Welcome to June 3, 2026
The Singularity has reached the stage where governments would rather benchmark it than license it. The White House issued an executive order, “Promoting Advanced Artificial Intelligence Innovation and Security,” directing agencies to build a classified test of AI cyber capabilities and to invite developers to voluntarily share “covered frontier models” for up to 30 days before release, while forbidding any mandatory licensing regime. Politico read the lighter touch as the AI industry’s latest win in dodging heavier federal oversight. The labs, freed from preclearance, are racing to feed the machines instead. Google is quietly buying code from Play Store developers to train its coding tools, and Microsoft launched its seven-model MAI family, including a reasoning model it claims beats Sonnet 4.6, a 5-billion-parameter coder cheaper than Haiku, an image model surpassing Nano Banana Pro, and the world’s fastest transcription engine across 43 languages.
Some mathematicians, meanwhile, are showing signs of a siege mentality. Sixteen of them, backed by the International Mathematical Union, published the Leiden Declaration on AI and Mathematics, asking the field to disclose AI use and keep humans accountable for correctness, arriving weeks after a model disproved an 80-year-old Erdős conjecture. The New York Times read it plainly, as a sign that even higher mathematics is now exposed to upheaval from AI.
Mathematics was only the leading indicator, and the rest of knowledge work is catching up. OpenAI’s Codex now ships “Sites,” a Lovable competitor that turns anyone’s prompt into a deployed app at a live URL. A Stanford blind study found law professors preferred AI answers to student legal questions in roughly 75% of 3,000 comparisons, flagging them as harmful a third as often as human ones. Preference measured in a study becomes behavior at planetary scale. ChatGPT became the fastest app ever to a billion monthly users, even as Claude’s smaller base compounds far faster at 640% a year. Codex itself passed 5 million weekly users, and OpenAI bolted on six role-specific plugins so analysts, marketers, salespeople, and bankers can all work without writing code. Microsoft answered from the operating system down, launching Scout, an always-on assistant across Outlook and Teams, Project Solara for agent-first devices, and Execution Containers, a Windows-level sandbox already adopted by OpenAI, Nvidia, Manus, and Nous Research. Anthropic is expanding Project Glasswing, opening its Claude Mythos Preview to roughly 150 organizations across power, water, healthcare, and other newly defended sectors.
All of this autonomy runs on borrowed silicon and borrowed money. Broadcom’s pledge to backstop a record $36 billion private-credit deal, structured to buy Google TPUs and lease them to Anthropic, has compressed yields on the senior tranche to about 5.75%, with the riskier unbacked slice paying 8 to 9%. A CoreWeave-linked data center joined the same rush, raising $900 million in junk notes at 7.5%, part of over $27 billion borrowed this year to pour concrete around GPUs. Microsoft, hunting cheaper compute at the physics layer, unveiled Majorana 2, a topological quantum chip designed with its own agentic AI that improves qubit reliability a thousandfold and pulls its scalable-quantum target forward to 2029.
Superintelligence is acquiring bodies and orbits. Barclays expects humanoid robots to become a $200 billion market within a decade, while SpaceX won FAA approval to test its Starfall capsules, reentry vehicles that will manufacture in orbit before splashing into the Pacific. Commentators noticed the obvious dual use, that a vehicle precise enough to land cargo is precise enough to deliver “rods from God” anywhere on Earth.
While the machines get new bodies, ours keep revealing undocumented features. GLP-1 drugs now appear to reduce the need for knee replacements, seemingly independent of the weight they take off.
Meanwhile, the economy is sorting out who, and what, actually did the work. New York Fed researchers found that remote work, not AI, explains nearly two-thirds of rising unemployment among young graduates, since employers stopped hiring juniors they could not mentor in person. Where AI is the cause, the bills arrive fast. Uber capped engineers at $1,500 a month per coding tool like Claude Code after burning a year’s budget in four months, even as Joshua Kushner’s Thrive Holdings bets $1 billion buying accounting firms to automate white-collar work. Workers are pushing back on the surveillance that automation requires, forcing Meta to roll back a tool that logged keystrokes and screens to train its agents. However, the era’s most audacious promises are mostly coming true, with a New York Times audit of 600-plus Musk claims finding he achieved 75% of his 2015 goals on time.
The future is already here, the last 25% just isn’t evenly distributed yet.
Here's are the four reasons behind my theory why Uber saw it's token costs spin out of control.
1. Users don't have a development plan. They say "Build this" -- and if successful will help the firm -- but with the vaguest of missions/goals, using the highest token consuming engine and then the model grinds and grinds and grinds, eventually providing some output that wasn't intended or was wildly off the mark the user imagined -- assuming the human actually KNEW before hand what was intended. So guess what happens next? The user yells "No, that's not what I want, I want this other thing"...which is equally poorly thought out.
And thus the cycle continues. And continues. And at the end of every conversation, which the user wrongly thinks is getting to the solution, the model says, "want me to do X?" And whether X is a good thing, not a good thing, or unnecessary, the user defaults to yes, and OFF WE GO!!
Vibe coding run amuck, in other words.
2. Users don't know how to write prompts and instructions using well reasoned, precisely phrased, thought through plain language English, and so the model infers, spits back stuff, stuff that isn't useful, and the cycle repeats over and over again.
3. Users employ the models on a whole bunch of non essential tasks, while the user cools his heels. Or they decide to do something with the engine that is 'cool" -- like spinning up a dozen agents to some non essential end, and the token cash register rings continuously.
4. Users like to chat with the models. The models are fun!! And so they chat. And chat. And chat some more. Water cooler conversations without the water cooler.
Other firms are sure to have the same experience, if they haven't already.
The NPR study about job loss being a partial mirage is not a well conducted study. They only looked at a single company, and they were looking at data from before the pandemic going forward it seemed. It appears that this unnamed company went through several internal political shifts about hiring and remote work, but this had nothing to do with AI.
The study is trying to say that companies change their hiring policies, and therefore that is possibly maybe kinda if-you-squint and don't look too closely, the reason why the tech job losses look so bad.
This makes no sense. The editor must have been as confused as the study to have printed the story.
Its practically white-washing. Job losses are a myth. Nothing to see here.