←
Interhuman AI57 posts
Interhuman AI avatar

Interhuman AI

@InterhumanAI

Social Intelligence for AI

◷ Joined November 2025▧ 27 Photos and videos
Interhuman AI avatar
Interhuman AI@InterhumanAI·Oct 14

Can someone appear confident and uncertain at the same time? We ran this clip of Elizabeth Holmes from the trailer of You Can See Everything, through Inter-2 to explore how social signals emerge, overlap and change throughout an exchange. What makes conversations interesting is rarely any single expression or gesture, but how these signals come together to shape the way we communicate and are perceived.

0:20
Interhuman AI avatar
Interhuman AI@InterhumanAI·Oct 7

Today, we’re introducing Inter-2 Audio, our audio-only model, built to detect social signals from the voice without requiring video. As conversational AI becomes part of more human conversations, from voice assistants and AI interviewers to sales agents, research platforms and customer experiences, we are increasingly speaking to AI rather than simply typing to it. Yet much of what these systems understand still comes down to the words themselves, even though the voice carries considerably more information about how something is being communicated. Inter-2 Audio makes more of that information available to AI products using only the audio stream, detecting a broad range of social signals even in noisier, real-world environments. Because it does not depend on visual input, Social Intelligence can now be integrated into many more of the products where conversations already happen through voice. It is the second model in the Inter-2 family, built for a world where talking to AI is quickly becoming normal.

Interhuman AI avatar
Interhuman AI@InterhumanAI·Sep 23

Today, we’re introducing Inter-2, our next step in Socially Intelligent AI. Inter-2 helps AI products understand how people communicate, beyond the words we use. It allows you to notice more by detecting social signals such as hesitation, agreement, confusion and skepticism in real time. For products built around human interaction, the model adds a layer of information that transcripts alone can’t capture allowing for a sales intelligence platform to detect how a customer is responding as a conversation unfolds; a user research platform to identify moments of hesitation or engagement; a coaching tool to understand not only the words being used, but how they are being delivered. Inter-2 is another step toward AI systems that can perceive and respond to the social dynamics of human interaction, helping products get more insight from every conversation and do more with it. Inter-2 is available today at interhuman.ai

0:24
Interhuman AI avatar
Interhuman AI@InterhumanAI·Jul 8

A robot arm you control with gestures — no joystick. But the interesting part: it watches *you* back. If you look distracted or confused, it stops moving. Built in a weekend at the Robotic Agents Hackathon in Milan, with Interhuman AI as the safety layer. Designed for wheelchair users with limited motor function. interhuman.ai/blog/wavearm-h…

Post attachment
Vinni avatar
⟳ Interhuman AI retweeted
Vinni@Vi20229·Jul 8

Love this! We built the same thing a few weeks ago with live social signals (@InterhumanAI reads confidence, hesitation, skepticism etc). AI Garry doesn't just talk back but actually reacts and provides feedback. You can try it out here: interhuman-yc-booth.vercel.a…

Post attachment
Interhuman AI avatar
Interhuman AI@InterhumanAI·Jul 7

"The experience was fine" — said with 12 seconds of hesitation and a frustration spike. AI-moderated research tools transcribe what respondents said. The behavioural layer — how they said it — gets thrown away with the video. We built a research-interview demo on Inter-1 Streaming that keeps it. One WebSocket, signals pinned to the exact phrase: interhuman.ai/blog/what-they…

Post attachment
Interhuman AI avatar
Interhuman AI@InterhumanAI·Jul 1

Your pitch deck can be flawless and still fall apart in the room. That's not a script problem. We built a demo that scores delivery in real time with Inter-1: confidence, hesitation, energy, each one flagged to the exact second it happened. Our own team's scores are already on the leaderboard, try and beat them: thepitchpractice.com Full blog write up: interhuman.ai/blog/pitch-pra…

Post attachment
Interhuman AI avatar
Interhuman AI@InterhumanAI·Jun 24

Two stacked priors, not one bug. We patched the prompt and shipped a guardrail for the easy half. The hard half — a model that would rather invent speech than admit "nothing was said" — is the real story, and we're still working on it. Full post: interhuman.ai/blog/goblin-ye…

Interhuman AI avatar
Interhuman AI@InterhumanAI·Jun 24

Then we held the prompt byte-identical and swapped the model instead. Larger variants + an earlier checkpoint barely hallucinated (0-2%). Only our further-finetuned model confabulated — ~12% of runs. The reflex to invent speech over silence is the fine-tune's.

Interhuman AI avatar
Interhuman AI@InterhumanAI·Jun 24

So we ran the obvious test: swap the prompt's example to a different day/time. Fabrication rate went UP (37%→50%), and the invented quote followed the swap exactly. Friday→Tuesday. The words are the prompt's.

Interhuman AI avatar
Interhuman AI@InterhumanAI·Jun 24

Found it somewhere dumber: our own system prompt. A worked example showing the model expected output format had that exact line sitting in context on every single call. Classic "in-context learning" — except the in-context example was leaking into hallucinated output.

Interhuman AI avatar
Interhuman AI@InterhumanAI·Jun 24

First instinct: it's gotta be in the training data. We scanned 30,960 datetime mentions, 4,603 video transcripts, ~800 inference logs, 584 storage buckets. Zero verbatim hits. Anywhere. The phrase didn't exist in our training data.

Interhuman AI avatar
Interhuman AI@InterhumanAI·Jun 24

Our model would sometimes "hear" a sentence that was never said. Feed it a video with zero audio. Ask what was said. It reports: "Yeah, Friday at five." Same exact line. Every time. On total silence. We spent weeks hunting this goblin. 🧵