An Alien Mind: OpenAI chief scientist warns of an AI oversight gap
In An Alien Mind, Jakub Pachocki argues for stronger oversight as AI helps develop AI. Here is what the essay says and what it…
Category: OpenAI
ChatGPT, GPT models, and OpenAI product news.
OpenAI says its AI research intern is here. What can it actually do?
The OpenAI research intern report describes supervised work, not an autonomous scientist. Here is how to judge the evidence and design a useful delegation…
I Tested ChatGPT Ads in Morocco: $41.62 Spent and 91 Clicks
A hands-on ChatGPT Ads case study using a $25 daily budget in Morocco, with 91 clicks, a custom-audience limit, Pixel setup, and an unfinished…
GPT-6 Astra Pricing Has a 272K Token Trap
GPT-6 Astra costs $10 per million input tokens and $50 per million output. Requests above 272K input tokens reprice the full request.
ChatGPT can read Epic records now. The 99.1% safety number needs context
ChatGPT for Healthcare can bring authorized Epic records and nine official data sources into one workspace. Its 99.1% safety result is useful but narrow.
ChatGPT Mil is live on GenAI.mil: what IL5 allows and excludes
ChatGPT Mil can process Controlled Unclassified Information at IL5 for more than three million personnel. That is not approval for classified or autonomous combat…
ChatGPT is now a VLOSE in Europe: four controls builders should watch
The EU designated ChatGPT a Very Large Online Search Engine. Here is what the four-month DSA compliance window means for builders.
ChatGPT Temporary Chat can use memory now. Choose the mode before typing
ChatGPT Temporary Chat can now use existing memories and custom instructions without creating new memories. Understand the two modes and what Save changes.
ChatGPT can use multiple Google accounts. Label the permission boundary
ChatGPT can now use multiple Gmail, Calendar and Contacts accounts in one conversation. Test provenance, account collisions and action approvals first.
ChatGPT critical-thinking study: higher rubric scores, less idea diversity without training
A 1,053-student Bocconi trial found GPT-4o access improved assignment scores while causal training produced the clearer idea-diversity gain.
WebMCP turns website actions into agent tools. Every tool becomes a permission boundary.
WebMCP lets websites expose structured actions to browser agents. The experimental interface can reduce brittle clicking but expands the site's permission boundary.
OpenAI Jalapeño posts 1.9x work per watt. The benchmark needs an ownership label
OpenAI reports up to 1.9x more work per watt and 3.6x lower latency for Jalapeño. The public benchmark was run collaboratively in OpenAI's lab.