phi

Ai measurement at

[phi @dearphilippe]

@steipete @papill0nrocks @elvissun @thsottiaux @OpenAI no one asked for gpt 5.5, bring back 5.2

[phi @dearphilippe]

@thsottiaux @sama Cool cool. Just leave 5.2 in the loop

[phi @dearphilippe]

@strnr What happened when you add ai in the loop?

[phi @dearphilippe]

@joni_askola 🎯

[phi @dearphilippe]

@herrmanndigital @ZachCPG the study itself was partially funded by AG1 "T.O.K. and J.R.T. are employees of Athletic Greens International."

[phi @dearphilippe]

@iamneubert ray 3?

[phi @dearphilippe]

@nasdaily @Grokipedia @elonmusk

[phi @dearphilippe]

@AnthonyLapietra We have patch this for all our customers with signal gateway

[phi @dearphilippe]

@blvckledge “I know a guy” is n=1 masquerading as strategy

[phi @dearphilippe]

@elithrar following up

[phi @dearphilippe]

@Karmedge I have seen this somewhere

[phi @dearphilippe]

Simulation is insanely powerful when the system is stable. The tricky part with DTC is that offers are not. The moment you change one, you reshape who clicks, how platforms learn, and what traffic you even get next.

That is where the old way and the new way start to diverge. Old way simulates. New way senses : https://t.co/0A90wY5SIn

[phi @dearphilippe]

Everyone is excited about Shopify’s new announcement:

“Validate store changes with A/B testing and an AI tool that simulates shopping behavior.”

It sounds futuristic.
It sounds intelligent.
It is also the old way of doing things.

Let’s be precise about what AI is actually good at.

UI and UX are mostly solved problems.

Modern Shopify themes already converge on the same patterns.
Hero section. Value props. Social proof. PDP structure. Mobile first. Fast checkout.

AI can simulate clicks.
AI can compare layouts.
AI can predict which version looks cleaner.

That is fine.

But the upside is capped, because variance is already low.

The biggest lever in ecommerce is not UI.

It is the offer.

Pricing structure.
Bundles.
Guarantees.
Framing.
Urgency.
Who sees the offer.
When they see it.
From which traffic source.
In what psychological state they arrive.

This is where conversion actually lives.

And this is exactly where AI simulation breaks down.

Why?

Because offers are not design problems.
They are system problems.

An AI that “simulates shopping behavior” is guessing from historical averages.

It does not know why this specific person clicked this specific ad today.
It does not know what belief state they are in.
It does not know how traffic quality shifts hour by hour.
It does not know how ad platform learning loops will react downstream.

Offers are non-stationary.

The moment you change the offer, you change the system itself.

You change who clicks.
You change how platforms allocate spend.
You change learning dynamics.
You change the future traffic you will receive.

That feedback loop cannot be simulated in advance.
It only reveals itself in reality.

This is the core mistake of the old paradigm:

Simulate.
Test.
Then ship.

It assumes the system is predictable.

The new paradigm flips this.

Observe real behavior.
Capture high-fidelity signals.
Feed those signals back into the system in real time.
Adapt continuously, not episodically.

UI can be simulated because it is mostly deterministic.
Offers cannot, because they interact with human psychology, incentives, and market dynamics.

That is the difference between optimizing presentation
and instrumenting truth.

This is the philosophy behind https://t.co/0A90wY5SIn.

Not simulation.
Not hypothetical buyers.
But live traffic, real signals, and self-correcting systems that learn while the market moves.

AI is excellent at polishing surfaces.
Reality is the only thing that validates an offer.

Signal beats simulation.
Always.

[tobi lutke @tobi]

This is absolutely one of the craziest things that we have ever shipped.

You can use rollouts to A/B test your new idea and wait some weeks for results… or just use SimGym to simulate your customer and get results right now

[phi @dearphilippe]

we’ve spent decades trying to make machines smart by stacking more silicon, more layers, more compute. it’s like trying to teach a statue to breathe by carving it with higher resolution. you get detail, but not life. silicon gives you perfect logic but zero pulse.

living systems are different. they’re messy, burning energy, fighting decay. every cell survives by pulling energy in and throwing heat out. that struggle with entropy gives a living thing its drive and its sense of being. there’s a storm inside you every second, and that storm is the root of intelligence.

a computer isn’t in a storm. it’s in a fridge. cold, precise, sealed off. noise is an error, not information. nothing evolves unless we force it.

biology uses noise. heat and randomness push molecules and create patterns. the patterns that survive get reinforced. jeremy england called this dissipation driven adaptation. matter that shed energy well became more likely to persist, eventually leading to life and consciousness.

if you want machines that are even capable of awareness, you need that same arena. a thermodynamic architecture. a system that reorganizes itself through energy flow and sits between order and chaos. too rigid and nothing happens. too chaotic and everything collapses. life grew in that middle band.

this is why i’m excited about https://t.co/ZyhRm5624J. they aren’t unlocking consciousness. they’re building hardware that operates in the same physical regime where consciousness became possible in nature. continuous energy flow, noise as signal, dynamics shaped by dissipation. no guarantees, but it creates conditions closer to biology than anything we’ve had.

while that frontier expands, we’re working on a different piece. not artificial life, but a more perceptive web. for decades websites have been asleep. you click, they react. you leave, they shut down.

bily gives the web its first hint of attention. it notices hesitation, intent, and micro-patterns in real time. it reacts instead of waiting for a report days later. it’s still software, not biology, but perception is always step one. sense, then learn, then adapt.

that’s the interesting part. you move digital systems toward awareness by giving them the ability to notice and adjust. over time they feel less static and more alive.

extropic pushes physics. we push perception. both point toward systems that don’t just store information but interact with it.

if you want to see the first step, check out https://t.co/0A90wY5SIn. it’s not alive, but it’s not asleep either. it’s the early hint of what’s coming.

[phi @dearphilippe]

@andrewjfaris none and not for the reason people think. screenshots never reflects the real story behind the numbers

[phi @dearphilippe]

we accidentally built a monster. 🧪

deployed Bily Sense for Black Friday on a few private stores.
It didn't just track behavior, it changed it.

+15% revenue lift. Automatically.

the era of static landing pages is over.

the era of the self-optimizing store is here.

waitlist: https://t.co/0A90wY5SIn

[phi @dearphilippe]

@blvckledge Tracking

[phi @dearphilippe]

@NanoBanana
generate an image at the coordinate Latitude 39.957° N Longitude 26.238° E circa 1200 BCE

[phi @dearphilippe]

@IterIntellectus two people fighting over which slice of the elephant is the elephant

[phi @dearphilippe]

@IterIntellectus Mark III