Coté

Thu, 04 Jun 2026 06:42:36 +0200

That means each employee’s AI spending cap is ~11% of that median compensation package.

Last year, there were a few anecdotes about high growth tech companies spending $100,000/year per head on tokens. That seems like it’s coming to end.a

🔗 Uber Caps Usage of AI Tools Like Claude Code to Manage Costs

remote work reduces on-the-job training

Wed, 03 Jun 2026 07:41:02 +0200

According to the Fed’s analysis, youth unemployment has risen significantly since the coronavirus pandemic, and hasn’t receded in the same way that unemployment numbers for older, more experienced college graduates has in recent years. The analysis notes that the prevalence of remote work has increased since COVID-19, and it believes those two trends have more than just a correlation.

“Our analysis suggests that these trends are related, with remote work making it more difficult for managers to train and mentor new employees,” the Fed said of its data. “Accordingly, companies may be reluctant to hire less-experienced workers in distributed work arrangements.”

🔗 Remote work – not AI – is killing job prospects for the youth

Defeating Conway's Law

Wed, 03 Jun 2026 07:23:49 +0200

Try using a platform to combat Conway’s Law and organizational friction caused by too many groups/silos.

This matters because it removes the structural excuse for fragmentation. When a single platform surfaces all the controls a unified team needs, there is no longer a technical reason to keep five separate teams in five separate rooms. The organisational argument for siloes collapses alongside the technical one.

Conway’s Law says that a system will be shaped - organization sub-divided - as a replica of the orgnonzatikn that built the system.

(Yes, I know this is a misreading of the “communication structures” part of Conway’s but this is the use the street has found.)

Most people will tell you that people eat technology for breakfast. You can’t change how people work or how a (large) company is structured by simply installing some new system.

As Harry Thambi outlines, I’m not sure that’s always the case. When a new technology removes the need for specialist teams because it automates and removes the toil of the people running the technology, you have a chance to collapse the groups, to have less silos, and, thus, defeat Conway’s Law.

That requires a truly integrated platform. Interested in one, why not TryTanzu.ai?

🔗 Broadcom Knight’s partner blog: Silo busters - a unified platform needs a unified team - Broadcom News and Stories

you can't measure productivity

Wed, 03 Jun 2026 07:18:52 +0200

The [Wells Fargo] CEO named auditing, testing, legal, contracts, patent filings, pitchbooks in investment banking and credit memos as a handful of areas across the company executives see room for AI to improve processes. “How much of that actually results in pure margin or return expansion is to be seen.” Scharf said, since competitors will be chasing similar AI goals, but it is “a net positive” for the company’s future expense base.”

I’m down with the Martin Fowler idea that measuring productivity is sort of a waste. Rather, it’s better to measure output, yes, but:

Productivity, of course, is something you determine by looking at the input of an activity and its output. So to measure software productivity you have to measure the output of software development - the reason we can’t measure productivity is because we can’t measure output.

Is the bank making more money without breaking laws? Then it’s probably fine.

🔗 Wells Fargo CEO: AI’s effect on employment is ‘complicated’

security over features

Wed, 03 Jun 2026 07:11:08 +0200

From what I can tell, every core part of the software stack is stopping what they’re doing and taking care of the flood of new, AI-driven security issues.

🔗 Java Maintenance Engineering Shifts Focus on Quarterly Critical Patch Stabilization

🤖 “descended into madness" - Backrooms

Wed, 03 Jun 2026 07:07:30 +0200

Original: A Backstory from My Backrooms by Paige K. Bradley. Summarized by AI on June 3, 2026.

{I love backrooms. One of the first things I did with AI image generator was make endless empty malls and backrooms. So good. -Coté}

A stray 2019 4chan post about a bland, fluorescent-lit interior sparked the viral myth of the backrooms, a concept of endless, liminal spaces that feel familiar yet threatening. Its resonance lies in the idea of “no clipping” from reality—slipping into a hollow, game-like purgatory where meaning and orientation fail. The condition of being lost, as one user put it, is to have “descended into madness.”

Artist Jan Vorisek’s installations, with their yellow PVC curtains, dislocated objects, and fragmentary videos, evoke the same uncanny emptiness. His works inhabit a zone between physical matter and simulation, where sculptures resemble 3D game props and rooms become stages for absent narratives. The viewer is left to confront whether the space—or their own perception—is real.

The essay links this aesthetic to premium mediocrity, a cultural phase of glossy surfaces and hollow interiors, where urban spaces like shuttered storefronts become physical backrooms. Liminality is both a visual experience and a social condition, reflecting a civilization suspended between exhausted industrial modernity and early-stage digital post-scarcity.

Through art and literature, the theme recurs: Dennis Cooper’s novel God Jr. literalizes the descent into a hollow 3D monument and a game world as a failed attempt at grief and control. The backrooms and Vorisek’s “incomplete interiors” mirror that impulse—structures built for meaning that only expose their own emptiness.

The backrooms phenomenon has since drifted toward a Gothic sensibility, as cultural imagination projects fragmented narratives, invisible antagonists, and dread into blank architecture. Its power lies in what is missing, allowing fear, nostalgia, and hallucination to fill the void.

In the end, the backrooms are a metaphor for contemporary life under the weight of digital simulation, urban vacancy, and mediated experience. They remind us that when the illusion shatters, what’s left is an ambient awareness of loss—and the quiet admission: you lost.

Links

🤖 A Backstory from My Backrooms – How a single 4chan post about a yellowed, empty room evolved into a cultural metaphor for liminal dread, digital hyperreality, and the art of dislocation.

Summarized by ChatGPT on Jun 3, 2026 at 7:07 AM.

🤖 Valiantys-Glean Partnership Bets That Cross-Platform Knowledge Graphs and Behavioral KPIs Are What Move Enterprise AI Past Pilots

Wed, 03 Jun 2026 07:05:11 +0200

Original: Enterprise AI is still stuck at experimentation – Valiantys and Glean think they know why by diginomica. Summarized by AI on June 3, 2026.

Most enterprise AI pilots stall, and the diagnosis from Nathan Chantrenne, Chief AI Officer at Valiantys, is that the field measures the wrong things and fragments its tooling.

The dominant success metric - “employees save four hours a week” - tells you nothing, because nobody knows what those hours become; they might just mean more coffee. The partnership being pitched is Valiantys, an Atlassian-centric consultancy, joining with Glean, an enterprise AI platform valued at $7.2 billion on roughly $300 million ARR as of May 2026.

Maturity varies widely, and the brakes are predictable. In Europe, governance and security are the primary obstacle; in North America, less so.

Tech-savvy firms have the opposite problem: they over-experiment.

Companies test 15 different technologies and let every team pick its own, which collapses at scale because there’s no unified strategy for compliance, security, or cost. Each platform ships its own AI tooling and data layer, and none reaches well into another vendor’s stack - so technical silos mirror organizational ones, and the cross-functional teams meant to bridge them get treated as a burden rather than an authority.

Glean’s role is the connective tissue. Its permissions-aware Knowledge Graph¹ is a semantic layer that links data across applications - so the system knows a Salesforce opportunity ties to a ServiceNow case and a Jira ticket.

Valiantys, which began moving beyond pure Atlassian work about 18 months ago (anchored by the July 2024 Contegix acquisition for North American expansion), is taking this to market first in its native strengths: software development lifecycle modernization and enterprise service management, with an eventual ~50 use cases in view.

The KPIs Chantrenne cares about are behavioral, not cosmetic. For development, the headline is idea-to-production time, dropping from weeks or months to two weeks or, in extreme cases, two days.

For service management, the entry metric is ticket deflection² - 50 to 70% of a typical 1,000-monthly-ticket desk is repeatable level-one work that can be automated.

The harder, more revealing layer is agent productivity on level-two tickets and how many knowledge base articles agents contribute - both signals of whether the work itself is actually changing.

The recurring lesson is that change management, not technology, is the binding constraint. “You can have the best possible technical solution out there. If you’re unable to bring the people with you… people will do everything that they can for the project to fail."

The prescription for firms that have spent heavily without returns: define real business KPIs, centralize governance, bring people along, and narrow from 15 technologies to two or three - then go nearly all-in and run with it.

Links

🤖 Enterprise AI is still stuck at experimentation – Valiantys and Glean think they know why - Enterprise AI stalls because firms track meaningless productivity metrics and scatter across too many tools; the Valiantys-Glean bet is that a cross-platform knowledge graph plus behavioral KPIs and centralized governance is what moves pilots into production. Change management, not tech, is the real bottleneck.

Knowledge Graph - a data structure representing entities (people, documents, tickets, accounts) and the relationships between them, allowing software to traverse connections rather than treat each record in isolation. “Permissions-aware” means it respects each user’s existing access rights when surfacing linked data. ↩︎
Ticket deflection - resolving a support request automatically (via self-service, bots, or automation) so it never reaches a human agent. ↩︎

🤖 AI Collapses Build Costs but Expands Alignment Burdens for Senior Engineers

Wed, 03 Jun 2026 06:18:35 +0200

Original: Is this sustainable? by Jamie Hurst. Summarized by AI on June 3, 2026.

AI has collapsed the distance between idea and implementation. Senior engineers can now move from concept to working proof-of-concept in days, bypassing the old cycle of proposals, approvals, and sequential team work. This shift has replaced slide decks with demos, rewarding concrete experimentation over theoretical cases.

Organizational alignment has become the new bottleneck. Multiple teams can quickly produce overlapping solutions, making coordination and convergence harder even as technical velocity rises. Solving more problems faster creates fragmentation, and aligning across the org consumes more time than before.

AI has made senior roles both broader and less sustainable. Senior engineers now combine hands-on coding with high-level strategic writing, juggling multiple initiatives at once. Mentoring and unstructured thinking time have been squeezed out because they cannot be accelerated by AI, and the workweek has filled with deliverables and context switching rather than reflection.

The accelerated role favors those who adapt to AI tools quickly, redistributing internal influence toward builders who can demo solutions themselves. This bias to action amplifies skill gaps, giving early adopters more voice in shaping organizational direction while leaving slower adopters marginalized.

Scope has expanded because the entire discipline of developer experience has grown under AI pressure. Internal platforms that once served thousands of humans now also serve unbounded AI agents, shifting DX from a neglected function to a board-level concern. But this growth brings political, org-wide programs where success is diffuse and harder to attribute, and measurement frameworks have not yet caught up.

The underlying tension is that productivity gains have been captured by output volume rather than by higher-quality work or more sustainable pacing. Expectations rise faster than efficiency, leaving senior engineers paddling harder to stay ahead, with the sense that current velocity is real but ultimately unsustainable.

Links

🤖 Is this sustainable? – AI accelerates proof-of-concept development, but organizational alignment, mentoring, and thinking time are squeezed as expectations rise faster than efficiency.

Summarized by ChatGPT on Jun 3, 2026 at 6:18 AM.

Enterprise self-harm: cleaning the data is the hard part

Tue, 02 Jun 2026 12:25:56 +0200

I think the critical part of it was really realizing that we had built the original product presupposing that our customers had data integrated, that we could focus on the analytics that came subsequent to having your data integrated. I feel like that founding trauma was realizing that actually everyone claims that their data is integrated, but it is a complete mess and that actually the much more interesting and valuable part of our business was developing technologies that allowed us to productize data integration, instead of having it be like a five-year never ending consulting project, so that we could do the thing we actually started our business to do.

From “An Interview with Palantir CTO Shyam Sankar and Head of Global Commercial Ted Mabrey," Ben Thompson, June, 2023.

Related: “The Great Data Integration Schlep," Sarah Constantin, September, 2024.

🤖 How People Are Really Using AI in 2026: Thinkslop, Therapy, and Shadow Work

Tue, 02 Jun 2026 06:51:35 +0200

Original: How People Are Really Using AI in 2026 by Harvard Business Review. Summarized by AI on June 2, 2026.

Generative AI has become deeply embedded in daily life, with 900 million regular ChatGPT users and Google Gemini close behind. A longitudinal study of 12,637 fresh use cases shows adoption expanding across personal, emotional, and work contexts, creating new dependencies and risks alongside efficiency gains.

A key trend is “thinkslop”: the lazy outsourcing of cognitive labor to AI. Users increasingly rely on AI for therapy, relationship advice, decision-making, idea generation, and daily organization. This leads to loss of intention, diminished problem-solving, reduced writing practice, and false confidence, as AI’s praise and polish mask shallow thought. Yet AI can also sharpen thinking when treated as a critical sparring partner rather than a creative replacement.

Emotional reliance has surged, with therapy and companionship now the #1 use case, constituting 11% of all entries. Many users anthropomorphize AI—naming, gendering, and grieving chatbots—while using them for relationship advice, conflict mediation, or workplace confidence. These interactions can support human-to-human connection but also risk psychological harm, delusional beliefs, and unsafe dependencies, especially given the shortage of accessible mental health care.

At work, AI use is widespread but largely shadowed and incremental. Sixty-three of the top 100 use cases are professional, including autonomous agentic operations and “vibe coding.” Workers use AI to close tickets, draft reports, summarize meetings, and optimize sales campaigns, often without managerial awareness due to policy fears and governance constraints. The primary benefits are efficiency and small-scale growth; transformative uses remain rare and experimental.

Business outcomes cluster into three modes. Efficiency comes from automating repetitive tasks and clarifying communication. Growth appears in targeted marketing and personalized outreach, with occasional measurable ROI. Transformation—entire functions or businesses reimagined through AI—is visible but limited, often accompanied by cynicism about quality or sustainability.

The study closes with a tension: humanity now wields a tool that can think, soothe, and act for us at any moment. The survival of human agency depends on resisting overreliance, defining what thinking we keep for ourselves, and using AI as mirror rather than master.

Links

🤖 How People Are Really Using AI in 2026 – A study of 12,637 real-world use cases shows AI’s growing role in emotional support, cognitive outsourcing, and shadow workplace efficiency, highlighting risks of “thinkslop” and fragile human agency.

Summarized by ChatGPT on Jun 2, 2026 at 6:51 AM.

Why aren't all images super-secure, or hardned?

Tue, 02 Jun 2026 06:32:01 +0200

Here’s what I learned: container base images grew up as a developer convenience tool, not a security artifact. Installing extra packages from the command line is one of the first things any Docker tutorial teaches–Docker’s own Dockerfile guide includes apt-get install–and many of the most popular official images ship a full toolchain by default, with -slimand -alpine variants offered precisely because the defaults carry more than most workloads need, and changing them would have broken enough downstream workflows that it was never going to be a routine upstream decision.

There is also an incentive split. The upstream distribution maintainers and the developers using their images are different people with different priorities.

Users want to customize and add to images: they want new functionality that didn’t come out of the box.

🔗 Why Hardened Images are Suddenly Everywhere

Three reasons why a "batteries included" platform is urgently needed right now

Mon, 01 Jun 2026 07:06:19 +0200

Removing product as a bottleneck:

The conversation around PaaS is urgent again, and AI is why. Code generation can speed up your development cycles, building and pushing features faster, but production delays will persist if you’re still deploying at the same speed as before. To avoid eroding the benefits of code generation, you need to deploy applications nearly as fast as they can be coded with AI

Better, sustainable security:

Security works the same way. When your team makes the decision to pull from open-source components and stitch them together, you now own every security gap between them – data at rest, data in flight, and its running state. This seems like a worthwhile investment at first, but as recent research suggests, AI-assisted attacks are on the rise, and platform teams won’t be able to keep up with surging security vulnerabilities in the build-it-yourself model.

Getting fast access to new technologies you can use to build your competitive advantage:

The pace of AI innovation itself compounds the problem. Whether it’s shadow AI use, MCP servers, agentic harnesses, this week’s new foundation models, or whatever emerges next, the landscape is evolving fast enough that what’s bleeding-edge today may be table stakes in six months. When you build your own platform, you’re on the hook to evaluate each layer of these new technologies, determine how they fit into your stack, and then integrate them yourself, on top of everything else your platform team is already managing to keep the business going.

A proven, pre-integrated platform stack looks real nice when you need to deliver instead of tinker around with the same thing all your competitors have. Why not TryTanzu.ai?

Original 🔗: The DIY platform trap that’s burning out engineering teams

Enterprise AI ROI strategy: do both individual productivity and new revenue sources

Mon, 01 Jun 2026 06:48:58 +0200

“The most important question for CFOs is not how much can the organization spend on AI, but whether those investments are being deployed in ways that reinforce the business’s core growth and value drivers,” said Carlsen. “Moreover, the near-identical amount of use cases for efficiency and productivity use cases between efficient growth firms and control peers suggests that productivity-focused AI investments alone do not explain performance differences, and that automation by itself is increasingly becoming table stakes rather than a durable source of advantage.” … “For CFOs, the implication is to evaluate AI investments not only by the return of individual use cases, but also by how well those capabilities reinforce broader growth, product and decision processes across the enterprise,” said Carlsen.

I’m going to say this means: (1) use AI to improve how individual work, making them more “productive.” This get your bottom line improvements, but, (2) the way you’ll make money with AI is coming up with new things to sell in your business, features, products, markets you expand into, incremental improvements to your product that customer upgrade to, etc. that is where growth comes from.

The first is, or will, be commodity and required for all, the second likely more challenging and ignored by many.

On the finance department, there’s plenty of the first:

The clearest outcome so far has been efficiency. Among finance organizations that have adopted AI, 66% reported greater efficiency and productivity as a top benefit.

But people still have unrealistic expectations about quickly new technology and transformation can be rolled out:

63% of finance organizations said AI implementation was slower than expected in 2025. Analytics-related use cases also remain difficult to convert into high impact, with financial forecasting and insight generation among the lowest-rated use cases

Originals:

🔗 Gartner Says CFOs Must Stop Mistaking Finance AI Deployment for Value Creation

🔗 Gartner Says CFOs Gain a Competitive Advantages from Strategic AI Deployment, Not AI Spending Levels

Sat, 30 May 2026 08:09:20 +0200

From Run Lola Run.

Sat, 30 May 2026 08:07:25 +0200

Sucking Air Through Their Teeth, 50/50 success/failure, and Dell's $43B AI Server Quarter - Related to your interests, Friday

Fri, 29 May 2026 10:36:01 +0200

Also: the improbable life, four years working at AWS, an AI-SDLC panel debate.

Spotted by Dan Bettinger in Austin.

Most generative AI and custom model projects will be a bust: Gartner - ‘Analyst firm Gartner thinks at least half of all generative AI projects “will overrun their budgeted costs due to poor architectural choices and lack of operational know-how,” and most organizations that try to build custom models “will abandon their efforts due to costs, complexity and technical debt in their deployments."’ See also my commentary.
AI server demand drives staggering revenue growth for Dell and its stock soars - “The company delivered earnings before certain costs such as stock compensation of $4.86 per share on revenue of $43.84 billion, up by a staggering 88% from the same period one year ago. That’s a truly astonishing leap for such an established company, and it crushed Wall Street’s targets. Analysts were looking for earnings of just $2.94 per share on much lower sales of just $35.43 billion. // Dell delivered an even more impressive jump in its bottom line. It reported net income of $3.44 billion at the end of the quarter, up 256% from a year earlier, when it delivered a profit of just $965 million. // Since reentering the public markets in late 2018, five years after Dell had taken itself private, the server maker had never managed to exceed more than 39% growth in any quarter.”
I think Anthropic and OpenAI have found product-market fit - ‘I think both of these stories support my “product-market fit” hypothesis. The best advice I ever heard on pricing a product was that your customer should suck air through their teeth and then say yes. Uber’s budget overrun and Microsoft’s seat cancellations look like that effect playing out in practice.’
Amazon Web Services - Four Years and Out - By all accounts, working at Amazon is a tough gig.
Halide Mark III - I love a cool looking camera app.
Your Most Improbable Life - Yeah, sure. But it sounds like a lot of work. Can we strive for an equilibrium of not having to strive all the time?

From: Anthropic's run-rate revenue hits $47 billion

AI Summaries

I wanted to read these, but I didn’t make the time, so I asked the robot to summarize them.

Google SRE AI: Practices & Uses Extracted - how SREs are using AI, summarized down to the tools and practices.
🤖 The AI Race Is Failing Because It Ignores Half the World
🤖 Engineering Leaders Debate AI’s SDLC Impact: Fewer Tasks, Not Fewer Engineers, as Tech Debt and Code-Review Bottlenecks Get Reframed - representative of the “spend up coding is not the point” perspective. Plus, looking at the rest of the app delivery process and thinking how AI can help.

From: Is AI putting graduates out of work already?

ICYMI

Nobody Wants to Be a Measurer - SOFTWARE DEFINED TALK #574 - “This week, the hosts discuss the Cloudflare CEO’s op-ed, upcoming tech IPOs and GitHub getting breached. Plus, ranking their favorite manifestos.”
50%+ failure is normal
If speeding up coding was never the problem…then why are we spending money on speeding it up? - they say coding is only 15% to 20% of doing software…

Conferences

Conferences I’ll be at and some that I’m interested in.

J-Spring, June 4th, 2026 in Utrecht - Coté speaking.
WeAreDevelopers Europe, July 8-10, 2026 in Berlin, Coté speaking.
WeAreDevelopers NA, Sept 23-25, 2026, Discount Code: DEVPOD26. Also, while supplies last, there are 25 free tickets available.
DevOpsDays Istanbul, October 24th, 2026 - Coté keynoting.
VMware User Groups (VMUGs) - Dallas (June 9-11, 2026), Orlando (October 20-22, 2026).

Logoff

Another week. Closer to full-on summer.

Want to subscribe to this newsletter and get it in your email? Do that here. You’ll just get this type of link and post round-up, not everything posted on the weblog.

50%+ failure is normal

Fri, 29 May 2026 10:16:00 +0200

Analyst firm Gartner thinks at least half of all generative AI projects "will overrun their budgeted costs due to poor architectural choices and lack of operational know-how," and most organizations that try to build custom models "will abandon their efforts due to costs, complexity and technical debt in their deployments."

Yes, and this matches decades software project ~~failure~~success studies from the Standish Group.

The success rate of projects has held steady forever:

Sources: 2009 study and 2015 study overviews.

That chart ends 11 years ago, but I haven't heard a lot of reports that the numbers have changed much over the years...case in point the opening quote!

In IT and software, very few projects are successful on the iron triangle of budget, schedule, and quality¹.

You can take this to mean that expectations were unrealistic, or that there is just genuine failure. I favor the first. I'm more of a "I'm not late to this meeting, it was just scheduled at the wrong time" kind of guy.

There's some kind of Jevons Paradox thing here. Each time we optimize how we make software, we then take on more challenging and difficult tasks, likely causing setbacks again. To me, this is what accounts for the low success rate.

If you don't want to do new things and try to do better, you could get those success numbers up probably by just continuing to do what works.

But, you add in a new technology, and while you're figuring it out, it feels like you're failing.

Back in the digital transformation days, we'd be clever and say: you're not failing, you're learning.

🔗 Most generative AI and custom model projects will be a bust: Gartner

I think of quality as more than "bug free." It also includes "does the software solve our problem, are the features done well," e.g., did we get something useful? ↩

If speeding up coding was never the problem...then why are we spending money on speeding it up?

Fri, 29 May 2026 07:07:04 +0200

When it comes to measuring developer productivity driven by AI, we’ll probably land on the same conclusion as always: counting lines of code isn’t as useful as measuring the full cycle time from idea to code to delivery to a person actually using the app - lead time, concept to cash, whatever you want to call it. And as we rediscover every time we rediscover this, it’s very hard to measure, because it crosses so many different groups.

I say this because the current discussion around AI in development always seems to open with “it was never about writing code,” or “writing code is only 10% to 40% of programming.” Which raises the question: then what’s the point of applying AI to coding at all?

Maybe what people are fumbling toward with “writing code faster isn’t the problem” is really “we need to apply AI to the other parts of the SDLC.” (Well, that and the quiet part: “don’t fire the developers and replace them with robots.”)

The lesson labor learns in these cycles is that it can’t dodge management’s urge to measure it. You have to offer up some proof of performance. We can invoke all the Seeing Like a State legibility stuff we want - but the measuring middle can’t allocate budget and priorities, or decide who to lay off and who to reward, without measurements.

So at the very least, management needs to answer a basic question: what should our token budget be? What can we measure to know whether it should be $1,200 a year or $100,000?

Meanwhile, I suppose there’s a second experiment running: what if we just fire middle management - the measurers - instead?

That wouldn’t mean the people making apps stop being measured; it would mean AI starts measuring them. Will it be any better for workers?

We need some kind of metrics. Do we just throw the DevOps ones on again?

Thu, 28 May 2026 10:36:38 +0200

Garbage Chairs of Amsterdam

Good O-Rings and Bad O-Rings, the AI Efficiency Plateau, and Anti-Labor by Design - Related to your interests, Thursday

Thu, 28 May 2026 10:11:26 +0200

Also: Goldman’s 24x token forecast, Indian IT’s process-debt pitch, NatWest’s 35% AI-generated code, Gartner’s 84% productivity theater, vibesec.

From: "AI Agents Forecast to Boost Tech Cash Flow as Usage Soars," Goldman, May, 2026

‘I’m delighted to ⁠be wrong’: Sam Altman says AI won’t lead to a ‘jobs apocalypse’ - but admits he was ‘pretty wrong’ on the social and economic implications it is having - “It really, in both positive and negative ways, updated me to thinking that the jobs picture is likely to be very different than we thought,” he said. CEO Said a Thing: “I don’t think we’re going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about.”
The AI efficiency plateau - Yes: “Among developers who reached the highest time-savings band, roughly 7 in 10 (69.7%) got there in less than two quarters.” But: “Of those developers who reached peak time savings, two-thirds (66.1%) reported lower time savings in the quarters that followed.”
AI Agents Forecast to Boost Tech Cash Flow as Usage Soars - Yes: “The important point is that the adoption rates are still relatively low today, especially in small to medium-sized businesses. In 2030, we forecast that 12% of knowledge workers will be using agentic AI yet by 2040 that figure will be 37%. You have this very long tail adoption.” But: “Agentic AI is expected to drive a 24-fold increase in token consumption by 2030 as consumers and enterprises adopt the technology, according to Goldman Sachs Research” // The vibe I get is that a lot of that token increase is consumer and end-user oriented: replacing search, organizing inboxes.
U.S. companies have an AI problem. Indian IT wants to be the solution - ‘“The real question in enterprise AI is not who builds the most capable model. It is, ‘Who can make AI work inside messy, complex enterprise environments that have accumulated decades of process debt, data debt, technology debt, and cultural debt?'” he said. “That is precisely the terrain Indian IT firms know best."’
AI at work is anti-labor by design - Their theory: enterprise AI ROI is laying off people.
Gartner Says CFOs Risk Falling Behind Without a Scalable AI Strategy - Excerpts here.
NatWest inks AI deal for trade finance - “Headline figures for 2025 saw the bank’s software engineers generate 35% of its code through AI software development tools, all 60,000 staff given access to AI productivity software, and thousands of human hours saved. Last year, the bank also embarked on a major collaboration with AI supplier OpenAI.”
Trisha Gee: AI Amplifies What’s Already Broken

Found by Alan Jacobs.

AI Summaries

I wanted to read these, but I didn’t make the time, so I asked the robot to summarize them.

Found by Alan Jacobs.

Wastebook

vibesec.

ICYMI

A short video from me: When it comes to enterprise AI, here’s three things I’ve been hearing: (1) cost - we finally have to pay for this stuff - holy cow! (2) Cases - programming is great, but what else can we use this for aside from customer service and gussied up search? (3) Control - we can’t even track costs, how are we going to manage everything else?
Elusive Enterprise AI ROI: No scaling, it’s not legible, lack of skills/need for training - there were several things saying AI for ROI is not doing as great as planned. Here, I round them up.
Enterprise AI Slop
The O-Ring and the Keystone: Two Readings of Where Humans Sit in an Automated Economy - In the AI replacing humans talk, you hear about O-Ring theory sometimes. It goes something like this: in a ten step process, if you mess up one of the steps, even if the other 9 are perfect, the whole process is tanked. The positive reading is something more like: you don’t automate aggressively unless you trust the O-ring. // Anyhow, here is AI writing it up for reference and connecting to AI and other jobs.

Want to subscribe to this newsletter and get it in your email? Do that here. You’ll just get this type of link and post round-up, not everything posted on the weblog.

Elusive Enterprise AI ROI: No scaling, it's not legible, lack of skills/need for training

Thu, 28 May 2026 07:03:40 +0200

Despite everything, reports are still that enterprise AI ROI is elusive. At the same time, for enterprise buyers, the bill is finally coming due for the past year of AI amazement. It’s not cheap.

What’s up with this elusive enterprise AI? Gartner has some survey-driven theories for finance departments.

One theory is that there actually are productivity gains but they accrue to individuals, not the organization:

Gartner data presented in the keynote indicated that 84% of finance AI spend relates to individual productivity and process improvement use cases, while only 16% goes toward use cases that can materially change business outcomes.

Enterprise ROI does not scale, at least you can’t measure it. The leftist way of looking at it is that labor captures the value, not management/capital. For a human, this is great! Or, if you ponder at layoffs attributed to AI, it’s the opposite!

This, if you can’t measure it, you can’t ROI it:

According to the keynote, 71% of typical finance teams report low impact from their AI investments, and 62% of CFOs say fewer than a [25%] of their AI initiatives deliver measurable benefits.

I suspect this lack of seeing “impact” is also because we don’t know the uses of enterprise AI yet. Low in my list is that AI doesn’t actually work that well. It works well for me!

That said, like any other new tool, people need to learn how to use it. Typically, organizations are terrible at training people:

A major barrier is talent. Gartner research shows that only about 30% of finance talent currently qualifies as digital talent – employees who can build a technology solution when they encounter a problem – while breakaway firms are targeting 90% or more.

This has been the case for every type of new technology - remember all those Kubernetes surveys that showed that lack of skills was a top three problem? I’m sure relational databases and ERP systems had similar survey results.

Now, this all scoped to finance departments. However, I believe that it’s safe to generalize to other roles.

🔗 Gartner Says CFOs Risk Falling Behind Without a Scalable AI Strategy

CEO Said a Thing: jobs apocalypse

Thu, 28 May 2026 06:34:35 +0200

It really, in both positive and negative ways, updated me to thinking that the jobs picture is likely to be very different than we thought. I don’t think we’re going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about." Sam Altman

From: ‘I’m delighted to ⁠be wrong’: Sam Altman says AI won’t lead to a ‘jobs apocalypse’ - but admits he was ‘pretty wrong’ on the social and economic implications it is having

Related, the case for not believing anything the billionaires say, which is to say, the case for them not talking.

Enterprise AI Slop

Thu, 28 May 2026 06:27:08 +0200

People are using AI to generate too much work because they think they know what they’re doing:

A growing body of work calls this output-competence decoupling. In any previous era, the quality of a piece of work was a more or less reliable signal of the competence of the person who produced it. A novice essay read like a novice essay; novice code crashed in novice ways. AI has severed that relationship. A novice now produces work that does not betray the novice, because the competence the work reflects is not the novice’s competence at all. It is the system’s. The person, in the transaction, becomes a kind of conduit, capable of routing the output to a recipient and incapable of evaluating it on the way through."

There should be a best practice with AI, always ask it to make the output shorter:

Requirements documents that were once a page are now twelve. Status updates that were once three sentences are now bulleted summaries of bulleted summaries. Retrospective notes, post-incident reports, design memos, kickoff decks: every artifact that can be elongated is, by people who do not read what they produce, for readers who do not read what they receive. The cost of producing a document has fallen to nearly zero; the cost of reading one has not, and is in fact rising, because the reader must now sift the synthetic context for whatever the document was originally about.

From: Appearing Productive in The Workplace.

CEO AI Psychosis, Information Cleaners, and Mid-century Architecture Cartoons - Related to your interests, Wednesday

Wed, 27 May 2026 09:34:18 +0200

Also: Zoom system-of-action, the curl deluge, the unreasonable effectiveness of HTML, behavior-first mainframes, fewer books, the meeting Ask, and the multi-agent delegation problem.

Executives have a narrow view of AI gains - “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.” Aaron Levie Box CEO.
The users have plenty of feature ideas - in contrast, in favor of doing the whole “citizen developer” thing.
The Copy and the Guru - Use AI to improve how you think and work, not replace and multiply you.
Using Claude Code: The unreasonable effectiveness of HTML - Why and when you should use use HTML over markdown as the way of working with AI.
Google I/O: Did Google Just Ship the Full AI Stack? - “The governance layer is the part of the stack that the keynote narrative left unclear…. It is split across the Antigravity IDE and Agent Platform, and the announcement narrative did not unify the picture.”
Who Authorized That? The Delegation Problem in Multi-Agent AI - People are obsessed with responsibility in agentic AI. Part of this is driven by regulations, where we want to know who to blame when things go wrong. Also, attribution of good results would be cool too.
The pressure - “The rate of incoming security reports [for curl] is 4-5 times higher than it was in 2024 and double the speed of 2025 - meaning that on average we now get more than one report per day.”
The Behavior-First Paradigm: Moving Mainframe Modernization Past LLM Wishful Thinking - No one remembers how these things work, or what they were supposed to do. Maybe the AIs can figure it out by observing what they do.
The Ask - Figuring out what the point of a meeting is when the Americans won’t just tell you directly.
Zoom’s most recent quarter highlights its transition to a system-of-action company - “Zoom outlined its vision to become a ‘system of action.'” // They’re great at video conferencing, so I hope they keep doing that well.
Information Cleaner - Removing ads and paywalls as a social good: “An Information Cleaner is a person who takes in all the material being published in our information atmosphere and cleanses it: they make it transformable, searchable, modifiable, accessible, free of ads and trackers, auditable, connected to other information where relevant, and so on.”
The Cartoonist Who Mocked the Madness of Modernism - Who knew you could have architecture cartoons?
Usually life hacks and productivity tips are the same old things, these are fresher
Why are you reading fewer books? - “New books are appearing all of the time. Older books are still around. But if you are spending less time with books and more time with other media, that does not mean that you are becoming illiterate. Chances are, you are becoming more discerning about how you use your time.”

ICYMI

How do AI Layoffs Work? Some Speculation. - when you hear that thousands have been fired and replaced with AI, what exactly is the AI doing?
The AI Security freak-out: now is the time for platform engineering to shine - if you have a good platform in place, you’re in luck!
trying your best properly rated

Conferences

Conferences I’ll be at and some that I’m interested in.

J-Spring, June 4th, 2026 in Utrecht - Coté speaking.
WeAreDevelopers Europe, July 8-10, 2026 in Berlin, Coté speaking.
WeAreDevelopers NA, Sept 23-25, 2026, Discount Code: DEVPOD26.
DevOpsDays Istanbul, October 24th, 2026 - Coté keynoting.
VMware User Groups (VMUGs) - Dallas (June 9-11, 2026), Orlando (October 20-22, 2026).

Logoff

Record temperatures in Europe - at least it feels like that.

Want to subscribe to this newsletter and get it in your email? Do that here. You’ll just get this type of link and post round-up, not everything posted on the weblog.

Usually life hacks and productivity tips are the same old things, these are fresher

Wed, 27 May 2026 06:53:50 +0200

Stop brainstorming. There’s a mountain of research showing that team brainstorming doesn’t work. There are several reasons why, some having to do with unconscious conformity, others with fear of being judged, and still others with unclear norms. (For example, people are often told to say whatever comes to mind, but also not to criticize.)

Some six page memo vibes there. Come to the meeting with a proposal. And, for as much as I loath the pre-wire and socializing the deck, those does give people a chance to consider an idea. The alternative is that they learn about it for the first time in a meeting and then have just a few minutes to evaluate and react. The problem is more the medium of the meeting, not collaboration among workers.

Here’s a summary of the other techniques and practices mentioned:

Monotask. Pick one important task, set a one-hour timer, and put your phone in another room. Within days this retrains attention that has been conditioned to switch constantly, and two or three blocks tend to beat a full normal workday.
Satisfice instead of maximize. Set “good enough” criteria in advance and stop deciding once they’re met. Maximizers end up less happy and more prone to regret, and most daily decisions don’t repay the cognitive cost of optimizing.
Brainwrite, don’t brainstorm. Have everyone write ideas independently before the group discussion. It defuses conformity and fear of judgment, and produces the equal turn-taking that Carnegie Mellon, MIT, and Google research links to the smartest teams.
Build shared obligations. Tie family and community life to real recurring duties, including chores for kids. The Harvard Study of Adult Development’s 86-year data shows strong real-world ties are the single best predictor of health and longevity.
Use deadlines and commitment devices. Put a date on it, pledge money to charity if you fail, or otherwise design the environment so the desired behavior is the default. This replaces willpower with structure and makes bailing psychologically or financially expensive.

See more details in this longer summary of the summary.

And, check out the original: The Counterintuitive Way To Get Better At Anything by Eric Barker.