maths, visualisations, conversational AI. VP Research @polyaivoice prev: @RekaAILabs, @Apple AI/ML, @GoogleAI, PhD @Cambridge_Eng
Edinburgh, UK
Joined February 2010
- Tweets6.7K
- Following2.7K
- Followers81.5K
- Likes8.6K
Pinned Tweet
a thread with some of my top animations 📍🧵
nitter.cf/matthen2/status/154322…
Matt Henderson retweeted
“Hello? Is anyone there?” 🤔
That’s what callers say when a voice agent takes a beat too long. AI can write production code, but many agents still can’t tell when to speak on a live call.
Dialog-RSN-1 hears the caller’s audio and handles the whole turn in a single pass, so callers hear a reply in under 1 second. Independent benchmarking by Coval found it the strongest realtime model they’ve tested on production-style calls.
@pmckee, @Addlesee_AI, and @matthen2 take you inside the model in Part 1 of Let’s Dialog: Building the Next Voice Agents.
Grab your spot: bit.ly/4dVUc6B
we are now officially a model company!
try out dialog-rsn-1 by using our openai realtime compatible api
Dialog-RSN-1 is now GA, and anyone can get an API key and build on it today.
Comparing real-time voice models? Start with the numbers. Median first-token latency is under 300ms, about the length of a blink. It has the highest quality, audio-understanding, and safety scores of any real-time-capable model, and it’s efficient enough to scale to every conversation you have.
It’s audio-native, so turn-taking, speech recognition, function calling, and the response all happen in a single pass, with a separate voice output you control.
Get started in three steps:
1. Pick your path: our Agent Development Kit (ADK), AWS SageMaker, or the public API, wherever your stack already lives.
2. Create your account and grab your API key.
3. Clone an example, add your key, and talk to it in your browser.
Start now here: studio.poly.ai/start-buildin…
some very nice explanations of our model dialog-rsn-1 from @TsungHsienWen -
it's an audio input LLM that takes control of turn-taking, and is specialised to task-oriented dialogue
Shawn Wen of @polyaivoice on how their new model Dialog-RSN-1 handles turn-taking (collab with PolyAI).
Out now on MLST
agents building agents
Meet Wren: the agent that looks after your agents.
Your dialog agents handle thousands of customer conversations a day, and keeping watch over all of them to make sure everything is working as well as possible is a lot for any team to stay on top of. Overnight, it reviews conversations, catches bugs like a validation rule rejecting valid customers, and has a fix ready before you log on. Live now.
Read all about it here: bit.ly/4iNF5ze
Moving my animation content somewhere nicer
bsky.app/profile/matthen.com
Matt Henderson retweeted
@inherent_labs' first research result is out! Our AI scientist Faraday learns to serve as a layer of scientific intuition on top of the advanced capabilities of coding agents. Check out the paper here arxiv.org/abs/2608.13331.
Very cool work from Inherent-
Learns to replicate the core results of ML papers more rigorously than Claude and Codex- using those as tools.
A neat task that exercises good science and creativity
What if our ears were like our eyes, and only had three receptors? Interestingly speech is still more or less intelligible, but not much else
too me this shows how much more discriminating our hearing is than our vision
To solve a maze, keep the entrance cold and the exit hot.
Once it settles, enter and always walk in the hottest direction.
At every step, each cell is updated to the mean temperature of its neighbours (except the start and end nodes held fixed)
So it is inefficient, but only requires local communication in the maze
are there gliders, spaceships, puffers?
nitter.cf/matthen2/status/208525…