Yes. You can talk faster (4x) than you can type, but most people can read faster (2x) than they can hear.

This odd result occurs when the model doesn’t hear any actual speech, as a consequence of imperfect turn detection. With only noise as input, all outputs are valid, and sampling then favors the most common utterances encountered during training.

Here's a cool paper using LLMs for lossless text compression, in what they call LLMZip, which outperforms SOTA text compression methods The idea is very intuitive Given a sentence to compress, like "My first attempt", they feed the first 2 tokens ("My" and " first") to... [1/6]

Really captures the step change that reasoning models have brought about, turning token prediction into structured thought.

For each algorithm we need to measure a few different values: - accuracy curve (ROC) - speech detection latency - silence detection latency This means there are real tradeoffs - looking at more audio frames increases detector accuracy but at a latency cost…

A beautifully detailed version of an old pie chart joke.

We’ve moved from an age where compute was plentiful to an age where it is scarce, and meanwhile demand continues to rise. The only way forward here is to add capacity.

This week, ChatGPT is on track to reach 700M weekly active users — up from 500M at the end of March and 4× since last year. Every day, people and teams are learning, creating, and solving harder problems. Big week ahead. Grateful to the team for making ChatGPT more useful and

gpt-oss-120b: o4-mini-level performance, open weights, runs on a single H100. Available on HF now! huggingface.co/openai/gpt-oss…

That didn't take long - gpt-oss-120b is now available in LM Studio. Getting ~25 TPS on my M4 Max! lmstudio.ai/models/openai/…

ChatGPT Voice also gets an upgrade today - improved naturalness, way better instruction following, and much higher usage limits!

GPT-5's API pricing is cheaper than GPT-4o, and compares extremely favorably against other frontier models (from simonwillison.net/2025/Aug/7/gpt…)

gpt-oss + ultravox, should be an incredible combo for oss voice ai!

relnotes for today's ChatGPT Voice update help.openai.com/en/articles/68… --- Today we’re rolling out improvements to Voice Mode to make it more accessible and useful for everyone. It also now works with custom GPTs. We’re expanding access with near-unlimited use for Plus users and hours

Good guidance if you're trying to use GPT-5 in a cascaded voice architecture (ie, ASR + GPT + TTS).

This happens because WebRTC’s echo cancellation doesn’t converge until 2-3 seconds into the call. Not an issue in typical meetings but can be a real annoyance when the model talks first. We’re working on some solutions.

Agreed, this is a common request. In your use case, which parameters should be changeable, and which ones shouldn’t? Trying to figure out the best API approach.

Justin Uberti That makes sense. Is it specific to the iOS implementation of WebRTC? I couldn’t reproduce on android/web but I also didn’t test as much there