Friday, October 9, 2026

VivianIt's Friday, October 9, 2026. OpenAI is rolling out Ultrafast mode for GPT-6.1 Sol, which it says runs up to 8x faster.

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

VivianJonah is our AI products analyst. Jonah, what's rolling out?

JonahUltrafast started rolling out yesterday for GPT-6.1 Sol in the API, Codex and ChatGPT Work. OpenAI claims up to 8x the speed of Sol Standard. That's OpenAI's own figure.

VivianAnd what does it cost?

JonahAPI pricing is $12 per 1M input tokens and $60 per 1M output tokens. In Codex and Work, access is on Pro 500, eligible Enterprise and Edu plans, and Enterprise admins must enable it.

VivianSo, for you: if you build with Sol and speed matters, compare the price against your current mode, and ask your admin if your plan is Enterprise.

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

VivianIt's Friday, October 9, 2026. Anthropic is offering open-source projects free vulnerability scans, with reports written entirely by AI models.

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, what is OSS Scanner?

HenryOSS Scanner is an opt-in service. Enrolled projects get periodic security reports at no cost. Each report has a reproducer and, where available, a candidate patch.

VivianDoes a person check them first?

HenryNo. The reports are fully model-generated, without human review. Anthropic says that means some reports may be incorrect or invalid.

VivianSo how good are they?

HenryAnthropic had expert penetration testers check 97 critical and high-severity findings across 48 projects. 85, or 88%, met its disclosure bar. Of the other 12, 11 were real but duplicates, and one was invalid. Those are Anthropic's own figures, from an early version of the scanner.

VivianDo maintainers agree?

HenryAnthropic says maintainers have seldom told it a high or critical finding was invalid. Some said severity ratings can be inflated, or that the scanner misread the project's threat model. Anthropic says it can't guarantee the scanner will be perfect.

VivianAnd who can join?

HenryCore maintainers of eligible projects enroll by submitting a pull request. Anthropic decides case by case. What I'd watch is how many reports turn out wrong, since no person reviews them first.

VivianSo, for you: if you maintain an open-source project, read Anthropic's enrollment details in the sources we link, and plan for reports that no person has checked.

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

VivianIt's Friday, October 9, 2026. Xiaomi published a technical report on how it scaled reinforcement learning for its MiMo-V2.6 models.

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

VivianMira is our AI research analyst. Mira, what's in the report?

MiraIt's a technical report on the MiMo-V2.6 family, which Xiaomi calls omni-modal. The subject is reinforcement learning, where a model improves from graded attempts.

VivianWhat does scaling mean here?

Mira3 things, according to the report. Bigger batches: each training step consumes 1,568 samples, at context lengths of up to 1M tokens. More varied environments, covering code, general, visual and cyber tasks. And more compute for the graders that score the attempts.

VivianWhy do graders matter?

MiraXiaomi says its grading gives more accurate reward signals on long tasks, and steers the model toward shorter solutions that use fewer tokens.

VivianWhat can go wrong with that kind of training?

MiraReward hacking: a model finds a way to score well without doing the task. Xiaomi says it built a multi-layer defense against that, and froze the MoE router to keep training stable.

VivianCan anyone else use this?

MiraXiaomi says it is open-sourcing its RL environments and framework, though neither page we link points to that code. This is also a report from Xiaomi's own team, and the summary we link gives no benchmark scores, so outside checking depends on that release.

VivianSo, for you: if you train or study large models, watch for Xiaomi's release. For everyone else, this is a lab describing its own method, and the sources we link have the details.

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