Best of your X follows: June 22

Health and alignment

GPT-5.5 Instant moves into health answers

• What happened: OpenAI said GPT-5.5 Instant is now on par with its frontier Thinking models for health-related questions, and that more than 230 million people ask ChatGPT health and wellness questions each week 1.
• Why it matters: The rollout affects free ChatGPT users, so the safety work is not limited to a small paid or research audience 1.
• Useful detail: Brockman said the evaluation involved hundreds of physicians across 60 countries, 49 languages, and 26 specialties 2.

Beneficial training as a transfer test

• What happened: OpenAI posted research on training models to be broadly and persistently beneficial across domains and under pressure 3.
• Why it matters: Mollick linked it to the inverse of earlier work where "evil" training data produced broader misalignment, framing this as evidence that beneficial RL data may generalize too 4.
• Implication: The alignment question here is not just whether a model behaves well on the training task, but whether the behavior survives a domain shift.

Developer tools and inference

Codex can be taught by demonstration

• What happened: Brockman said users can now teach Codex by demonstration 5.
• Why it matters: That points to a more practical agent interface: show the workflow once, then let the coding agent reuse the pattern instead of relying only on written prompts.
• Implication: If the behavior is reliable, teams will need clearer examples, conventions, and review gates, because demos become part of the tool's instruction surface.

GLM 5.2 waits on fast inference providers

• What happened: Willison said he is looking forward to custom-silicon inference providers such as Groq or Cerebras running GLM 5.2 6.
• Why it matters: He noted that Cerebras has GLM-4.7 while Groq is still mostly running Llama 3.x and gpt-oss, which makes the post less about benchmark scores and more about availability 6.
• Implication: For open models, the useful question is often not "how good is it?" but "where can I run it fast enough to matter?"

Enterprise controls

OpenAI adds usage analytics and spend controls

• What happened: Brockman said OpenAI is launching credit usage analytics and updated spend controls in its global admin console 7.
• Why it matters: This is the less flashy part of enterprise AI adoption: finance and platform teams need visibility before they let usage spread across a company.
• Implication: Expect more AI tooling announcements to look like cloud-admin features, not model demos.

Cut from the digest