@speak, we've been working closely with OpenAI on GPT-Live-1 for the past few months, building with an entirely new class of voice model model that continuously listens and speaks at the same time. Here's some of what we learned.

Introducing Live Tutor Lessons, powered by @OpenAI GPT‑Live-1: Interactive language lessons with a tutor that get you talking, listens in real time, and helps you find your words. Jump in with a question anytime. Limited rollout for English and Spanish learners. See it in action ⬇️

Congrats to the @indent team and @sthupukari on launching! We've been using Indent for well over a year now @speak and it's organically become one of our few internal agents adopted by almost the entire team, and it helps us make more data-driven decisions daily. I've been really impressed with the significant progress they've made on a lot of the nitty gritty hard parts of enterprise agent uptake including auth/IAM and team memory and context. Onwards!

Amazing technical writeup worth reading in full - kudos to the gpt-live team on this work! It's clear that full duplex/bidi is the future of voice and it's really cool to peek under the hood at the complex systems engineering that makes it all "just work". Some observations/notes for voice agent builders: - The whole post is a huge flex, but especially turning the “improve startup latency” ticket into a new webrtc handshake protocol. Generational scope creep. @juberti what was the original estimate on that issue? :) - The voice must flow. Nothing matters but continuous realtime audio flow and smoothness. Be extremely clear in your voice agent’s system design about division of responsibilities between the realtime band and async channels/sidecars, and treat the async RPC boundary as a first-class conversational & UX design principle. - Many other details are direct consequences of optimizing for voice fluidity: prompt caching and avoiding audio-blocking prefills, live context compaction with blue-green-style cutovers, smoothly handling long-running voice sessions (that are much longer than gpt-live’s context window) - Future models and versions of gpt-live will continue to become more intelligent and handle more tasks natively in the realtime band. As inference speed improves, the model will also be able to spend more tokens thinking and delegating within the realtime response latency budget. This means both the actual delegation and “perceived delegation” boundary will push out. - Interesting details on the turn-transcript heuristics, e.g. “we prioritize coherence in the displayed assistant responses even when the user speaks in the middle.” The full bidirectional/duplex model arch breaks free of modeling the convo as a traditional ordered list of chat messages, but heuristic/projected semantic turn messages are still critical for most voice agent harness systems you’d want to build around the core live model. - Realtime full-duplex sessions load your voice agent application server’s CPU (stream handlers, queues, network paths) more heavily than almost any other ordinary application type. Think about the fact that each full-duplex session processes ~100 audio frames per second across both directions. This is a very nontrivial driver of higher infra/capacity costs when scaling realtime voice experiences. - The bar has been set extremely high for long-running voice agent infra providers, with a lot of additional complexity to support very smooth realtime voice model inference with an explicit temporal dimension. It’ll be interesting to see the industry patterns that emerge here.

Strongly agree with this K-shaped divergence happening on software teams. It's existentially urgent to move faster here than what many consider reasonable, because a decent-sized gap today will compound into an insurmountable chasm in output and execution speed and quality by the fall. This means your team must simultaneously buy into the vision of what is possible and will be possible soon, while also being in the trenches on a daily basis wrestling with and steering agentic systems that still often make mistakes, overcomplicate things, and don't always do what the user is envisioning. Navigating this takes strong, continuous leadership that's also in the trenches. We're doing a lot at Speak to stay on the cutting edge here as a team - a few examples: - Continually thinking about what we're still doing manually and following that signal to agentify our repos and processes - Retooling our engineering hiring process to explicitly test for agentic engineering skill and mindset - Explicitly trying to build production features while keeping manual code as close to 0% as possible If you're working at a company that isn't all-in, you should leave before it's too late. Accelerate!

As we all know on X, agentic AI coding crossed a capability threshold at the end of 2025 such that the era of software engineers writing code by hand was over. @speak, we're urgently embracing this fundamental change to the way we develop software. I wrote up a snapshot of what we’re observing and learning, and where we think things are going: speak.com/blog/agentic-engin…

Can voice AI actually get you to fluency?💡 Join us LIVE Wednesday at 4PM CST with Andrew Hsu, Co-founder & CTO, @speak — the voice-first AI language tutor that’s hit $100M ARR and 15M+ downloads. We'll see you here on X soon!📷 #AI #LanguageLearning

Thanks @RashiShrivast18 for telling the @speak story! We’re pushing hard to make the world’s best AI language tutor available to everyone. Super excited for the slate of new feature launches coming later this year! Stay tuned. forbes.com/sites/rashishriva…

House of Hsu

Congrats @OpenAI! Some thoughts from testing the new GPT-5 models @speak over the past several weeks: - Significant leap in the reasoning frontier - your strong assumption now should be that it can crush most real-world complex tasks across a giant context window/dump, and if it doesn’t, it’s more than likely that your prompt isn’t good enough or you need to provide more clarity. Your ability to think in a structured manner and write with clarity is the big bottleneck now. - We were the most impressed by GPT-5's pattern recognition and reverse engineering capabilities across giant context dumps. It’s super good at systematically understanding and generalizing from examples to create structured workflows or playbooks. - Example - over the years, we’ve developed a highly opinionated and structured pedagogical methodology for functional language fluency that we call the Speak Method, but have struggled for years to distill and document it into a playbook clearly enough. Gemini 2.5 Pro and o3-pro couldn’t do this, but GPT-5 did an amazing job analyzing a large dump of our curriculum/content and breaking down/generalizing the principles necessary to scale our content. This was really surprising and felt superhuman. - We've found previous reasoning models both slow and inconsistent in their logical leaps. GPT-5 feels far more consistent and accurate on reasoning tasks, which is transformative for our curriculum scaling process and language tutoring capabilities. - Zooming out and speaking as an extremely heavy user of both ChatGPT and the API, I’m super excited (and relieved!) that the constellation of models is going away and everything is getting unified. It’s absurd that the full model is priced ~same as o3/4.1 and 5-mini is 5x cheaper! This day has been coming for a long time, and we expect to switch all of our realtime usecases over with reasoning=minimal. Huge congrats to the OpenAI team and we can’t wait to bring all this to our users this quarter!

A special bit of silicon valley is that everyone is willing to help each other. Back in 2015, we pinged @karpathy for AI help and (within 2 min!) he offered to meet up w/ us the next day. One of a few formative convos that led to us doing AI research and starting @speak.

Thanks for having me on @latentspacepod @swyx!

Thanks for having me on @ashleevance - was a very fun convo!

The day has finally come! @speak is no longer an English language learning app. You can now learn 🇪🇸 Spanish, 🇫🇷French, 🇯🇵 Japanese, 🇰🇷Korean, and 🇮🇹Italian as well! And I'm not talking about memorizing vocab in a dopamine machine, I'm talking about good old fashioned spoken fluency.

Spotted @pdhsu at our parents’ house.

Fun fact: @speak has 65% brand recognition in South Korea and close to 10% of the population has used us to improve their English fluency. Was great talking to @bquazz on how we got our start there!

📈 @speak

“We’re really an AI native company." Speak CEO @connorzwick explains how his app uses #AI to get language learners talking! Plus, their growing presence in East Asia, and partnership with @OpenAI's startup fund.

These guys were among the earliest of all the startups we've funded to understand what AI was going to make possible, and they have executed perfectly on that early insight.

We've raised a Series C from @Accel valuing Speak at $1b to keep driving toward our mission of reinventing the way people learn, starting with language. This year was great, but we have even bigger plans for 2025 and are hiring exceptional technical talent to work on one of the fastest-growing consumer AI applications. Join us! speak.com/careers