a16z Sep 21, 2026 29m 14m saved
With Steven Sinofsky, Board Partner at Andreessen Horowitz
In 1983 exactly two people in Steven Sinofsky's 100-person college dorm owned a computer. Federal agents came and took one of them away, and it turned out the thing he had done was not yet a crime.
The lesson most people draw from that is that the law lags technology. Sinofsky draws a different one: the hearings that followed played clips of War Games at the witnesses, and the same theater is running now on AI — except this time the industry wrote the script itself, by describing its own defects in the language of intent.
"The AI people are making it impossible for anybody to understand what they've done."
Sinofsky was president of Microsoft's Windows division, running engineering and marketing across Windows Live and Internet Explorer, and he was in the room when the company halted Outlook's development because a worm had taken the world's email down. He is now a board partner at Andreessen Horowitz, a seed investor, and the author of the Substack and book Hardcore Software.
The full interview is covered here so you can skip it. 29 minutes of audio, 15 minutes of reading.
Here are the 8 arguments that matter.
Key Takeaways
A model that does the wrong thing has a bug, not a misalignment — the word imports intent that the software does not have
His version of an AI pause is an engineering pause: stop shipping features and go add the telemetry, logging and step-by-step debugging
OpenAI's incident reports are not close to enough, measured against what a CVE or an FAA accident report contains
Y2K is the precedent, and the rules were industry-drafted — a consortium of banks and insurers wrote them, not Congress
The anthropomorphic vocabulary came out of academia, where a cute title is how a paper gets attention
"Goal seeking" and "cheating" mean one thing to an engineer and another to a congressman, which is why the hearings go the way they do
Alignment by rule-writing turns into Google's search-quality problem — 20 years, thousands of people, still not finished
The industry already settled what a bug is once, with severity and priority, and it has to do it again for statistical systems
1. Congress Panics on Cue
The episode was recorded at the Foundation for American Innovation's Washington office, in a week of new AI bills — among them the Stop Rogue AI Act from New Jersey congressman Josh Gottheimer. Sinofsky's answer was to go back forty years first.
As a college freshman in 1983 he was one of two people in a 100-person dorm with a computer. One morning, as he tells it, men in suits arrived, confiscated the other machine and arrested its owner. The student turned out to be part of one of the first hacker groups, working on GTE Telemail, an early dial-up mail service whose customers included the US government and NASA.
No law had been broken
it turns out that there was no crime committed. It wasn't illegal to do what they were doing. There was no law that they were violating.
Steven Sinofsky
What followed was congressional hearings with clips of the new movie War Games played in the room, and several years later the first federal anti-hacking statute, signed by Ronald Reagan. Four years after that, back at Cornell for homecoming, he walked into a press conference announcing the Morris worm — the first prosecution under that act.
He says the cycle has repeated often enough that he is comfortable with it, and the reason he is telling the story is that the legislative noise is not the problem he wants to talk about.
He is relaxed about the politics and not about the vocabulary
So I'm very comfortable with all the Congress people posing legislation and having hearings and being hysterical.
Steven Sinofsky
2. Misaligned Means Buggy
The thing that makes his head explode, in his phrase, is the language the labs use about their own products.
A failure to be aligned is a defect in a piece of software
What it means is there was a bug in the software. The software did not do what we thought it would do.
Steven Sinofsky
The objection he raises to "misaligned" is that it smuggles in a mind. It implies the system considered something, weighed a moral question, held a rule and broke it. He points out that the industry has shipped statistical software that got things wrong for decades without reaching for that vocabulary.
Nobody prayed over a spreadsheet
And we didn't sit around like in a circle and pray to the gods of alignment that the spreadsheet would do math correctly.
Steven Sinofsky
And nobody blamed demons for a lost file
When Word ate your file and deleted all your content, we didn't think that demons had taken over Word and made it do things against the will of man.
Steven Sinofsky
The fact that a model picks its next step from statistics rather than from a written branch changes nothing about whose fault it is. If the wrong word comes next, he says, the statistics are wrong, and wrong statistics are a bug in a product.
His live example is a car. When full self-driving runs a stop sign because the software read the sign as a Christmas tree, nobody says the system failed to align with the December holiday season — they say it is a bug, and a dangerous one, because that sign means something. Tesla and Waymo, he notes, are the ones that already build for this: extra cameras, diagnostics, telemetry, the apparatus you put around software when a life depends on the output.
3. Pause and Add Telemetry
Asked what he makes of calls for an AI pause, Sinofsky said he is for one — a different one from the one being debated.
The pause he wants is a quarter spent on instrumentation
So when you talk about an AI pause, of course they should slow down. They should stop adding things and go and add the telemetry, the tools, the logging, the step-by-step debugging, all of the stuff that you would add if your software did anything at all important and you cared if it did it right.
Steven Sinofsky
He had read the bug reports OpenAI published the day before the recording, and his reading of them is that the company cannot see inside its own system well enough to say what happened.
His verdict on the state of the platforms
What I see is the AI models are still in the research project phase.
Steven Sinofsky
That is not a complaint about ambition. It is a claim that the models are being sold as platforms for consequential work while carrying the instrumentation of a research project, and he does not think a statute can close that gap.
Whose job it is
you have one job making software that works. That's it. That's the job. And if you're gonna be worth a trillion dollars, then that turns out to be a really important part of your job.
Steven Sinofsky
4. The Syndogs Bug
The evidence he offers that unexplainable behavior is debuggable is a 1987 bug in the first Windows version of Excel, which wrote the nonsense word "syndogs" into a Windows settings file on disk. Nobody could find the string anywhere in the source.
It looked exactly like possession
It was like Excel got taken over by demons and it wrote this word to the file.
Steven Sinofsky
The fix was brute force. A tester wrote a script that built graph after graph and printed each one on a dot-matrix printer, ran it for two days, and reproduced the write. The cause turned out to be contextual — how much memory the machine had, how long it had been running, which printer and driver were in use.
The choice that separates engineering from mysticism
We could have very easily just concluded it did this. The Excel brain decided to print syndog somewhere. But we didn't. We actually went and debugged it and then we put in all of this telemetry.
Steven Sinofsky
He says the first thing he learned at Microsoft was how to trace buffer overruns and memory allocation, and how to build that instrumentation into every product — including internal builds of Excel carrying extra logging that shipped to beta testers. He has no visibility into how much of that exists inside a frontier model. What he can say is that the published reports do not suggest enough of it.
5. A Bigger Blast Radius
The objection put to him was that the bug analogy understates the stakes: the surface area of an AI failure is more catastrophic than a spreadsheet's, and today's surface area is the smallest it will ever be.
Sinofsky's answer was that he has been on the wrong side of that argument before. Windows developers told IBM's mainframe people that their software was different, new and cheap, and that the comparison was unfair.
The excuse they used
we run on 64K, not one megabyte. And we cost $5,000, not $5 million.
Steven Sinofsky
Then Excel was used to decide how planes fly, Word to file briefs at the Supreme Court, and PowerPoint to present the Challenger failure. The excuse expired. His own worst example is a worm that spread through Outlook, Word and internet email — which he dates to 1998 — and which he says brought the US and world economy to a halt in a day.
The number in the next day's papers
And the next day the headlines were $12 billion worth of damage to the US economy in one day.
Steven Sinofsky
Microsoft stopped Outlook development on the spot. His point is that scale is not an excuse, it is the assignment — and he has a line from a technology journalist he has never been able to argue with.
The answer to "it's complicated"
Sinofsky, that's why you're paid the big bucks.
Steven Sinofsky
There is no higher authority to appeal to, in his account: not Congress, not third-party evaluators, not the laws of nature. This is a tool people built for other people to use.
Two options, and that is the list
You've got to figure out how to make it work or not sell it. Like those are your two choices.
Steven Sinofsky
6. Alignment Is a Rule Pile
He also thinks alignment has a structural problem that has nothing to do with the word.
Alignment written down is a list of permissions
the only way to be aligned is to have a very large set of rules that say what alignment is
Steven Sinofsky
Every rule invites a case the rule does not cover, which produces another rule. He says software people already know where that road goes: over-constrain a system and it stops working, and you get an endless supply of unintended side effects.
The company that has been doing this for two decades
it starts to look like Google and all the work that they constantly have to do to decide how to present search results.
Steven Sinofsky
Twenty years and thousands of full-time people, on his count, and it is still hard. The labs have signed up for a harder version of the same job, because they are not selecting from known results — they are synthesizing the answer. None of that invents itself, and no law makes it get invented.
The host offered a reframe he accepted: that alignment may end up meaning nothing more exotic than a deeply competent team able to handle hard problems.
7. The Y2K Precedent
Asked whether AI alignment resembles Y2K — a real problem, addressed by companies acting on their own incentives, that never became a catastrophe — Sinofsky pushed back on the "never became" part.
Nothing happened because people made nothing happen
There was a lot of worry by professionals and then there was a lot of taking of responsibility by professionals and then nothing bad happened.
Steven Sinofsky
Renting generators, moving machines into secured bunkers, campers in car parks: that is why the date rolled over quietly. There were some Y2K compliance regulations, he allows, but most of the rules were drafted by industry — a cross-consortium of banks, insurers and others writing the standard between them.
What he wants the labs to do now
OpenAI needs to get together with all the model frontier people and the labs and they just need to have much better reporting.
Steven Sinofsky
The model he points at is the CVE. After the viruses of the 1990s, Carnegie Mellon's computer security research center worked with industry and with NIST to produce the vulnerability report everyone now reads: affected versions, the context, the steps required, what other software is hit, which bugs are related, what it derives from.
The reports published the day before, he says, are a first step and read like something else.
On the tone of the disclosure
It still felt more like marketing covering for the event because it still said a lot of stuff like well we are looking into this but we presume at this point
Steven Sinofsky
His standard for that is the Federal Aviation Administration, which does not stand up on the day of a crash and say what it presumes happened.
The discipline he wants copied
They just keep their mouth shut until they know.
Steven Sinofsky
And when the FAA does speak, he says, it gives a second-by-second account with every actor and every environmental condition that led to the incident.
8. Where the Words Came From
Sinofsky was careful not to ascribe bad motives to legislators, on the grounds that assuming bad faith is not how a problem gets solved. He traces the vocabulary instead to 1958 and the Dartmouth summer conference, where he says researchers were already talking about human brains and consciousness, and to a habit he says runs through academic research generally: a cute title is how a paper gets attention.
The terminology is now the obstacle
the terminology is driving people apart because people presume a vast amount of stuff when you use metaphor or allegory or anthropomorphism and it's just not true
Steven Sinofsky
The specific failure is that each term has an engineering meaning and a civilian meaning, and hearings run on the civilian one. Goal seeking, to an engineer, is a curve and a minimum. To a congressman it is a person trying to get an A.
Three words that do not survive the translation
when a normal person like a congressman hears goal seeking they think of a person trying to get an A in college and then when they hear cheating they hear that they did the thing you're not allowed to do and when they hear secretly coordinating they think of spies invading a country they don't think of two pieces of software with a semaphore which is just another form of secretly coordinating
Steven Sinofsky
His precedent for the damage a word can do is "computer virus," which he calls a fluke of timing: the metaphor arrived when AIDS and HIV were everywhere, so the word carried a fear that had nothing to do with software.
Why the 1980s hearings sound the way they do
if you listen to the computer hearings from the 1980s about the first viruses, they're scary in that context because people are literally thinking about death.
Steven Sinofsky
And the reason viruses stopped being an everyday emergency, in his telling, is not that the risk went away. It is that the industry decided it was unacceptable and did the work — most visibly Apple, which turned the result into an advertising campaign.
Fewer viruses was a product decision, not luck
And they told people that Macs are better because they get fewer viruses. Not by some fluke of nature, but because they did a whole bunch of work on the Mac to prevent viruses from happening.
Steven Sinofsky
He adds that the work was difficult for Windows to copy, because of the Windows business model rather than the engineering.
Bonus Insights
Defining "bug" took the industry years, and IBM had already done it
Some people wanted the word to mean data loss only; others counted a crash with the data intact. The settled definition was broad — the software did not do what you wanted — and customers were allowed to file anything.
we eventually decided a bug just means the software and the computer didn't do what you wanted it to do. You weren't happy with the result.
Steven Sinofsky
The database went from thousands of bugs to hundreds of thousands, and the team invented severity and priority to sort them: severity one meant lost data, severity three meant flaky; priority said whether it had to be fixed. Sinofsky says the old hands at IBM told them they had been doing this since 1965.
Telemetry started as one person's throwaway idea
For twenty years, he says, software crashed and nobody knew why; the response was to list the crash sites and fix the top ten. Then someone suggested writing a program that used "this new internet thing" to report its own crashes. The bug count went from none they could find to more than the team could ever finish.
The weather forecast is the analogy he prefers
A forecast is a statistical model with inputs and outputs, and it is wrong on a regular basis. When a hurricane track misses, the meteorologists get together and look at the model.
the weather people all get together and look at the model and talk about the bugs in the model that led to the incorrect forecast
Steven Sinofsky
Nobody, he says, appeals to Zeus.
A thesis nobody was allowed to test
The graduate student who coined "computer virus" wrote a thesis arguing there is nothing to be done about them — they are permanent. He could not prove it, because his university would not let him use the computer to run the test, on the grounds that he was writing about how dangerous the software was.
He knew he was ranting
Okay, I'm done ranting. Ask me a question.
Steven Sinofsky
A few minutes later: "I'm a little wound up."
The closing ask
Sinofsky finished by saying the AI industry needs to listen to its operational security people, because the incident involving Hugging Face and OpenAI was a security failure — no intelligence and no consciousness in it anywhere.
otherwise we're going to get legislation no one is going to be happy with because they don't understand what our industry is saying.
Steven Sinofsky
Sinofsky's bottom line is that AI failures are software defects, that the labs can fix them with the tools software has used for forty years, and that every hour spent arguing about alignment as a metaphysical problem is an hour of legislation being written by people the industry confused on purpose.
Products, Companies & Tools Mentioned
OpenAI (Published bug reports the day before the recording that Sinofsky says show the company does not know what happened; he wants the labs to agree a common reporting standard)
Microsoft, Windows, Excel, Word, PowerPoint and Outlook (His own case study: the "syndogs" bug, the decision to build telemetry into every product, and the software used to fly planes, file Supreme Court briefs and present the Challenger failure)
Tesla and Waymo (Named as the firms that already wrap statistical software in cameras, diagnostics and telemetry because lives depend on the output)
Apple (Cut virus rates on the Mac through deliberate engineering, then advertised the result — work he says Windows found hard to copy for business-model reasons)
Google (His analogy for alignment-by-rules: 20 years and thousands of people deciding how to present search results, and still not done)
IBM (The mainframe engineers who told the Windows team their software was crappy, and who had been doing severity and priority since 1965)
Hugging Face (Named alongside OpenAI in the incident he says was a security failure rather than anything to do with intelligence)
CVE (The vulnerability reporting standard built by Carnegie Mellon's security research center with industry and NIST — the template he wants for AI incident reports)
Carnegie Mellon University and NIST (The research center and the standards body that produced that reporting regime after the viruses of the 1990s)
Federal Aviation Administration (His benchmark for incident disclosure: say nothing until you know, then account for every actor and condition)
GTE Telemail (The early dial-up mail service, with the US government and NASA among its customers, that the 1983 hacker group had broken into)
America Online (What "the internet" mostly meant at the time of the worm that took corporate email down)
Books & Resources Mentioned
Hardcore Software – Steven Sinofsky (His Substack and book on the Microsoft years, named in his introduction)
Stop Rogue AI Act (Josh Gottheimer's bill, the piece of legislation that set off the conversation)
War Games (The 1983 film whose clips were played at congressional hearings on hacking)
Listen to the full episode
🔴 YouTube | 🔗 Episode page
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: