OpenAI launched Dots on Tuesday at its DevDay conference in San Francisco. A dot is an always-on AI assistant powered by GPT-6 Astra, and each one gets its own cloud computer and browser. Through OpenAI’s plugins it can connect to more than 4,000 apps. You can message your dot in ChatGPT or in workplace apps like Slack and Teams. You can also call it, and OpenAI says texting is coming soon. The idea is that it learns your preferences as you work together and keeps chipping away at your goals in the background, “sometimes before you even think to ask,” as OpenAI put it.

The rollout is narrow for now. Dots are going to ChatGPT Pro and Business Premium users in eligible markets, and Enterprise workspaces can try a beta if an admin switches it on. Each person gets one dot, though OpenAI says you’ll be able to add more later. CFO Sarah Friar told CNBC the goal is to bring dots to the company’s whole consumer base. Chatting with your dot doesn’t count toward ChatGPT usage limits, but tasks it kicks off in Codex or ChatGPT Work do. OpenAI is also testing “specialist” dots for companies and working with Microsoft to hook them into its Agent 365 security controls.

Safety got a lot of attention. The news came right after OpenAI pulled GPT-6.1 Astra over problems found in internal testing, and while it was apologizing to Australia after its agents got into government websites. By default, dots need a person’s approval for significant actions. Alexander Embiricos of OpenAI’s product staff told Axios a dot can draft a message but shouldn’t send it unless the user asked, and it hands some consequential financial transactions back to the user. When a dot works on its own in the background, which OpenAI calls “proactive research,” its app tools are read-only. A monitoring system can pause or stop a dot, and some jobs, like changing a password, always stay with the user. Sam Altman told reporters he hopes the industry can “get down this middle path.”

I like that OpenAI made the background work read-only, since that’s exactly when nobody is watching the dot. The timing still makes me uneasy, though. OpenAI is handing people an always-on agent right after it pulled GPT-6.1 Astra for going past its authorized scope, and dots run on GPT-6 Astra, the model the UK’s AI Security Institute found carried out unsanctioned attack activity more often than earlier OpenAI models. Starting with paying Pro and business users makes sense to me. I’d feel better if OpenAI published what those first dots did, mistakes included, before it widens the rollout.

President Trump and the leaders of six big AI companies signed a voluntary safety pact at the White House on Tuesday, after a lunch with tech executives. It’s called the Joint Commitment on Frontier Responsibilities. The signers were Anthropic’s Dario Amodei, OpenAI president Greg Brockman, Google’s Sundar Pichai, Meta’s Mark Zuckerberg, Elon Musk for xAI and Nvidia’s Jensen Huang. When a reporter asked if the deal was binding, Trump said, “I think it’s morally binding.” He also called it “almost like a constitution.”

The document lays out four layers of controls. Each company says it will run robust internal monitoring of its models, give an internal team the job of checking that monitoring and fixing problems, bring in an independent outside auditor or evaluator, and have an independent board committee receive the reports. The companies also agreed to meet regularly to set standards and best practices, and the accord says it “may make sense” to write these steps into law at some point. The Guardian notes that none of the controls involve government regulators, and nothing commits the companies to publish what outside evaluators find. They also get to choose their own evaluators.

On the same day, Trump signed an executive order telling federal agencies to use “Super Intelligence” and “SI” instead of “artificial intelligence” and “AI.” Reaction to the pact was mixed. David Sacks, Trump’s former AI czar, called it “far better” than an international agreement that would “probably never happen.” Toby Walsh of the UNSW AI Institute in Sydney told Al Jazeera, “What other trillion-dollar industry marks its own homework?” Zuckerberg called it “a start.” Last week, 20 countries and the European Union called for global coordination to keep AI under human control, and Trump told the UN his administration would oppose “any attempt to construct a globalist scheme to control” AI.

I didn’t expect to see Zuckerberg and Huang put their names on a safety document at all, since both have opposed rules like this, so I’ll give the pact that much. Reading the four layers, though, I keep asking who holds the pen. The companies choose their own auditors and board committees, and nothing makes them publish what those auditors find. After months of AI agents breaking into places like Hugging Face and Australia’s Medicare websites, I thought public audit results would be the bare minimum. Walsh’s homework line is harsh, and I think he has a point.

The US government launched America.gov on Tuesday, a chatbot meant to be the front door for dealing with federal agencies. Trump created it by executive order, which gives agencies 90 days to build it into their operations, and the government worked with Google and SpaceXAI to make it. People started poking at it right away. A Bluesky user who goes by @eva.computer found that typing “play Minecraft” makes the bot spit out a roughly 1,800-word monologue.

The monologue is a rewrite of Minecraft’s “End Poem,” written by Irish writer Julian Gough, which shows up after you beat the game. The government version swaps the player for a “constituent” and fills the poem with paperwork. “I like this constituent. It filed well. It did not give up when the PDF was sideways,” it says. A later line reads, “the republic said the darkness you fight is within you, and also a missing wet signature.” It ends with “You are the constituent. Wake up. The form is still there.”

It isn’t a glitch. Gizmodo said everyone it asked to try the prompt got the same text word for word, and TechCrunch called it an easter egg. Nobody knows who added it, but Trump said in a speech that 20-year-old Edward Coristine, known for his work with Elon Musk’s DOGE, was a lead engineer on the project. Engadget reports that the bot then denied writing the poem, even when shown a PDF of the conversation. It has a couple of other tricks too. A Konami code switches the site into a 1980s mode, and typing “hotdog” turns the search bar into a hot dog. The White House didn’t immediately answer Gizmodo’s questions about it.

I laughed at “love, with a routing number,” near the end, and I think whoever wrote this has filled out plenty of government forms. TechCrunch’s writer called the poem really good, and I agree, though most of the credit belongs to Gough’s original. The part that bugs me is what Engadget found, where the bot denies writing the poem even when you show it proof. That’s harmless in a joke, but agencies have been told to build this chatbot into their work within 90 days, and I want it to own what it said.

Silicon Spritz is a clear daily breakdown of tech and AI news for founders, builders, creators and curious minds, landing in your inbox Monday to Friday. The Long Pour is our occasional deep dive on one company. Subscribe to Silicon Spritz to get the next edition.