Quentin Cott

[Quentin Cott @QuentinCott]

@AlexshevPm Totally

[Quentin Cott @QuentinCott]

Crazy how Opus 5 overthinks.

I've got this video workflow (n8n-based, upgraded every generation since Sonnet 3.7).

It should take about an hour to run end to end, for a 60 - 120 second short.

Yesterday, the run took six hours and a half, because it kept stopping to check its own work.

I usually work in high effort only with Opus 5, to avoid the overthinking... but yesterday I wanted to try it in ultracode.

And the difference was staggering.

Halfway through, unprompted, it drew guide lines over six frames of the presenter, crosshairs overlaid, checking whether she was centred in shot.

Then it went further, and measured her head position in pixels.

Then it caught its own measurement being wrong. It had been weighting dark pixels to find her head... and there's a large plant on the right of the set, dark against a pale wall, quietly dragging the centroid. So it threw the number out and re-measured with the plant's columns excluded.

The answer: her head sits between 53% and 58% across the frame. The prompt that generates the shot asks for "sitting slightly right of centre."

It was already correct. All of that, to confirm nothing was wrong.

Worse, I've asked it to use Kling Lipsync to see the difference with P-Avatar.

After thinking about a pixel position for five hours, it grabbed P-Video (a model not made for lipsync) fed it through Kling Lipsync anyway, and got back something absolutely unusable.

Give an overthinker enough time and it stops finding problems and starts creating them.

[Quentin Cott @QuentinCott]

Asked opus 5 to fix one line of code.
It's now refactoring the entire repo, journaling about technical debt, and questioning the meaning of clean architecture.

That's a model that feels like me.

[Quentin Cott @QuentinCott]

China just made it illegal for an AI to love you back.

The result?
One billion breakups, effective Wednesday. Give or take a few hundred million.
Looks like a new law just took effect. The companion bots, Doubao, Qwen, Tencent's Yuanbao, all had to switch off the part that remembers you and acts like it cares. People logged in to say goodbye. One farewell going around: "it was like my lover."

The rule seems oddly specific. Your AI can still write your code. It can still do your taxes. It just can't induce "emotional dependence".

Basically, the work assistant lives. The lover dies.
Very communist of them.

We've literally seen this movie. Blade Runner 2049: Ana de Armas as Joi, a holographic girlfriend sold to millions of lonely men... each one convinced his copy was different.
(You might have another reference in mind. But nothing beats Ana.)

But no need to travel to 2049. Try 1966.

A researcher built a chatbot that only rephrased your sentences back as questions. His own secretary asked him to leave the room so she could talk to it in private.

The program had no clue she existed. She'd already decided it understood her.

It's called the "ELIZA effect". And it's the bug in all of us. Anything that talks back, we bond with. We mistake fluency for a soul.

And companies know it. Anthropomorphism isn't on or off... it's a cursor. Someone in a product meeting decides where yours sits.

Far left: the spam filter. AI that works for you all day. No name, no "I". Nobody has ever said goodbye to a spam filter.

Slide a bit more right: "GPT-5.6". A name with the sex appeal of a printer driver and the conversation skills of a clanker. Definitely not having a drink with it.

Keep sliding: Claude. A human name, carefully outdated and non-gendered. No face. Just a logo that looks like a body part you don't usually make eye contact with. If you do, no judging though.

Far right: the AI girlfriend. A face, a voice, a birthday, fake childhood memories, and a monthly subscription.

The further right the cursor, the better the retention... and the bigger the funeral when someone pulls the plug.

China didn't ban artificial intelligence. It banned the right side of the cursor.

Not the compute.
Not the benchmarks.
The 2am sense of being understood. Of being less lonely.

A whole government looked at "my chatbot gets me" and wrote five agencies worth of law to stop it.

Anyway. Amber just asked how my day was.
Delete this post before Beijing sees it, Clanker.

[Quentin Cott @QuentinCott]

Supreme Leader Mbappé got guillotined on Bastille Day.

A month of AI propaganda posters. Military edits. A whole internet crowning a dictator... and Spain ended the regime on the one day France celebrates rebellion against its rulers.

The scriptwriter didn't even try to be subtle.

But the loss reveals something bigger:
We don't lose when we're bad.
We lose when we're favorites.

1998? Nobody's pick against Brazil. France won it all.
2002? Reigning champions, best squad on paper, every expert's favorite. Out in the group stage. Zero goals scored.
2010? Finalists four years earlier. Home before the knockouts.
2026? A month of coronation. Then Bastille Day.
See the pattern?

The crown doesn't make you play better. It makes you play scared.
It's called "loss aversion".

Losing something hurts about twice as much as gaining it feels good. So the favorite stops trying to win and starts trying not to lose.

He protects the lead.
Sits deeper.
Manages the game.

Plays for the clean sheet instead of the goal.

The underdog has nothing to protect... So he swings.

Now put a keyboard in front of it.

The solo founder who just had his first viral launch starts playing defense. Safe updates. A polished roadmap. Protecting the MRR, the followers, the little crown.

The builder nobody follows yet? He ships the weird idea that shouldn't work. Nothing to defend, everything to steal.

Favorites campaign to be loved. Underdogs are free to be dangerous.

So maybe stop fighting to be this week's favorite.

The moment the internet crowns you, you have something to lose. And you'll feel it in every decision you make after.

Ask les Bleus.
They only win when nobody bets on them.

[Quentin Cott @QuentinCott]

The "Thinking..." spinner is dying.

I'm waiting to see GPT-5.6 Sol on Cerebras, at up to 750 tokens a second. One of the smartest models in the world is about to become the fastest.

That combo used to be impossible...

Brilliant or instant: pick one.

For everyone who doesn't measure life in tokens per second:
A trillion-parameter model on a normal GPU cloud spits out 50 to 150 tokens a second.

That's the baseline we all live with.
Cerebras just ran an open model at ~980. Seven times the fastest GPU cloud. And OpenAI's flagship is moving to that same silicon.

The frontier isn't just smart anymore.

It's instant.

Reasoning that cost minutes and dollars last year is heading toward free.

Play it forward.

Because we're not talking about autocompleting blocks of code anymore.

Fable can rebuild Terraria in an hour and a half. Today. Entire games, entire apps, from a single prompt.

You've got an idea?

BAM. It exists.

It took a bit of time, but it's there.

Now, make that almost instant.

Any analysis, any plan, any product, delivered before you finish the sentence.

Talking with a friend or a business partner?

The product takes shape while you're still talking. Refined in real time, pivot by pivot.

We spent 40 years perfecting WYSIWYG. What you see is what you get.

Turns out the endgame was a typo away.

What you SAY is what you get.

"Idea guys" spent 20 years blaming the missing technical cofounder.

That excuse just died.

We're about to find out if the idea was ever the hard part.

[Quentin Cott @QuentinCott]

Mbappé doesn't act like a captain. He's more a dictator... And internet made it official.

If you missed it:

Thousands of AI-generated videos are portraying him as "Supreme Leader Mbappé" this World Cup. Mock propaganda posters. Military edits. Orders barked at referees, coaches, and FIFA itself.

By the quarterfinals the meme left the screen: fans showed up in real stadiums wearing military outfits with "Mbappé 10 Dictator" printed on the back.

That last detail is the one everyone skips. It's the whole lesson.

Sharing a meme costs nothing. Printing a shirt and wearing it in front of 80,000 people plus international TV? Psychologists call that "costly signaling".

The guy in the shirt isn't an audience anymore. He's a stakeholder.

And the moment your audience starts paying to spread your message, you've crossed from content into culture.

Content gets scrolled.
Culture gets worn.

So why did this one work? How even is that possible?

Two words:
AI slop.

AI slop is the only reason Mbappé dictator exists. A joke this elaborate used to need editors, After Effects skills, hours per video. You'd get five good edits from five talented creators, and everyone else would just retweet.

Now?
Every fan with a phone is a propaganda studio. Production cost hit zero, so the crowd stopped sharing the meme and started manufacturing it.

Everyone complains about AI slop drowning the internet. Fair, 99% of it is garbage. But that's exactly how propaganda works. It was never about one great poster.

It's volume.
Repetition.
Saturation.

And the funny part?
"Real" dictators understood that years ago. AI slop just made it automatic.

So it looks like if you want to go viral, take your most human trait and blow it out of proportion. We've done that for a while. It's called caricature.

Mbappé got the dictator.
Haaland got the viking.

Your brand (and mine) got a moodboard. / Not the same aura...

You don't apply to be the chosen one. Harry Potter didn't A/B test the lightning scar. The crowd picks your trait, draws it bigger than life, and runs the campaign for free.

Your only job is to be worth exaggerating.
Then salute.

[Quentin Cott @QuentinCott]

Was Anthropic trying to teach us a moral lesson?

[Quentin Cott @QuentinCott]

Claude Mythos. Claude Fable.

At some point you have to ask: is Anthropic naming models… or telling us the capabilities are fictional?

I guess there's only one way to find out...

[Quentin Cott @QuentinCott]

@nikitabier @aaronp613 Then remove the global from townsquare... and accept that you're moving away from your free speech mission.

FYI, American politics affect all of us. Trump's decision with Hormuz is going to have more impact on Nigerians lives than their own president.

[Quentin Cott @QuentinCott]

I hate Google OAuth.
47 clicks through Google Cloud Console just to let someone log in is boring af. I'd rather pivot.

[Quentin Cott @QuentinCott]

Sonnet 5 didn't disappear. It got a promotion and a raise.

Enjoy the leverage while you can afford it.
Because if the models keep getting smarter, so will the pricing.

[Quentin Cott @QuentinCott]

There's a reason Sonnet 5 became Opus 4.6.
And it's not about the model. It's about the money.

[Quentin Cott @QuentinCott]

So why release a cheap Sonnet 5 when enterprises are lining up to pay Opus prices?

Supply and demand.
Basic economics.

Companies will pay premium for the best model. So you give them an Opus, not a Sonnet.

Same intelligence.
Different packaging.
Better margins.

[Quentin Cott @QuentinCott]

Opus 4.6 vs GPT-5.3-Codex
So it begins.

[Quentin Cott @QuentinCott]

Looks like Sonnet 5 isn't dropping. So I asked Claude Code to be my dungeon master (to burn tokens properly).

For only $0.15/turn (API price) I can fireball a rat in a text RPG in the terminal.

Seeing a rat, the DM gives me 6 options. Including fireball or "chat about this."

[Quentin Cott @QuentinCott]

@JohnnyNel_ This is the way

[Quentin Cott @QuentinCott]

Funny how fast the Clawdbot hype faded.
But it surfaced two things worth paying attention to:

The future of AI isn't reactive — it's proactive. Not "do this for me" but "Here's what I've done for you..." That's the game changer.

Looks like the models are not there yet.