Tally Is Open
HTTPS://tallyy.org
I’m getting tired of watching people talk about AI like it’s a person.
Like it remembers.
Like it “knows” them.
Like there’s continuity behind the interface.
There isn’t.
A large language model is stateless.
Every call is a fresh invocation.
No memory.
No continuity.
No accumulated understanding unless you explicitly pass it in.
That “conversation” you think you’re having is just context reconstruction… over and over again.
So Let’s Be Honest About What That Means
If every call is independent…
Then every call is a decision.
And one of the most important decisions is:
Which model should handle this specific piece of work?
Treating model selection as a one-time architectural choice is just wrong.
Different calls have:
different complexity
different structure
different tolerance for failure
different cost sensitivity
So wiring one model into everything…
isn’t a strategy.
But We Haven’t Solved This
Let’s be clear about that.
We don’t think we’ve nailed this.
What we’ve done is introduce a way to think about the problem.
Right now, that looks like question shapes.
Trying to describe a task in a structured way:
what kind of work is this
how complex is it
what kind of output is expected
And using that to make a decision about which model should handle it.
Maybe that’s right.
Maybe it’s incomplete.
Maybe it’s wrong in ways we haven’t seen yet.
That’s exactly why Tally exists.
This Needs To Be Figured Out Together
You can’t solve this in isolation.
Not across real systems.
Not across real workloads.
Because the moment this touches production:
assumptions break
edge cases show up
costs behave differently than expected
The only way to get this right is through shared usage and feedback over time.
What Tally Does
Tally gives you a way to make model selection a per-call decision…
and learn from it.
You:
describe the task
ask for a routing recommendation
run the call the way you normally would
report back what actually happened
Success. Failure. Cost.
Tally learns from that.
That’s the loop.
No proxy.
No rewrite of your system.
No pretending this is magic.
Just a better way to make a decision… and improve it.
What We Want From You
We’re opening this up because we need signal.
Not interest.
Not compliments.
Signal.
run real workloads
describe tasks honestly
send back real outcomes
push it until it breaks
Watch:
where it saves you money
where it makes bad calls
how quickly it adapts
Then tell us.
How To Get Involved
Sign up:
https://tallyy.org
Install the SDK.
Run it on something real.
Then join the Discord and tell us where it’s wrong… where it’s useful… and where it needs to go next.
One Last Thing
Yes… it’s tallyy.org.
Not tally.
We know.
If this works, maybe we earn the second “y” back… or maybe we can finally afford to buy the one we actually wanted.
Tally is a business.
But if we’re honest… it’s also a tool we built because we needed it.
Now we find out if anyone else does.