AI isn't one thing
Same person. Different AI. Different starting point.
Models have different defaults. Yu is building the layer that carries how you work across them.
Model behaviour changes by version, system instructions, product surface and release. The comparisons below are illustrative working defaults, not permanent measurements.
The models
Where each one starts
Working comparison. Not a benchmark.
They're not the same machine.
Every model was trained on different material and then shaped by different people. That's why the same question gets a different answer depending on where you ask it.
You can't change this layer. You can pick which one you're in, and you can tell any of them how you work.
These are defaults and they move with every release.
None of these is better than another. They are different starting points, and the positions above are drafted defaults kept for illustration, not scores from a benchmark run. ChatGPT, Claude, Gemini and Grok are products of their own companies and are not affiliated with Yu.
What this means for a person
You end up being the translation layer.
Switch tool and you relearn all of it: how much it pushes back, how long it answers, whether it starts or asks first, how formal it sounds, how readily it acts. Nothing you taught the last one carries over, so you explain yourself again from the top.
Today
- You
- explain yourself
- model A
- You
- explain yourself again
- model B
- You
- explain yourself again
- model C
What Yu is for
- You
- your Yu
- whichever AI you choose
One description of how you work, carried in front of whichever model you happen to be using.
Baseline and with Yu
One directive, four different starting points.
Preview. Direction, not live
Your directive, sitting with you only
While I'm exploring, help expand and test the idea. When I'm about to commit, switch into challenge mode.
Starts from its own defaults.
Starts from its own defaults.
Starts from its own defaults.
Starts from its own defaults.
Each model starts where its own defaults sit. Nothing knows how you work, so what you get depends on which tool you opened.
Yu doesn't make the model smarter. It changes the interaction policy around the person.
This section is product direction, not a shipped integration. Yu does not currently apply your directive automatically inside ChatGPT, Claude, Gemini or Grok. Today you carry your instructions across yourself. Automatic cross-provider application is what we are building toward.
The eventual round trip
A profile that learns from what actually happened.
The point of writing down how you work is not a one off personality prompt. It is a loop, so a condition that fires at the wrong moment can be corrected and the description gets closer to you.
Your Yu
How you work, held in one place.
Portable instructions
The part a model can actually read.
AI provider
Whichever tool you chose that day.
Response
What it gave you back.
You use it
In real work, not in a test.
Correction or misfire
It pushed back while you were still exploring.
Yu refines the condition
The when, not just the what.
Back to your Yu, one round more accurate than it was.
Why context matters
A correct instruction can still fire at the wrong moment.
Not enough
Challenge me.
Always on. It arrives while you are still finding the idea and kills it before it has a shape.
Better
Challenge me when I'm about to commit. While I'm exploring, help me expand and test.
The same preference, with a condition attached. This is the part Yu's research is about: not what you want, but when you want it.
The whole build is on the record.
The Live Lab holds what exists, what the evidence supports and what is still an open question.