Ary

agents + ai platform & funmaxxing @tryramp | founding team @Microsoft Copilot | growth @Meta

[Ary @aryg18]

It’s never been clearer that every product is a wrapper around a data funnel. The surface exists to create value for the user; the structure beneath exists to capture the data that makes that value self-improving.

Most people still treat “data” as an input: something you collect once to train a model. In reality, the winning architectures are loops, not datasets. Capture proprietary interaction data, structure it, and feed it back into the system that generated it. Each turn of that loop makes the underlying system better – and, crucially, makes the user more likely to keep contributing.

This is the invisible flywheel behind every enduring AI business. When OpenAI or Anthropic releases a new model, it triggers a feedback event: every query, rating, and correction expands its private corpus of behavioral data – how humans ask, reason, and refine. Tesla does it with its fleet, Amazon through commerce, Netflix through watch time, Stripe through transactions. Factory through (code) Signals.

What separates winners from the rest is the tightness of that loop: how efficiently a system turns raw interaction into signal, and signal into improvement.

It’s not the size of the model, the polish of the interface, or even the cleverness of the idea. It’s the gradient between usage and intelligence – the rate at which a product learns from its own operation.

This is an uncomfortable truth for most “AI products” today, many of which are simple front-ends without a loop. They route data into someone else’s model, improving the upstream provider while standing still themselves.

Nonetheless, the few that master the loop – that own their feedback, align it with their users, and reinvest it continuously– will compound indefinitely. Each new user makes the system smarter for every existing one.

In the end, every product is temporary, every model replaceable, every interface imitable. The only enduring advantage is the loop.

It is the network effect of the AI age.

[Eno Reyes @EnoReyes]

Today we're sharing Signals: our closed-loop system for recursive self-improvement in software development agents.

Droid learns from interactions with hundreds of thousands of users - detecting improvement opportunities and autonomously implementing them in it’s own codebase.

[Ary @aryg18]

@adisingh Diminished Serendipity

https://t.co/jJ3vVkNZ0E

[Ary @aryg18]

The more time I spend with the latest generation of models, the more I think we’re observing a change in character, beyond a mere increase in capability.

They are undeniably more capable. They reason better, execute better, follow instructions more faithfully, and fail less often. Across almost every benchmarkable dimension, they have improved.

Yet in making models better specialists, we may have diminished their polymathic qualities.

Whilst earlier frontier models were less reliable, they possessed a certain cognitive breadth. They wandered more freely through idea space. They drew stranger connections, explored less obvious paths, and occasionally arrived at genuinely novel insights.

In a way, we’ve asked them to grow up.

The child-like curiosity that once led them down strange and unexpected paths has given way to a more disciplined form of intelligence.

The irony is that some of the qualities we associate with creativity emerge from behaviours that are often difficult to distinguish from inefficiency. Exploration is inefficient. Wandering is costly. Strange associations often look like mistakes until they appear as genius.

Increasingly, progress at the frontier may be defined by our ability to balance exploitation and exploration.

[Ary @aryg18]

The more time I spend with the latest generation of models, the more I think we’re observing a change in character, beyond a mere increase in capability.

They are undeniably more capable. They reason better, execute better, follow instructions more faithfully, and fail less often. Across almost every benchmarkable dimension, they have improved.

Yet in making models better specialists, we may have diminished their polymathic qualities.

Whilst earlier frontier models were less reliable, they possessed a certain cognitive breadth. They wandered more freely through idea space. They drew stranger connections, explored less obvious paths, and occasionally arrived at genuinely novel insights.

In a way, we’ve asked them to grow up.

The child-like curiosity that once led them down strange and unexpected paths has given way to a more disciplined form of intelligence.

The irony is that some of the qualities we associate with creativity emerge from behaviours that are often difficult to distinguish from inefficiency. Exploration is inefficient. Wandering is costly. Strange associations often look like mistakes until they appear as genius.

Increasingly, progress at the frontier may be defined by our ability to balance exploitation and exploration.

[Ary @aryg18]

@JackStettner The General Standard Company of Applied Intelligence

[Adam Shuaib @adamshuaib]

Ultra successful people typically cancelled all backup career options when they were younger.

These people will quit a high-paying job, drop out, sell their apartment or sacrifice life's comforts in the pursuit of success. Society will judge you for making a reckless decision, but in our experience, the ones who kept the backup option to become a consultant or join a big company almost always took it, usually 2yrs in when the work got boring.

People with a backup option live as two people: the version doing the work and the version that exists if you quit. This split personality means you only put 50% of your maximum mental effort into the risky option.

To be world-class at anything, you have to set fire to the boat and bridge behind you.

[Ary @aryg18]

@ian_dot_so these last 12 months

[Ary @aryg18]

@venturetwins Meet users where they are.

https://t.co/w39dewKAIn

[Ary @aryg18]

The history of technology is marked by disjointed leaps - moments when the prevailing approach no longer fits the emerging reality.

We are on the cusp of another such shift: from app-centric experiences to products & services that embed seamlessly into the workflows, platforms, and environments people already trust.

Users don’t want to juggle dozens of apps when most of what they need can be accomplished inside a handful of core platforms — just as @krandiash describes from his own experience. Success lies not in building another destination, but in embedding value seamlessly into the everyday.

Expanded thoughts below.

[Ary @aryg18]

There is far greater asymmetric upside in running toward the lights that have not yet been turned on than in chasing the ones already glowing.

The crowd follows illuminated paths. Once visible, much of the upside has already been captured.

Alpha is the reward for seeing before others see, believing before others believe, and building before others understand.

[Ary @aryg18]

@sebgoddijn correct

[Ary @aryg18]

@diegozaks @iankar_ “a product built with joy always wins”

❤️

[Ary @aryg18]

@ian_dot_so “you’re a world class expert… think step by step”

[Ary @aryg18]

@eglyman tokenmaxxing, “lights on, no one home”

[Veeral Patel @vral]

the best product we built at ramp was the culture - this is the best articulation i've seen yet from krish @krrishd

[Ary @aryg18]

@imjaredz

[Ary @aryg18]

@sebgoddijn Seb, are you, or are you not, in the process of becoming a professional saxophonist

[Ary @aryg18]

@zack_field 👨‍🍳

[Ary @aryg18]

@chongz My boy playing 4D chess

[Ary @aryg18]

@matanSF @ClementDelangue Feynman?

[Ary @aryg18]

A mental model that’s been useful for me is that we’re building for a new species: agents.

They are simultaneously coming online and evolving in real time.

Our convictions, then, emerge from observing, guiding, and learning alongside the species we’re helping bring into existence.

[Ary @aryg18]

@paulodoestech @simoncorry and @tryramp

[Ary @aryg18]

@martin_casado