ChatGPT adds a PIN lock for your conversations

Welcome to The Signal from Superpower Daily with Maya and Theo

So picture this

You hand someone your unlocked phone to show them a funny meme

and they accidentally swipe over to your ChatGPT app A

very modern panic attack Right

Suddenly

they're reading your prompts about you

know how to ask for a raise or something Well

ChatGPT

is rolling out a new pin lock for conversations today But

the real story is what that lock actually protects and well what

it completely ignores So

starting October 20th 2026 OpenAI

is pushing out this Locked Chats feature It's

hitting free Go

and

Pro users across all their apps You just

tap a chat and lock it right Exactly

It

completely vanishes from your history your search

your pinned list You can

only get back into it with a four digit pin or your device's biometric security like Face ID On

the surface I mean that sounds amazing like

a digital safe But

I'm always pretty skeptical when we talk about cloud privacy My

Face ID doesn't magically stop OpenAI from reading the data on their end It absolutely

does not This is

purely

a device level shield It just

stops the person physically

holding your phone Yeah not the servers

Right The data still

travels

to OpenAI's data centers There is no end to end encryption here at all So what

about the training data If I

lock a chat does that opt me out of the improve the model for everyone setting No

it doesn't Okay

That

is the biggest caveat of this entire update Locking a

chat and data sharing are completely separate controls Wow Yeah

if

your main setting allows training data your locked

chats are still perfectly fair game to train future models Okay so

it's

really just a modesty screen What happens

with the memory feature then Because the

AI learns facts about you over time

Does it keep those walled off It does

but it's

tricky They built

a bifurcated memory pool If you

start a brand new chat lock it

and tell the AI a secret that new

fact stays trapped inside the locked sandbox Okay well

that makes sense But here's

the trap door If

you take an old unlocked

conversation that already has sensitive stuff in it and you move it to the locked folder today the

AI does not forget what it already extracted Oh

so the horse already left the barn Exactly

You

just hid the original transcript from yourself The

memory is still sitting in your main profile unless you manually go in and delete it That

is going to trip a lot of people up And

speaking of tripping up what

if I just forget my PIN Then

you lose the data Permanently

Wait

really

Yeah

Resetting

your PIN literally incinerates all your locked chats and the files attached to them Just

completely wipes the vault I

guess that makes it secure against someone brute forcing it It's

a scorcher security protocol It treats

the data as highly disposable if authentication fails So why

do we even need this if we have temporary chats that don't save at all Temporary chats

are for you know

quick disposable questions Locked chats

are for ongoing sensitive projects where you actually want the AI to remember the context over a few days or weeks just in an isolated way Got it

Though I

did notice you can't lock group chats or anything from a shared link Right No

collaborative

features are allowed And what

we really need to watch next is the enterprise rollout TechCrunch is

reporting it's coming to business and enterprise plans soon That's going

to be an absolute nightmare for IT admins managing shared corporate iPads It will

be a massive compliance headache yeah That

wraps

up the new privacy controls Let's shift gears Next up

let's talk

about a major milestone in AI logic Sid Chatterjee

a computer scientist at UT Austin built a

program called Darmok And it

just beat the human champion at the Times Cryptic Crossword Final It beat

Liam Hughes by about three minutes which is

incredible if you know anything about cryptic crosswords Right This

isn't

your standard American crossword where it's basically just trivia Cryptics are

basically hostile riddles Exactly American

crosswords are just massive database retrieval problems Cryptics require

you to parse two overlapping grammars simultaneously I want

to talk about the clue that actually won it for Darmok because it's wild The clue

was wolf trap missing its lowest cross piece five

letters Now

to a human that sounds like a literal question about a broken animal snare Right

a

physical trap Even

a 14 time human champion Mark Goodlife stumbled on this exact clue But

Darmok answered snarf S N

A R F Let's unpack that logic

because it's brilliant The definition part of the clue is just

the word wolf as

in the verb to wolf

down

your food Yeah to snarf And the second part of the clue is trap missing

its lowest cross piece So a synonym

for

trap

is snare Right S N

A R E But here's the crazy part The AI had

to visualize the physical printed capital letter E If you take away the lowest horizontal line

the lowest cross piece it

turns

the E into an F It's pure geometric wordplay Exactly The AI had to basically act like

a detective looking at the font of a

ransom note completely ignoring the meaning of the words themselves

It changed snare to snarf just based on physical shape Which is an insane

leap for a language model that

literally doesn't have eyes

How did Chatterjee actually achieve this

Because just last year Darmok

failed

miserably at this tournament He made two big changes

First he threw an absurd amount of compute

at it He uses an ensemble of four different models now

Gemini 3 1 Pro and Clod Fable 5 do

the heavy lifting

And he keeps OpenAI's GPT 5 and GPT 6 Astra as

backups right Yeah exactly If the main models hit

a deadlock the backups step in

But the real breakthrough was teaching it delayed gratification

Before if it guessed an answer it would just lock it into

the grid immediately Which

completely

poisons the intersecting words if that first guess is wrong It creates cascading errors Right So the new protocol forces Darmok to attempt every single

clue in the entire puzzle before it is

allowed to write down a single letter It maps

out the whole probability space first

That is a

brilliant application

of search space pruning Are there any blind spots left It still struggles

seemingly at random but specifically with one human puzzle maker who goes

by the pseudonym Enigmatist His British

cultural references

and syntax are just too complex right now So what's the

next frontier here Generating them Writing a cryptic

clue that functions flawlessly as a logic

puzzle but also reads like perfect natural English prose That's the ultimate test of semantic

control Well from the rigid logic of a crossword we are moving to something incredibly fluid The Boston Symphony Orchestra is

premiering a symphony that treats an

AI not as a synthesizer

but as a live collaborator

It's called Already and

Not Yet written by composer Todd Mockover

It premieres on October 22nd 2026 at Symphony Hall And this isn't just triggering

pre

recorded sequences The AI is actually listening and interacting with the human

musicians in real time Right The system was built by

Manaswi Mishra at the

MIT Media Lab And

the AI's role actively shifts as the piece

progresses The structure is fascinating In the second movement the AI's job is

to isolate individual players It finds a specific

instrument in the chaotic noise and actively spotlights it But then in the third movement it flips It stops highlighting the individual and tries

to find common sonic ground between

all those isolated solos

It tries to weave them back together It acts like a shadow

conductor I love the concept but I have to push back a

little on the autonomy here Is

this algorithm actually making creative decisions on

its own up there That is a

very fair skepticism And no it is not running wild Anna Schon who

is another MIT researcher is physically sitting at a console

mixing

the sound live during every performance So the human safety

net is very much in

place

Plus Makover himself admits the AI doesn't actually understand the music It just recognizes audio patterns Right A massive spike

in frequency doesn't tell

the AI if the violinist is playing with intense anger or

profound sorrow It lacks human intent But Makover actually uses that lack of

intuition to his advantage The weird alien collisions it causes fit

the theme of

the piece perfectly The whole symphony is essentially a metaphor for the

strain of democracy Exactly The three

movements are smash up wake up

and stand up Smash up is just chaotic polarization The core melody

tries to emerge but it just gets completely crushed by aggressive brass and

electronic noise Then wake up

is that second movement where the AI

highlights the individual voices And stand up is

the climax It's

the attempt to rebuild the community finding a new harmony

between individual expression and collective order It's a huge conceptual weight to put on an orchestra Listeners should definitely

track that core melody to see if

it

survives the chaos and stabilizes by the end It sounds

like an amazing live experience Meanwhile let's pivot from abstract art to cold hard economic reality

Bloomberg just reported that former Google CEO Eric Schmidt holds a massive

personal

stake in Anthropic We're talking about a position potentially worth up

to 20 billion Yeah TechCrunch corroborated

it Schmidt apparently owns roughly 1 of the

company To understand how he got that you have to look back

at the origins He participated in the Anthropic Series A

funding round way back in 2021 And he put

in

tens of millions of dollars

which for frontier

AI is basically seed money Right Building foundation

models is incredibly capital intensive And his early bet paid off

phenomenally well Private investors are currently valuing Anthropic between 1

5 and 2 trillion So a 1 stake at

2 trillion is

20 billion

That is just an absurd economy shifting return on investment It really is For context

Anthropic's co

founder and CEO

Dario Amadei is estimated to have about

a 2 ownership stake So

Schmidt an external

investor has half

the

ownership of the founder Though we definitely

should clarify that's just equity not

voting power Correct Anthropic uses a long term benefit trust structure Equity does not dictate

board control

And we also need to

throw a heavy dose of reality at these multi billion dollar figures Yeah that's

a crucial caveat 20 billion on a private valuation spreadsheet is not the

same

as liquid cash We don't even know

if he's quietly sold shares on secondary markets

already Exactly It's paper money until

an

exit happens And that's what we need to watch next When do they actually go public Well the

Times of India reported that Anthropic

confidentially filed draft IPO papers on June 1st 2026 So the regulatory

clock is ticking TechCrunch thinks a listing could happen before the

end

of the year We will definitely

be keeping a

close eye on that ticker Now let's jump into our quick reads Let's

do it First up Meta just disabled the sharing feature for its Muse

AI agent This happened after security researchers at Cleroc found

what they call ghost shares Which were essentially

searchable links to highly sensitive private chats The crazy part is the architecture choice here Muse was actually embedding

the specific titles of user conversations

directly into the web subdomains

It was under cf metaaiusercontent com So

passive DNS records which just monitor internet traffic swept up all these private chat

titles Cleroc found over 6

800 exposed links in a single

database query They

sampled a few hundred and

found confidential executive search notes bank fraud alerts and full company

vacation schedules Meta completely shut down the

feature once the vulnerability was exposed Next up Anthropic released 211 Claude

designed proteins And 17 of them successfully

bind to scorpion toxins in lab

tests This was a

challenge from Citrini Research Claude ran for 64 straight hours used 10

000 GPU hours and designed synthetic proteins that actually latch onto

the CN2 scorpion toxin Independent labs like AdaptiveBio and

Twist Bioscience verified the results The binding

affinity was in the single digit

nanomolar range which means it's an incredibly

tight connection But the

caveat here is massive Right This was purely in vitro testing in a glass

dish This does not prove it neutralizes venom in a living cell or

an

animal It's a research starting point not an actual

anti venom you can use today Finally Sakana ai

just released a new system

called QDREC It's a recommendation engine that uses quality diversity search Instead of just

ranking options it evolves

a diverse set of proposals and generates written reasoning for each one It actually mutates ideas Like taking a

corporate training hypothesis meant for an educator and successfully adapting its logic for

a clinical nurse They used a multi armed bandit algorithm to balance safe ideas with wild mutations But there is a severe hallucination risk built into this because the

system is specifically optimized to sound highly

persuasive It frequently makes up data to support its arguments Yeah During a test on corporate acquisitions the AI confidently claimed

a fictional

airline loyalty program had 100 million members just

to make the pitch sound better Totally fabricated So human expert verification remains

an

absolute bottleneck All right We are moving to our three takeaways from today's deep dive First AI

privacy features are increasingly operating at the device level rather than at the data

center

level meaning user data is still being ingested even

if it's hidden from the screen AI

is moving past simple answers into complex

synthesis

whether that's interpreting visual puns and crosswords orchestrating live musicians

or evolving business hypotheses The gap between simulated

success and physical reality remains wide Whether it's

paper valuations for private AI companies

unverified business data or unproven scorpion anti venoms human validation is the ultimate bottleneck And looking

ahead to tomorrow watch how the enterprise markets respond to

ChatGPT's new local locks Given the high stakes of shared

devices in corporate environments this feature

could trigger rapid shifts

in internal security policies