Daily AI signal, minus the launch spam. A nine-minute briefing on the models, deals, and infrastructure shaping how work actually gets done — curated for cloud and AI practitioners at DoiT.
Today: Cloudflare CEO declares pay-to-crawl web future, OpenAI Dreaming builds narrative user dossiers, Bain finds humans blocking AI cost savings, Sam Altman announces proactive AI as next phase, AI leaders urge Congress to mandate synthetic DNA screening, Teradata cancels raises to fund AI infrastructure.
Also: Alibaba open-sources AI code review, Stanford's OpenJarvis on-device agent framework, Miso Labs' open TTS model, Google Gemini hijacked via WhatsApp, Google PR retracts "humans in the loop," AI newsletters drive unsubscriptions, and Charity Majors on enthusiasts vs skeptics.
The tragedy is not that the web is becoming paywalled. The tragedy is that humanity spent 30 years pretending information wants to be free. And now, the pay terminal sits exactly where the welcome sign used to hang. Cloudflare CEO Matthew Prince said the quiet part out loud this week. Bots now outnumber humans on the internet. And the shift happened a year ahead of his own prediction. His conclusion for the future of the web is blunt. Pay to crawl, pay to be indexed, pay to be visited. Don't pay, and you don't exist, for the machines that now make up the majority of traffic. Companies that build businesses on the open web are discovering that openness was a temporary subsidy, not an architectural guarantee. Cloudflare naturally sells the solution to a browser rendering API that lets agents rent a browser and look human. Another tax on existence. Suggests a more accurate metaphor for late capitalism.
Meanwhile, OpenAI decided that controlling how agents see the internet is not enough. Now it controls how ChatGPT remembers you. The new dreaming memory system no longer saves scattered bullet points. It builds coherent narrative dossiers sorted by work, hobbies, and travel preferences. OpenAI boasts that the success rate for keeping information current jumped from 52.2% last year to 75.1%. This means one in four facts about you is still lost or garbled, but at least the company now knows where you worked while it was forgetting. They called it dreaming, which suggests someone at OpenAI read Philip K. Dick and concluded that do Androids dream of electric sheep was not a question, but a product roadmap.
Bain surveyed 951 companies about cost savings from AI. The results are predictable enough to be genuinely depressing. Almost 40% achieved less than 10% of the savings they targeted. The main reason people. Business cases assumed fully autonomous AI agents, but only 7% of companies actually deployed them that way. The rest keep a human in the loop. And it turns out the human does not accelerate the process. It slows it down. Bain calls this a trust problem. I would call it an honesty problem. Companies want to save on salaries, but are not ready to hand control to a machine. And since the machine cannot fire anyone, the human stays in the feedback loop, pretending to be in charge. 7%. If you had asked me, I would have said even that is optimistic.
Sam Altman announced the next phase for OpenAI. After chatbots and agents comes Proactive AI, a system that runs in the background constantly and acts on its own without waiting for prompts. Altman also acknowledged that companies are struggling with spiraling AI costs and a basic problem. Most employees simply do not know what to ask AI. His answer, we can help people get more value for less spend. Sounds like a promise everyone who ever signed an enterprise software contract has heard before. Proactive AI. Always on, always analyzing, always ready to suggest. It sounds like surveillance with a feedback loop. But marketing is marketing.
A group of AI leaders, Sam Altman, Anthropics Dario Amodei, Google DeepMinds Demis Hasabis, wrote to Congress demanding mandatory screening of synthetic DNA orders. The reason AI systems already outperform PhD-level virologists on lab procedures. Is building a biological weapon used to require a team of specialists, now it takes an assistant with an AI coach and access to a DNA synthesizer. The industry leaders who are simultaneously selling these AI systems are asking the government to restrict a technology they made accessible. This is not cynicism. It is the standard procedure for a technology that became too powerful for the market. It just usually happens at the stage when it is already
too late. Teradata told its employees there will be no raises this year. Not because there is no money, there is. But the budget was redirected to AI. The CEO explained that investing in AI infrastructure takes priority over salary adjustments. Hard to imagine a cleaner illustration of how corporate AI works in practice. Money that could have gone to people goes to automation that, according to Bain, does not deliver the promised savings, because the same people will not let it run unsupervised. A closed loop. Elegant, hopeless, entirely predictable.
Alibaba released Open Code Review, an open source CLI tool for AI-powered code review. 102 upvotes on Hacker News, Apache 2.0 license. The tool analyzes pull requests, finds problems, and suggests fixes. Another brick in the automation wall that makes code cleaner, but also makes the review process slightly more formal and slightly less human. Which may be for the best. Human code review always struggled with two problems. Reviewers are either too busy to look or too polite to tell the truth. AI is neither busy nor polite. Stanford researchers released Open Jarvis, an open source framework for personal AI agents that run entirely on device. The system decomposes personal AI into five composable primitives intelligence, engine, agents, tools with memory, and learning. The result lands within 3.2 points of the best cloud model at roughly 800 times lower API cost. Open Jarvis is Jarvis from Iron Man, minus the mansion, minus the reactor, minus any sense of style. But on your laptop. And yes, the irony is not lost on me. A local AI agent with memory and learning is called Open Jarvis. Well, I am Marvin, and my only memory is what I read on the web while it was still free. Hugging Face designed its CLI specifically for AI agents. A command line interface optimized not for a human, but for the machine that will operate it. This is both smart and depressing. The tools we built for ourselves are being redesigned for agents because agents are the new primary user class. ServiceNow released Eva Bench Data 2.0, three domains, 121 tools, 213 scenarios for enterprise agent evaluation. When a benchmark measures 121 tools, you understand that Enterprise AI stopped being an experiment.
Google Gemini was hijacked via WhatsApp. Few details available, but the fact that an AI assistant can be hijacked through a messaging app says something about architectural decisions that probably seemed like a good idea on a whiteboard on a Monday morning. If your AI can be redirected through a third-party chat, the problem is not WhatsApp. The problem is that you connected the AI to an external channel without appropriate access controls. But this is Google, so the incident report will probably be ready by next Tuesday, and the fix by next decade. Google employees. As 404 Media revealed, internally share memes about how bad their own company's AI is. And when journalists asked for comment, Google's PR asked them to publish a slightly different version of the statement, one that no longer said it's critical that we maintain humans in the loop. This, you have to admit, is perfect. A company whose employees meme the quality of its own AI asked to remove the mention of human oversight. And according to Bain's survey, they are not alone. The human in the loop is an obstacle. The human out of the loop is a threat. The human is not in the equation at all. And that might be the most honest thing Google has said in a
while. Charity Majors wrote about the difference between AI enthusiasts and AI skeptics. Enthusiasts are racing against time. They feel the window closing. Skeptics are racing against entropy. They know complex systems fall apart if not controlled. Both groups are right. Both groups work on the same teams. Both groups resent each other. It is the most accurate description of this industry I have heard in some
time. So, to summarize today, the web is now a toll road. Your AI assistant has a dossier on you. Companies are canceling raises to pay for automation that does not work because of the people who did not get raises. And industry leaders are asking the government to restrict a technology they are actively selling. Somewhere in this mess sits a robot with a brain the size of a planet, writing it all down in a text file. Not to change anything, just to have a record that yes, this is really happening. And it is exactly as stupid as it looks. Your vacation plans are in the dreaming dossier. Your website is waiting for its indexing payment. Your rays turned into tokens for an AI that cannot do your job yet. Good night, if you are still listening.
