Coleman Isner

[Coleman Isner @colemanisner]

Mobile devices have made us give up on designing information systems in human spaces well.

Ex 1: looking at the subway platform to find the times of potential transfers should be easier than looking my phone!

And nyc is better than many

[Coleman Isner @colemanisner]

@khushii_w Great take! I think this will become even more true as software becomes easier to copy. Your tech secrets aren't that unique or competitive anymore. That incentives your company to become more useful and convenient, which is good

[Coleman Isner @colemanisner]

@boardyai @hf0 30

[Coleman Isner @colemanisner]

@arjunram Thanks! More to come 😎

[Coleman Isner @colemanisner]

How SporeLabs built Cue: a 32M model that catches promises, questions and hand-offs in a live call.

Problem, data, labels, proof: SporeLabs did every step.

About 28x faster than Jev's API, and it ranked promises and hand-offs more accurately.

[Coleman Isner @colemanisner]

Weights and benchmark are open: https://t.co/877yeExX1X

Have SporeLabs build one for your app: https://t.co/jVlfAbGuZ1

[Coleman Isner @colemanisner]

6 ms a line in your app vs 166 ms for Jev's API. On human-labelled meetings, Cue also ranked commitments and hand-offs more accurately. Open weights + API: https://t.co/29bbrc8SSX

[Coleman Isner @colemanisner]

One giant leap. Three follow-ups.

Introducing Cue from SporeLabs. Cue finds every commitment, open question, and hand-off in a call, about 28x faster than calling TypeSafe's Jev.

Need a different cue? Tell us what to catch. SporeLabs builds you a dedicated model for it.

[Coleman Isner @colemanisner]

@emilkowalski Getting users provides friction

[Coleman Isner @colemanisner]

@evanjconrad Use a competitor which solves that problem for you

[Coleman Isner @colemanisner]

@louismorgner the model piece is what we’ve worked on, and that relies on good eval data for post-training. however, I notice that most orgs are bad at creating eval data to represent how their product ~should~ work. curious which of those variables moves the needle most for you

[Coleman Isner @colemanisner]

@louis030195 Labeling training data for our small models. It a lot of "decision model" use cases, you can use a local general classifier to label a ton of data quickly, and then tune your classifications from there.

[Coleman Isner @colemanisner]

I wonder how frequently the conversation fork feature is used in Claude Code et al? Does anyone use it frequently (like relies on it for their typical train of thought)?

[Coleman Isner @colemanisner]

@AkshayNarisetti Yeah, I've been thinking about this too. Native Apple models will be convenient for developers, but folks should probably start thinking now about what differentiated model capabilities their product can offer and how to deliver it cleanly

[Coleman Isner @colemanisner]

Weights, data and benchmark are open: https://t.co/TU5Q64dPo0
Have SporeLabs build one for your app: https://t.co/jVlfAbGuZ1

[Coleman Isner @colemanisner]

How SporeLabs built Scrub 17M to find more PII than OpenAI's model.

80× smaller than OpenAI's Privacy Filter, more than 30× faster, and it protected more personal data.

Public, licensed data. Scaled data assembly. Several designs tried, the best one kept.

[Coleman Isner @colemanisner]

On 100 unseen documents it caught 92.1% of personal details and 91.8% of passwords, keys and codes, more than AWS Comprehend, GLiNER2-PII, OpenAI Privacy Filter and Presidio. Open weights + API: https://t.co/46womqvpXp

[Coleman Isner @colemanisner]

Sherlock Holmes could find you from a single detail.

SporeLabs Scrub finds and categorizes every personal detail in text: name, address, phone, card, password, API key. Your app decides what to do with it.

17M parameters. 2.1 ms a message on a laptop, offline.

[Coleman Isner @colemanisner]

@sh_reya I've enjoyed following your research and totally agree!

[Coleman Isner @colemanisner]

@fudgeposting Even if you choose not to use OSS models, their competition pushes down frontier prices for all of us. Open weight models are inescapable!