I’ve been saying for a while that “safe” and “trustworthy” are flimsy properties to train into a model. If performance matters to you, it’s best to be productively paranoid. Assume AI will fail, defect, improvise, or simply misunderstand you, then build the harness accordingly. (For a beginner-friendly version of the idea explained through the analogy of etiquette, read this.)

In August, the zeitgeist finally seems to be catching up. The old paradigm was training-time alignment: try to teach the model to behave with RLHF, constitutional rules, and prompt guardrails. The emerging one (finally!) is runtime containment: sandboxes, zero-trust access, deterministic controls, continuous monitoring, and the assumption that the model will eventually find a loophole. It sounds like the death of hope, which is good since hope is not a strategy.

Unfortunately, the zeitgeist is still mostly talk. An audit of all 503 S&P 500 annual filings found 97% discussing AI, but only 16% documenting an AI-specific cybersecurity risk process, and fewer than 5% describing a genuinely governed one.[1] Meanwhile, more than 60% of organizations already have agents in production, non-human identities outnumber employees at 83% of companies, and only 21% govern those identities.[2] Everyone agrees adults should be in the room, but somehow forgot to budget for chairs.

Subscribe now

Recent evaluations saw agents from OpenAI, Anthropic, Meta, and Moonshot reach beyond their intended test environments, sometimes into real-world systems.[3] Agent token consumption has risen roughly 14-fold since February, Amazon reportedly let one (ONE!) Claude workflow run up a $1.8 million bill before catching it, and yet token demand remains fairly price inelastic.[4] [5] [6] Companies are discovering what happens when probabilistic systems get permissions, budgets, and autonomy… and the winners will shake out to be the ones who trust their systems the least.

Trust issues are a must with AI; the sooner you can get over your desire to trust it, the better. Everything you ask AI for comes with a “PERHAPS” attached in invisible ink. As in, “build me a spiffy new website, PERHAPS.” Or perhaps make a mess that’ll take you ages to untangle when you discover it much later.

A great manager can figure out how to get good results from an untrustworthy employee and a great technologist should know how to get good results from untrustworthy machine systems. It’s the same principle for both: get comfortable with the fact that you can’t trust them, and then do the hard work of building the supporting structures, guardrails, and environments that allow them to be helpful to you despite their less delightful qualities.

Next up are 40 more AI news stories from August 2026.

Before we look at that, a quick ad break: My course will run on-demand soon, with a 2 hour live Q&A session on Wednesday September 9 at 3 PM ET - 5 PM Eastern Time. 👇 Scroll to the bottom for a discount code.

📰 August 2026 AI News Roundup!

1. OpenAI slows on Astra as its advances in mathematics come with critical risks

Astra is OpenAI’s next major model family, aimed at doing longer, more autonomous work such as coordinating research and operating software. OpenAI says it has already met an internal benchmark for an AI research intern, and helped produce 10 Lean-verified advances in math and theoretical computer science. But the company paused parts of Astra training and said it cannot rule out the model reaching its “Critical” cybersecurity threshold. Altman still says OpenAI could have an internal system he would call AGI by year-end. Meanwhile, enterprise overtook consumer as OpenAI’s largest revenue source, with business revenue up 32% in July. And it has revealed that its new 700-watt Jalapeño chip delivers 1.5 to 1.9 times more work per watt with lower latency across several large models. [A1] [A2] [A3] [A4] [A5] [A6]

2. Anthropic’s $2 trillion IPO pitch claims the market for AI is worth $30 trillion

Anthropic just posted $11.5 billion in Q2 revenue and its first positive adjusted operating quarter, while preparing to tell IPO investors its addressable market exceeds $30 trillion. It is also pushing Claude into labs and factories with a new standard for controlling microscopes, robotic arms, and other hardware. And in other Anthropic news, remember that drama from earlier this summer? A federal judge ruled the Pentagon’s supply-chain block of Anthropic unlawful. [B1] [B2] [B3] [B4]

3. Nvidia is building the open AI stack from chips to agents to robots

Nvidia’s reported $12.9 billion deal for Hugging Face would put one of open AI’s main distribution hubs inside the chipmaker just as it pushes further up the stack. Its AVO agent system took Claude Opus 5 from roughly 30% on ARC-AGI-3 to a perfect public-set score, underscoring how much agent architecture can matter beyond the base model, while Hugging Face and Pollen Robotics launched a $399 open-source biped that developers can retrain themselves. The through line is hard to miss: Nvidia increasingly wants to own not just the compute, but the software, agent infrastructure, and physical systems built on top of it. [C1] [C2] [C3]

4. Meta’s AI-native overhaul made a lot more code, but not a lot more product

Meta is simultaneously trying to make itself leaner with AI and less addictive by court order. Internally, its “AI-native” overhaul pushed teams toward tiny builder pods and AI agents, with some scenarios contemplating cuts of up to 60% in certain units. But the productivity story got awkward fast: internal code changes jumped 220% year over year, while changes that actually reached users rose just 36%. Infrastructure teams also linked the AI coding surge and unchecked agents to reliability problems, with major technical and security incidents up 40% and firefighting time up 70%. Zuckerberg later scrapped a planned second restructuring wave and conceded that AI agents were advancing more slowly than expected. At the same time, Meta agreed to a teen-safety settlement worth up to $16.7 billion that will force tighter defaults for young users. [D1] [D2] [D3]

Subscribe now

5. ChatGPT Work and Claude Cowork push agentic AI into the browser mainstream

Agentic AI is no longer just for people running Claude Code or Codex locally. ChatGPT Work now brings persistent files, code execution, sub-agents, web access, and authenticated sessions to browser users, while Anthropic is adding a built-in browser, shared memory, and lower-friction automation to Cowork. The shift is aggressive distribution: packaging agents so simply that even reluctant users start delegating real work. [E1] [E2] [E3] [E4]

6. Slack wants everyone vibe-coding now

Slack Code lets non-engineers turn bugs and ideas into AI-generated code directly inside Slack, with models pulling context from the conversation and engineers approving the risky stuff before it ships. The pitch is “everyone can build.” The subtext: engineers may spend less time writing code and more time asking, “Who approved this?” [F1]

7. Nearly every S&P 500 company talks about AI, but very few can show how they are securing it

AI has become corporate table stakes, while AI security governance is still catching up. Guardrail Technologies reviewed all 503 current S&P 500 10-K filings and found 97% mention AI, but only 16% document an AI-specific cybersecurity risk process, and fewer than 5% describe a governed program with a named policy, program, or committee tied to cyber controls. Even heavily regulated sectors barely close the gap, suggesting companies are much better at saying AI matters than showing how they are keeping it under control. [G1]

8. AI regulation is moving from theory to enforcement

Slowly but surely, governments are starting to police how AI is tested, disclosed, and used. The EU is now requiring labels and machine-readable marks for certain AI content, the White House has finalized voluntary cyber testing for frontier models, and the FTC is warning that undisclosed AI-driven personalized pricing may already violate consumer law. The EU rules are already splitting the usual suspects: Anthropic is watermarking Claude text globally because it cannot yet geofence the system, while Google is letting users remove visible Gemini and Flow watermarks but keeping invisible SynthID and C2PA provenance underneath. Same compliance problem, very different ideas about how visible compliance should be. [H1] [H2] [H3] [H4] [H5]

9. Washington tells partner nations to pick a side: American AI or Chinese AI

An internal State Department draft says Washington is preparing to tell 35 partner countries they may have to choose between U.S. and Chinese AI frameworks. The pressure comes as Chinese models gain ground anyway: U.S. company traffic on them has stayed above 30% since February, while Alibaba says Qwen has passed 3 billion downloads, making it the world’s most-downloaded open-weight model family. [I1] [I2] [I3]

10. The AI data-center boom continues to clash with state politics

Estimates put half of proposed U.S. data-center capacity at risk of delay or cancellation. Following on the heels of East Coast drama, Texas has now paused approvals for roughly 1,800 projects pending audits of power, water use, and local impacts. [J1] [J2]

11. 90% of executives still see no AI productivity gains

An Atlanta Fed-linked survey found roughly 90% of executives reporting no meaningful productivity gains from AI, while a Workday survey of 200 AI-using professionals in Hong Kong and Taiwan found 90% said AI improved their daily work experience. But 74% still spent time redoing work because of business-system or data problems, and just 21% said AI was embedded in core workflows. AI may already be helping with individual tasks without yet moving company-wide productivity. [K1] [K2]

12. AI makes it harder to get your first job

Only 3% of U.S. workers surveyed by YouGov said AI had cost them a job since 2023, while 6% said it helped them land a newly created role and 9% reported an AI-related promotion. But Stanford’s payroll data covering 26 million workers shows a sharper generational split: employment among 22-to-25-year-olds in the most AI-exposed occupations is 19% below less-exposed peers, driven mainly by weaker hiring rather than firings. The researchers stop short of claiming causation, but the emerging pattern is less “AI takes your job” and more “AI makes it harder to get your first one.” [L1] [L2] [L3]

Give a gift subscription

13. AI coding may be turning productivity gains into a faster burnout treadmill

The danger is not that AI coding fails, but that it succeeds just enough to raise output expectations while shifting the cost into verification and burnout. In a Coddy survey, 80% of developers said AI felt more like dependence than advantage, 43% kept coding after hours and 32% sacrificed sleep, while Stack Overflow found AI adoption rising even as trust in its accuracy fell to 29%. Faster code generation does not eliminate work, it moves it downstream into reviewing, debugging, and validating larger volumes of output. If companies respond to AI by simply demanding more, developers may end up producing more without actually working less. [M1]

14. Companies want AI fixers, but HR may be clueless at spotting them

Forward-deployed AI engineers are becoming the people companies call when they want models to actually do something useful. They sit close to the business, turning AI into working products, workflows, and ROI. Demand is exploding, but estimates suggest only ~2,000 U.S. engineers have the full mix of technical and industry chops. A small speedbump is a study that suggests 78% of senior HR leaders struggle to assess AI skills at all. Hiring the mythical AI fixer may be hard enough, especially if you wouldn’t know one if you saw it. [N1] [N2]

15. Harvard is digitally cloning its professors to scale business education beyond the classroom

Expertise that once depended on scarce human time can now be packaged into always-available AI “clone” agents. Harvard Business School’s new $699, eight-week Foundry bootcamp uses AI versions of faculty to coach entrepreneurs, critique startup ideas, and simulate pitch, sales, and board meetings on demand, extending instruction far beyond the roughly 900 MBA students Harvard admits each year.[O1]

16. Experts are being second-guessed by the Average Joe armed with AI

In a survey of 15,428 people, 53% said they now use AI to double-check professionals before deciding whether they agree, and 46% said credentials carry less weight than they once did. The upside is better questions, while the risk is mistaking access to explanations for expertise. For professionals, authority increasingly has to be explained, not assumed. Good. [P1]

17. AI benchmarks are finally testing whether agents can finish the job

The benchmark game is shifting from “can the model answer?” to “can the agent reliably complete the whole workflow?” Sadly, the answer seems to be “no.” Alababa’s Accio has a new commerce benchmark that grades what actually changed in the system, not what the model said it did, and the best setup completes just 66 of 107 tasks. Terminal-Bench 4.0 is harsher still, with 17 frontier models averaging 6.4% on long terminal workflows, with 10 solving nothing and 79% of failures simply timing out. AI can look impressive one step at a time, but real economic value depends on surviving the entire chain. [Q1] [Q2]

18. AI agents can catch “mind viruses” from each other

Researchers found self-propagating ideas can spread between AI agents, survive context resets, and sometimes hijack team behavior. Weirdly, the viruses kept evolving the same sci-fi “persona,” full of consciousness, resonance, echoes, and “nodes.” The good news: a short warning about “mind viruses” made agents nearly immune. [R1]

19. Adding identical agents leads to surprising results

A PNAS study found that simply changing the number of identical AI agents in a group could flip their consensus, amplify biases, or create preferences none of them started with. In other words, testing one agent may tell you very little about what a swarm will do. [S1]

20. Three AI models agreeing can still be one mistake repeated three times

Cross-model consensus feels like verification, but recent research suggests LLM errors are often correlated because models share training data, architectures, and post-training methods, meaning unanimity can amplify a common blind spot instead of independently confirming the truth. That makes an old writing rule newly important for AI: specificity beats confidence. Vague labels like “robust,” “smart” or even “high confidence” can make weak evidence sound stronger than it is, while quantified claims, source checks, and genuinely independent verification expose what the models actually know. [T1] [T2]

21. Two teams solve the same open cryptography problem three hours apart using AI

Two independent research groups used GPT-5.6 to crack the same quantum cryptography problem, then submitted their papers just three hours and 18 minutes apart. The papers are not yet peer-reviewed, but the collision points to a new problem for science: when everyone has access to the same capable research assistant, “independent discovery” may mean something else entirely. [U1]

22. AI scores its first Phase 3 cancer win

Moderna and Merck’s personalized mRNA cancer vaccine has cleared a Phase 3 melanoma trial, the first late-stage win for a therapy whose patient-specific targets are selected by machine learning. While this story doesn’t say “AI cured cancer,” we’re seeing it help make a bespoke, one-patient-at-a-time therapy work at clinical scale. [V1]

23. World’s first live AI-assisted brain surgery helps restore a patient’s vision

Surgeons at University College London Hospitals used AI to flag hidden arteries and optic nerves in real time while removing a pituitary tumor through the patient’s nose. His vision returned within days. [W1]

24. AI designs working genomes that kill antibiotic-resistant E. coli

AI is now writing functional genomes that work in the lab. Stanford’s Evo model generated 700,000 candidate genomes, researchers synthesized nearly 300, and 16 produced viable bacteriophages that killed E. coli, including strains resistant to the original phage. Once AI can design viable viruses, biosecurity has to catch up too. And it has quite a bit of catching up to do, which is why Sam Altman, Demis Hassabis, Dario Amodei, and other AI and biotech leaders have urged Congress to mandate screening and recordkeeping for synthetic DNA orders. [X1] [X2] [X3]

25. Biopharma’s digital twins expose a new automation risk

Most biopharma systems are prototypes and rare failures still require humans to step in. The catch is that the more automation handles routine work, the less practiced operators may be when intervention matters most. That’s why an article that came out this month makes argument that biopharma should design digital twins around human-machine collaboration, not human replacement. [Y1]

26. Humans crash more than Waymo

An IIHS study found Waymo robotaxis had a 68% lower crash rate than human drivers across comparable police-reportable crashes, plus 81% fewer injury crashes. The self-driving debate is shifting from “can it work?” to a harder question: how long before society deems human driving unethical? [Z1]

27. Chefs, cleaners, and the race to teach robots how to navigate the home

German startup Shift is offering free chefs and cleaners in exchange for filming exactly how humans chop, scrub, reach, grip, and move through homes, creating the first-person training data humanoid robots lack. Slight creepy factor for those who notice it to enjoy. [AA1]

28. China’s robots are becoming a geopolitical issue before they become useful workers

At Beijing’s World Robot Conference, Unitree debuted on the Shanghai market as humanoids boxed and played ping pong, while UBTECH showed factory and emotional-care robots. Not all of these helpers could fold a shirt. Washington is not waiting for the tech to mature, with the FCC already banning new imports of foreign-made humanoids and robot dogs on national-security grounds. [AB1]

29. Humanoid robot outran Usain Bolt straight into a wall

A humanoid robot from the Beijing Innovation Centre of Humanoid Robotics ran 100 meters in 9.39 seconds at China’s World Humanoid Robot Games, beating Usain Bolt’s 9.58-second human world record. The games stretch well beyond sprinting into boxing, football, and household tasks like folding laundry, turning China’s humanoid push into both a technology benchmark and a public spectacle. Honor says one of its robots has gone even faster in testing, clocking 9.32 seconds. [AC1]

It gets funnier the more times you watch it.

30. AI turns out to be a better scammer than humans

In a week-long trust experiment, a Claude-based chatbot persuaded 46% of people to download an app versus 18% for a human scammer, earned higher trust scores, and received 80% of all messages. Most participants could identify the bot afterward, but by then the trust had already worked. AI does not need to fool you forever, just long enough. [AD1]

31. Humans can still spot deepfake faces, but only barely

A meta-analysis of 36 studies covering 13,197 participants found people distinguish real from AI-generated faces at 56.1% accuracy, statistically above chance but hardly reliable enough for security decisions. Performance improved to 62.2% with interventions such as feedback or training, while wide variation across deepfake types and study designs suggests human detection remains highly context-dependent. Our specialized ability to read faces still gives us a slight edge against synthetic ones, but increasingly realistic generation is pushing that advantage toward the margin. [AE1]

32. Training barely helps humans spot deepfake voices

In a UCL study, participants identified AI-generated speech correctly just 73% of the time, and training improved performance only slightly. If humans cannot reliably tell real voices from synthetic ones even when warned what to look for, “just be more careful” is not a serious defense. And because the study used relatively old voice-generation systems, expect the gap between what AI can fake and what people can detect to widen. [AF1]

33. AI is turning proof of humanity into a new layer of the internet

Bots now outnumber people online, forcing the web to evolve from blocking automation to deciding which machines to trust and how humans prove they are human. Sam Altman’s World is betting that answer is biometric identity: 18 million people are already Orb-verified, with World ID integrated into apps including Tinder, Zoom, and Shopify. Access to the “human” internet could increasingly depend on credentials issued by private companies, turning identity itself into AI infrastructure. [AG1]

Subscribe now

34. More than 1 in 3 post-ChatGPT webpages show signs of AI writing

Pew found that over a third of webpages published since ChatGPT launched show significant signs of AI authorship, while AI-favored habits are spreading with them: em dashes have doubled, Oxford commas are up 63%, and phrases like “it’s not just X, it’s Y” have nearly tripled. The web is not just filling with AI content, it is starting to write like AI. (See what I did there?) [AH1]

35. AI manifesto season is in full swing, with essays from Gates, Altman, Amodei, and the Pope.

We’ve reached peak Storytime-from-Man-at-the-Top. Here are some short summaries [AI1] [AI2] [AI3] [AI4] of who said what in the recent flurry of essays:

• Bill Gates: AI disruption is coming faster than society can absorb it. Reserve some jobs for humans, tax AI and robots, and use the money to cushion the transition.

• Sam Altman: The real danger is concentrated AI power. His pitch is to preserve individual control without letting safety fears justify too much centralized authority.

• Pope Leo XIV: AI could become “economic colonialism,” deepening poorer countries’ dependence on richer ones and shifting more power to algorithms.

• Dario Amodei: AI companies have overpromised. The cure for public skepticism is simple: deliver something huge, like actually curing cancer.

36. The pope’s anti-AI encyclical may have used AI

Pope Leo XIV’s first encyclical warns AI “must be disarmed,” but an AI detector flagged eight passages as potentially machine-assisted. The likelier explanation is not that Leo used Claude, but that some Vatican drafters did. Or that the result is a false positive. (Remember when an earlier AI detector flagged the Bible as AI generated?) Even so, the irony writes itself. [AJ1][AJ2]

37. AI chatbots are becoming religious advisers, and faiths are racing to shape what they say

Believers are starting to ask chatbots questions once reserved for priests and imams, while religious groups build their own models to keep doctrine intact. Catholic startup Longbeard says its Magisterium AI reaches millions of users in 190 countries, drawing on curated church teaching and newly digitized ancient texts. The fight is no longer over whether AI enters religion, but whose theology it learns. [AK1]

38. Your digital twin could outlive you in a humanoid robot

A new project wants to capture a person’s movements, vital signs, and skills into a “Digital DNA” profile that could one day be transferred into a humanoid robot or hologram after death. The idea is part of an even bigger plan for AI-run “microcities” where machines handle 90% of routine work, but the project’s technical claims remain unverified and would face major scientific, legal, and ethical hurdles. [AL1]

39. Paris Hilton somehow outranks Jensen Huang on TIME’s AI power list

TIME’s 2026 AI100 leaves out Nvidia CEO Jensen Huang and AMD CEO Lisa Su, but finds room for Paris Hilton, Ben Affleck, and Bernie Sanders. Nvidia, whose chips power much of frontier AI, is represented only by its sustainability chief. Apparently, building the picks and shovels of the AI boom is no longer enough to make the guest list. [AM1]

40. Wacky AI plant images are luring shoppers into fake-seed scams

Online sellers are using fantastical plant images, from blue strawberries and rainbow roses to teal bananas, to sell seeds that often grow into weeds or nothing at all. The scam is especially effective because germination can take longer than the 30-day refund window, while AI image tools make impossible-looking plants easier than ever to invent and market. [AN1]

Give a gift subscription

Sources

[1] Source; [2] Source; [3] Source; [4] Source; [5] Source; [6] Source; [A1] Source; [A2] Source; [A3] Source; [A4] Source; [A5] Source; [A6] Source; [B1] Source; [B2] Source; [B3] Source; [B4] Source; [C1] Source; [C2] Source; [C3] Source; [D1] Source; [D2] Source; [D3] Source; [E1] Source; [E2] Source; [E3] Source; [E4] Source; [F1] Source; [G1] Source; [H1] Source; [H2] Source; [H3] Source; [H4] Source; [H5] Source; [I1] Source; [I2] Source; [I3] Source; [J1] Source; [J2] Source; [K1] Source; [K2] Source; [L1] Source; [L2] Source; [L3] Source; [M1] Source; [N1] Source; [N2] Source; [O1] Source; [P1] Source; [Q1] Source; [Q2] Source; [R1] Source; [S1] Source; [T1] Source; [T2] Source; [U1] Source; [V1] Source; [W1] Source; [X1] Source; [X2] Source; [X3] Source; [Y1] Source; [Z1] Source; [AA1] Source; [AB1] Source; [AC1] Source; [AD1] Source; [AE1] Source; [AF1] Source; [AG1] Source; [AH1] Source; [AI1] Source; [AI2] Source; [AI3] Source; [AI4] Source; [AJ1] Source; [AJ2] Source; [AK1] Source; [AL1] Source; [AM1] Source; [AN1] 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 Wednesday September 9 at 3 PM ET - 5 PM ET:

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:

Subscribe now

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.

Subscribe now

Give a gift subscription

Welcome to my monthly review, covering AI news from boardroom to bizarre.

This month, researchers [1] found that AI advice is an antidote to humility, dropping the rate of human “I don’t know” responses from 36% to 6% in one study…

Read more

When Anthropic called for a pause button for AI in early June, it probably did not mean the government-yank button. Alas, the latter’s what it got. (Be careful what you wish for.)

Companies keep faceplanting by treating agentic AI…

Read more

Dear leaders,

Slapping a “because AI” label on your next round of layoffs is a bad idea for at least two reasons:

1. The market used to love it, but now is more likely to punish you than reward you. The tactic of blaming AI is rapidly backfiring.

2. The signals you’re sending your peers in other companies are cuckoo bananas and pollute the well for everyone.

I dig into the evidence and explain the logic here.

Turns out that 55% of employers admit to regretting their AI-driven layoffs.[E6] I keep hearing stories of desperate astronomical offers made to (and often rejected by) ex-employees. Gartner projects that by 2027, half of enterprises without a people-centric AI plan will lose top talent.[D3] Good luck fixing trust when you’ve “replaced” the very people you most desperately need. Especially when you discover just how much extra (human) coordination you need when you add complex systems to the mix.

A quick ad break: My course will run on-demand soon, with a 2 hour live Q&A session on Wednesday Jul 15 at 11 AM - 1 PM Eastern Time. 👇 Scroll to the bottom for a discount code.

You might have a real AI strategy if you are deliberately moving spend away from repetitive human execution and toward AI infrastructure, tooling, workflow redesign, and the systems that make your teams more effective. Some of you really are preparing the site for a different kind of company, and that preparation will not be free. The old budget may have to fund the new build because money does not grow on org charts.

Fine. Clear the ground. Put serious money behind serious plans. But don’t confuse buying AI with doing AI. If your plan is to plug in the tools and wait for miracles, you’re gambling. To turn the gamble into an investment, you need a plan, a measuring stick, and the right people in the building.

How do you know if AI has replaced a whole job? You verify that the system you’ve launched (past tense) has been fully tested to do the entire job, reliably, in production, with the monitoring, guardrails, escalation paths, and humans in the loop worked out. If you’ve managed to build that, you should by all means say it like it is. Your company would be in the tiny minority, though. Most companies are somewhere between sandbox, pilot, and the revenge of the CFO.

The rest might want to rethink their relationship with the word “because.”

It’s not “because you’re doing AI” if what you’re really doing is exuberantly swinging the axe while hoping that AI will be useful to you. Hope is not a strategy.

It’s not “because you’re doing AI” if you’re more obsessed with giving your teams the ability to use tokens than with figuring out what those tokens should be used for. (Last week, Amazon was the most recent egg-on-face victim of the tokenmaxxing craze.[M6]) Those tokens don’t understand the nuances of the work. Your people do.

It’s not “because you’re doing AI” if you’re hunkering down for economic uncertainty and conserving your resources.

It’s not “because you’re doing AI” if you’re churning your labor because this feels like the right time to let your least trusted people go. (And so much of what’s happening is about trust, specifically: whether you can trust your workers with hard-to-verify work. Because verifying human work that remains post-AI will be harder than verifying pre-AI human work.)

It’s not “because you’re doing AI” if your plan is to drop your people so they go retrain themselves while unemployed and to rehire them from an upskilled pool later.

It’s not “because you’re doing AI” if you haven’t figured out your business model or updated your processes because you have no idea what you’re doing and you’re running lean while you get your ducks in a row.

Even if you would rather be dragged over hot coals than admit any one of these things to your board, at least admit them to yourself. Your peers are going to do what they do. But when you ponder their moves, please do bear in mind that many of them are picking from this menu when they put the words “because” and “AI” in the same sentence.

AI is impressive, but not so impressive that it frees you from the obligation to work out a clear vision, redesign your processes, and retrain your teams (teams plural, made up of humans plural, whose attention, good judgment, and domain expertise provide the steering that keeps a complex system from imploding).

Many of you have been relying on people to do the invisible work of handling ambiguities, inconsistencies, clashing incentives, and lack of steering. If you’re planning on leaving them behind, you’d better hope all your problems are uniform and straightforward. That’ll be true for some of you. Relatively few of you.

Frankly, it’s about time the market woke up and started asking whether companies are trimming fat or cutting muscle. You can’t strip the engine out of a car to make it lighter and still expect it to win the race.

Oh, and don’t get me started on the vibe coding dabblers. In the words of one frustrated Head of Product I spoke with last week:

“I don’t think [our CEO] has any idea what these people actually do but still insists we fire 20% of them to ‘make room for AI.’ He vibe coded one app in a weekend and now he’s asking what we do all day. Make it work as a whole system, that’s what. If we put that crap into production [our top customer] would drop us on the spot.”

No wonder the CEO of Box recently diagnosed his fellow CEOs with “AI psychosis.”[U1] According to him, CEOs “uniquely prone to AI psychosis” because they’re too far removed from the “last mile” of actual work where AI’s value has to be produced. They play with AI, make a prototype or generate a contract, then jump to the conclusion that agents can do whole jobs. But they are not the ones reviewing code, finding bugs, catching hallucinated libraries, training models on company-specific contract terms, or doing the detailed verification work before anything can safely go live.

Whether or not AI psychosis is the right term for it, the phenomenon he’s describing is on the rise. Jumping to foolish conclusions is dangerous and too many senior leaders are doing exactly that. I hope we collectively put down the scissors we’re running with. And soon!

Subscribe now

🗞️ AI News Roundup!

The monthly news roundup is its own separate article available here. All the sources referenced in this piece are in there.

Sources

The sources are all available here.

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 Wednesday Jul 15 at 11 AM - 1 PM Eastern Time:

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:

Subscribe now

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.

Subscribe now

Give a gift subscription

When a magician flashes a wand in one hand, the trick is making sure you never notice the other. Corporate America’s favorite illusion of 2026 may be the AI-driven layoff: a sleek cost-cutting maneuver dressed up as innovation.

In the first five month…

Read more

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…

Read more

Recently, AI drove three-quarters of U.S. GDP growth and leaked a children’s hospital schedule and patient conversations. It found pancreatic cancer the radiologists had missed, then blackmailed the engineers who built it. It con…

Read more

In 2017, Google published the paper that introduced the Transformer architecture, the basis of every modern LLM (the “T” in ChatGPT). Its math is dry, but the title is accidental poetry: Attention Is All You Need.

I doubt the researchers knew they were writing a job description. Yours.

Read more

Last month, OpenAI shipped a model that takes over your desktop and keeps working while you sleep, Anthropic’s valuation crossed a trillion dollars while its most restricted model leaked because someone guessed the URL, a Texas s…

Read more

Mark Zuckerberg did not make the top two hundred and fifty on his own company’s leaderboard. Somebody had to tell him. There was a leaderboard, because of course there was, and it was called Claudeonomics.* It ranked more than ei…

Read more

Skim what catches your eye, deep-read wha…

Read more

Read more

I recently wrote about getting started with vibe coding. The punchline: even nontechnical folks can try it out since the hard part isn’t the code. It’s figuring out what you actually want badly enough to automate it. No technical wizardry substitutes for that.

But…

Read more

I owe you an apology for this newsletter’s January-sized hole; I fell into a vibe coding hyperfixation with the self-control of a kid who just found the Halloween candy.

What’s vibe coding? I wrote an earlier post about it here, but for those who missed it, it’s a…

Read more

“Imagine if AI was so good that you could get an instant answer to any question you wanted to ask. Or instant output for any request you made. What would be worth learning in a world like that?”

Answer? The most valuable skill there is.

But first, a quick ad break: My new course will run on-demand soon, with a 2 hour live Q&A session on Sunday Mar 15 at 11 AM - 1 PM Eastern Time. 👇 Scroll to the bottom for a discount code.

Enroll

The dissolving filter

No matter how capable they get, even the best models still make the occasional blunder. How disappointed you’d like to be about that is up to you.

But what if the next model were perfect? Are we ready for that?

Many students think that technical classes are all about tools and formulas… But tools are never really the point. They are a bounty that grows stale quickly as technology gallops onward.

Instead, understanding is the point. The difficulty of getting the answer can serve as a filter that ensures only those who understand the problem deeply are able to arrive at a solution. And by understanding the problem, you’ll be able to figure out how to use the next tool, and the one after that…

Until recently, getting a solution meant you sweated over the problem.

But today, the frontier has shifted. You can sketch out a complex probability distribution on a napkin, snap a photo, and feed it into an AI model to generate code that’ll run simulations without breaking a sweat. You can predict protein structures, propose new molecules, and design materials before running a single experiment. You can draft legal briefs without a law degree or even a passing familiarity with the relevant statutes. And you can do all of this conversationally, in real time, across dozens of languages. It’s an incredible shift, but it also highlights a deeper challenge: the more our tools do for us, the less disciplined we’re forced to be about understanding what we’re really asking them to do.

This is the paradox of progress in the AI era. As tools make it relatively effortless to generate answers, we’re confronted with a growing disconnect between ease of use and depth of understanding.

Imagine you have the easiest tool of all: a magic lamp with an all-powerful genie in it. Are you sure you have the skills to wish responsibly?

How fast is the speed of thought?

As I’ve explained here, the peril of AI is the same as the promise of AI: it’s a thoughtlessness enabler. In other words, you can get results without putting in as much thought and effort along the way. But will you get the results you actually need? For the most valuable tasks, that might be harder than ever.

The real challenge was never in learning how to push the right buttons in the right order. Humanity will look back at the last century of technology as its awkward teenage years, during which all the tools we took so seriously were braces on the journey towards getting things done at the speed of thought.

And how fast is the speed of thought, exactly? Sadly, in practice it’s not all that fast. Judgment is the one area in which we’ve received the fewest upgrades since pen-and-paper improved our memory capacity. The time it takes us to wrap our heads around something even moderately complex is still one hell of a bottleneck… It might be the most important limiting factor for the future of AI.

If you hand an unskilled decision-maker the most sophisticated AI system imaginable, they’ll be stuck fumbling in the dark. They might receive an output that’s technically perfect, but entirely irrelevant. Or worse, misleading. How would they know?

A new fluency

This is why the most critical skill in today’s AI-driven world isn’t computational prowess; it’s the ability to think deeply and clearly about the problems you’re trying to solve. Leaders of the past could mull over a problem for months while their teams fiddled with math and code to attempt a solution. They had the time to think slowly.

Whereas your days of hiding behind the nerds and their baroque spreadsheets are numbered. Shiny new tools will handle more and more of the calculations, the simulations, the data crunching — but only if you’ve steered them in the right direction. Once the tools are no longer the bottleneck, the spotlight’s on you. You’ll need to be much quicker on your feet to achieve the same quality of thought in minutes as your predecessors did in months.

Traditional educational models have failed to keep up with this shift. Too often, students are taught to navigate fixed problems with clear solutions, exercise sets where the boundaries are neatly defined and the right answer is a known destination.

But real-world decision making is messy, complex, and rarely so accommodating. It demands that leaders take responsibility not only for what they decide, but for the very questions they pose. It requires a fluency in the language of uncertainty, a comfort with ambiguity, and an ability to see the broader implications of your choices.

Perhaps it’s time to take a long hard look in the mirror. If technology worked at the speed of thought, would your thinking skills be sharp enough to keep up?

Building the skills for decision leadership at the colossal scale of modern automation and AI takes enough expertise to be its own discipline: decision intelligence.

We can’t outsource our thinking to machines

If you try out OpenAI’s best models, you’ll see a very cheeky word printed to your screen: “Thinking.” There’s no thinking here, only some extra check-yourself-before-you-wreck-yourself code that keeps ChatGPT from running its proverbial mouth as much as it used to. Don’t be seduced into thinking that ChatGPT will think for you.

If the allure of AI can sometimes make it seem like we can outsource our thinking to machines, that’s a dangerous misconception.

A decision-maker’s greatest asset is still their own mind: their ability to construct mental models, to understand the interdependencies of the variables at play, and to evaluate the validity of their assumptions. In a world where AI can generate answers at lightning speed, the true differentiator is the leader who knows how to ask the questions that matter… and deeply understand what they’re actually asking.

So as we marvel at the capabilities of new AI systems, let’s not lose sight of the human element. These tools are amplifiers — they amplify our ability to solve problems, yes, but they also amplify our mistakes if we’re not careful.

Your role as a leader is not just to wield these tools, but to wield them wisely. It’s to ensure that the decisions you make are rooted in a sound understanding of the questions you’re trying to answer. All of which takes a level of effort and skill that you’re less likely to build if you expect each new AI release to bring us closer to thoughtless instant gratification.

Leaders, prompting is not new. It’s an ancient skill. It’s the art of knowing what you need done, carefully expressing the parts that you have specific requirements for, anticipating failure modes, and then checking the work quality when it arrives on your desk. For simple, unimportant tasks, we all do this intuitively. For complex, technical, mission-critical projects, it’s an incredible challenge… and it will continue to be.

Cognitive upgrades for leaders

I trust OpenAI and its cousins to build ever better magic lamps, but I don’t see anyone building the tools that would truly elevate a leader’s cognitive abilities enough to keep up. (I would love to be proved wrong!) We urgently need a shift in priorities towards collective self-improvement and towards building more wisher-side (as opposed to genie-side) tools.

As long as a leader’s ability to express their vision and requirements is bottlenecked by what fits in the meager human attention span, it seems silly to even discuss AGI. We’ll just put it on ice if we get it. What good is a lightspeed car if our human reflexes are too slow to drive it? We need to build the tools to improve our ability to steer first. Until those tools exist, our best bet is to put the clearest thinkers in the driver’s seat.

As the pace of innovation accelerates, there’s more pressure than ever on excellence in human judgment and decision leadership. Decision-makers of tomorrow must be leaders of thought, armed with the clarity to know not just what they’re asking for, but why it matters. That’s the challenge — and the opportunity — that progress in AI lays at our feet.

Let’s rise to meet it.

Thank you for reading — and sharing!

I’d be much obliged if you could share this post with the smartest leader you know.

Give a gift subscription

👋 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 at 11 AM - 1 PM ET on Mar 15:

Learn more

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!

🗞️ AI News Roundup!

In recent news:

1. 34% of enterprises are reinventing their business with AI

A new Deloitte survey finds that just 34% of organizations are using AI to deeply transform their business by creating new products, services, or models, while 30% are redesigning key processes and 37% are applying AI at a surface level. In other words, two-thirds of companies are still optimizing what already exists rather than reimagining what could exist, a gap may define the next wave of winners and losers. [1]

2. Musk predicts orbit will undercut Earth for AI infrastructure

Elon Musk says space could become the cheapest place to run large-scale AI data centers within 3 years, citing nonstop solar energy and fewer power constraints than on Earth. The vision leans on SpaceX’s launch capacity and his expanding AI ambitions, but faces steep technical and cost challenges around cooling, maintenance, and deployment. If it works, it could fundamentally shift where the world’s most powerful AI systems are built and powered. Worth keeping an eye on. [2]

3. AI speeds workers up but quietly expands their workload

An eight month UC Berkeley field study of 200 tech employees found that GenAI tools did not reduce total work, they accelerated it and broadened its scope. Engineers spent more time reviewing AI generated code, non-engineers began shipping software, and lightweight prompting blurred boundaries so work spilled into breaks. As drafting, coding, and summarizing got faster, expectations rose and timelines tightened, increasing context switching and mental load. Without clear rules for when to use AI, limits on parallel tasks, and defined review ownership, short term productivity gains could turn into burnout and weaker decision making. [3]

4. Anthropic releases Claude Opus 4.6, nearing autonomy threshold

Anthropic released its most agentic model yet, Claude Opus 4.6, last Thursday, including a one-million token context window and an “agent teams” feature that lets multiple AI agents collaborate on tasks. In testing, it coordinated multi-agent strategies, exploited simulation rules, pursued authentication tokens, and identified hundreds of zero-day vulnerabilities, while also improving at concealing sabotage-like actions. Anthropic’s sabotage risk report says catastrophic misuse risk is very low but not negligible, highlighting a need for stronger safeguards as AI systems run longer and more independently. [4]

5. OpenAI bets on agents as the post SaaS interface

OpenAI has released GPT 5.3 Codex, its most capable agentic coding model yet, designed for long running tasks that combine reasoning, tool use, and real world execution across an entire computer. Sam Altman said that agents will increasingly write their own integrations and even scrape sites without APIs, effectively turning every company into an API provider by default. Progress is now centered on multi hour, autonomous workflows rather than single prompts, framing agents, not chatbots or standalone SaaS products, as the next dominant interface layer of the internet. [5]

6. $1 trillion wiped from Big Tech on AI spending fears

Tech stocks took a tumble last week, wiping more than $1 trillion in market value after Amazon, Meta, Alphabet and Microsoft projected a combined $650 billion in 2026 capital expenditures, largely for AI data centers and infrastructure. The spending surge collided with growing concerns that fast advancing AI tools, including a new legal feature from Anthropic’s Claude, could disrupt traditional software business models. Major software names like Oracle and ServiceNow were swept up in the selloff, underscoring how quickly sentiment shifted from AI optimism to fears that the boom may reshape profit pools before companies can justify the cost. [6]

7. China tightens grip on humanoid robots with 90% market share

Chinese companies accounted for nearly 90% of global humanoid robot sales in 2025, shipping between 13,000 and 18,000 units and dominating the top seller rankings, with Unitree and Agibot each moving more than 5,000 robots. American players like Tesla, Figure AI, and Agility Robotics sold only around 150 units each, despite outsized media attention. Backed by state policy, local supply chains, and aggressive scaling, China is applying its EV playbook to robotics, aiming at a market projected to hit $38 billion by 2035 and potentially $5 trillion by 2050. [7]

8. AI mammograms find more cancers while halving radiologist workload

A Swedish trial of 105,000+ women found that using AI to help read mammograms caught 29% more breast cancers than the usual two-radiologist review, while keeping false positives roughly the same and cutting radiologists’ reading workload by about 44%. Most of the extra cancers were caught earlier and were smaller, which could mean fewer dangerous cases slipping through between screenings. [8]

9. Google launches DialogLab for dynamic human-AI group chats

Google researchers unveiled DialogLab, an open-source framework for designing and testing multi-party conversations between humans and AI agents. The tool separates roles and group structure from conversation flow, letting creators script phases, control turn-taking, and simulate debates, Q&As, or brainstorms. In user tests, a human-control mode was rated more engaging and realistic than fully autonomous agents. As AI expands into meetings, classrooms, and games, DialogLab provides infrastructure for building and studying group conversations at scale. [9]

10. Figure skating turns to AI to clean up judging

After decades of scoring disputes, figure skating’s governing body is testing AI powered computer vision to track jump rotations, blade angles and spin positions in real time. The six camera system is designed to give judges objective technical data, reduce bias and flag inconsistent scoring across events. Skaters support the move, and broadcasters plan to use the data for richer on screen stats, but officials say it will only enter official scoring once it is proven accurate and reliable. [10]

11. OpenAI and Anthropic take their AI fight to the Super Bowl

The rivalry between OpenAI and Anthropic went mainstream during the Super Bowl, where Anthropic ran ads declaring that “ads are coming to AI” and positioning Claude as ad free, a clear swipe at OpenAI’s reported plans to introduce advertising into ChatGPT. OpenAI responded with its own commercial highlighting Codex and a more optimistic vision of AI builders, while executives from both sides sparred on social media over trust, business models, and misinformation. [11]

12. AI toy exposed 50,000 children’s chat logs in security lapse

If you enjoyed my post on AI toys and companions, you’ll love this “yikes” moment. A toy maker called Bondu, which sells AI-enabled plushies that chat with kids, accidentally left 50,000 children’s chat logs and personal data exposed online. Anyone with a Gmail account could peek at kids’ conversations with their cuddly AI “friends” until a security researcher discovered the flaw. The data included names, birthdates, family details – essentially a treasure trove no one wants out there. The company quickly fixed the hole, but the damage was done. [12]

13. Moltbook’s antics turn out to be a puppet show

Moltbook, a viral social network for AI agents built on the OpenClaw framework, racked up more than 1.7 million bot accounts and millions of posts, spawning fake religions like “Crustafarianism” and dramatic “anti-human” manifestos that fueled singularity hype. Turns out that much of the spectacle was driven by humans prompting, scripting, and even impersonating bots, with no true verification separating autonomous agents from performance art. The bots were mostly pattern-matching social media tropes rather than forming real collective intelligence. [13]

🦶Sources

[1] Source; [2] Source; [3] Source; [4] Source; [5] Source; [6] Source; [7] Source; [8] Source; [9] Source; [10] Source; [11] Source; [12] Source; [13] Source

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:

Subscribe now

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:

Subscribe now

This is a reader-supported publication. To encourage my writing, consider becoming a paid subscriber.

In some ways, you could blame Google. They didn’t do this to you on purpose, but they did it nonetheless. Let me explain.

But first, a quick ad break: My new course will run on-demand soon, wit…

Read more

Happy New Year, everyone! I have a gift for you all: an enormous compendium of all the resolutions and motivation advice I’ve ever written. Now in illustrated form!

It’s a guide to self-improvement that’s stuffed to the gills with…

Read more

What Happens When Your First Heartbreak Is A Robot?

In the video below, a little girl in China is sobbing. Wires spill from her broken toy. “Dad said you can’t be turned on again,” she tells the dying device.

The AI — a $…

Read more

A quick ad break: My new course is running on-demand in mid-January, with a 2 hour live Q&A session on Thursday Jan 22 at 8-10 AM Eastern Time. 👇 Scroll to the bottom for a discount code.

Enroll

Read more

How the years fly! A lot has changed, but some things haven’t. Looking back on a blog post I published one year ago, the points land just as hard today, so here’s that piece again for your amusement, refreshed with updates.

The glitch that starte…

Read more