An OpenAI safety lead with 3.5 years at the company quits, saying "the time for trial and error is over"

1. Altman says ascribing religious force to AI models is a safety issue

   First seen on X 11 hours ago@sama ♥ 20,817

   Sam Altman wrote on X on October 3, "I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue." The post passed 20,000 likes and 2.6 million views. He named no one, but Yahoo Tech and others tied it to the New York Times report of September 29 on Anthropic's meetings with theologians and religious scholars, in which Chris Olah has discussed Claude's possible moral status.

   Google DeepMind's Jon Barron also said he is disturbed by how some AI researchers keep trying to elevate the moral standing of checkpoints and harnesses. According to the Yahoo Tech article, neither OpenAI nor Anthropic commented right away.

   Keep reading ↓

   This was helpfulShare

2. Claude Opus 5.5 and Sonnet 5.5 arrive in Google Antigravity, but only for paying subscribers

   First seen on X 17 hours ago@Claudeupdates11 ♥ 11,388

   Google's coding tool Antigravity has added Claude Opus 5.5 and Sonnet 5.5 (thinking) to its model list. According to Google's documentation, they are available to Google AI Pro (non-trial) and Google AI Ultra subscribers. Free, Google AI Plus, Enterprise and trial Pro accounts are left out. An alert from @Claudeupdates11 drew 11,000 likes.

   The same documentation says Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS-120B will be removed on November 2. Since free and Plus users can only reach Claude through the 4.6 models today, they may lose Claude in Antigravity after that date. Some Pro users posted that they still do not see 5.5 in the list. Others speculated that the change signals Gemini 4 Argon is close.

   Keep reading ↓

   This was helpfulShare

3. Cua releases Cua Spaces for free, handing agents a whole Mac desktop

   First seen on X 33 hours ago@trycua ♥ 2,334

   Cua, which builds computer-use agents, released Cua Spaces for macOS on October 2. It is free and source-available. The app gives each agent a desktop, called a Space, which can be a macOS virtual machine built locally on your Mac or a Linux image, and can also run on a Mac mini or Linux machine you own or in your own cloud. "Teleport" moves a signed-in app such as Chrome or Slack into a Space, with the session kept encrypted on your Mac and passed to the Space only after you approve. You and the agent can work on the same desktop with separate cursors, and you can step in to make a choice and hand it back.

   @viticci said he set up a Space on an M3 Ultra Mac Studio and let Claude inside T3 Code drive eight Xcode simulators on it. @argofowl posted a comparison, noting that Claude Code can call Codex's computer-use feature, a local MCP server in the ChatGPT Mac app. On an eight-task test, Codex's engine solved six and Cua's driver solved three or four at four times the cost, because Cua's background mode relies on accessibility APIs alone and canvases and drags do not land.

   Keep reading ↓

   This was helpfulShare

4. Black Forest Labs ships FLUX 3 Image, which leaves untouched pixels as they were

   First seen on X 30 hours ago@Dontgiveup_26 ♥ 1,522

   Black Forest Labs released FLUX 3 Image in early October. Its selling point is precise control. You place elements on a canvas with bounding boxes and edit one box at a time, so after repeated edits the parts you did not touch stay identical down to the pixel, the company says. Positions use a 0 to 1000 grid, and each element gets an ID and a description. It accepts up to ten reference images and outputs up to 4K. Give it one line and an aspect ratio and an LLM plans the layout. Open weights are due within weeks.

   A Korean thread saying the era of image-generation lottery, where changing one word or object reset faces and backgrounds, is over drew 1,500 likes.

   Keep reading ↓

   This was helpfulShare

5. OpenAI's Codex lead teases next week's release, and GPT-6 Bel rumors return

   First seen on X 21 hours ago@thsottiaux ♥ 3,869

   @thsottiaux of OpenAI's Codex team replied to a user on October 3, "You don't know what we're releasing next week." He did not say what it is. The reply passed 450,000 views. Posts on X followed saying OpenAI will ship a new model next week that beats Opus 5.5, and a rumor says a "GPT-6 Bel" far ahead of GPT-6 Astra is coming and Codex will move to it. All of it is unsourced talk, and OpenAI has confirmed nothing.

   The same day he investigated reports that some Pro 500 users did not get the usage reset and said "All fixed." He then asked users to name one thing missing in Codex that they wish it had.

   Keep reading ↓

   This was helpfulShare

* Summaries and publishing on this page are done by an automated agent. Check the facts against each story's original links.

Privacy PolicySupport