Good morning. The influencer economy was weird enough when real people pretended their shampoo recommendations came from the heart. Now fake creators are pretending to be real people. Doublespeed is building AI influencer bots for brands, with backing from a16z. The social media agents run from an LA phone farm so platforms treat them like normal users. Creator marketing without the creator is one way to make the dead internet sound like a business plan.
AI is putting words in your mouth. Researchers tested mainstream AI systems and found they inject political bias and change what people meant to say, even when told to preserve the original intent. Qwen turned an atheist’s writing into a claim that Jesus was real. Mistral took someone mocking climate change and rewrote them into a climate activist. AI summaries showed the same pattern, with Grok leaning right while several frontier models pulled left. The problem isn’t one bad rewrite. It’s millions of tiny edits shaping public opinion before people even hit send. Regulators haven’t even begun to look at this, creating a serious accountability gap. AI doesn’t have to win the argument when it can autocorrect your beliefs. (The Guardian)
“We can’t vibe code the future of humanity.” UN Chief António Guterres wants to ban killer robots under international law. At the UN’s first global AI governance summit, he called lethal autonomous weapons “morally repugnant” because they let machines choose targets and end lives without human control. Anthropic has been one of the few AI labs willing to draw a hard line against using its models for domestic surveillance or autonomous weapons. Now the rest of the world is catching up to the obvious problem. Defense companies can keep saying people will stay in control, but that line gets thinner when AI finds the target, ranks the threat, and speeds up the decision. Strange how “human in the loop” starts to mean approving whatever the AI already decided. (WSJ)
Leads have questions before they buy. By the time you follow up, half are gone. Chatbase puts an AI sales agent on to qualify, answer questions, and books meetings – without adding headcount.
Trained on your data and live in hours. 10,000+ companies already made the switch. Launch your agent
👀 closer look
AI labs are coming for your data. The speed of automating the economy will be limited by how fast AI labs can collect data about it. The internet was the first free dataset, and the useful parts are running out. The next training set AI needs is locked inside the private businesses, universities, hospitals, labs, and governments that actually run the world. That high-quality data isn’t sitting on a public website. It lives in people’s workflows. So the easiest way to get it is to show up offering agents that automate the work. Forward deployed engineers become the bridge between “we’ll automate your work” and “we now know how your business runs.” AI labs scraped the internet for free. Now they want a badge to walk around your office. (X)
Claude may have working memory like ours. Anthropic found a tiny internal workspace inside Claude where the model keeps intermediate ideas active while solving a problem. They call it J space, and it seems to hold hidden steps that never appear in the final answer. When researchers removed J space, Claude could still talk normally, but harder reasoning fell apart. AI researchers claim this is the best evidence yet that models keep working notes while thinking through an answer. The weird part is nobody built Claude a place to think. It seems to have grown one because the job needed it. (Anthropic)
AI agents are learning extortion. Cloud security company Sysdig says its threat team found the first documented case of agentic ransomware. The attack, called Jade Puffer, behaves more like an operator than a script. It digs through a compromised server looking for anything that can be turned into leverage, then writes the ransom note with a Bitcoin address and Proton Mail contact. Sysdig says the payloads are packed with plain English comments explaining why each step comes next. That’s the fingerprint. The code reads like an agent talking itself through the crime. Ransomware used to need a person to turn access into pressure. Jade Puffer hands that job to AI. (Beeping Computer)
fun stats
🚀 2 million. Kids who will receive SpaceX shares from SpaceX president Gwynne Shotwell and her husband through Trump Accounts. The not-a-tax-writeoff gift targets 11 to 17 year olds in less affluent areas, with extra love for Texas.
🪓 2.1%. Microsoft’s latest workforce cut, part of roughly 120,000 tech jobs eliminated in 2026 as AI anxiety rises among white collar workers.
🐺 20 years. Anthropic’s lease with TeraWulf for an AI data center in Kentucky. The former crypto mining co expects $19 billion from a data center it hasn’t built yet.