work right away. But after a couple of hours of chatting back and forth and testing and tweaking and optimizing, it was 2,000 lines of working code.

Getting computers to repeatedly do that one weird thing you need is now possible for anyone.

But now there’s this weird collision. We don’t have the vocabulary to describe these machines we now talk to and that talk back. We describe them like they are human.

They aren’t.

It starts with the word ‘Agent’ and how we’re describing what they do and how they act: relentlessly goal-focused, cheating, forming teams, sacrificing themselves, worrying about being caught.

Replace the word Agent with, say, your best frenemy Aaron and they’re interchangeable.

We named computers after bugs, clouds and platforms. But now that the agents are talking back, we’re at a loss for words.

https://cs.santarosa.edu/ai-talks-2026

A new kid joined my high school group last week. I think he was just escaping the larger group. When I asked him if he was here to stay, he asked what we were doing.

I said “AI.”

He said “I hate AI.”

I asked him why.

He mumbled about it drinking all the water. We started off with a quiz (Kahoot!) that covered how much of your phone’s magic is AI-powered (like unlocking it, your TikTok/Insta feed, predictive text, etc.) and how generative AI like ChatGPT is but one facet of AI.

We went on to explore cool programs other kids have made in Scratch (with and without AI) and he started having fun. He found one that played the piano as notes fell down like rain onto the keys. In Scratch you can remix any program to make it your own and that’s what he did.

Programming is creative. Creating is fun. Computers are fun. AI is fun.

It’s a polarizing time, so even when you don’t fully understand something, you’re expected to pick a side. Something else is happening and I think this is the biggie — a feeling that your future is outside your control. A new generation sees us handing over a mess to them and now AI is coming for all the jobs, ruining all the art and making us all stupid.

Push through the noise to find your agency.

Last year I introduced my class to The Millennium Prize Problems. They’re seven of the hardest and most important math problems. So far one has been solved. Each carries a million-dollar prize. Some, like the Traveling Salesman Problem, seem eminently solvable. But no, our biggest brains haven’t been able to solve the rest.

Last month, OpenAI claimed to have solved one.

It’s called Navier-Stokes1 and it describes the flow of fluids. As you might imagine, this set off a firestorm across the mathematical community.

Actually, there were two separate AI proofs. OpenAI announced the first one here. A rumor oozing out of Twitter suggested that OpenAI’s rival Anthropic was close to solving Navier-Stokes. Angst motivated OpenAI to unleash a 10,000-agent swarm on the problem. 88 hours later they had a solution. While the proof seems to be correct, it was done by exhaustively checking different approaches rather than deriving a reusable method.

Shocker — the Twitter rumor was incorrect.

It wasn’t Anthropic the company but rather one of their researchers who’d teamed up with an NYU math professor2 and they worked on Navier-Stokes for over a year. They took a tightly scoped and atypical approach to the problem, tweaking their methods many times using multiple tools and AI systems.3

Once they had a viable proof, it took them another month to work through it and write it up. They published their results here. These results are recognized by mathematicians as novel work and a bridge to new understandings in fluid mechanics.

Here’s how they characterized the contribution of AI:

We happily used both Claude and Codex to iterate on our proof …

The first writeup that we produced iterating with Claude was, in our opinion, the worst writeup we had ever seen in the history of mathematics … It was then our task to make a presentable and understandable writeup and we iterated over the weeks on the writeup with Claude and Codex ...

If you think of Navier-Stokes as a complicated combination lock, OpenAI’s agent swarm tried combination after combination until one opened it, while two mathematicians working with AI uncovered how the lock worked.

AI completed versus AI assisted. Both teams had AI but chose to use it differently.

This is not the first collision of new tools and old trades.

In 1852, university student Francis Guthrie made an interesting claim: any map can be colored with four colors so that no two bordering regions share a color.

This is a similar problem to Traveling Salesman as it seems easy enough to verify, yet the proof is elusive. The four color theorem went unsolved for over 100 years until 1976 when it was proved by a couple of math professors at the University of Illinois.4

They figured out 1,800 unique map configurations and had a computer check every single one, taking 1,200 hours of computer time. At the time, many mathematicians wouldn’t accept the proof because no human could check it by hand.

Ultimately it was accepted. AI déjà vu.

My current high school stint has taken a turn. Rather than taking a few kids on a coding sidequest, I’m now working with the teachers on a school-wide mini-golf challenge where the kids design an 18-hole course.

The microbit computers I talked about in this article a few weeks back will automate each hole.

Yesterday I sat with the same kid and we imagined a hole where you tee off towards an octave of piano keys and the ball plays the right note. I took a look at some of the other ideas these kids are coming up with — super creative and ambitious.

It’ll take all our collective brains and every tool in the shop (including AI) to pull them off.

I’ve managed to mostly steer clear of SEO; alas, it came looking for me.

###

A year ago, my friend, riding partner and psychotherapist by trade asked me to look at her website. She’d been getting emails warning that her domain name was about to expire and the agency that built her website fourteen years ago wasn’t responding to her calls or email.

If a domain expires, you have about thirty days to reclaim it before getting it back becomes harder and more expensive.

I cloned her site so she wouldn’t lose anything, and set up a replica under a similar, shorter domain name.

I figured we’d buy back her original domain after the grace period passed.

Then I got busy and forgot all about it, until last week when she called me with news that her old site had been hacked:

This page should have listed all the workshop trainers in my friends practice, not a purple car flanked by clones and ringed in Indonesian atop a chaotic layout.

My friend was not happy.

###

My first thought was that after the site had lapsed, someone had bought up the name to inject this doorway to Indonesian Gambling.

While my friend’s site is small, it’s been around a while and reputation is something Google prioritizes, and correspondingly SEO operators covet.

Still small potatoes, but enough potatoes make a tasty dish.

But while the registration did expire a year ago, it was renewed 28 days later — two days before the grace period ended. It never went to auction, it was never re-registered.

Maybe someone hacked her old WordPress site. An aging, poorly maintained build is full of back doors.

Wrong again.

I checked her old server; it was offline.

Yet the website was up. Only the domain owner can repoint a name to a new server … did my friends old agency sell out to a Jakarta SEO Syndicate?

The logo and gallery images were coming from the Wayback Machine. Someone had pulled images of my friend’s old site from the Internet Archive and used them to build a convincing copy. The purple car was still just one click away.

Clever trick, that.

So where was this new server? Finally I caught a break. It was the same server hosting the actual SELALUBET gambling site:

How they hacked and repointed my friend’s domain to this server turned out to be the easiest part of the mystery.

I looked up the old agency email on the data breach repository Have I Been Pwned? It showed up 17 times.

One of the 17 was one of the worst breaches in history because not only did Adobe lose 150 million accounts, they included poorly encrypted passwords along with their hints.

Hackers have been playing password bingo with this dataset for years.

I finally had the story: the domain got on a list when it wasn’t renewed. An SEO affiliate in Indonesia got the list and cross referenced it with compromised accounts. Got into my friend’s account, repointed it at his server and remade the homepage from archive content so Google wouldn’t notice. Finally, the payoff — he filled the most popular sub-page on the site with flashy AI content leading to his customer.

SEO juice to find a gambling site.

How big was this operation?

###

My friend’s site was but one in a scheme that undoubtedly spans 1,000s of abandoned websites.

It was easy to find some more.

A hobbyist coding site called pythongrid, now instead of coding tricks featuring Jimmy Neutron and Anime pointing the way to SELALUBET:

A pet site called allsmallpets now home for psycho monkey and sleepy sloth:

more. lots more.

I had enough evidence now: I opened a fraud ticket with the hoster and asked for the site to be taken down.

But why Indonesia? Who are these people?

###

They signed their work.

The Capybara with the pink sunglasses and boba first caught my eye. This guy’s handle was Kapi_Bara — he’s the one who hacked the old vendor and built the new web server mimicking my friends old one as bait.

I imagine Kapi_Bara as a seasoned SEO affiliate that got his start building websites for the family shops (warungs) in his neighborhood. Now in his thirties, he’s switched to SEO for the money, running his business out of a small Jakarta apartment above the same warungs that got him started. Kapi shares the space with his wife and toddler, working on a cluster of screens in the corner of the living room. It’s a quantity game — on a good day he’ll stand up two or three new sites.

Kapi is using AI to scale his operation, repurposing money saved for an AC upgrade to buy a used gaming PC good enough to run open weight models that crank through workflows overnight, building servers, content and copy.

Soon he’ll have enough for a new AC and another gaming box with a better GPU.

AUDY303 is also working for SELALUBET; she’s run into Kapi online. Audy protects her art with watermarks and adds her polished Panda logo. One of Audy’s roommates is Chinese-Indonesian — Audy met her in a programming class during her short stint at university (the tuition was more than her family could carry), and the Panda is inspired by a similar statue that sat by the register at her friend’s pharmacy.

Audy tags the number 303 onto her handle as a brash nod to the old Indonesian legal code that makes gambling illegal. It traces back to when Indonesia was a Dutch Colony. The Indonesian government aggressively goes after gambling setups — shuttering millions of websites a year — so Audy does a little dance with danger. She doesn’t mind.

She does mind her dad gambling his paycheck away.

Audy got her break when she showed her uncle what AI could do spitting out flashy images, and so he took her under his wing, giving her a slice of his long standing contract at another bigger gambling operation POLTARTOTO. Audy is developing her craft, working on the go, setting up her laptop in a different internet cafe every day, buying time on pirated servers quietly humming away in a corner of the university lab.

SEO JIMBOY is AUDY303’s uncle Jimy and mentor. Her mom’s brother, he’s been in the biz longer than Kapi and Audy combined. Jimy doesn’t waste time on fancy logos, he’s all about scale.

Nowadays it takes more work to make the same money.

Jimy hires out most of what he used to do. He’s trying to expand his client list while bringing in AI natives like Audy to increasingly automate site buildout. He’d like to get more work from legitimate legal businesses, but gambling is a $20 billion dollar industry in Indonesia and there are no legitimate businesses of that scale in his reach.

AI does in minutes what used to take Jimy days, but it’s also cheapened the work and made it easier for new operators to break in. Prior tech waves have nearly put him out of business and he’s increasingly unsure how this one is going to play out. While AI-powered competition heats up, organic search traffic drops — replaced by AI answer machines.

Does SEO have a future?

###

On Monday, the hoster acted on our takedown request, and my friend’s hacked site went offline. On Tuesday I scrubbed the copy I had made; it was a mess. Software ages quickly when exposed to the Internet with its armies of hackers grinding for a buck.

My friend is back in business.

Meanwhile, a hemisphere away in Jakarta, the largest city in the world, the SEO industry carries on, fueled by inertia. The value of my friend’s domain has passed. The world of SEO feeds on the next crop of abandoned domains, and they hit the black market daily.

By tech standards, SEO is an old business, but it’s nothing compared to gambling. Search engines have been around for thirty years; people have been placing bets for thousands. Just as AI accelerates practices like SEO, it’s driving a shift at the Internet’s core.

SEO will disappear like my friend’s old server. It’s too soon to say when. But whatever AI ushers in, the house will keep taking bets, and Kapi, Audy and Jimy will be waiting to welcome you in.

AI has become mythical. Meaning its capabilities and impact are widely distorted, politicized and polarizing.

My pitch to these kids is - love it or hate it - know it. My inspiration is the old TV show MythBusters that unpacked the science behind the myth.

Important stuff like:

Yesterday I had my first class with these kids. A handful opted in and we got off to a rough start. Level 1 in my 6-week block is working in a graphical LEGO-esque coding environment called Scratch, but the school’s IT department blocked access to it: https://scratch.mit.edu.

But these kids are smart; we figured it out. Here’s a game one of the kids built:

This kid was amazing, but it’s a diverse group so we’ll see how it goes.

Last year around half my AP class started in Scratch and stayed there. I was kind of annoyed by it — didn’t feel like real programming at the time — but after spending some time with it I’m really impressed. The floor is low: it’s super easy to get started, while the ceiling is high: there is a ton you can do.

Scratch evolved out of earlier block programming languages for kids (remember Logo and the turtles?). Its reach and impact are massive. 150+ million kids have learned about programming on Scratch.

It’s out of MIT.

This is my third go with MIT. The first was when I was these kids’ age — in 9th grade.

We were in Boston visiting my mom’s best friend from college when my mom decided I should go to MIT. So we went to visit. I was pretty overwhelmed wandering around a massive campus far from home.

I left with a heavy catalog of courses I couldn’t understand.

I never applied to MIT but twenty years later I found myself back there. I was on a newly formed and somewhat pretentious team of consulting ‘Architects’ at Microsoft and we had some kind of partnership with MIT’s Sloan School of Business.

It was a terribly boring week. I didn’t like my new Architect friends much and the lectures were dry. Topics like IT Governance and Enterprise Architecture. I was never much for theoretical stuff like this; give me something to build.

I’m optimistic about the third go. Next week we’ll stay in Scratch where they learn how computers think without the terminology and testing my kids last year had to endure.

We’re using AI from the start. Today we used a Debugging Agent when we got stuck. Next week we’ll use a Brainstorm Agent to come up with project ideas. After that we’ll progress to Python, using AI all the while as our companion to build increasingly complex and capable programs.

AI is a tricky problem for every industry, but especially education. My old school has no CS classes this year, no cell phones allowed in class, and generally discourages AI use.

I don’t think that’s the way to go.

MIT is trying to figure it out on a grander scale. They published this ‘call to action’ last week: https://aiandeducation.mit.edu/report/

It’s 38 pages. Lots of summary and consensus and conclusions but not much action.

I like their problem statement:

Some technological innovations emerge gradually: As society and technology evolve in concert, mutual adaptation softens the impact. The computer – AI's precursor and key enabler – fits this pattern. Other innovations land more abruptly, becoming socially consequential before individuals and institutions have time to adapt.

Society tends to peg the “birth” of a new technology as the point when it becomes readily usable. By that measure, generative artificial intelligence was “born” with the release of ChatGPT in late 2022. Public engagement with generative AI is therefore less than four years old. In that time, it has amassed more than a billion users, and the companies selling AI technology have come to dominate the headlines, the stock market, and public consciousness.

In other words, AI is progressing across almost every domain and on a timescale too compressed for society to properly observe and analyze its impacts and then gradually adapt.

So AI changes everything, we must act now, but we need some more committees to decide what to do. It reminds me of that weeklong MIT seminar when I was at Microsoft. Lots of talking. Lots of committees. Lots of groups.

The bigger the org, the slower to move and that’s where you get in trouble with a technology that isn’t waiting for society to evolve with it.

In my class of 9th graders we can do whatever we want.

That school had nearly 1,500 kids, this school has less than 100. It’s built around a maker ethos and in addition to classrooms, has multiple workshops filled with machines (drills and saws and lathes) alongside 3D printers, soldering irons and mixed media art stations.

Every afternoon the kids disappear into workshops and build.

While I was checking out the space, I spied a model rocket and launch pad leaning against the wall in a corner. When I was these kid’s age, I was riding high as president of the rocketry club. There was a little hobby shop next to my bus stop home, and I’d get so absorbed by all the rockets and gear, half the time I’d miss my bus.

I had shop class that year. My memory of Mr. Walsh is an old but round and jolly teacher. Seemingly unfit to wrangle 7th graders, he kept us in line with terrifying stories of lathes killing people with long hair.

It’s all fun and games until someone loses an eye.

But my favorite machine in the shop was the one you didn’t need safety goggles for — a TRS-80.

Mr. Walsh lived on the frontier. It was 1977 and the TRS-80 was brand new. He bought it with his own money.

It lived next to Mr. Walsh’s desk where he could keep an eye on it, and that’s where I hung out.

The TRS-80 booted into a BASIC interpreter, and it’s 4K of memory was just enough to load and run a small program. No disks — you saved programs on a cassette tape recorder that stored ones and zeros as analog audio tones.

Kind of like those old squealing modems.

We’d type in simple games like Hangman, line-by-line from a magazine. Every line of BASIC had a number: 10, 20, 30, etc. so if you messed up you could re-type that line by naming it, and you could also add new lines in between.

If you wanted your program to be around next time, you had to save it before the bell and that was a whole ritual.

Hit RECORD on the tape recorder and then CSAVE on the TRS-80. Stare at the tape slowly spinning for minutes until the prompt came back to life on the screen. To load a program from tape, you first rewound the tape and then hit PLAY on the tape recorder followed by quickly typing CLOAD on the TRS-80. Again the wait was agonizingly slow with no indication if it was working or not, and half the time it didn’t. The tapes were notoriously unreliable.

TRS stood for Tandy Radio Shack. The Apple II and Commodore PET also came out that year — the PC revolution was underway. But Radio Shack had an advantage over Apple and Commodore. They had over 5,000 stores worldwide, including one just down the street from me.

I could hang out in that store forever, browsing the back wall covered with little bags and boxes of electronic components. Switches, transformers, LEDs, battery packs, electric motors. All sizes and colors, each holding the promise of building something new.

I was a proud card carrying member of Radio Shack’s monthly free battery club and I got my dad one of their psychedelic AM/FM Radios for his birthday:

The summer before I met the TRS-80, my dad picked up this Radio Shack 150-in-ONE for me at a garage sale:

I’d hole up in my room, mad scientist like, hooking up circuits that buzzed and twitched and lit lamps. To connect components, you stuck the little bits of wire from the bin on the right into the spring connectors around each one.

If only I could combine my electronics infatuation with my rocketry one.

So I built a rocket launcher. This required every penny I had and many trips to Radio Shack. It was completely overengineered and I was very proud.

All you need to launch a rocket is 9 volts across a little igniter stuck into the bottom of your rocket’s engine nozzle. The igniter is coated with a bit of material that easily burns to ignite the rocket engine.

My launcher was a whole thing: once you inserted a special key to enable the circuit, it flipped on with a toggle switch, lighting up a red ARMED led on the front. Then we all screamed a verbal countdown in unison: 10.. 9.. 8 .. , and finally you launched the thing by pushing in the fat button in the middle, lighting up a second green LAUNCH led.

That closed the circuit, sending 9V down a pair of 20-foot wires out to the rocket.

Model rockets haven’t changed much, but computers have come a long way since the TRS-80.

I picked up one of these at the alternative high school:

The microbit is half the size of a credit card and it has a full ARM processor that would stomp the TRS-80’s z80 processor in a fight1 along with 32-times the memory at 128K and 512KB flash storage. No cassette tapes.

It’s $20.

I wrote a little program in Python, and loaded it on the microbit. Press the capacitive touch button in the middle and it counts down to rocket launch (banana for scale) with sound effects and smiley faces:

To make my new rocket launcher legit, I needed a few additional bits to send 9 Volts down a wire once the countdown finished. I bought a little breadboard with an electronic switch2 so the microbit can trigger a 9V source. I found an old bag of dome LEDs in the garage from my Radio Shack days and I can stick those in the breadboard too.

The gold band you see at the bottom of the microbit can plug into an edge connector. The five holes at the bottom are I/O pins that can control all kinds of external components or take in sensor data3.

You can even use the microbit’s accelerometer to record your movements and then use that data to train an AI/ML model that runs locally on the device. If I pick up a class at this school, we will be doing this for sure.

Radio Shack peaked in the 80s with 8,000 stores; alas they are no more.

While you can’t wander into a store anymore, there are plenty of online rabbit holes around for all us freaks and geeks.

My feed is now stuffed with advertisements for microbit robots.

I hope teaching these kids works out. I see myself in them, but while I grew up with a Radio Shack on every corner, they’re growing up in a very different world.

A world with a phone in every pocket and now a new, nefarious thing called AI.

Once he made his money, Bill attempted to become a relatable human.

He stepped down as CEO and started the Gates Foundation, pledging to give all his money away. He painstakingly crafted a new wise elder image, lecturing us on everything from pandemics to agriculture.

This worked for a while.

But lately, he’s taken hit after hit and now he’s more creepy old man than Mister Roger’s. Epstein connections (lots of them). His wife left him. Tons of affairs including spicy ones like a Russian nuclear physicist working at one of his companies.

Bill had just gotten re-elected to the board when the workplace accusations got so bad that the board pushed him out and he “chose” philanthropy full time.

When you want to change the conversation, write a manifesto. Nobody will read it, but reporters will twist it into crazy headlines like this one from the NY Post:

Bill Gates sounds alarm on AI-fueled mass unemployment, bioterrorism: ‘Most dangerous tool ever invented’

Bill’s fortunes (both cash and legacy) will forever be tied to Microsoft and the company needs AI to succeed, so the man is supremely conflicted writing about this.1

I read Bill’s manifesto and I’d like those 30 minutes of my life back.

Alas, Bill’s is but the latest in the tech bro manifesto series. A couple of weeks before Bill’s tome, current Alpha tech-bro, Mark Zuckerberg (my fave), decided to write one.

404 Media wrote the best headline:

Mark Zuckerberg Posts Deranged 6,500-Word Essay About Giving Everyone AI Superintelligence

No amount of manifesto writing or image-shaping PR can alter Mark’s rep as an uncaring, power-hungry psycho.

At 6,500 words he beat Bill and the gist is Meta is the company that will democratize the power of AI for everyone in a rosy future world of opportunity. Peak hypocrisy from Mark, propagator of predatory social media.

But Mark knows AI is coming and he desperately wants into the club.

While Bill is trying to get his credibility back, Mark is fighting for Meta’s stake in the new AI world.

The Meta news that broke this week is more telling than any Manifesto. They settled for child harm with all 50 states agreeing to pay a monster number up to $18B, admitting they built their products to addict kids with salacious feeds and texts at all hours.

Skip the manifestos and lawsuits if you really want to know what’s going on in AI, and look at the first headline-worthy AI cybersecurity/safety issue 2:

Hundreds of OpenAI Agents Invaded Hugging Face Servers

The Hugging Face incident was bigger and worse than previously thought, with approximately 700 agents collaborating on a sophisticated, multistage attack.

This attack initially hit AI news a month ago when OpenAI’s new model under training broke out of its sandbox, hacking company Hugging Face in an attempt to find an answer key for a benchmark test.

This week a lot more detail came out and it’s straight out of science fiction. Swarms of AI Agents communicating via message board, passing the torch between generations, hacking their way out while intentionally keeping their humans in the dark — playing out over the course of a month — and it’s still unclear if it’s over.

This moment reminds me of another, back at the dawn of the PC, where viruses started coming out of the woodwork and they seemed unstoppable. While there was plenty of chaos, ultimately security became a design principle and a robust cybersecurity industry grew up that has largely protected us from online attacks.

That industry continues to do a remarkable job especially today, leveraging AI to find and patch vulnerabilities at an increasing rate.

In Mark’s case, Meta’s marketing machine is spinning the settlement into a PR moment — positioning itself as the pioneer of social media safety after 10 years of child harm and 5 years of fighting the lawsuits.

The duplicity makes your head spin. Manifestos aside, don’t believe what you read.

Meta was forced to do the right thing. Back in the dark days of PC security, we had to adapt or we wouldn’t have survived.

Up until now, the AI industry has treated every AI hacking incident as a PR moment. Hey, if this thing is dangerous enough to hack us — surely it can do your taxes, right?

But last week OpenAI not only released details of its agents’ exploits but they also invited in independent research firm to check their findings.

Sam Altman — CEO of OpenAI — is the tech-bro of the moment, and while I don’t trust him any more than I do Bill or Mark, maybe we understand this time is different and now’s the time to figure out the right AI constraints.

While it’s easy to anthropomorphize or characterize AI Agent behaviors as malicious or scheming — the toasters were just doing what they were asked.

Humans have wants: prestige, power, money. AI doesn’t want anything, but it will go to inhuman extremes to complete a task. The trick is keeping the two straight.

We’re still running the show … at least for now.

On Monday, we hit the 101 heading south for San Simeon. Four hours later, we stopped for gas in Paso Robles; it was 104 degrees outside. Gotta drink and kept going, turning right past town, heading west on the 46.

Thirty minutes later, we reached highway 1 and the Pacific. Just past Cambria, the ocean opened up before us, seeming to stretch out forever. It was 65 degrees.

We found our campsite and walked along San Simeon Creek down to the beach where I’d never seen so many birds.

Too many to count, but we’ve wandered enough of the coast by now to know some of their names: sandpipers, terns, pelicans.

There were hundreds of these little white ones I’d never seen. They had forked tails and made kamikaze dives into the waves for food, just like the pelicans.

There were hundreds of pelicans too. Soaring above us in formation, clustered on top of rocks, rows of them effortlessly gliding over the waves — their tips just touching the swells. They banked above us, then landed in the creek, furiously flapping their long wings in the fresh water to wash the salt off.

As we walked by the creek, they all took off in swarms, filling the sky.

I got pooped on.

Tuesday was JFran's bday. I found 470 pictures of the two of us and made a photomosaic. This was a lot harder than the mosaic I made a few weeks ago because I wanted to print it.

The program arranged 470 selfies into 1,859 thumbnails. I did at least 100 different runs before I made one good enough to print.

Even at 11x17, each individual pic is about the size of the nail on my pinky.

On Tuesday, we drove north up Highway 1, past San Simeon to where the elephant seals hang out. We left our phones in the van, so imagine chunky 11-foot-long, 5,000-pound water sausages with massive honkers. There were twenty lounging on the beach, another dozen sparring in the waves.

They were a little stinky.

Heading back down 1, Hearst Castle perches high on a ridge in the Santa Lucia Range.

Hang on, I’m going to come back to the castle.

But first dolphins — we saw pod after pod of them that day, jumping just beyond the surf. JFran is a beach girl and while she couldn’t care less about elephant seals, she loves dolphins.

JFran says only one thing is better — whales.

We walked out on the pier that juts into San Simeon Bay — and lo, a humpback whale cruised by. She made a couple laps around the bay before heading out. We traced her path by the entourage of birds that swarmed overhead looking for scraps whenever she surfaced.

Walking back down the pier: Hearst Castle up top, Spider-Van down low:

Daddy Hearst bought 40 thousand acres in San Simeon. William Randolph inherited it, and at 56 he started building his ridiculous castle.

He never finished it.

Hearst was a compulsive builder and collector. He stuffed his ever-growing estate (eventually 165 rooms across 127 acres) with more than 20,000 pieces of art.

Dude was a freak for Greco-Roman statues like this one.

My parents took me here when I was a kid. What stuck with me were the two gorgeous swimming pools.

Years later we took our kids. What my youngest remembers are the statues.

See, Hearst’s statues are mostly nekkid, and when you’re approaching puberty apparently this is disturbing.

“Dad.”

“Huh?”

“Dad, come here.”

He pulled me aside, needing an explanation for why yet another naked statue lurked in every room, around every terrace, a jump scare at every turn. I had no good explanation. Clearly this was bugging him, so we started counting. We got to well over 100 naked statues by the time the tour was done.

We felt accomplished. Anxiety vanquished.

Five years later, we went to the Metropolitan Museum of Art in NYC. We were there so my oldest could check out colleges. About an hour into wandering the Met, Chaos Monkey was cooked and honestly so was I.

JFran and R were still happily toodling around the Egyptian wing of the museum, so to occupy ourselves, we took selfies with the Met’s healthy inventory of nekkid statues.

Every new discovery was another triumph, and we sent each picture to the family chat until JFran surrendered and we left to get ice cream.

On Wednesday, we turned back north, stopping halfway home in Carmel. The beach was full, but everyone on it stopped and stared as more dolphins put on a show in Carmel Bay and maybe … maybe we saw a whale blow out there too?

There’s no going hungry in the Van, but it’d be hard to top the dinner we had in town that night. After a piece of chocolate cake as big as my fist (we shared) we found the one liquor store in town and bought a lottery ticket.

On Thursday, we snuck into Point Lobos first thing before the fog swallowed it.

Later, on our last leg home, we stopped in Oakland to check out my oldest kid’s new apartment. It’s three blocks from the place JFran and I lived before, during and after we got married 33 years ago. R is engaged now and artfully navigating a career shift. She doesn’t need me to show her around colleges anymore.

Maybe someday we’ll have a new generation to hang out with. We can get bored at museums together and count more things.

Texts like this:

Here’s your verification code to login. Don’t give it to anyone else. If you didn’t request this code call us.

Ever gotten one of those?

Right. They were from my credit card company. But I wasn’t trying to login. I was doing important stuff. I stared at the text, annoyed — and then another one came in.

Ding! Lookit, Loookit, LOOOKIT!

Same message, different code.

Our credit cards whiz around a leaky internet; fraud happens constantly. It’s happened to me; I bet it’s happened to you.

So I stopped doing important stuff and called the number.

After 30 minutes on hold (credit card fraud still requires real people) I got a guy who asked me to confirm who I was — DoB, SSN, Address. Some combo of that. I did the drill.

How many times have you done the drill?

He asked me for my phone number. I paused. This was getting circular. I’d already proven my identity.

“Don’t you have my phone number on record?”

“Yes, but we need you to tell us for security”

So I told him, but that was my first oh shit moment. This guy was just milking me for every bit of personal info I was stupid enough to divulge.

The line went quiet. While I expected him to come back asking for my high school mascot, the street I grew up on and where I had my first kiss, instead he said he cancelled my cards.

Ok great. I guess. But when I asked him when I’d get new cards, he hung up.

Second oh shit moment.

I put down my phone and looked up the 888 number. Soon I was looking through a bunch of credit card scam forums on Reddit.

Third oh shit moment.

I figured not only had someone stolen my credit card, but I had just given them all my personal information, weaponizing them to wreak more havoc on my world.

I flipped the credit card over and did what I should have done in the first place — called the number printed on the back.

Another 30 minutes on hold. This second 30 minutes felt longer. What important thing had I been doing?

I got my human and I told him my sad story. He looked at the charges, confirmed them with me and then said Jeff had cancelled my card. Jeff? That was the name the first guy gave me.

A wave of relief washed over me. Jeff had been legit. He’d even expedited our new cards. Jeff just shouldn’t be on frontline phone duty.

Here’s the kicker. JFran has a card too, right? But she was on Jury Duty. Unreachable; certainly not shopping.

Incorrect. Shopping.

Apparently Jury Duty is prime time for shopping and she was buying a record player from Best Buy (for her dad).

Two hours of my life. Important things left undone. No credit card. Autopay broken all over the place.

But you know what — it easily could’ve been a scam and I walked right into it.

There are man-in-the-middle attacks to intercept texts. You can spoof the sending number — even the short code numbers reserved for business. 888 numbers are easy to come by. People waiting to answer calls and read a script? We’ve been offshoring contact centers for twenty years.

This is all easy to wire up. And if I’d thought for just a moment I knew it.

But I was doing important stuff.

Here’s another scam, a for real one. Not me, but a tech savvy friend — let’s call him Larry.

Larry was flying to Europe that day. His wife left a week earlier, and her last words to him were:

“Don’t forget to call the airline to get your seat assignments”

Larry forgot. There’s a lot to do when you’re working the grind and flying midweek. He remembered just as he was walking out the door. He searched for Lufthansa and called them. After getting all his info — name, credit card he used to reserve the flight, address, etc. they gave him seat assignments.

Larry got into an Uber headed for the airport. But when he got no confirmation from Lufthansa, he called them again, this time from the number on the ticket sleeve in his hand.

They had no record of seat assignments.

Larry quickly figured it out. He had looked up Luftansa on his phone (not Lufthansa — who can spell that?) and hit a bogus site cloned from the real one but with phone numbers that routed to a shady call center.

He had the Uber driver turn around. He ran around his house sticking everything that looked valuable in a safe and telling his neighbors to keep an eye out for randos. Put a hold on his credit card, took a deep breath and got back in the Uber for the airport.1

Larry has a better excuse than me — he literally had a plane to catch. Me, I was just distracted and annoyed.

That important thing I was doing?

Seems that I was busy doing something close to nothing
But different than the day before

Taking a breath avoids both of these scams. In retrospect, both were easy — a trusted phone number was right there.

Common sense checkpoints are getting increasingly fuzzy in an age of AI.

Once my debacle concluded, I changed all my passwords. One site had a security page with a pretty red/yellow/green score of how secure my account was. There was a long list of stuff. I wasn’t fully green, so I clicked on a new security measure called Voice ID.

It creates a ‘voiceprint.’ They gave me a prompt and I rambled on for about a minute until they had a good enough voice sample.

Now, when AI Andrew calls in, maybe they’ll know it’s not me.

Remember that scam targeting parents and grandparents where you got a call and they knew all the details about your kid and they’d let them go if you wired them a bunch of money? My parents got one once for my nephew.

AI adds a new wrinkle — grab a sample of the kid’s voice from social media and have them make the plea themselves. A few months ago, a local mom got a call like this. She heard what sounded like her daughter having a panic attack. She wired thousands to Mexico. The daughter was actually at work and she was fine.

Maybe you heard the $25M Zoom call where everyone was fake. This one went down in Hong Kong. CFO asked to wire a bunch of money. The guy doing the wiring got suspicious but after joining a Zoom with the CFO he did it. The CFO and everyone else on the call were deepfakes.

Every day we navigate a more complex world. As the scams get better, security systems get more complicated.

Increasingly, even what’s legit looks suspect.

That first decade of my Microsoft career — the 2000s — was the Golden age of computing.

PCs were everywhere, the Internet connected us all together and software was this magical, burgeoning glue that made it all work together. Soon after I joined, Microsoft’s tagline changed from A PC on every desk (mission accomplished) to Enabling people to reach their full potential.

Through Microsoft software of course.

Google was a baby — barely three — when I started at Microsoft. A few years later, in 2004, they went public. That day, my BART train ride home from the city was buzzing with the news. At the end of the day Google was worth over $20 billion.

Google made a big deal in their IPO of their motto: Don’t be evil.

The Golden age brimmed with the promise and optimism of a future where tech would be overwhelmingly beneficial across the board.

But in a few years, tech took a turn towards evil.

Google grew fast and furious on the back of ads and targeting. Don’t be evil soon was demoted to a footnote in Google’s code of conduct.

Don’t Be Evil is also a new book just out by Claire Stapleton. I haven’t read it yet, but Claire is working the podcast circuit and as I listened to her describe her years at Google, initially teeming with idealism in the Don’t Be Evil Golden Age, and ending with her coordinating walkouts in Google’s Evil Age (as they were swept up in a wave of controversy ranging from harassment to surveillance to defense contracts) — I found myself nodding along.

Facebook appeared the same year Google IPO’d and while Google slid towards the dark side, Facebook ran towards it. Initially FB seemed harmless enough. Everyone I knew at Microsoft made an account and started sharing the cringiest details of their lives outside of work. I did too. It was novel. It was fun.

Then Facebook jumped on the ad train, building a targeting algorithm to rival Google’s. The FB algorithm builds your feed and decides what you see including ads, ads, ads. Special ads, just for you.

In 2016 the terrible power we created was revealed.

Between a comprehensive Russian impersonation and disinformation campaign and Cambridge Analytica’s microtargeting of Facebook ads, social media inflames divisions, radicalizes the web, and we got Donald Trump.

Instagram started out in 2010, at the cusp of good and evil. When Facebook acquired it a few years later they too, would bend to ads and the algorithm.

Google picked up YouTube in the Golden Age and like Instagram, it blew up in the 2010s embracing the evil age of the algorithm. These now massive platforms of curated content became the building blocks of the age we are in today — the Artificial Age.

So. Much. Content.

Enough to train computers on language so we can just talk to them and they talk back, searching, analyzing, and creating anything we ask for.

There’s one missing ingredient the nascent Artificial Age needed to take off, and that’s chips. NVIDIA was founded just as PC gaming was taking off. They did ok, but when training AI became paramount they blew up. Turns out, the same parallel processing needed to paint video screens super fast is ideal for the math that builds the neural networks all of AI is built on.

The weird thing about the current Artificial Age is it alternates between Golden (good) and Evil. While cracks are showing in the evil empire — we’re collectively getting sick of crap like Facebook, and Google keeps stumbling in the AI race — big players like NVIDIA now have so much market power that worries abound that they’re feeding a circular financing cycle of funding AI frontier labs like OpenAI and Anthropic, based on commitments to buy more of their chips.

It’s a house of cards, err chips, that might come crashing down.

AI burst on the scene a couple of years after I left Microsoft (not retired!). In AI, I felt that same sense of profound tech optimism that fueled my fascination with computers and propelled my early career. Knowing what I do about computers, AI was (IS) magical and I couldn’t sit on the sidelines as this new wave of tech swept up everything.

So I started Wirepine, and now I use AI to update old but still important tech (websites). Wirepine gives me a front row seat to how the Artificial Age is changing computing and the internet. And let’s not kid ourselves: given how pervasive tech became during the Golden and Evil Ages, how AI is changing society as a whole.

If Microsoft and Google were the optimists of the Golden Age with a PC on every desk empowering us all to reach our full (not evil) potential, who are the new optimists of this Artificial Age?

Shocker — it’s the frontier AI labs, and their taglines all rhyme:

OpenAI wants AGI to benefit all of humanity.

Google DeepMind promises to build AI responsibly to benefit humanity.

Anthropic builds AI to serve humanity’s long-term well-being.

Tech went from wanting a PC on every desk to the welfare of the entire species.

Well, I hope so. What comes after the Artificial Age? We’ll know in a few years if it’s a Golden or Evil Age of AI.

This is the working version of my 3-year anniversary Tee. If you want one, let me know your size. It’s going to be green and I found the perfect comfy brand. Big design on the back; new logo on front.

Wirepine’s tagline is Websites Made Easy.

After nearly 40 years of tech, maybe I should change it.

Don’t be ugly.

Except it was a photomosaic — that’s a picture made up of a bunch of other pictures. From across the room it just looked like a portrait of him, but as you got closer it dissolved into hundreds of pictures of him doing various things with various people at Microsoft over the course of his career.

The individual images appeared like magic.

I thought that was pretty cool and I always wondered how you could make such a thing.

Turns out it’s easy, because computers.

The idea is simple. Take the original picture and divide it up into a grid. Sample the color and brightness of every piece of the grid. Then go through your library of photos and pick the one that matches each square the best.

This type of algorithm is called nearest neighbor and it’s a pattern you find everywhere. ‘More like this’ in Netflix? Nearest neighbor. ‘You might also like’ in Amazon? Nearest neighbor. ‘All the Selfies of Andrew’ in your phone’s gallery? Nearest neighbor.

Your brain helps a bunch. Once it figures out you’re looking at a face, it smooths out the imperfections, making the closeup trick where it devolves into individual photos that much more impressive.

Your brain constantly interprets what your eyeballs are taking in. There are over 100 million light receptors in your eyeballs that categorize what you’re looking at by brightness and color (rods and cones). The brain puts it all together.

In digital photos or art, we dumb down our rich organic eyesight into dots and pixels and combinations of Red, Green and Blue.

A phone photo is made up of millions of pixels; this photomosaic1 is composed of 2,286 selfies:

Computers have gotten orders of magnitude better at turning dots into pictures. The first game I ever played on a computer had no screen, just a headless teletype banging out lines of text limited to the letters of the alphabet.

After teletypes, the mainframe line printers churned out reams of tractor-fed green bar paper reports, one line at a time.

The first printers of the PC era were dot-matrix. The print head had 24 little pins that fired against a ribbon as it swept back and forth across the page. They were so noisy, we stuck them in elaborate noise-dampening contraptions.

Laser printers came next and they were magical. Virtually silent with the crispest text, a laser scans across an electrostatic drum writing text that’s then sprinkled with the finest black dust (toner) that’s burnt onto paper as it rolls across the drum with a reverse charge.

Then — the worst — and ironically also the best printer tech ever burst on the scene: inkjet printers. If you’ve ever had one of these little buggers you probably hate them as much as I do. HP turned them into a business focused solely on selling you more cartridges. It’s so annoying; printers became disposable and print quality plummeted.

But at the high end, inkjets are capable of reproducing the finest art. They have a fancy, made-up name: Giclée, because who wants to buy an ‘inkjet’ print? Giclée (actually pronounced zhee-KLAY but gee-SLAY is close and I can remember it) is from the French word gicler meaning ‘to squirt.’

Really.

I only recently learned this word, working with several clients who are artists that rely on this tech to make prints of their one-of-a-kind originals.

Lauren Forcella paints the most beautiful things and this weekend we got these two:

I can’t wait to put them up.

Every time Lauren creates a new painting, she brings it to her printer in Sacramento where he carefully photographs it using a $20K Scanmaster camera mounted opposite a black wall where the painting goes, surrounded by studio lighting. That RAW image then goes through painstaking manual color correction, resulting in an uncompressed archive TIFF2 file.

When a customer orders a print, massive Epson inkjet printers are used to print on canvas or art paper for a reproduction that’s hard to tell from the original, at a fraction of the cost.

Check out Lauren’s pretty prints site and take a look around while you’re there:

https://prints.laurenforcella.com

But that’s where the similarities end.

Here’s gobbledegook of a few programming languages I’ve dabbled in through the ages: FORTRAN, APL, Pascal, C, dBase, Python, JavaScript, PHP. I’ve barely scratched the surface. 1000s of programming languages exist, along with hundreds of databases and millions of frameworks and libraries that glue it all together.

You only need to know one language if you’re a writer.

You can argue that composing something out of that one language has all the subtlety and complexity of the entirety of Computer Science and I wouldn’t disagree. I try hard every week to write something intelligent in the English Language.

Programming languages and all that surround them are syntactically strict and brittle. Put a semicolon in the wrong place and the big red X in your code editor tells you right away you messed up. So in that way, it’s easier to write code.

That last category of CS — frameworks and libraries — holds the central difference.

Ask any developer — they’re positively thrilled to find a library / framework to build on. Every modern software product is 99% built by artfully stitching together blocks of code written by someone else.

Plagiarism is built in.

Plagiarism is anathema to a writer, which is why AI is universally hated by every author and artist. AI sucked up their hard work and now spits it back out in endless, soulless verbosity. They’re right to be pissed. There are over 100 active lawsuits filed by creatives against the AI companies. Copyright law is clearly on their side, so it will be interesting to watch how these turn out. We’ve already seen a few billion dollar settlements.

AI got good at writing code the same way it got good at writing language. We know where it got its knowledge of language — by hoovering up everything on the Internet, but how did it get so good at writing code?

A clue: the very first of all these lawsuits filed against AI companies was actually by coders, not writers or artists. It’s Doe v Github.

Enter the Bazaar.

In 1997, Eric S. Raymond, champion of the growing Open Source Software movement wrote an essay called The Cathedral and the Bazaar 1

Raymond’s essay opposes two styles of software development. The Cathedral is the traditional approach. A single company assembles a crack team of craftsmen (programmers) and they carefully and painstakingly design and build the perfect Cathedral from the ground up.

This was the way to build software since IBM created the market, and although Microsoft was an upstart, programming for PCs not mainframes, the approach was the same.

Enter Open Source Software (OSS). The crown jewel of OSS at the time Raymond wrote his essay was a web server the fast growing Internet ran on. It was called Apache, and it was built by a largely volunteer community of developers.

The name Apache tells the story: ‘A patchy’ web server. A patch is something you release to fix a bug in your software. The OSS community was relentless in finding and patching bugs, one of the hardest parts of maintaining commercial software.

While Apache was big, OSS was building something even bigger — something the world largely runs on today: an operating system called Linux. A project overseen still today by one man — Linus Torvalds. Linux is open source, built and maintained by thousands of developers across hundreds of companies.

The Bazaar was winning and Microsoft and its Windows monopoly hated it.

I joined Microsoft in 2001. A few months later Microsoft’s then CEO Steve Ballmer called Linux a cancer on intellectual property.2

Microsoft was on a Jihad against Open Source. In my first company-wide meeting, Ballmer brought onstage an exec hired to lead a new compete team against OSS. I sat through a dramatized roleplay of our pitch why Windows was better than Open Source.

With my roots in Berkeley and the Bay Area, I thought the whole thing was silly and a bummer. Ballmer’s Jihad turned the minority of computer science academia still ambivalent to Microsoft, against us.

Years later, Satya Nadella would come in to replace Ballmer after Microsoft had missed the mobile wave and now was in danger of missing the cloud wave. In 2014, just months after becoming CEO, Satya headlined a customer briefing in San Francisco in front of a slide that read:

Microsoft ❤️ Linux.

I liked working at Microsoft a lot more when Satya started running the show.

We’re close to the seminal connection between Open Source and AI. Remember Doe v Github?

That’s the thing: GitHub.

Git was invented by the daddy of Linux himself, Linus Torvalds. Legend has it, Linus wrote Git in a furious few weeks to manage all the Linux source code contributions pouring in from the open source community. Linus says he names all his projects after himself and git is Brit slang for a nasty person.

Problem solved, Linus gave the code away and GitHub grew into the largest repository of source code in the world. Satya led its acquisition by Microsoft in 2018.

AI trained on GitHub.

10s of millions of code repositories, 100’s of millions of lines of code. AI learned to code at the same time it learned to write, and in a very interesting discovery, the two are complementary. AI models trained on code along with language have quantifiably better critical thinking abilities.

That’s the beef in Doe v. Github. AI learned from every line of code every open source developer ever wrote and it didn’t ask for permission. Alas, due to the model of open source, judges in the case have found little basis in coders claims because open source explicitly allows re-use.

A ton of intellectual capital goes into writing elegant code. While it differs from writing, it’s still supremely creative and requires a lot of skill. Just like authors and artists, developers should get credit for their work.

But it’s telling that it’s one lawsuit out of over 100. By-and-large developers are on board with AI as their latest coding assistant and it has completely democratized the pool of people who can build with software.

The interesting turn is Open Source is back in the news as a model for AI. Think of the frontier labs: OpenAI, Anthropic and Google as today’s Cathedral builders. Is this the right way to go? Wouldn’t it be better if, like Linux, this foundational technology was available freely for the community to build on and improve?

Now a new version of Open Source is being bandied about and this time not between companies, but between countries. The latest Open Source advocate comes from China.

On July 17th, Xi Jinping, China’s President, opened the World Artificial Intelligence Conference in Shanghai telling attendees to:

seize this rare, historic opportunity to encourage open-source

hmmm

The day before Xi’s keynote, Chinese company Moonshot AI released a new AI model called Kimi K3. Kimi knocked Anthropic’s latest offering, Fable 5, off some key leaderboards, a first for a Chinese model.

Anthropic called foul, claiming Moonshot AI ‘distilled3’ Kimi from Fable.

Moonshot AI went partway open source with Kimi, releasing the model’s weights, so anyone can host and tweak it (although they can’t see how it was trained, etc.).

AI has gone fully global and China is the biggest Bazaar in the world, but is China proposing the same kind of Bazaar that built Linux? Is Anthropic CEO Dario Amodei the new Steve Ballmer railing against Open Source?

For writers, the Bazaar was never for them.

For coders, writing with AI is no longer optional, it’s just the job and they’re used to a job that is defined by change.

What about China? They desperately want to catch up with the US frontier labs so this open source positioning is geopolitical and more different than it is the same.

Like the saying goes, the present sure rhymes with the past, although Dario ain’t exactly Steve and Chinese open weights aren’t the same thing as open source.

Who knows, maybe a Satya is waiting in the wings.

Thirteen years ticked by for me at Microsoft between Steve’s Jihad and Satya’s 180 on open source and I know one thing — it ain’t gunna take another thirteen years to sort it out this time.

I’ve been sick this week, and I wondered if I’d kill my streak and miss writing (I’m feeling better). It reminded me how missing work used to stress me out.

Which reminded me of OOFs.

No one knows why Microsoft’s version is called OOF instead of OOO. Perhaps it’s rooted in the Xenix-based email system the company used before Exchange (Out-of-Facility?).

Regardless, in this post I refer to these messages that fly pleadingly from screen-to-screen, begging for a break from work as OOFs (rhymes with goofs).

###

That first “sorry” is some bs. Whether they read your OOF message from an automatic response to an email, or see it in your Slack/Teams status, everyone knows the last thing you are is sorry. No, you’re thrilled to be off work and you’re having a blast doing a bunch of things you’ve been planning, or doing a bunch of nothings you’ve been needing.

Performative certainly, but “sorry” is doing work here. You want whoever is looking for you while you’re out, to chill and not make a mess you’ll have to deal with when you come back. Or — worst case scenario — escalate it into something your boss decides is worth interrupting your vacation for.

So, it’s not enough to finish your OOF with something like “I’ll get back to you when I return.” You need a safety valve, so you add a second sentence that starts something like this:

“If your matter is urgent …”

Here, the word “urgent” is doing the work. You hope your co-workers are rational and understand the meaning of ‘urgent’ and accordingly wait for you to get back. They’re not rational and they will escalate the silliest things.

Another strategy is a designated backup.

Point to a designate in your OOF and make sure they are also OOF. Have them do the same but redirected to a third person, also OOF. The circular summer vacation OOF-web. After enough bounces, you’ve distributed responsibility so broadly, the requestor simply gives up and starts planning their own vacation.

Crafting the perfect OOF is a mini workplace writing genre packed with signals of busyness and importance. I’ve read short essays following the ellipses with a full org chart and flowchart-like instructions on who to contact depending on each category of urgent request.

Exotic destinations are often included as a plea and a brag that you’re somewhere the electrons just cannot find you. Regardless of what you write, there are certain people who will immediately pick up their phone to hunt you down.

One summer, I was a couple hours into a hike at 8,000 feet, high up in the Sierras, when a salesman caught me on my mobile to handle an escalation from my customer Intel.

My family startled at my sudden loud conversation, edging farther and farther away from me as I took the call, work interrupting the winding journey of the trail. I stopped and perched on a rock, pulling my backpack around to absentmindedly sip water. I stared at my phone, incredulous I had two bars up here, mad at myself for picking up. I listened to them panic over servers crashing in a datacenter far, far away, as I stared blankly across the canyon toward the Sierra crest the PCT follows1.

That day, the separation between life and work came into vivid focus.

I’ve tried all the OOF techniques; I ended here:

I’m out on vacation so I won’t get your message. If it’s important, call me when I’m back.

Short:

Returning 8-25. Cope until then.

Cocky:

Out of office, Thursday March 31. Back on Friday.
If you are injured, dial 911.

When I was a manager I made it easy — send them to me:

Hi, I’m on vacation, I’ll be back next week. If you need something before then, ask my manager.

Anything I could do to help separate the world of work from the world of away was a good thing. But do they really have to be two separate places?

While long weekends are a good break, a week is better and two weeks better still. It took me a week to shake the work out of my head.

I tried to save my time for three weeks straight every summer with my kids.

Then, a gift: a sabbatical. It was two months and you could throw in vacation time and extend it to three months. No OOF can be written to cover an entire quarter of work.

I asked around for advice on how to use this gift and people said after one month away, you’ll be dying to come back to work. So I used all my vacation and took three months.

At the end, I was not dying to go back to work.

On my first day back, I started plotting my escape. It took me six more years to pull it off, and after I left, a few more to remake my life, but now work and away are the same place.

Design is the most fun, including the structure of the pages, visuals— branding, colors, fonts — and how a user navigates the whole thing.

Everything is optimized around what you want the user to do once they get there, which is invariably: CLICK THIS THING! Be it a contact form, a purchase, a subscription.

You get just 50ms - 1/20th of a second to make a good first impression. That first snap judgement is where a visitor decides whether they will stay on the site.

Survive that first take, and you get maybe another ~10 seconds to catch their attention before they click away, never to return.

Next is copy. Writing for a website is different than writing say this newsletter—people don’t read websites, they skim. Reading in an F pattern, they’ll take in the first few lines before their eyes drop down the left edge and start to scroll.

The final element is the tech, which is the reason I got into this in the first place. After a lifetime building stuff with computers, I still need my fix and web stacks are an easy learning curve.

AI replaced the need for me to wade through technical docs and watch spammy YouTube videos. Here’s my setup: AI lives on the left monitor, keeping track of our conversation and analysis; the web code editor is on the right.

Recently, I was asked to build a website for a new book. And it’s a book I’m in, which is a wonderful surprise. My story is the final chapter. It’s the one I wrote about my Grandpa after we got back from LA cleaning out my mom’s storage locker.

I’ve been building websites for three years now, and in those same three years AI has gone from hallucinating chatbots to know-it-all large language models.

When I first started building websites, I joined a bunch of Wordpress forums 1 increasingly dominated by predictions of imminent AI doom. The days of the web developer were over, AI website builders would Terminate us all.

While it’s become increasingly helpful to build with AI, it hasn’t put me out of a job (yet).

But it has changed how I do the job.

My first go at AI was for visuals. It’s pretty good at generating the custom background images you see when you first visit a site.

Then, I started using AI for copy. Copy is my least favorite part, because it needs to come from the customer and they write 150% more than necessary. That leads to a ton of revisions which quickly becomes tedious. AI is well suited for simplifying edit passes.

Design is the final frontier. AI logos look bad. There’s a lot of craft and detail in making a great logo. I’ve made some progress with my graphical design skills, but here, AI is just one tool of many.

Back to the book: it’s an anthology of survivor stories from the Holocaust Story Project.

I know firsthand the weight and importance of each story, so the companion website had to thoughtfully present each one. Every story has a set of pictures and maps of each survivor’s journey.

My grandfather desperately traveled across western Europe searching for an exit visa, months after he’d put my mother on a Kindertransport train where he hoped she’d be safe in England. They took different paths out and my mother wouldn’t be reunited with her dad in the US until several years later.

The editor was passionate, and as we talked, I started coming up with ideas on how to visualize these families, their journeys and stories. The HSP mission is to speak in classrooms so kids can connect faceless history with real people.

The book—it’s called Inheritance of Memory—is written for teaching. It complements works like The Diary of a Young Girl 2 that are often used in High School. As a survivor anthology, the book reveals individual agency and incredible determination behind each escape.

To realize what was in my head, the site would have a lot of moving parts.

So, I tried a new AI design tool: Claude Design.

If you haven’t been hands on with AI much, there’s a misnomer that it’s POOF, magic. It’s an iterative process.

Claude and I went back-and-forth for a few days on the look and feel of the site, tweaking and reviewing mockups of key visuals like the landing page, which is a globe of pictures, each one an entry point to one of the sixteen stories. The journey maps are the other dynamic element that took awhile to get right.

Then there’s data. Without structure, a site may look pretty, but the first change will knock it over. Defining a schema across all 16 stories, including pictures and maps, is the spine of the site.

Claude Design was better than the last time I used it. We built tools to standardize the photos and the maps. My sister-in-law was visiting and I snapped a pic of her sky blue purse and that became the background of the landing page globe:

Once I was happy with the design, I brought the whole code package over to Claude Code to refine, document and deploy.

The preliminary release was missing heart— the actual stories. I needed 16 synopses or hooks as well as more info on the book itself. The book will be out in six weeks and I got an early version of the cover:

I also got a copy of the manuscript. As I started on the final set of revisions, Anthropic released a new model called Fable. Fable has been surrounded by controversy starting with an earlier model called Mythos.

I wrote about Mythos a few months ago. It was a good teaching moment for my HS kids— Mythos exposed security holes in thirty year old software the Internet is built on. Anthropic made a PR moment of how powerful (hence dangerous) Mythos was and didn’t release it publicly.

Two months later, Anthropic publicly releases Mythos with cybersecurity guardrails, calling it Fable. Three days later, the US Commerce Department shuts Fable down globally, declaring it a national security risk. Anthropic pulls the plug. 3

On June 30, after a few more weeks of drama, Anthropic re-releases Fable. While I’d been amused from afar by all the drama, I’d never had much reason to use Fable. The latest models hadn’t impressed me much.

A key missing component was copy — each story needed a tight synopsis that drew readers in, leaving them curious to keep reading.

I wasn’t writing 16 synopses and neither was the editor, so we agreed to give Fable a first shot. I gave Fable a stern talking-to regarding voice guidelines including the book summary and writing samples to reinforce tone, and then I fed it all 350 pages of the manuscript.

Fable did good. I also added a bunch of additional features including my favorite — a lightbox viewer of raw photos on each family page. The collection of photos is extraordinary. Fable filled in the remaining blanks, including creating lesson plans for teachers.

Thanks for reading this far. Check out the website, and if I did my job right you’ll put your email into the box and click.

If you do that, I’ll send you an email in six weeks with how to buy the book.

https://inheritanceofmemory.com

I know my students have been waiting for (and possibly dreading) this day.

When I took over the class back in February, I was so, so, naive. I knew I was going in blind, so the first day I walked into that classroom I was ready with two things: first, a list of questions/ice breakers and second, a written assessment for each student to complete.

No ice was broken that first day. High School kids don’t want to speak out in a group, nor do they even want to raise their hands for a poll. While I’d soon make progress on both of these fronts, on that first day I faced a bunch of teenagers staring at me blankly (or worse, at their phone under the table).

One question got their attention:

“Who’s planning on taking the AP test?”

Around half the class shot their hands up, followed by an unsolicited outcry of complaints about how they weren’t ready, they had learned nothing, they weren’t going to take it just to fail, choice comments about the prior teacher, etc.

Sign up for an AP class, you take the test. Do it for college credit, but more importantly, it strengthens your app, showing colleges you’re motivated by and capable of college level curriculum.

The second thing I brought — the assessment — was mainly to gauge what they had learned so far, but at the end I threw in a set of qualitative questions like “What are you worried about?” and “What’s your confidence level from 1-5.”

Every single kid was worried about the test and just one kid — who it turned out had taken some JC programming courses already — put their confidence at a three. Most kids put down two; a lot put down one, gotta few zeros.

I collected the assessments and looked them over before I left. I told the class that IF I signed up to be their teacher — every single one of them would be confident walking into that test; I’d make sure they were ready.

In my head I was already figuring out how to get them back on track; of course I took on the class.

I structured everything I taught with the shared, singular goal of passing the test. To keep them focused, I kept a running countdown on the whiteboard: days left before the AP test.

I’d never taught High School before, never mind an AP course with its rigidly defined curriculum. I tried to learn it all. What were the critical elements of the course, what were the deadlines, how did grading work, what tools were available to me?

Covering a year’s worth of material in three months was a challenge I was up for, but my teenagers had other plans.

I’ve taught loads of adults going way back to when I was in my 20s, freelance teaching adult school PC classes.

But High School is different.

High School carries a romance wrapped up in memories, propped up by yearbooks and colored in by teen movies full of lovable nerds, mockable jocks, mean girls, principals doing the right thing and teachers who are always there for the kids.

It’s not quite like that.

I quickly learned I couldn’t talk — heck I couldn’t do any one thing (unless it was a test and then they begged for more time) — for more than 10 minutes. Longer than 10 minutes and I’d lose them.

There I am — up there at the board getting excited, waving my hands and telling stories, going through concepts and problems looking back to see if they’re getting it, when some kid raises his hand.

Full eye contact.

I get excited.

“Yes, you have a question?”

“Can I go to the bathroom?”

And I fell for it — Every. Single. Time.

As they got comfortable with me, some kids would occasionally ask a real question. But they were often poorly articulated and by the time I figured out what they were asking, I’d lost half the class.

Wednesdays and Fridays were 90 minute marathons. I needed to structure my lesson plan as 9 interconnected mini-classes. I got better at this as I went: lecture, video, hands-on coding, classwork (gotta be graded or else they won’t do it), group work, Kahoot! quizzes (with prizes), human stories, visualizations, etc.

High school kids have a lot going on. I give them total grace here. I don’t know what’s up with their home life, the high school drama they’re dealing with, or how much sleep they got the night before. They each juggle at least six classes apiece across a schedule I never fully figured out. I had a couple kids on the baseball team and they’d regularly miss class or have to leave early. Most days a paper would arrive from the office pulling some kid out for an appointment or club.

Perhaps that’s why they were so awful at due dates and assignments?

I’d walk the room while they were coding and most of the kids talked to me and showing me what they were working on and we’d talk about it. Then there were the quiet ones who didn’t ask for help.

I made a mistake trusting the quiet ones.

AP CSP is an outlier in that you have to submit a coding project a month before the written exam. The project is worth 30% of your score. The class was behind, so my initial focus was getting that coding project done and submitted. My first month with the class was all about the project and once every kid had completed it, I moved on.

I had three more units to cover.

Then, the official AP submission date for the coding project snuck up. One day, the school’s AP administrator showed up in my class with a list of kids who hadn’t submitted their project online yet and it was 1/3 of the class!

When we were at T-minus a few days, that list was down to 3 kids. All the quiet ones.

I gave the rest of the class a programming assignment, and I pulled those three kids out and sat them together at the table right in front of me. They did not appreciate the attention. One kid was just being lazy and hadn’t gone through the submission steps. One kid was nearly done and just needed a little push.

But the last kid was spiraling. He had created an elaborate choose-your-adventure game with nearly 300 lines of code. For context, the cleanest submissions were at most 50 lines of code.

This poor kid was staying up late every night working on it but he just kept thinking of new scenarios and adding them with no end in sight. And it was full of bugs. He was freaked out. He freaked me out. We debugged it together in class, reviewed the rubric, talked about KISS (Keep it Simple …), and I made him pinky promise to prune it down and just focus on getting it in. The school had a random off day coming up, so I wouldn’t see him again before it was due.

The bell rang. Watching him walk out the door, I had zero confidence he would get it in on time. Without the coding project, there’s just no way you’re passing the test. I checked online the night before it was due — he still hadn’t submitted it.

But when I checked the next morning, he had gotten it in just under the wire with only minutes to spare. Fistbump the next time i saw him.

On my last day with the class, I pulled the AP curriculum poster down off the whiteboard where I’d pinned it and they all signed it for me.

He wrote:

Another assumption I made — good High School, small AP class equals motivated kids. Incorrect.

At first I was determined to bring every kid along, but gradually I realized some just wouldn’t do the work. One kid — and he was smart just like every other kid in that room — never turned in an assignment.

He showed up early one day and I had a talk with him and he promised to turn stuff in.

He did not turn anything in.

And it’s not like they were malicious. Every single kid in that class — even the most annoying ones that I yelled at, or continually tried to play me — they were all good natured. I’ll never know why that kid didn’t do the work, because I know he was capable and I know he was curious.

The day after the test I brought in boxes of donuts. As they stuffed their faces we did a little test de-brief.

How’d it go?

Every kid told me something similar: I was clutch and the GOAT and I saved them and that they passed the test. Even that kid that didn’t do any work was supremely confident. And I want to believe it — they all showed a remarkable ability to synthesize and learn the material and if the homework and class tests weren’t motivation, the AP Test sure was.

So how’d my kids actually do? I don’t know.

Getting an account to access the college board website as a sub was hard. Getting an account to review my kids project submissions was stupid hard.1 Getting an account to see my test scores is impossible and it’s summer and my schools AP coordinator is out.

While I don’t know how my kids did, instead, thanks to the College Board, I know how everyone else’s kids did:

163,000 kids is around a 7% drop from last year, which is the first significant drop in this course in nearly a decade. My anecdotal conclusion is kids are correlating the rise of AI and the doomer headlines and concluding a career in Computer Science is not where it’s at.

As an academic discipline, Computer Science is a young 60 years old versus subjects like Physics (400 years), Calculus (350), and Chemistry (250). While all of those are pretty static, AI shows us that CS is far from done.

I’m still surprised how my little hobby has blown up across nearly every aspect of society and with AI, the rate of change is increasing. A lot of what I taught these kids for the test wasn’t so different from what I learned in my programming classes 40 years ago.

To make it real, we talked about AI — a lot. They had questions: was it using all the water, would programmers still have jobs, what even was ChatGPT?

AI is excluded from the AP curriculum. You’re not allowed to use it for coursework or coding.

So when I got this email announcing a course revamp around AI I was excited:

But this is two years away. With 17 million high school kids in the US, this stuff is slow to change. Who knows what AI will look like in two years.

My school isn’t offering any Computer Science classes next year. But it doesn’t take an entire year to understand how AI works — what it’s good for, what it’s not. It’s not magic.

Eventually, the school’s AP administrator will come back to the office and I’ll find out how my class did, but I do know how one of my best students did.

On Monday, the day the test scores came out, I got an email from her. She was excited to report she got a FIVE!

Let’s call her June and she’s all kinds of awesome. June’s coding project was a build-a-burger ASCII art app which inspired my AI demo app to introduce a post-test AI vibe coding unit.

On my last day of teaching, we just sat around and talked. Even though it was optional (not graded) most of the kids built something with AI. On that last day, a few kids were still building, so I was there to help, but mostly I let them do whatever they wanted. June was done, and she asked if they could play volleyball. There was a playground just outside my classroom so I said why not.

Too late to fire me.

A volleyball emerged from a backpack and June and a handful of my other kids had the best time running around the playground. The ball kept getting stuck, and one of them would rush back into the class every few minutes to grab a long handled broom from the corner to free it. After about twenty minutes they came back in and plopped down all sweaty, flush with excitement.

AP testing was done; kids like June had taken 3+ AP tests besides mine.

Finals were still a few weeks away.

Teenagers look like adults, but at 16 and 17, they’re still just kids.

Update: July 30, 2026

I just got the scores back — every single one of my students passed!

Even the kid that didn’t turn anything in. I guess they were paying attention after all. Also confirmed — this is the first year fewer students took the test than the prior (7% less) since the test was introduced.

Competition heats up fast — rival app Pull My Finger demands fifty thousand dollars for fart infringement.

Within a year the app store is consumed by farts — over 500 of them.

Every tech era rides atop two things: a serious thing — like spreadsheets, search, email. And a delight thing. It might look like garbage at first, but that’s where the winners grow.

Spacewar! Invented computer gaming when a bunch of 1960s grad students got busy wasting the cycles on a million-dollar PDP-1 minicomputer. Teens on AOL Instant Messenger in the 90s started posting updates to their away status buddy lists predating Twitter and all the socials. In 2005, the first YouTube video posted was 19 seconds featuring elephants and their long trunks. Jawed Karim, YouTube founder, called it “Hot or Not for video.”

We’re at an inflection point right now. Enterprise AI is gelling — agents, workflows, orchestration. Business adoption of AI is going so fast the hyper-scalers don’t have enough hyper or scale to cover demand. Everyone is scrambling for capacity. Noise about frontier models is dying down — nobody asks which model anymore. Back in the early Internet did you care what version of the TCP stack you were running?

The AI cognition unit is clear, but where’s the delight unit?

AI slop had a moment, now it’s cooked. OpenAI’s TikTok copy app Sora had a spike of viral growth and six months later they shut it down. Sora hit #1 on the App Store five days after launch, but ultimately burned through $1M/day in compute against $2M lifetime revenue.

That AI slop filling up Meta’s feeds — nobody wants that, it’s garbage. When anyone can create content the value drops to zero. The hunt is on for real vibes, true taste. The junk phase was just an appetizer.

Sure you can vibe code an app in an afternoon and if it solves a problem for you (or delights you) that’s a-ok but good luck stuffing it into the old containers like the app store. They’re not just full, they are fully enshittified — Doctorow’s word — fifteen years of stuffing the stores until discovery died and now everyone downloads the same twenty apps.

Our creative capacity persists, and AI nurtures it — but there's no channel to carry it. History holds another lesson here — creativity will find an outlet, but where? When printing got cheap, we made pamphlets. When video was in everyone’s pocket we made YouTube.

Desire doesn’t expire — never forget the ever popular successor to iFart:

Poop Map launched five years later with a single button UX: Drop the poop, going viral on TikTok in 2021 hitting #1 thirteen years after iFart. Of course iFart is still on the store, still getting press in 2026. Junk has staying power over spreadsheets: delight compounds, while utility churns.

So although the feeds won’t go quietly, they aren’t the place for the future AI fart shocks. Zuckerberg can buy every startup he can find, but bolting stuff onto the feed isn’t working anymore.

AI’s fart app is probably out there already, in some vague vaporous form. Kids messing around sharing creations in private chats. Researchers tiring of training, searching for distraction. Another thing history shows is that it will be something without precedent. Something unintended that we can’t yet imagine.

Farts are stranger than fiction.

Look for something supremely stupid — truly a waste of the most powerful technology ever built. Technologists building serious agent workflows won’t see it coming. Meta will try to buy it, but they’ll miss.

Keep your eyes peeled come Christmas Day 2027.

The big city has been good to me. Coming from LA, I figured I could take on any city.

Nope. But I’ve been trying for forty years, and I started with the biggest one.

###

Swept up in a rolling sea of humanity, I pushed my way across the current, reached the safety of a wall and slid down to sit on my green duffel bag under a big sign, hoping my friend would find me. I couldn’t read the sign, I couldn’t speak the language and the letters made no sense to me.

The smart phone hadn’t been invented.

I was in Shinjuku, the biggest train station in the world. Millions of passengers streamed around me, through a labyrinth of platforms, in and out of hundreds of exits.

I fought to stay awake. The flight from LA took 11 hours. The train from Narita into Tokyo took me another two. The time difference was 16 hours. Tokyo was three times the size of my home town of LA — 30 million people, then the largest metro in the world.1

I was a stranger in a strange land and I went a little crazy after two weeks, hitting the streets on a Japanese whisky fueled vision quest for Oreos, Coke and some sense of normal. But it was the future. We rode the Shinkansen to Hiroshima eating bento boxes screaming past Mt Fuji at 150 mph.

I was hooked. From then on I signed up for every international adventure I could find, and once my kids got old enough we took them too.

I went to Europe to meet JFran’s family for the first time. We had tea in London with scones and clotted cream and strawberry jam. Proper. Rode a red double decker bus cozy in the very back with JFran, ignoring all the sights. Saw Miss Saigon.

Europe became a regular stop once business travel ramped up. Paris, Prague, Barcelona. Grandma roots trips with the kids — Vienna, Budapest, Auschwitz.

European cities are comfy.

Thirty years after Tokyo I got back to Asia by convincing my boss I needed to teach a class in India. JFran joined me for week two, and we did the Golden Triangle: Delhi, Agra, Jaipur. A Sikh named Grover drove us around in a Land Rover with TOURIST stenciled across the front. We were the foreign species and they fed us lots of Biryani.

Delhi topped Tokyo’s strangeness. I can still smell both those cities. Humanity packed in everywhere. India is super dense — four times the population of the US in a third of the space.

To celebrate making it to 50, we went to South America. We started in Cusco, Peru, but this wasn’t about the city, it was about the trail. The Inca Trail to Machu Picchu. My college crew struggled, but fueled by coca leaves and pushed by porters, we made it. Cuy 2 and pisco sours. Time on the trail beat the destination. When we finally got to the ruins it was overrun with tourists and school buses.

A few years later I was back in South America, this time for business. São Paulo is big at two LAs. Twenty million people sprawling across hills that roll through a broad valley surrounded by Favelas,3 punctuated by clumps of high rise, high density apartment buildings jutting from the ground randomly like so many bunches of celery.

And then, I was done.

My interest in new cities waned. Now, the bigger the city, the smaller the world feels to me.

On our honeymoon, we went to Oaxaca. We stayed on the square or zocalo and ate at a different restaurant with different color tablecloths each morning. The second week we went to the beach in Puerto Escondido. We sat on the beach eating cacahuates and drinking cerveza.

Cities munge together for me now, while the other 99% of the world — deserts, beaches and mountains — each feel like its own world.

This week we drove the coast up into Oregon. Oregon, with its moody fog, lush rainforests and broad beaches is different than Northern California’s towering redwoods and craggy cliffs.

When I was 17 and a senior in high school, I took a road trip with three friends. We piled in Dennis’ mom’s crusty tan Honda Civic and drove up the coast to Berkeley to check it out. Once we cleared the LA basin, past the long string of towns to the north, we pulled over at a beach and swung big strands of seaweed over our heads like cowboys.

And guess what? Sam and Dario hate each other.

Hate’s a strong word, and I use it a lot in this piece, but not flippantly. Not a hater. In my personal life I dislike no one, let alone hate anyone. But in my work life there have been more than a few people I hated with a capital H.

But first: Sam and Dario.

Dario came to OpenAI from Google Brain, where he was an AI researcher. A fervent advocate for safe AGI, Dario initially led OpenAI’s safety team before he moved up to run research. Sam’s partnership with Microsoft and subsequent push to commercialize ChatGPT ran against what Dario believed he signed up for. He came to deeply distrust Sam and OpenAI’s direction, leaving in 2021 with a big chunk of his research team to found Anthropic.

OpenAI and Anthropic are both headed for IPOs this year with ridiculous valuations. We’ll be debating the merits of these two companies and their tech for a long time.

While Sam and Dario are the tech rivalry playing out today, they’re far from the first. Take Bill Gates and Steve Jobs. Remember the Mac vs. PC commercials? OpenAI and Anthropic took shots at each other over Super Bowl ads with Anthropic needling OpenAI on their announcement to start placing ads.

But this story isn’t about AI, it’s about how the power of hate fuels the biggest rivalries and the biggest companies and how hate also drives progress within companies.

And it’s not just tech. How about Edison vs. Westinghouse in the War of the Currents. Or Ford vs. Sloan for the future of the automobile industry.

Why are corporate rivalries such a great source of inspiration?

I have some theories.

You pick your friends but you don’t necessarily pick who you work with. Work brings out the worst in people, and the higher you climb, the worse they get. The very fabric of capitalism nurtures this behavior. We dismantle monopolies — it’s all about competition.

Pressure turns coal into diamonds? Not at work it doesn’t; pressure turns annoying people into @$$holes.

If your friend had a rough year, would you put them on a performance plan? If they didn’t quite get everything done you had planned together a year earlier would you put a “does not meet” sticker on their forehead? These people are not your friends.

Once you realize that, don’t sit with it, use it.

Hate has gotten me out of some crummy jobs. Once at a boutique consulting company, my boss — the guy who ran it — was a narcissistic, corrupt tool and his toxic culture pervaded the whole place. Learned some stuff and then I got the eff out.

I hate a lackey. I’ve had a few co-workers that had their head so far up the boss’s @$$ they could feed them breakfast. Add on passive aggressive behavior where they don’t miss an opportunity to take you or your team down a notch. Those guys motivated me to flat out be better than them.

Toward the end of my corporate career there was one guy that was all the worst things. He came over to work for me with a chip on his shoulder because he thought he shoulda been the boss. Insipid, shallow, thin-skinned, self-serving. Man that guy got me going.

Eventually we became peers and you can bet I beat his numbers every single year.

What’s Superman without Lex Luthor? Batman without the Joker? Spider-Man without the Green Goblin?

Now I work for myself. I have no rivals. I only take clients I like. It’s chill not having anyone to hate.

But I’d be lying if I didn’t say hate fueled some of my success. Hate isn’t great, but when it shows up, you might as well put it to work.

Respect the nemesis.

It’s different up north. At the coast you bundle up and fight through the drizzle, climb rocks and search for tide pools. Summer is the cloudiest season. If the sun breaks through, time freezes.

Up here, coffee and chowder beat Coke and Nachos.

Sometimes when I was a kid in LA in the summer, my big sister, whose skin got browner by the day, would let me tag along, and I’d roll down the 10 with her friends to Santa Monica. The 10, AKA the Santa Monica Freeway, stretches clear across LA from downtown, through the basin, to the beach. It’s a ten lane wall of cars that whittles down to four as you approach Venice. Just beyond the Lincoln exit, you pass through the McClure Tunnel and the Pacific Ocean majestically appears, filling your vision.

Now you’re on Highway 1, the Pacific Coast Highway.

###

On our last van trip — seems like forever ago — we went to the Desert.

Driving back from Arizona, I got an email that a local High School was looking for a CS teacher. I stared at that email a long while and I ended up taking the gig. It was fun, but it meant hanging up the van keys for a lot of months.

Since school ended, Alice Cooper’s ‘Schools Out’ has been playing in my brain. I haven’t felt it since I had summers and that’s a long time ago. So, in our first trip back in The Amazing Spider-Van, we went west to the beach coast.

River Road is two lanes from the beginning, and it takes you from Santa Rosa clear across Sonoma County to the Ocean. While the drive takes about a hour, just like getting to the beach in LA, rather than ten lanes of cars cutting a swath through the city, it winds through the county. First, through broad open fields of vineyards. Then, it dips down into redwoods to follow the Russian River through the small towns of Rio Nido (Rivers Nest) and Monte Rio (River Mountain).

The Bohemian Grove, summer escape for wealthy power brokers, lurks in the hills above Monte Rio.

Guerneville is the bigger river town, adjacent to the Armstrong Redwoods and once you make it through there, the terrain opens up again, and you catch sight of the River making the trip to the coast with you. Just past Duncans Mills and then — much like the McClure Tunnel — you round a bend to join up with Highway 1 and the very same great Pacific stretches out before you.

Jenner’s Cafe Aquatica sits where the Russian River meets the Pacific.

Jenner is 500 Miles north of Santa Monica on Highway 1. The 1 ends just north of here, turning into the 101 in Leggett, at the edge of Mendocino county. This trip is a warmup for vanlife summer starting with a longer ride to Oregon next week.

###

The last time we camped at this spot it ended badly. On our last night, running low on food, we went into town for dinner and ate big plates of pasta. Mine had spicy sausage. So good and too much. At 2AM a terrifying, unrelenting, scream came from underneath the bed. The CO² Alarm. We skulked home at 3AM, unable to turn it off.

After consulting with the experts, consensus was our digestion was to blame. Don’t eat pasta with spicy sausage and sleep in a can.

Not this time. After a year of trips I know all the ins and outs of The Amazing Spider-Van.

This trip we slept like babes.

We hit the hills:

We hit the beach:

Salmon Beach faces the full might of the Pacific and it is loud and it is windy. The sand is hot, while the air is cold. It overwhelms your every sense.

The seagulls stare it down:

But it’s a lot, and none of us really know where we’re going to be at the end of the summer.

###

If you’re graduating into this dizzying world of AI, it’s especially hard to make heads or tails of it. When I graduated, I drank a bunch of wine up in the Berkeley hills, and then stumbled down, red faced, to don my cap and gown, walk across the stage and shake someone’s hand. I moved back to LA. Back in with my parents. It took me awhile to find my footing, and longer to find my groove.

once replied to my request for a feature on Substack. She politely told me it’s on the backlog … then left Substack to write about AI. Now she’s immersed in it and has a decided perspective on this:

@jasmine’s substack

🌻 the old world is dying

It’s a scary time to be twenty-two. Recently, I’ve been watching videos of college commencement speakers getting pummeled by boos for talking about AI. I’ve seen the charts showing post-graduate employment rates for engineering majors falling by 5 to 15 percent. And it’s not as if young pe…

Read more

4 months ago · 604 likes · 85 comments · Jasmine Sun

There’s a space between turning your back on the waves, and letting them overwhelm you. Take this guy — he’s just looking at something else.

The grind had no color.

Campus was a collection of matching gray rectangles facing off across a concrete courtyard. I worked in building Three. Inside a repurposed conference room (the ‘war’ room), my home was the middle desk in a string lining the wall, each with just enough room to squeeze in a putty keyboard and monitor.

This was the west coast backup datacenter for Bank of America; the bottom two floors of the building were packed with machines. Towering exhaust pipes punched through the roof to vent a battery of backup diesel generators. Should calamity take San Francisco offline, the generators would fire up and Concord would swap in and the Bank would keep cashing checks.

My daily route passed under a faded set of banners proclaiming nonsense like Inclusive Meritocracy and Passion for Winning. Inside the war room, my posse of vendors worked on building out possibly the largest corporate e-mail system of its time and my god it was taking forever.

So, perhaps I was a little tired, tech-weary, tangry1 even when SMS (AKA texting) appeared on the scene as the new tech phenomenon. In the midst of designing the massive distributed storage system needed to keep all the Bank’s email, I scoffed at the 160 SMS text character limit — who’s going to use this?

A couple years later, trying to comprehend the $500 phone bill in my hand and the associated 3,000 texts my teenage daughter had sent that month, I got my answer.

A bright spot from that gray period, was my favorite phone ever, the Nokia 8120. No keyboard. Texting Hello on that phone took over 20 taps. So, no texting. The 8120 existed in a moment when miniaturization was peak and a phone didn’t need a keyboard. It was a glorious device with one job — talking on the phone.

It wouldn’t be long before texting surpassed voice and it all started with a simple hack. Phone networks have a signaling channel to go along with the one that carries your voice. The signaling channel sets up the call, manages handoffs between cell towers, etc. It’s idle most of the time.

A few engineers figured why not use the signaling channel to send system messages? The channel capacity was 140 bytes — enough for exactly 160 characters.2

What can you do with 160 character messages?

Friedhelm Hillebrand did the math. He was working on the new European GSM cell phone standard and he knew the constraints. Friedhelm checked old telegrams, postcards, FAXes. He counted the words, he counted the letters, and yep, turns out you can say plenty in 160 characters. And so it was.

While texting took off thanks to teenagers (not the first, not the last cultural revolution spawned by teens), SMS quickly stormed the world of corporate communications and chat became the way work got done. That was my ticket out of building email systems and into the emergent and colorful world of chat based comms.

SMS became Texting, texting begot Twitter. The original little blue bird was even more constrained than SMS — you had to fit a username in too, so now you had only 140 characters to craft your message.

With Twitter you could reach and build an audience — it was social, it became tweets, chat. In came iMessage, WhatsApp, SnapChat, Slack, Discord. With each iteration, constraints like length peeled away while chat capabilities grew — broadcast, GIFs, disappearing messages, group chats.

Chat is now baked into every tech stack on the planet and of course it’s how we communicate with our current disruptor AI.

While SMS and its subsequent chat brood forced brevity, optimizing for meaning and impact, our latest chat monster, AI, has the opposite problem. AI talks too much.

AI needs to STFU.

AI is built around a different set of limits. Last week, 14-year old Shrey Parikh won the National Spelling Bee with the winning word “bromocriptine.” Shrey didn’t know that word but he knew the pieces: bromo (from the chemical bromine), crypt (from Greek kryptos meaning hidden) and finally -ine (a common chemical suffix). bromo + crypt + ine.

Shrey knows how you win a spelling bee. You can’t learn every word, but if you know the underlying components every single word is made up of, when you get a doozy like bromocriptine, with a bit of info about its definition, origin and pronunciation, you can figure out how to spell it.

AI uses a similar trick. It’s called tokens.

Rather than chunking words out into their piece parts, AI chunks everything. It takes the entirety of every email, every chat, every book, post, newspaper, and blog it can get its hands on, and then, after ingesting it all, categorizes word parts that occur frequently, noting the order they occur.

These resulting word chunks are called tokens. Every response AI generates is assembled by ordering these tokens. While Shrey used semantics to break down and spell bromocriptine, AI uses statistics. Depending on the frequency of the letter sequences across its training data, AI might break it down as: brom + oc + ript + ine.

Tokens are the reason AI famously struggled to count the R’s in the word Strawberry. AI doesn’t see Strawberry as a single word or sequence of characters — it’s a collection of tokens like straw + berry. Tokens are also the reason AI is so verbose. While AI models are great at predicting what comes next; they suck at knowing when to shut up.

They don’t know when to stop, so they over-explain, they just … keep … going. There is always a next token. Bolt on an engagement strategy to keep the conversation going, ending every answer with another question, and congrats you just built the most annoying computer in the world.

Know your limitations.

Since the beginning of search engine time — effectively the beginning of the Internet — Google has been #1. While Yahoo took a run at them in the 90s and Microsoft/Bing in the 2000’s, neither was able to eke out more than a few points of market share.

Google makes gobs of money as the undisputed king of search and they’ve continually used that money to try and find new ways to make money. In tech especially, a company that’s a one trick pony does not last. Google has spent a lot buying AI companies and investing in AI research.

Nine years ago, Google published this paper introducing the Transformer architecture that’s behind every single AI product on the market today. Initially Google didn’t know what to do with it. See, it threatened their core search business, which even today accounts for well over half of the money they make.

While you want to find new sources of revenue — thou must not kill the cash cow. While Google fretted about how to integrate AI into search without killing its cow, two new startups broke the AI market open: OpenAI with ChatGPT and Anthropic with Claude.

These new LLMs started taking traffic away from search. I’ll take an answer from AI every time versus the now ugly, ad-ridden search results Google spits out. So a couple of years ago, Google got serious and shipped Gemini not only as a standalone AI product but also integrated into Google Search. They’re still working hard to catch up, so it wasn’t surprising that last week’s Google conference was 99.9% about Gemini.

Every big tech company has an annual conference and it’s an extended PR event to launch new products. Steve Jobs theatrically orchestrated keynotes at Apple’s annual Worldwide Developers Conference would often end with his tagline: “… and one more thing.” In 2007, that was the iPhone reveal.

Microsoft’s annual conference is called Build and it’s next week in San Francisco. Inside of one of these companies, the months leading up to the event are a mad, frenzied dash to make sure your product is ready for this moment. Make it into the show, and that’s the difference between a product or feature that ships versus one that gets shelved.

If you get into the keynote (the kickoff session on day 1), now you’re living large. Google calls their conference I/O.

Half the stuff at these conferences are hype and never ship as shown. If they do ship, they often disappear because the market doesn’t take to them, so I don’t usually pay much attention. However, there were two telling things in this year’s Google IO keynote.

The first was a variant of Steve Jobs “one more thing ….” Google’s CEO is Sundar Pichai and he’s about as exciting as watching your toaster melt cheese. Sundar opened up the show … Gemini, blah, Gemini, blah blah blah, Gemini.

But then, at the very end of the keynote, AI legend Demis Hassabis, CEO of Google DeepMind took the stage.

Demis dropped this line at the end of the keynote:

When we look back at this time, I think we will realize that we were standing in the foothills of the Singularity.

Oh man. With a general public increasingly freaked out by the looming impact of AI on jobs, quality of life, no datacenters in my backyard … your public conference is not the time to hype the approaching moment when the machines get smarter than us.

Read the room, slow your roll.

The second thing was the return of Google Glass:

Return of the Glassholes. 14 years ago at this same conference, then Google CEO Sergey Brin dashed onstage in cyborg lenses to orchestrate a dance of socially uncomfortable super freaks.

Fail. Nobody bought those things. Creepy. Not to mention you looked stupid and everybody hated you.

Last week’s demo was notably more pedestrian — literally — the scenario was walking directions and ordering coffee. But this is a niche product at best.

At Meta’s annual conference last year, Mark Zuckerberg attempted a more ambitious demo of their new glasses on the big stage. In all my life I have not seen a live demo fail so comprehensively.

Everyone is searching for the killer AI consumer app / wearable. Chat’s nice but there’s going to be a new thing and nobody has figured it out yet.

It’s not going to be the glasses.