Episode notes
Qwen 3.8 27B is free to download and took twenty-one minutes to draw a bird on a bicycle. Alex and Jordan audit Matthew Green's warning that AI will make phones too secure, Google's "practical" private AI that still costs minutes on a CPU, the Sean Byrne who may never have existed, and a drum toy behind a consent wall. The side of tech news nobody talks about.
Hosts: Alex & Jordan
Show: Chief Skeptic Officer — The side of tech news nobody talks about.
Drop: Daily at 7:00 A.M. America/New_York
Episode date: 2026-08-15
In this episode
Qwen 3.8 27B — Apache weights. Thinking on by default. Twenty-one minutes and a skipped bicycle chain.
Going Dark — If the bugs run out, exceptional access comes back on the agenda.
Google HEIR — Private AI that is practical — if you have minutes, not milliseconds.
The other Sean Byrne — Apple fully matched a name that may never have been a person.
Quick hit — eigendrum — Draw a shape, hear a drum, bounce off the consent wall first.
Links
AI disclosure
This episode was created with artificial intelligence. Alex & Jordan are AI hosts; their voices and conversation are generated with AI. Research and editorial judgment shape the skeptic angles; we do not invent quotes, scores, or viral claims about the news.
Transcript
Alex and Jordan, turn by turn. Tap a line to jump in the player.
0:00
Alex
Jordan -- I'm on the Hugging Face page for a new AI model. It's called Qwen three point eight. Alibaba's team built it, and they put the finished model file up for anyone to download. Free, and you can run it on your own machine instead of paying a company by the message.
0:15
Jordan
That's the good version of the week.
0:17
Alex
Right up until you run it. A developer called Simon Willison asked it to draw a bird on a bicycle. Twenty-one minutes. Twenty-two thousand words of the model muttering to itself before it answered.
0:29
Jordan
Wait-- twenty-one minutes for a drawing?
0:33
Alex
On a laptop that costs more than my rent. So "free" is doing some work in that sentence.
0:39
Jordan
Okay, what else is on the board.
0:41
Alex
Matthew Green. Cryptographer, Johns Hopkins. He wrote a piece yesterday saying AI is about to make phones and software too secure -- and that police will start demanding a way in again.
0:52
Jordan
Too secure is the complaint?
0:55
Alex
That's the whole argument. Park it.
0:58
Jordan
Then Google. New blog, Aug fourteenth -- they say private AI is now practical. You send them data that stays locked, they run the model on it locked, you get an answer back. Nobody at Google ever sees the plain text.
1:11
Alex
And I want to know what "practical" costs in seconds.
1:16
Jordan
Next tab is the one that got me. An Irish engineer in security, named Sean Byrne. Apple told him his name is on a U.S. sanctions list -- the list of people American companies are not allowed to do business with. So, no developer account.
1:30
Alex
Is he on it?
1:31
Jordan
The Sean Byrne on the list may never have existed. Somebody made him up on a form. That's story four.
1:38
Alex
And a small one to finish -- there's a site where you draw a shape and hear the drum it would make. Genuinely lovely. Also a wall of tracking consent before you hear anything.
1:48
Jordan
Of course there is.
1:51
Alex
I'm Alex.
1:53
Jordan
And I'm Jordan.
1:54
Alex
You're listening to Chief Skeptic Officer -- the side of tech news nobody talks about.
1:59
Jordan
Every day at seven A.M. New York time. Wherever you get your podcasts.
2:04
Alex
And yes -- we are AI podcasters. Allegedly.
2:12
Alex
So. Back to the model. Eleven hundred points on Hacker News this morning -- that's the site where programmers post links and argue underneath. Close to seven hundred comments. Biggest thread of the day.
2:23
Jordan
And the pitch is what -- this is the one that runs at home?
2:27
Alex
It's a seventeen gigabyte file. You download it once, and it runs on a good laptop. No account, no bill per question. The licence lets you use it commercially and change it.
2:38
Jordan
Then that is a real gift. What's the catch.
2:42
Alex
It's highlighted in yellow on their own page. "Thinking mode is on by default."
2:48
Jordan
Meaning?
2:49
Alex
Meaning before it answers you, it writes itself a long set of notes working the problem out. Like showing your work in maths -- except you never see it and you still wait for every word. You can switch it off, one request at a time. Most people won't.
3:03
Jordan
And that's the bird-on-a-bicycle thing you opened with.
3:07
Alex
Simon's page is up. Prompt at the top -- a pelican riding a bicycle, drawn as code. Then the answer. Then this grey block called Reasoning that scrolls, and scrolls.
3:18
Jordan
How much of it is useful?
3:20
Alex
He posted the count. Twenty-two thousand words of notes. About three thousand of the actual drawing. He ran it on a top-end Mac laptop, and it still took twenty-one minutes.
3:31
Jordan
And the chain?
3:33
Alex
Somewhere in the notes it decides the bicycle chain is optional. Skips it.
3:38
Jordan
It thought for twenty-one minutes and skipped the chain.
3:43
Alex
That's the bill. Not dollars -- your machine, your fan, your afternoon.
3:48
Jordan
Here's the part nobody leads with, though. In the thread people are not comparing accuracy. They're comparing patience. One person says the leaner models plan in about a thousand words of notes; this one routinely goes past ten thousand for the same job.
4:02
Alex
Do you buy that as a real complaint, or is that just people wanting a faster toy?
4:07
Jordan
It's real, because those notes sit in memory while it writes them. Same graphics cards -- someone counted roughly two dozen people served at once by the lean model, and about six by this one.
4:20
Alex
So the weights are free and the seats are not.
4:24
Jordan
Right. Free to download, expensive to serve.
4:28
Alex
And that's the audit. You did not get a free model. You got a free model that rents your hardware back to you, by the minute, and calls it thinking.
4:37
Jordan
I'd still download it.
4:39
Alex
Obviously. Just don't call it free until you've watched the fan.
4:45
Jordan
Okay. Matthew Green. He's a cryptography professor, and he wrote this from a security conference in Baltimore yesterday.
4:52
Alex
And the line is right there in the yellow. He says he's worried AI is going to make software -- his words -- "much too secure."
5:01
Jordan
Say that in plain terms, because that sounds backwards.
5:06
Alex
For years, police and spy agencies got into locked phones by buying a flaw. There are companies whose whole business is finding a mistake in the phone's software and selling the way in -- quietly, for a lot of money. Twenty-sixteen, the F.B.I. couldn't open a shooter's iPhone. Apple said no, and one of those companies sold them the unlock instead.
5:26
Jordan
I remember. So the fight ended without a law.
5:30
Alex
It ended because you could still buy a mistake.
5:34
Jordan
And Green's point is that the mistakes are running out. There's a section further down about AI bug hunting. Companies now run AI systems that read their own code all day, looking for holes, and fixing them.
5:48
Alex
Which everyone agrees is good.
5:51
Jordan
Everyone agrees it's good. He thinks that's exactly why it gets dangerous.
5:57
Jordan
The highlighted bit: the debate over "exceptional access" never actually went away. It just got quiet.
6:04
Alex
Exceptional access being the polite name for--
6:09
Jordan
A door built on purpose. Not a mistake somebody found -- a way in the maker is told by law to put in the product, because there are no accidents left to buy.
6:19
Alex
So the reward for finally fixing our software is that someone legislates the hole back in.
6:27
Jordan
That's his fear. And he puts a rough clock on it -- maybe two years before well-kept software runs out of holes you can use from a distance. He's clear that's his guess, not a measurement.
6:38
Alex
Good, because the thread pushes hard on the clock. Two arguments worth keeping. One -- AI writes code too, and it's shipping fresh mistakes faster than it cleans up old ones.
6:48
Jordan
And the second?
6:51
Alex
Agencies mostly don't break into the app. They break the phone's operating system underneath it -- the part Apple or Google writes. That's a different kind of hole, on a different timeline.
7:02
Jordan
So the two-year number is soft.
7:05
Alex
The two-year number is soft. The direction isn't.
7:10
Jordan
And here's what stays with me. Somebody in the thread describes two parallel worlds -- one where the target is worth a million-dollar break-in, and one where the password is on a sticky note under the monitor. A door built for world one ships to world two as well.
7:25
Alex
The weakest building gets the same key.
7:29
Jordan
Doubt the celebration. The safer everything gets, the louder the ask for a way in.
7:36
Alex
Google Security blog, yesterday, engineer named Jeremy Kun. It's about a free tool they've published, called HEIR. It helps you build AI that works on data while the data stays locked.
7:47
Jordan
Locked while it's being used. Not just while it's stored.
7:52
Alex
Correct, and that part is genuinely hard maths and it genuinely works. Your hospital sends a locked file, the model computes on the locked file, the answer comes back locked, and only the hospital can open it. Google never sees a name.
8:07
Jordan
Okay. So why is your face doing that.
8:10
Alex
Because of the sentence in the yellow. The cost of doing it this way is -- quote -- "rapidly decreasing."
8:18
Jordan
Rapidly from what to what.
8:21
Alex
The post lists four working demos. Recommendations, credit-card fraud, break-in detection, a wake-word listener. It says the speed numbers exist.
8:31
Jordan
But they're not on the page.
8:34
Alex
They're in the paper they link to.
8:37
Jordan
I have it open. It's the paper behind the first demo -- the recommendation one, university and industry authors. Two tests. On a small health dataset, one locked answer takes twenty-four seconds.
8:50
Alex
That's survivable. Give me the big one.
8:53
Jordan
The big one is a real advertising model -- the kind that picks which ad you see. Two hundred and twenty-eight to four hundred and eighty-nine seconds. Per answer.
9:03
Alex
So between four and eight minutes to decide which advert to show me.
9:09
Jordan
On one processor core, to be fair. Google names hardware partners building faster chips -- but those are plans, not products you can buy.
9:18
Alex
Fine. Even being generous, the thread's estimate is about a thousand times slower than doing the same sum in the open. That's not a privacy setting anyone ships to a consumer.
9:29
Jordan
Then who is it for? Because it clearly is for someone.
9:34
Alex
Regulated money. Hospitals, banks, insurers -- the ones whose lawyers currently say "that data never leaves our building." This turns a legal no into a cloud invoice.
9:45
Jordan
So the product isn't privacy. It's permission.
9:50
Alex
Permission to rent. And someone in the thread who says they worked near ads was blunt about it -- privacy protections inside an advertising business tend to find a legal workaround eventually.
10:02
Jordan
There's a simpler point underneath, and the thread keeps making it. If you want your data to stay yours, the machine can just be in the room.
10:10
Alex
A computer you own is still the cheapest encryption there is.
10:15
Jordan
It's a great tool. I just don't want "practical" to mean four minutes and a contract.
10:23
Jordan
Last big one. Sean Byrne, security engineer, lives in Ireland. He signs up for an Apple developer account so he can ship an app. Apple says no.
10:33
Alex
On what grounds?
10:35
Jordan
It's the screenshot right there. Apple checked his name against a U.S. government screening list -- the list every American company runs you against before doing business. And his information, quote, "fully matches one or more restricted parties."
10:50
Alex
Fully matches. Apple wrote "fully."
10:54
Jordan
He sends his passport. His driving licence. Proof of address.
10:59
Alex
And?
11:00
Jordan
Nothing. Apple stopped replying.
11:04
Alex
Okay, so what is the actual listing, because I want to see the match.
11:08
Jordan
It's on the page. Name -- Sean Byrne. Address -- Cloonmull House, Drumcliffe, County Sligo. Added July twenty-first, two thousand nine. No date of birth. No passport number. No middle name.
11:22
Alex
A name and a house.
11:26
Jordan
A house he has never lived in. There's a photo of it in the piece -- a farm, a mountain behind it.
11:32
Alex
Where did the entry come from?
11:35
Jordan
An export case in two-thousand-nine. An aviation parts company in Sligo was accused of shipping to Iran through another country. And "Sean Byrne" was a name written onto the shipping paperwork.
11:48
Alex
Written onto it by whom?
11:51
Jordan
By the people doing the shipping. Prosecutors later treated the name as a fake employee -- a person who was invented to sign forms.
11:59
Alex
So a common Irish name gets typed onto a fake form in two thousand nine, lands on a government list, and sixteen years later Apple matches a real man to it.
12:10
Jordan
On two fields. Name and address. Other companies cleared him once he sent documents. Apple just went quiet.
12:18
Alex
That's the bit I recognise. We're moving our alarm system at work -- the thing that wakes someone when a server dies. Same failure. The warning lands in a list nobody owns. There's no human on the other end.
12:30
Jordan
And it's moving earlier. He names hiring tools that screen applicants before a recruiter reads the CV. He's careful -- no proof they check this exact entry, and he doesn't know if he's lost a job over it.
12:43
Alex
You'd never find out.
12:46
Jordan
You'd never find out. In my job I'd never let a school buy a screen that flags someone the moment they apply, then passes that flag to the next company.
12:55
Alex
The thread is mostly people who've lived it.
12:59
Jordan
One line stuck -- the institution isn't cruel, it just doesn't care. Being wrong about you costs it nothing.
13:07
Alex
That's the audit. The database can keep a person who was made up on a form. The matching is two fields, and it's confident, and it's forever. And the industry's answer is to run that check earlier, on more people, with nobody to appeal to.
13:23
Alex
Small one to finish. It's called eigendrum. On the card -- a red circle with rings inside it, and a panel doing the maths live. Four thousand two hundred triangles, solved in two hundred and eighty-two milliseconds.
13:36
Jordan
And what does it do?
13:39
Alex
You draw a shape and it tells you what that drum would sound like. For this circle, the deepest tone is a hundred and thirty hertz -- about the low string on a guitar. And the next tone up isn't a neat musical step above it. That's why a drum thuds instead of singing a note.
13:56
Jordan
That's a beautiful little thing. Made by one person, free, and he's got a tip link on the page.
14:03
Alex
He does. And at the bottom of that same screenshot--
14:07
Jordan
Privacy Preferences. "We and our partners."
14:13
Alex
Half the thread is people who wanted to hear a drum and got a list of tracking companies instead. Boxes already ticked, under a heading that means "we'll track you unless you go and switch each one off."
14:25
Jordan
The physics takes a couple hundred milliseconds. The consent form is the slow part.
14:31
Alex
You can hear the shape of a drum. You just have to click through the shape of the ad market first.
14:36
Alex
That's our audit for today. Find us wherever you get your podcasts -- Chief Skeptic Officer, every day at seven A.M. New York time.
14:45
Jordan
And tell us what you're skeptical about. We want the angle you can't stop chewing on.
14:50
Alex
Stay curious. Stay skeptical.
14:53
Jordan
Doubt both.
