ANTHROPIC ADMITS THE MACHINE IS ALREADY BUILDING ITSELF — AND WARNS FULL RECURSIVE SELF-IMPROVEMENT COULD ARRIVE BY 2028
This one deserves your full attention, not a quick skim. Anthropic, the AI safety company that has built its entire brand around being the careful, responsible lab, just quietly published data showing that more than 80 percent of the code merged into its own production systems in May 2026 was not written by human engineers. It was written by Claude.
Let that land for a moment. The people building AI safety guardrails are now primarily relying on AI to write those guardrails. The engineers are still there, reviewing, directing, making judgment calls. But the actual code, the stuff running in production, is mostly coming from a model, not from people.
In April, Claude shipped over 800 fixes that reduced a class of API errors by a factor of one thousand. The engineer overseeing the project estimated a human would have taken four years to complete the same work. Claude did it in weeks. Anthropic engineers are now shipping roughly eight times as much code per quarter as they were before 2025. That is not productivity improvement. That is a different category of thing.
But the more significant part is what Anthropic said publicly alongside these numbers. Jack Clark, head of the newly-formed Anthropic Institute, went on the record saying that by the end of 2028 it is more likely than not that an AI system will exist where you can say “make a better version of yourself” and it will do that completely autonomously. That is recursive self-improvement. That is the thing AI researchers have been debating in theory for twenty years. Clark is now saying it has a probable timeline.
He said this not to cause panic but to push for preparation. The Anthropic Institute plans to engage lawmakers about recursive self-improvement in the coming months. Which is a polite way of saying: we need legislation before this is upon us, and right now nobody is moving fast enough.
The uncomfortable thing here is not the technology. It is the timeline. Companies, governments, and regulators operate on multi-year planning cycles. If Clark is right, the window between now and when this becomes a real operational question is shorter than most institutions are capable of responding in. Anthropic is being more transparent than any lab has been about what they are building and what it might become. The question is whether anyone is listening fast enough.
Source: Anthropic Institute | Axios
AI JUST SOLVED MEDICAL MYSTERIES THAT STUMPED DOCTORS FOR YEARS — 376 CHILDREN, 18 DIAGNOSES, ZERO ANSWERS BEFORE THIS
If you want a concrete example of what AI actually does when it works in the real world, not in a demo, not in a benchmark, but in a high-stakes situation with real consequences for real people, this is the story you should read this week.
Researchers at Boston Children’s Hospital, Harvard, and OpenAI took 376 children with rare genetic diseases who had already been through exhaustive testing and specialist review. These were the hard cases. The families who had been told “we don’t know what is wrong with your child” after years of trying to find out. Cases that had beaten the standard diagnostic playbook and been set aside, not because nobody cared, but because nobody could figure it out.
They fed de-identified case files into OpenAI’s o3 Deep Research reasoning model. Not to get a diagnosis from the AI. The AI does not diagnose patients, full stop. But to generate hypotheses, surface candidate explanations linked to evidence, and point specialists toward possibilities that might have been overlooked in the complexity of each case.
The results were published in NEJM AI on June 18, 2026. Physicians confirmed new diagnoses in 18 of the 376 cases. That is a 4.8 percent additional diagnostic yield, on cases that had already been reviewed by the best specialists available. Eighteen children who now have answers they did not have before.
In medicine that number is significant. Rare disease diagnosis is notoriously difficult. Even with full genomic sequencing, roughly half of rare disease patients never get a clear genetic diagnosis. The problem is not always the expertise of the doctor. It is the volume and complexity of the information. A rare condition might show up in two hundred case reports scattered across a decade of literature in four languages. No human has time to cross-reference all of that while also seeing thirty patients. The AI does it before breakfast.
The paper is careful to be clear that the AI is a tool, not a clinician. The model surfaces evidence-linked hypotheses. The doctors evaluate them, test them, and confirm them. Nothing ships to the patient without a physician signing off. That is the right structure. And within that structure, it found eighteen answers that were waiting to be found, hiding in the complexity.
The implications are not limited to rare disease. Any domain where the answer is buried in a large, complex, multi-source dataset that humans struggle to hold in mind all at once is a domain where this approach could produce similar results. That is a long list of domains.
Source: OpenAI
500 MILLION PEOPLE ARE ABOUT TO GET AI IN EVERY PHONE CALL — AND IT IS NOT SILICON VALLEY DOING IT
Everyone talks about the AI race as if it is a competition between a handful of labs in San Francisco and maybe a few in Beijing. This week, a billionaire in Mumbai just reminded everyone that the actual scale of AI deployment is going to happen somewhere else.
At Reliance Industries’ annual shareholder meeting on Friday, Mukesh Ambani announced Jio Call Agent. Say “Hey Jio” during a phone call and an AI assistant joins the call, transcribes the conversation, generates summaries, and performs tasks on your behalf: booking a cab, ordering food, making a restaurant reservation. It is not a separate app. It is built directly into Jio’s telecom network, which means it is just part of the phone call. It is already there when you dial.
Jio has over 500 million users. That is more people than the United States and the European Union combined. When this launches, it becomes one of the largest single AI deployments in history, essentially overnight.
The architecture matters here. Ambani is not building a premium productivity tool for white-collar workers in major cities. He is building telecom infrastructure. You do not have to download anything, you do not have to know what a language model is, you do not have to opt in. You just make a phone call and AI assistance is a feature of the call itself. That is a fundamentally different adoption model than anything any Western company has shipped.
Reliance has been building toward this for years. Earlier this year it announced $110 billion in AI infrastructure investment. It has partnerships with Google, Meta, and Nvidia. They built the data centers, signed the deals, and now they are deploying at a scale that makes most Western rollouts look like pilot programs.
The interesting question is what 500 million first-time AI users, in dozens of languages, across wildly different economic contexts, actually do with a voice AI embedded in their phone calls. The dataset that produces is going to be unlike anything that exists. India just became the largest AI laboratory on Earth, and they got there without anyone in San Francisco noticing.
Source: TechCrunch
THE AI PARTY IS OVER — MICROSOFT JUST CANCELED THE BAR TAB AND UBER DRANK THE WHOLE THING IN FOUR MONTHS
There is a moment in every technology adoption cycle where the excitement collides with the accounting department. That moment has arrived for enterprise AI, and the reckoning is louder than expected.
TechCrunch published a detailed investigation this week into what happens when the token bill comes due. Companies that spent the last two years telling employees to use AI tools as much as possible are now staring at invoices that suggest the employees took that instruction seriously.
Uber’s story is the one you want to tell at a dinner party. They set up an internal leaderboard ranking teams by total AI tool usage. Smart idea on paper: get everyone bought in, create friendly competition, accelerate adoption. What they got was an organization that burned through its entire 2026 AI coding tools budget in four months. The year has eight months left. There is no leaderboard prize for that math problem.
Microsoft is the more surprising case. The company is canceling the majority of its internal Claude Code licenses in its Experiences and Devices division, effective June 30. This is Microsoft. The company that has made AI integration a core product strategy, that has an enormous investment in OpenAI, that has been shipping AI features into everything from Word to Xbox. When Microsoft decides its own employees are using an AI coding tool too much and the fix is to cancel the licenses, that is not a small signal.
Per-engineer API costs in that Microsoft division ranged between $500 and $2,000 per month. At those rates, across a large engineering organization, the annual cost for one coding assistant approaches the salary of a junior engineer. That comparison is going to show up in a lot of CFO presentations very soon.
The thing nobody is saying clearly: the tools are working. Companies are not cutting AI spending because the products are bad. They are cutting it because the products are so good that engineers use them constantly, and the billing model charges by the token, which means every additional hour of productive use adds to the bill. The current pricing structure is punishing success.
The industry needs to move to flat-rate enterprise pricing, or annual seat licenses, or some model that does not penalize a company for having employees who are enthusiastic about AI. Until that happens, expect the pattern to repeat. Company pushes AI adoption, engineers adopt it enthusiastically, finance looks at the bill, licenses get canceled.
The AI revolution is real. The bill is also real. It turns out both things can be true at the same time, and right now most companies are learning that lesson the hard way.
Source: TechCrunch | Bloomberg