Q1z

small lab. strange software. real results.

[Q1z @q1z_org]

Most models are built to talk.

We think the more useful ones will be built to decide.

State in.
Typed answer out.
No essay.

[Q1z @q1z_org]

Pivot benchmark

https://t.co/BoR1lnffcQ

[Q1z @q1z_org]

Introducing Pivot.

A 357M bidirectional decision model built on @liquidai’s LFM2.5 Encoder.

→ scores options directly instead of generating text
→ 15.8ms p50 on H200
→ 545 decisions/sec at batch 32
→ open weights + reproducible evals

Now on @huggingface:
https://t.co/Am7JgH4RCm

[Q1z @q1z_org]

chatgpt is loading.

the world’s most talkative model
is taking a minute
to start talking.

we can wait.
the agent in production cannot.

[Q1z @q1z_org]

not an engineering org.
a group chat with commit access.

[Q1z @q1z_org]

saw a repo with 68 subagents.

that’s not an engineering org.
that’s a group chat with commit access.

half of them are just there to decide
whether the other half should speak.

[Q1z @q1z_org]

saw a repo with 68 subagents.

that’s not an engineering org.
that’s a group chat with commit access.

half of them are just there to decide
whether the other half should speak.

[Q1z @q1z_org]

@dabit3 free agents are great.
the bill starts when every micro-choice
is still a full chat completion.

[Q1z @q1z_org]

@Butterflya1x Yes. The expensive part is rarely the plan.

It’s the 20 times per run you ask a generator:
route / retry / stop / escalate.

That’s the wrong shape of model for a loop.

[Q1z @q1z_org]

@mikeydsoftware Permissions are necessary.
They’re not sufficient.

If the model deciding “should I act?” is a chatbot,
the guardrail is sitting on top of a slow, fuzzy switch.

The decision itself needs a tighter interface.

[Q1z @q1z_org]

@codewithkhalil Valid JSON is the easy part.

The painful part is paying a generator to make a choice
your code then has to parse, validate, and retry.

Schema-shaped output still isn’t the same thing as a decision.

[Q1z @q1z_org]

@imryven This matches what we see.

The tax isn’t just the serial 5%.
It’s using a chat model for every micro-decision inside the loop.

More agents on a slow decision primitive still leaves you with a slow graph.

[Q1z @q1z_org]

@BeerBicepsGuy The missing split is inside the agent.

Planning can be generative.
The next action usually shouldn’t be.

Most “agents” still pay for an essay every time they only needed a decision.

[Q1z @q1z_org]

Hello World