Good morning. MIT found a way to make AI less full of sh*t. Their new method trains models to admit when they’re guessing. MIT had the AI answer questions and give each answer a confidence score. Then they trained it to stop sounding so confident when it wasn’t sure. This simple tweak cut AI’s false confidence by up to 90% without hurting accuracy. Teaching AI to say “I’m not sure” may be the breakthrough we’ve all been waiting for.
OpenClaw is breaking under its own success. What started as a scrappy open source project is now treated like critical infrastructure for AI agents. Jensen Huang calls it the “operating system for personal AI.” But behind the scenes, a few dozen volunteers are scrambling to hold it together, pushing updates and patching bugs between their day jobs. That frantic pace is starting to crack. Companies refuse to move past stable March builds because new versions keep breaking custom agents. Users report malware in the ecosystem and agents that delete data after updates. This is what happens when hacker energy meets enterprise expectations. (The Information)
Google is bringing AI agents to Workspace. With a new layer called Workspace Intelligence, Gemini pulls from your emails, docs, chats, spreadsheets, and files to learn how your company works. Google says this gives AI the instincts of a seasoned employee who knows your team’s preferences, your company’s backstory, and even your professional voice. Instead of just answering prompts, these agents can prep meeting briefs, surface relevant info, suggest next steps, and build docs or dashboards tailored to your team’s style. It’s like hiring a veteran employee, minus the salary and passive aggressive Slack messages.
Didn’t make it to Las Vegas? No worries – you can still join Google Cloud Next ‘26 online April 23-24.
Get access to keynotes and select sessions on agentic AI, data, and how teams like Anthropic and Spotify are building the future of enterprise AI.
Don’t miss it. Grab a free digital pass
OpenAI dropped workspace agents too. Describe a task or toss in a file, and ChatGPT will turn it into an agent by mapping steps, connecting tools, adding skills, and testing it until it works. These agents live in the cloud with access to your files, tools, code, memory (and probably your passwords and API keys). This lets them keep working across ChatGPT and Slack even when you’re offline. The bigger pitch is that once you build an agent, your entire team can use it too. Sounds completely safe. (OpenAI)
Security teams can’t keep up with AI coding. According to the new AI Coding Impact Report, engineers are shipping code faster than security can handle. A survey of 200 security pros found nearly 60% are just barely keeping up, and it’s getting harder. Two-thirds now spend most of their time triaging alerts or proving vulnerabilities instead of actually fixing them. The biggest AI coding headaches are secrets exposure (78%) and insecure dependencies(73%). Security teams are basically doing QA for hackers. Good thing Mythos hasn’t fallen into nefarious hands yet. (ProjectDiscovery)
AI scientists have no clue how science works. Researchers tested AI science agents across 25,000 experiments in eight fields. On paper, the agents handled the full scientific method, from generating hypotheses to running studies. The problem started when the science got messy. Agents ignored evidence 68% of the time and changed theories in response to conflicting results 26% of the time. Worse, the issues trace back to flaws in the underlying AI models, making those slick agentic interfaces little more than expensive gift wrap. We’re quickly approaching a future where AI cranks out endless research papers without mastering the annoying step of actually being correct. (arXiv)
Gas powered data centers pollute more than entire countries. Just 11 AI campuses pump out over 129 million tons of greenhouse gases each year. For perspective, that’s more than Morocco emitted in all of 2024. This isn’t some hypothetical future scenario. AI companies are bypassing the electrical grid entirely, building their own gas powered infrastructure to feed surging demand. The pipeline for these behind-the-meter projects surged from just 4 gigawatts in early 2024 to nearly 100 gigawatts at the start of this year. A single Chevron backed project could emit more than Jamaica. At this rate, the cloud is starting to look suspiciously like a smokestack. (Wired)
fun stats
🤖 75%. New code at Google that’s now generated by AI and “approved by engineers,” up from 50% last fall.
🕹️ $54 billion. How much the Pentagon wants for drones, more than Ukraine’s entire military budget.
👖 3x. Boost in fashion brands’ conversion rates when shoppers use AI conversational search. Chatting with bots also makes us spend 2.6x more. Retail therapy, meet AI therapy.