Introducing RenRead more

Secnd is an AI research lab building decision models, and products on top of them.

Models that pick an option, say how sure they are, and ask when they aren't.

Read the researchGet the weights

Connects to (planned)

So much of what makes your work yours is the way you decide: what you approve, what you push back on, what you let wait. It takes years to learn, and it's the hardest part of you to hand to anyone.

We believe a model can learn it. Software should be able to carry your judgment, act on it when it's sure, and come back to you when it isn't.

Knowing when to ask matters as much as knowing the answer. A model that can't say "I don't know" can't be left alone, however often it's right. Trust has to be earned one decision at a time, with proof you can check.

Diagram. Three inputs, situation, options and examples, feed Ren inside a frame labelled Secnd. Ren reads, weighs, picks and checks, then passes its answer to the gate. The gate sends it to one of two outputs: act or ask. A return wire runs back to Ren, labelled: with consent, real decisions train the next Ren.

Figure 1. Ren reads a situation and its options, and a gate decides whether it can act alone or must ask.

What we do

Models built to decide.
A decision model takes a situation and a list of options, picks one and says how sure it is.
Proof before acting.
We measure how many decisions a model can handle alone, with proof, per dollar, including where it fails.
Products on top.
Secnd, our first product, will connect to your work tools, learn how you decide, act when it's sure and ask when it isn't. It's in development.
Read the research

In the open

We publish our results, including the weak ones.

SpeedOne decision
Time for one decision
GPUNot counting the network26ms
EndpointMedian, India to one GPU in the US424ms
CPUOn CPU2.18s
Training data12 licensed public sets and 26 written ones422,906 rows38 sets
ComputeTo train RenAbout $129
Acting aloneAt our 99% gate

MTOP 21.0% answered alone, 1 error in 923 answers, lower confidence bound 99.40%.

BANKING77 0% answered alone, because it couldn't prove it was safe.

Typed decisionsMixed test, 2,000 decisionsA tie with nano.
  1. 75.4%OpenDecider's run of Jev 1.13
  2. 79.6%OpenDecider-nano
  3. 79.7%Ren
Test accuracyFixed official test sets
Data table
Test accuracy, in percent
SetSeriesNo examples8 nearest32 nearest
MTOPRen v1.1.075.99%92.38%93.59%
MTOPSame training, seed 078.64%92.93%93.23%
BANKING77Ren v1.1.068.99%89.90%86.23%
BANKING77Same training, seed 068.90%90.62%87.44%
Limits

One seed with full history. It doesn't certify a new business. No real business has run Ren yet. Jev scores 79.7% on BANKING77 with no examples and costs about $0.008 per 1,000 decisions, against Ren's $0.083.

Run it where you want.

  1. On a GPU

    001

  2. On a CPU

    002

  3. In your own cloud

    003

Weights are public under CC BY-SA 4.0. With consent, real decisions from Secnd will train the next Ren.

Writing

Empty

Nothing published yet.
Writing starts soon.

The research is already public.

Read the research

Careers

No openings

No open roles yet.
We'll post them here.

We're a small lab, and we'll list roles when they're real.

See what we're building