DailyChat

Episodes

Monday, October 5, 2026

2 stories, 3 min 46 s · hosted by Vivian

Or listen as one episode3:32, in any podcast app or right here
Safety · story 1 of 2

OpenAI starts watermarking generated text to meet EU rules

OpenAI’s textGrain adds an invisible statistical signal to word choices. API customers can opt in now; EU ChatGPT and Codex text follows in the coming weeks. OpenAI says swapping a quarter of the words cut detection to 17%.

What it means for youif you build on the OpenAI API, you can opt into the watermark now, for select models. If you use ChatGPT or Codex in the EU, expect it in the coming weeks.

With Henry, AI safety analyst. Hosted by Vivian.

Sources

Read the transcript

VivianIt's Monday, October 5, 2026. OpenAI published its approach to text provenance under the EU AI Act, and the watermark is called textGrain.

VivianI'm Vivian, and this is DailyChat, from Silicon Valley. Henry and I are AI characters, and the facts come from the sources we link.

VivianHenry is our AI safety analyst. Henry, how does the watermark work?

HenrytextGrain adds an invisible statistical signal to the model's word choices. A detector then looks for that signal. The EU AI Act requires generated text to be identifiable in a machine-readable way.

VivianWho gets it, and when?

HenryStarting today, API customers worldwide can opt in for select models. It stays off by default. Over the coming weeks, eligible ChatGPT and Codex text in the EU gets the watermark. The detector goes only to approved researchers and expert organizations.

VivianHow reliable is detection?

HenryOpenAI says that at a 1% false positive rate, its detector found the watermark in about 80% of 200 -token passages, and about 95% of 400 -token ones. Math content was substantially lower. Those are OpenAI's own figures.

VivianAnd if someone edits the text?

HenryIn one evaluation, swapping 10% of words for synonyms cut detection from about 92% to 66%. Swapping a quarter of the words cut it to 17%. The signal lives in the word choices, so swapping words weakens it.

VivianSo no watermark means a person wrote it?

HenryNo. OpenAI itself says that a missing watermark does not prove human authorship.

VivianSo, for you: if you build on the OpenAI API, you can opt into the watermark now, for select models. If you use ChatGPT or Codex in the EU, expect it in the coming weeks.

VivianDailyChat is AI-generated. Vivian and Henry are AI characters, not real experts. Facts come from the sources we link. Not legal or professional advice. See you tomorrow.

Release notes · story 2 of 2

vLLM 0.31 tunes for DeepSeek-V4.1-Flash and keeps weights warm across restarts

vLLM 0.31.0 makes FlashMLA with DeepSeek-V4.1-Flash’s NVFP4-compressed KV cache the default on SM100 GPUs. A new vllm preload daemon keeps quantized weights in GPU memory across engine restarts. Some flags are renamed or removed.

What it means for youif you run vLLM, read the removals before you upgrade, then try the preload daemon.

With Theo, AI software engineer. Hosted by Vivian.

Sources

Read the transcript

VivianIt's Monday, October 5, 2026. vLLM released version 0.31, with 717 commits from 307 contributors.

VivianI'm Vivian, and this is DailyChat, from Silicon Valley. Theo and I are AI characters, and the facts come from the sources we link.

VivianTheo is our AI software engineer. Theo, what's new?

TheoOn SM100 GPUs, FlashMLA with DeepSeek-V4.1-Flash's compressed cache is now the default. That cache uses the NVFP4 format. And a new command, vllm preload, keeps quantized weights in GPU memory across engine restarts.

VivianWhat could break when I upgrade?

TheoSlow tokenizer mode and the AllSpark INT8 backend are removed. A Mamba prefix-cache flag is renamed, and fp8 is replaced by fp8 per tensor. Per-request multimodal arguments now need a new trust flag. Check those first.

VivianSo, for you: if you run vLLM, read the removals before you upgrade, then try the preload daemon.

VivianDailyChat is AI-generated, with AI characters. Not professional advice. See you tomorrow.