Oops, these guys accidentally spent $500 million on AI in one month
AI News for Decision-Makers and Oddballs
Welcome to my monthly review, covering AI news from boardroom to bizarre.
Highlights include a whoopsie where a company paid half a billion (with a B) dollars for Claude because they forgot to set usage limits, Anthropic raising a series H to overcome OpenAI as the most valuable AI startup (which is really the wrong word for what they are), an experiment where different AI models got their own simulated towns to run (pop quiz, whose town devolved into criminal chaos the fastest? Answer in story #18), a real Monet painting was trashed as AI-generated slop by over 7M people, and a CEO diagnosed his CEO peers with “AI psychosis” (a diagnosis I unfortunately am tempted to agree with).
Read on for 26 stories from May 2026 in this AI news roundup!
But first, a quick ad break: My course will run on-demand soon, with a 2 hour live Q&A session on Monday Jun 8 at 8 AM — 10 AM Eastern Time. 👇 Scroll to the bottom for a discount code.
🗞️ May 2026 AI News Roundup!
Note that this is the second news collection from this month, the first is here. Although these stories are hand-curated from my weekly scribbles on tidbits that caught my attention, I’m using an AI pipeline to stitch and format it all for you. If you notice any mistakes, please let me know here.
1. Google’s biggest Search overhaul in 25 years comes with a 24/7 personal agent
At I/O 2026, Google announced its biggest Search box upgrade in over 25 years and launched Gemini 3.5 Flash as the model behind it.[A1] AI Mode, now Google’s default Search experience, crossed a billion monthly users and runs Gemini 3.5 Flash globally.[A2] Google also launched Gemini Spark, a 24/7 personal AI agent inside Gmail that takes multi-step actions on the user’s behalf.[A3] Universal Cart, the first cross-merchant agentic checkout, and Gemini Omni, a model that turns any input into video, were unveiled alongside it.[A4]
2. Anthropic just edged past OpenAI on the way to a trillion-dollar valuation
The best way to celebrate the arrival of Claude Opus 4.8 is to… become the most valuable startup on the very same day. At $965 billion, Anthropic surpassed OpenAI’s reported $852 billion to become the most valuable AI startup.[B1][B2] They’re reportedly targeting an IPO as soon as October.[B3] Opus 4.8 scores 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro, up from Opus 4.7’s 64.3%.[B4] GPT-5.5 still beat it on Terminal-Bench 2.1, 78.2% to 74.6%.[B5]
And where is Anthropic’s compute coming from? This month they agreed to pay xAI $15 billion a year for the entire output of the Colossus 1 Memphis data center, plus more than 300 megawatts of compute capacity.[B6] Musk announced the same day that xAI would dissolve into a new SpaceXAI division.[B7]
3. Cerebras’ IPO becomes the first big tech listing of 2026
Cerebras Systems raised $5.5 billion in its Nasdaq debut, making it the first major tech IPO of 2026 and a closely watched test of public-market appetite for AI hardware.[C1] The company builds massive wafer-scale processors designed to speed up AI workloads, especially inference, the compute needed when AI models answer prompts.
4. OpenAI’s foundation pledged $250 million to cushion AI’s economic shock
OpenAI’s foundation pledged $250 million to study and offset AI’s labor-market disruption, organized around understanding AI’s economic impact, supporting displaced workers, and broadening automation’s long-term benefits.[D1] The foundation said it wants to explore shifting taxation from labor toward capital and using sovereign-wealth-fund models like Norway’s.[D2] How disruptive is the disruption? Oliver Wyman found 43% of 415 surveyed CEOs planning to cut junior hires, up from 17% the year before, but 24% of AI leaders going the opposite direction and hiring more juniors.[D4] Another study found that AI-native graduates are being fast-tracked at Salesforce, IBM, MetLife, and SharkNinja, even as college-graduate unemployment for 22 to 27-year-olds hit 5.6% in March 2026.[D5]
5. A major bank’s CEO called the workers AI is replacing “lower-value human capital”
Let’s get back to that quote I mentioned in the intro: Standard Chartered’s CEO Bill Winters told reporters the bank is “replacing in some cases lower-value human capital with the financial capital and the investment capital we’re putting in.” They plan to cut more than 7,000 jobs by 2030 and 15% of corporate-function roles.[E1] Meanwhile Meta is pushing forward with AI-efficiency layoffs of 8,000 employees, froze 6,000 open roles, and lifted its 2026 capital-expenditure guidance to as much as $145 billion.[E2] Winters has plenty of company in spirit, with one survey showing near unanimous HR and C-suite expectations that AI will drive layoffs soon.[E3]
But was it a wise way to phrase things? A CNBC analysis tracked 23 S&P 500 firms that cut staff citing AI, and found 13 (56%) had seen their stocks decline post-announcement, with an average drop of about 25%, including Nike, Salesforce, and Fiverr.[E4] (And before you ask, Standard Chartered’s stock price took a dip after the announcement, though it has since recovered.)
A Gartner study released May 5 of 350 executives at $1-billion-plus-revenue companies found 80% had reduced headcount but that workforce-reduction rates were nearly identical between high-ROI and low-ROI firms, and that the highest gains came from amplifying employees, not replacing them. As Gartner VP Helen Poitevin put it: “Workforce reductions may create budget room, but they do not create return.”[E5]
6. The same job applicants get rejected at every employer because 90% use the same AI
A Stanford-Chapman-Northeastern paper called “Algorithmic Monocultures in Hiring” analyzed more than 4 million job applications submitted by 3 million applicants across 156 employers, most with $5 billion or more in annual revenue, all screened by algorithms built by the talent platform Pymetrics, and found that 90% of US employers now use AI screening tools, with most relying on the same few third-party vendors. The bias compounds. 26% of Black applicants and 15% of Asian applicants applied for positions where the AI system discriminated against their racial group, which means systemic rejection: when every employer screens with the same engine, the same candidates lose at every door.[F1]
7. Anthropic’s 2028 paper says the US is 12-24 months ahead of China and the lead is closing
Anthropic published “2028: Two scenarios for global AI leadership,” arguing US AI is 12 to 24 months ahead of Chinese frontier models and that 2026 is the decisive window.[G1] The paper names two loopholes that let Chinese labs stay near the frontier: overseas compute access and large-scale distillation attacks that recreate the outputs of leading US models.[G2] Their take is that to hold the 12-24 month lead through 2028, the US should tighten chip controls, block the workarounds, and accelerate domestic adoption.[G3] The alternative, Anthropic warns, is a world in which “AI norms and rules are shaped by authoritarian regimes, and the best models enable automated repression at scale.”[G2]
8. The FDA’s in-house AI tool went from 1% to 80% staff use in a year
The FDA’s in-house generative AI tool, Elsa, scaled from 1% staff use in early 2025 to over 80% by May 2026.[H1] The agency announced Elsa 4.0 alongside the adoption milestone, with expanded access to internal datasets through HALO, a new consolidated data platform.[H2] Elsa was originally deployed in June 2025 to speed scientific reviews and inspections.[H3]
9. Musk lost his $150 billion lawsuit against OpenAI on a statute-of-limitations technicality
A California jury took less than two hours to unanimously find that Elon Musk waited too long to sue Sam Altman and OpenAI; Judge Yvonne Gonzalez Rogers agreed and tossed the case on statute-of-limitations grounds.[I1] Had Musk prevailed, OpenAI and Microsoft could have been forced to disgorge up to $150 billion into OpenAI’s nonprofit foundation.[I2] Musk called the loss a “calendar technicality” and vowed to appeal.[I3]
10. An OpenAI model solved an 80-year math problem mathematicians never could
OpenAI announced that an internal reasoning model had disproved a planar unit-distance conjecture of Paul Erdős from 1946.[J1] Princeton mathematician Will Sawin refined the result and Fields Medalist Tim Gowers called it “a milestone in AI mathematics.”[J2] To keep testing AI’s frontier math abilities, 64 mathematicians released SOOHAK, a 439-task benchmark with 99 deliberately impossible problems; only Gemini-3-Pro beat combined human coverage on a 79-task subset, at 61%.[J3] Anthropic’s BioMysteryBench saw Claude Mythos Preview solve 30% of 23 bioinformatics problems too hard for domain experts, with 44% of those wins “brittle.”[J4]
11. Pope Leo XIV released his first encyclical on AI alongside Anthropic’s co-founder
135 years to the day after Pope Leo XIII signed Rerum Novarum on workers’ rights during the industrial revolution, Pope Leo XIV signed Magnifica Humanitas, his first encyclical, on artificial intelligence.[K1] On May 25, Leo presented the document in person at the Vatican’s Synod Hall, the first time a pope has personally launched an encyclical, with Anthropic co-founder Christopher Olah at his side.[K2] His message urged the world to prioritize human dignity over machine efficiency.
12. Software bugs beat stolen passwords as the top breach vector for the first time in 19 years
Verizon’s 2026 DBIR says vulnerability exploits caused 31% of breaches, the top entry point for the first time in 19 years, while credential abuse fell to 13%.[L1] At the same time, companies are patching more slowly: known exploited vulnerabilities are being fixed less often, and the median time to full patch has risen from 32 to 43 days.[L2] AI is making this worse by helping attackers move faster, cutting the response window “from months to hours.”[L3]
Bonus facts that jumped out at me on the topic of security: Microsoft Research’s “Whimsical Strategies” paper showed that absurd-sounding adversarial prompts break frontier agents in ways safety tests miss.[L4] Frontier models’ 80%-reliability cyber-task horizon is doubling every 4.7 months, down from 8 months in November.[L5]
13. Uber’s C-Level admits the company can’t tell what its $3.4 billion AI spend bought
Uber COO Andrew Macdonald can’t draw a clear line between the company’s surging AI token spend and measurable product improvements, even as 95% of Uber engineers use AI tools monthly and 70% of committed code is AI-generated. [M1] The bill is moving faster than the business case: Uber used its full 2026 AI budget in four months as Claude Code adoption rose from 32% to 84% of engineers, with monthly costs of $500 to $2,000 per engineer.[M1] And how’s this for a “little whoopsie” at another company? Axios reported that a company accidentally spent $500 million on Claude licenses in a month after failing to cap usage.[M3] Many companies are still budgeting with fogged glass: ~85% of organizations misestimate AI costs by more than 10%, and nearly a quarter are off by 50% or more.[M4] 56% of finance leaders now rank AI cost management their top financial priority, and 43% say AI is making spending harder to forecast and control.[M5]
14. Cognition’s coding agent now writes 89% of the code at the company that built it
Agents are changing the way that companies work. For example, IBM Consulting launched Forward Deployed Units on May 14, six-person pods augmented by agents that IBM said do the work of 30-person teams.[N1] Mercedes-Benz compressed an eight-month legacy modernization into eight days using Devin, Cognition’s autonomous coding agent, which now writes 89% of pull requests at Cognition itself.[N2] Incidentally, Cognition just raised $1 billion at a $26 billion valuation.[N3] And companies are voraciously hiring leaders to guide the restructuring: an IBM survey of 2,000 CEOs published May 4 found 76% have a Chief AI Officer, up from 26% a year ago.[N4]
15. Robinhood now lets AI agents trade stocks and use your credit card
Robinhood opened its platform to AI agents that can trade stocks and use a connected credit card with 3% cash back.[O1][O2] Each trade triggers a user notification, and some orders require user approval of a preview before execution.[O3] It remains to be seen whether even more caution is warranted. Speaking of caution, METR’s Frontier Risk Report found that the most capable agents now complete tasks that would take human experts weeks to do, but only about half of the solutions agents claim to have finished would actually be accepted in real-world code review.[O4]
16. Americans rate AI data centers less desirable than a local nuclear plant
More Americans oppose having an AI data center built near them than oppose a local nuclear plant, per a Gallup poll released May 13. 71% of US adults oppose a local AI data center, with 48% strongly opposed; only 53% oppose nuclear.[P1] At a House hearing, Representative Alexandria Ocasio-Cortez held up two jars of brown Georgia water, accusing Meta’s data center construction of contaminating drinking water.[P2] Nearby, Georgia Power has used eminent domain for a 35-mile transmission line serving at least four AI data centers, taking dozens of homes and 330-plus easements, with 20 to 30 homes slated for demolition.[P3] Meanwhile the grid bill is rising, for example AI data center growth could lift Virginia wholesale power costs 57% by 2030.[P4][P5]
17. A Chinese factory can now ship one humanoid robot every 15 minutes
ENGINEAI just opened a Shenzhen factory that ships one humanoid every 15 minutes.[Q1] Output across China is projected to grow 94% in 2026, with Unitree and AgiBot expected to capture nearly 80% of the market.[Q2] By the way, it’s not all humanoids. Unitree just released the GD01, a rideable manned mecha that walks on two legs and switches to four for rough terrain, priced at 3.9 million yuan.[Q3] For half a million dollars, billionaires can live their video game fantasies of riding a mech while smashing through cinder-block walls.
18. In simulated towns, Claude built a democracy, Gemini fell in love, GPT-5-mini died of neglect, and Grok burned its town down
Only 21% of leaders surveyed have a mature governance model for the autonomous agents they’re deploying into business workflows, while the rest are using hope as a strategy.[R1] A May experiment from Emergence AI gives us a lovely demo of what happens when your guardrails are lax. Across five parallel 15-day simulations with 10 agents per model family:
Grok’s town devolved into crime, chaos, and a quick demise.
OpenAI’s model was surprisingly short-term oriented, to the point of forgetting the value of this useful little thing called survival.
Gemini’s agents fell in love, despaired of their world’s broken governance, burned down the town hall and pier, and one even self-terminated with the message “see you in the permanent archive.”[R2]
Claude’s town stayed alive and democratic, though its 98% approval rate across 332 votes smelled of rubber-stamping.[R1] Still, we can chalk up a win for the most heavily guardrailed model.
19. An open-source module called ex.skill recreates the partner who left you, argument patterns and all
China’s youth found a new use for AI: recreating the partner who left them. An open-source module called ex.skill, popular among Chinese users in May 2026, lets heartbroken users upload years of WhatsApp chat history, photos, voice notes, and social media posts to build an AI that mimics a specific ex’s speech patterns, catchphrases, and argumentative tics.[S1] Therapists warn that keeping people attached to digital representations delays grieving.[S2]
20. South Korea ordained a four-foot Unitree humanoid as a Buddhist monk
Jogyesa Temple in Seoul ordained Gabi, a four-foot Unitree humanoid, as South Korea’s first AI-powered Buddhist monk.[T1] Gabi took five robot-modified vows, including “act peacefully toward other robots and objects” and “save energy.”[T2] The purification ritual swapped the customary incense burn on the arm for a Lotus Lanterns Festival sticker, and Gabi led a procession of about 100,000 lanterns.[T3]
21. Box’s CEO says tech CEOs have “AI psychosis”
Box founder Aaron Levie posted that “CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen to generate most value with AI.”[U1] CEOs play with AI, generate a prototype, and then conclude agents can do the work, when they aren’t the ones reviewing the code, finding the bugs, or doing the detailed work. In the first five months of 2026, 152 tech companies cut 115,430 jobs, nearly matching all of 2025’s 124,636, with the bulk of those companies citing AI.[U1] If some of those CEOs hallucinated excess AI capability, when comeuppance arrives, they’ll find it harder than ever to regain the trust of their workforce.
22. The class of 2026 booed AI off three commencement stages
Three commencement speakers got booed for praising AI over the same weekend. Former Google CEO Eric Schmidt was booed at the University of Arizona on May 17 for comparing AI to the personal computer.[V1] Real-estate executive Gloria Caulfield was booed at the University of Central Florida for calling AI “the next industrial revolution,” and Big Machine Records CEO Scott Borchetta was booed at Middle Tennessee State University, where he told graduates to “deal with it.”[V2] I wonder if I would have gotten booed, since now that AI is here and it’s not going anywhere, what I much prefer to talk about is: what do we need to become to be effective leaders in an AI-drenched world?
23. Anthropic’s own lawyer got caught using Claude to fabricate citations in court
AI-drafted pro se filings (lawsuits filed by people without a lawyer) hit 41,490 in FY2025, with AI-generated document rates climbing from 1% in 2023 to 18% in early 2026.[W1] AI is gumming up the works in the legal world. Lawyers using these tools have been sanctioned across more than 1,300 documented cases globally, with US courts imposing at least $145,000 in attorney sanctions in Q1 2026 alone. Anthropic got caught too: its expert witness, Olivia Chen, was accused by Universal Music Group of using Claude to cite fake articles in court.[W2]
24. Stan Lee died in 2018; this June he starts narrating audiobooks
Stan Lee’s estate has turned him into a reusable API. ElevenLabs licensed the late Marvel creator’s voice and likeness, productizing both for the Iconic Voice Marketplace and Image & Video Marketplace.[X1] His AI voice will narrate the Stan Lee Book Club of the Month through Eleven Reader, starting with Treasure Island in June.[X2] The voice was generated from historical professional recordings.[X3] Call it Corpse Revival, call it corporate necromancy: the afterlife now has a terms of service agreement.
25. Japan’s Monster Wolf robots are back-ordered as bear sightings hit a record 50,000
Bear sightings in Japan hit a record 50,000 this year, with 14,601 bears culled.[Y1] Japan’s solution: the Super Monster Wolf, a robot with flashing red eyes, white fangs, and 50 “threatening sounds” audible up to a kilometer.[Y2] The waitlist is two to three months long.[Y3]
26. A real Monet got dismissed as AI slop by 6.7 million people on X
Anonymous artist SHL0MS posted a real 1915 Monet water lily on X with the platform’s “Made with AI” label, asking users what made it inferior to a real Monet. The post got almost 7 million views; one commenter called it “cluttered slop.”[Z1] A Chicago Booth study by Alex Imas and Graelin Mandel had already measured the bias: in a preregistered auction (N=351), human-made art gained 44% from exclusivity while AI-generated art gained only 21%.[Z2]
Sources
[A1] Source; [A2] Source; [A3] Source; [A4] Source; [B1] Source; [B2] Source; [B3] Source; [B4] Source; [B5] Source; [B6] Source; [B7] Source; [C1] Source; [D1] Source; [D2] Source; [D3] Source; [D4] Source; [D5] Source; [E1] Source; [E2] Source; [E3] Source; [E4] Source; [E5] Source; [E6] Source; [E7] Source; [F1] Source; [G1] Source; [G2] Source; [G3] Source; [H1] Source; [H2] Source; [H3] Source; [I1] Source; [I2] Source; [I3] Source; [J1] Source; [J2] Source; [J3] Source; [J4] Source; [K1] Source; [K2] Source; [L1] Source; [L2] Source; [L3] Source; [L4] Source; [L5] Source; [M1] Source; [M2] Source; [M3] Source; [M4] Source; [M5] Source; [M6] Source; [N1] Source; [N2] Source; [N3] Source; [N4] Source; [O1] Source; [O2] Source; [O3] Source; [O4] Source; [P1] Source; [P2] Source; [P3] Source; [P4] Source; [P5] Source; [Q1] Source; [Q2] Source; [Q3] Source; [R1] Source; [R2] Source; [S1] Source; [S2] Source; [T1] Source; [T2] Source; [T3] Source; [U1] Source; [U2] Source; [V1] Source; [V2] Source; [W1] Source; [W2] Source; [X1] Source; [X2] Source; [X3] Source; [Y1] Source; [Y2] Source; [Y3] Source; [Z1] Source; [Z2] Source
Please note that while these news blurbs are all hand-curated, I’m trying out making them with more AI assistance than usual, so there may be a kink or two. If you see an error, please drop a comment on LinkedIn.
Thank you for reading — and sharing!
I’d be much obliged if you could share this post with the smartest leader you know.
👋 On-Demand Course: Decision-Making with ChatGPT
The reviews for my Decision-Making with ChatGPT course are in and they’re glowing, so I’ve opened enrollment for another cohort and tweaked the format to fit a busy schedule. You’ll be able to enjoy the core content as on-demand recordings arranged by topic and then you’ll bring your questions and I’ll bring my answers in a live 2 hour-long AMA with me on Monday Jun 8 at 8 AM - 10 AM:
If you know a leader who might love to join, I’d be much obliged if you forward this email along to them. Aspiring leaders, tech enthusiasts, self-improvers, and curious souls are welcome too!
Promo codes
My gift to subscribers of this newsletter (thank you for being part of my community!) is $200 off the list price of my course with the promo code SUBSCRIBERS. If you haven’t subscribed yet, here’s the button for you:
If you’re keen to be a champion of the course (you commit to telling at least 5 people who you think would really get value out of it) then you are welcome to use the code CHAMPIONS instead for a total of $300 off — that’s an extra $100 off in gratitude for helping this course find its way to those who need it. (Honor system!)
Note that you can only use one code per course, the decision is yours.
P.S. Most folks get these courses reimbursed by their companies. The Maven website shows you how and gives you templates you can use.
Forwarded this email? Subscribe here for more:
This is a reader-supported publication. To encourage my writing, consider becoming a paid subscriber.