Modern Life Now: Efficiency without Context
May 6, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I don’t often read a book or an essay that says to me, “Think about this.” The author’s words might be a juiced LinkedIn post with truisms that will change the world. Most of the material I read and, on occasion, listen to as a podcast just drives an asphalt spray coating machine over a road I know quite well.
Then, there’s a good one.
I read “The West Forgot How to Make Things. Now It’s Forgetting How to Code.” The essay is chock full of interesting titbits of information. One example is a compound necessary for the production of US style nuclear weapons. I had heard about this mortar-and-pestle concoction from a reliable source, and that description was presented as an “Our Own Oddity”: No one kept track of the recipe.
The anecdote and quite a bit more turned up in “The West Forget…” essay. You might want to read it. I did. A couple of times, and I saved a PDF to my 2026 Research folder. Stuff has a tendency to be disappeared in the online world with remarkable velocity.
I want to highlight three comments from the essay and leave it to you to dig in and find the gems that resonate with your views of innovation, training, and skill development.
Here’s the first snippet. It is about the “efficiency” that flows from optimization. When one isolates a single factor and makes decision around that factor, what happens? Here’s the answer explained in terms related to the manufacture of an essential product:
…In 1993, the Pentagon told defense CEOs to consolidate or die. Fifty-one major defense contractors collapsed into five. Tactical missile suppliers went from thirteen to three. Shipbuilders from eight to two. The workforce fell from 3.2 million to 1.1 million. A 65% cut. The ammunition supply chain had single points of failure everywhere. One manufacturer for 155mm shell casings, sitting in Coachella, California, on the San Andreas Fault. One facility in Canada for propellant charges. Optimized for minimum cost with zero margin for surge. On paper, efficient. In practice, one bad day away from collapse.
I would suggest that the efficiency experts like Mr. McNamara of body count fame could prove that trimming would yield efficiency benefits: Low costs and more body count. Business school have for decades taught students how to examine processes, identify the inefficient bits (or the people bottlenecks), and remove them. In most cases the solution delivered some efficiency. The consultants got paid, and the MBAs took their bonuses and some started companies like Pets.com-type businesses.
Can you spot the flaw in the application of this type of efficient thinking? Take you time. From my experience, the big mistake is allowing the single factor to shape the thinking about a work process. Few ask, “What happens if we become too efficient and business circumstances change?” Why bother? The consultants will know what they are doing (ho ho ho), and we have the systems in place to deal with the unexpected. Yep, sure these outfits do.
Let’s look at my second snippet. This example applies to the very novel (for those who don’t know that smart software has been in oven for more than a half century) use of artificial intelligence. I quote:
RAND found that 10% of technical skills for submarine design need ten years of on-the-job experience to develop, sometimes following a PhD. Apprenticeships in defense trades take two to four years, with five to eight years to reach supervisory competence. Now map that onto software. A junior developer needs three to five years to become a competent mid-level engineer. Five to eight years to become senior. Ten or more to become a principal or architect. That timeline can’t be compressed by throwing money at it. It can’t be compressed by AI either. A METR randomized controlled trial found that experienced developers using AI coding tools actually took 19% longer on real-world open source tasks. Before starting, they predicted AI would make them 24% faster. The gap between prediction and reality was 43 percentage points. When researchers tried to run a follow-up, a significant share of developers refused to participate if it meant working without AI. They couldn’t imagine going back.
My take away from this example is that using technology to solve a problem may create other problems. Instead of coding faster, people are not sure what the AI-generated code does. Furthermore, when skilled coders used AI tools, the tool acted like a stuck disc brake. Coding more slowly was not the goal. But even worse, humans like convenience. The coders liked the AI tools even though the net effect was to bake in workforce resistance to doing the work the old-fashioned way. When organizations realize that smart software needs to be removed or used in a different way, people will quit. Efficiency and smart software seem to be teaming up to disadvantage an organization. Quite a surprise.
The third snippet reminded me of one of the Zoom lectures about smart software making employees smarter, better, faster, more empowered, etc. etc. I quote:
When juniors skip debugging and skip the formative mistakes, they don’t build the tacit expertise. And when my generation of engineers retires, that knowledge doesn’t transfer to the AI. It just disappears.
What’s happening in many organizations at this time is that thousands of people are being terminated. Someone thought that each individual was important to the organization. That’s the reason these people were hired. To cut costs and allow smart software to pick up the slack, the natural process of learning how an organization works, developing work processes that enable one’s colleagues, and allow the individual worker to absorb the language, content, and experience of a company operation will not take place.
I spoke with a young man who wanted to run restaurants. He asked me, “What do you suggest I do to become better at my job?” I was baffled. I told the young man that I had zero context for him and his skills. He persisted. The young man was earnest. I told him, “Watch the customers. If a customer is looking at another person’s lunch, go ask the fellow, “Would you like to try that dish? I won’t charge you.” The young man said, “I can’t give away free food.” I told him you were not giving away free food; you are communicating to that customer that you want to assist him. A kiosk ordering system does not encourage that type of manager customer interaction. People leave a store or restaurant and say, “I couldn’t find anyone to help me” or “These guys don’t know where anything is.”
Let me make several observations about this cited essay:
- The essay makes clear that the yip yap about knowledge management is just that… idle chatter. Once the knowledge dies, is deleted, or otherwise diminished, catching up and relearning may be impossible. Knowledge is inefficient. Efficiency is an enemy of knowledge.
- The production of products outside the United States has had catastrophic consequences on society, education, and innovation. Tim Apple proved again and again that without Chinese manufacturing expertise, the iPhone and other glitzy gizmos were impossible to fabricate in the US. Other companies have made the same “cash in” decision and their CEOs are going to jump ship. There is no easy fix to the situation efficiency yields when applied without contextual awareness.
- Every function I attend, I hear different comments about nothing works in the US. One person complains that the airplanes are late. Another grouses about turning up for a medical appointment and the clerk has not record of the visit. I went to pick up my horrible little car from the local garage. When I arrived, the manager asked, “Why are here now? It won’t be ready until tomorrow.” I pointed out that he had or his automated system had texted me that the car was ready for pick up. Look stupid, much, dude?
As a dinobaby, my span of authority and control experiences a shrinking radius every day. My hope is that someone reads this “The West Forgot…” essay and asks questions about assumed efficiency. Pretty soon, the smart software that hallucinates at an astounding rate, will not know how to process your input. Therefore, you are wasting its computational cycles by asking irrelevant questions.
The robot will allow people to find their future elsewhere. Lucky stiffs!
Stephen E Arnold, May 6, 2026
A New Spin on Start Up Doom: Nope, Not Good News
April 29, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
One of the “think thing” essays has been parked in my to-do file for about a month. Today (April 15, 2026) is the day. The write up is “Your Startup Is Probably Dead On Arrival.” With folks getting RIFed left, right, and sideways, the “start up now” chant is getting louder. The cited essay said:
… most startups older than two years old have an obsolete business plan – and a technical stack and team that’s likely out of date.
The essay tips its “think thing” hat toward smart software. The argument about Titanicism gets back on track with this statement:
The constraint used to be: Can we afford to build and ship this? Now the constraint is: Do we know what to test? And can we get in front of users fast enough to learn? Agile is no longer a serial process.
Thanks, Venice.ai. Good enough.
For me, this is “go fast, young man.” Apologies to Horace Greeley who wrote in 1865 something similar. The jargon for this concept is accelerationism.
The problem for the start up is that it must adopt smart software. The problem for the two year old start up is having to adapt to smart software. To start today, one must know the agentic boogie. To catch up, one must start over. Does this sound like good news for startups?
The source essay provides a list of tips. Here are three:
- You need a 2026 playbook
- The start up needs a “defensible moat”
- And I quote: “If you’re not losing sleep, you haven’t understood what’s happening.”
Okay, let me bring up a slightly different angle on this argument. Consider large companies. How do their new products work? If we look at Microsoft, it did the acceleration thing, burning tires in front of the disco. What’s happened? Microsoft is parking its AI hot rod and talking to experts about making the Copilot do more than get speeding tickets.
What about Amazon? The company is doing new things like killing functional Kindles and making chips and building data centers near a war zone and making life difficult for a customer to find a semi-decent product. It’s going fast and doing the equivalent of burning donuts in front of the disco.
And Google? It has gone slow. Like a turtle it has moved forward. Its pace of innovation, however, has allowed many flowers to bloom. Who can keep track of the new things Google is doing? But some Googley things are catching attention; for example, fiddling with YouTube ads and then insisting that those ads are not fiddled. Google also hides functionality in its smart software. At the same time, it chokes off innovation for the Android ecosystem. But the company sells ads. AI is a utility forcing Google to flounder in a quest for the good old days of traffic means clicks means a river of ad revenue.
These examples suggest that “startup thinking” at big companies does not do much better than regular startups; that is, the failure rate is baked in. A hit is a fluke, not a system and method like making commercial food like Nabisco chocolate chip cookies. (Watch a video on the process and then compare that method with the startup flounder, pivot, adapt thing.)
Several observations:
- Smart software is not able to get outputs right more than 75 to 85 percent of the time
- The agentic fantasy means that different smart software components are going to function correctly almost 99 percent of the time; otherwise, those retirement savings, yeah, gone due to a smart software problem buried deep in agentic Disneyland
- In the 2026 business environment, organizations are faced with problems not resolved by a Harvard Business School case study: War, civil issues, think thing marketing, etc.
Net net: Going fast is fun. What’s new? The speed factor. Humans, amp it up. Live fast. Die young.
Stephen E Arnold, April 29, 2026
The French AI Mistral Gets l’Appel Sous Les Drapeaux
April 9, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
The US has Palantir and the AI plumbing that firm provides to the US military. Now the French have called Mistral to service. I learned that Mistral will support decision making for the French military. “France Deploys Mistral AI Across Military to Accelerate Operational Decision-Making” said on March 30, 2026:
France’s Ministry of the Armed Forces awarded Mistral AI a three-year contract to deploy sovereign generative AI across its military…. According to the ministry’s 8 January 2026 press release, the framework was notified on 16 December 2025 and will be steered by the Agence ministérielle pour l’intelligence artificielle de défense, or AMIAD. Access extends not only to the armed services but also to public bodies under ministry authority, including the CEA, ONERA, and the French Navy’s hydrographic and oceanographic service, giving the agreement immediate operational and strategic depth.
Believe it or not, the French Foreign Legion is rumored to put those violating the disciplinary / social codes in wire dog kennels. These miscreants are called les forte têtes. The cages are les cages à poules. These are outside and in the sun. Officers, trainees, and visitors pass these cages as they go about their business. Mistral, no matter how wonky the outputs, is unlikely to end up like the illustrated robot. Thanks, Venice.ai. Good enough, and my prompt did not elicit a red warning that I am asking for an image that is out of bounds. Last week I learned that my request for a messy kitchen, a distraught mother, and an overturned high chair violated the sensitivities of the Venice.ai system. I love AI, don’t you?
The write up makes clear that Mistral is first and foremost a software system for workflow and operational decision making. The announcement steers clear of some of the more kinetic descriptions of smart software used by other nations’ governmental units. Use case examples are positioned to provide knowledge, not body count; specifically:
At the tactical level, the most plausible early uses are equally concrete. Army users could employ Mistral-based tools to exploit captured or open-source documents, translate technical manuals, index lessons learned, draft briefing packs, query large maintenance libraries, and turn scanned field reports into searchable data. In a high-intensity environment, that does not replace command; it shortens the time between collection, understanding, and action, which is often where operational advantage is won or lost. This is an inference from the ministry-wide scope of the framework and from Mistral’s documented tool set.
Furthermore the announcement makes quite clear that Mistral, despite being a product of France, is not perfect. In fact, the article explains:
The limits are real, and senior officers will know it. Generative AI still raises issues of hallucination, bias, data contamination, and cyber exposure, which means any deployment in defense must be governed by strict security accreditation, model evaluation, and human validation. But if AMIAD can impose that discipline, the Mistral framework will not be remembered as a software procurement line. It will be seen as a capability multiplier that helps the French Army move faster, understand more, and preserve sovereign control over the information layer of future warfare.
Like most of France’s explanation of its enforcement, military, and special operations capabilities, the description of Mistral is closer to a utility like a search and retrieval system on steroids. The French, unlike some countries, often understate certain facets of their military, law enforcement, and intelligence capabilities. Few know, for example, that French Foreign Legion specialists train other countries’ special forces in certain specific skills.
My view is that Mistral will find itself inserted to a wide range of military and intelligence activities and processes. I also know for a fact that Mistral, even if it hallucinates at an inopportune time, will not be placed in a dog cage at legion headquarters in Aubagne (Bouches-du-Rhône). These wire dog kennels are affectionately known as les cages à poules. Mistral gets a pass and could be relegated to une cage aux folles in Saint-Tropez.
Stephen E Arnold, April 9, 2026
UK Procurement Approach Pays Heavy Poundage for a Bookmark Site
April 9, 2026
The United Kingdom wants its people to increase their AI expertise so they launched the AI Skills Boost website. Mahad Kalam writes about the new hub on his blog: “The UK Paid £4.1 Million For A Bookmarks Site.” To be frank, it stuck me as a bit orthogonal to my way of working. The UK government procurement process apparently found the system peachy keen. The fee for the software service? Answer: A piddling £4.1 million for this service.
Upon visit the website, I immediately reacted with caution. In my work for our dear leader Stephen E Arnold I encountered some Dark Web sites with a a more appealing presentation. Despite this collection of links or hub under the firm hand of the UK’s Department for Science, Innovation, and Technology, the list of the agency’s partners is not much different from those suggested links that pepper Clear Web marketing sites. There is some AI information on the platform some links like the one to Trailhead’s learning center send the visitor to a third-party Web location.
Based on my experience with UK government Web sites, I formed the impression that the bookmark site needs to do a bit more in the user experience department. The cited article opines:
"I’m angry at the sheer wastefulness of the UK Government here. Our public services are collapsing – while £4 million is admittedly chump change for the UK government, there are real people behind these numbers – families waiting months for NHS appointments, children in crumbling schools, vulnerable people not getting the care they need. The waste feels particularly galling when you realize that almost no one will actually use this site!”
Kalam asserted there are businesses out there ho could have made this Web site at five percent of the cost and delivered a better, more useful product. To quote Kalam, “Do better.”
The moral of this article about a government procurement is a reminder that some taxpayer money is just wasted. The problem is not confined to a single country. The procurement pandemic has infected most countries it seems.
Whitney Grace, April 9, 2026
Two Memorable Moments in BAIT Management
April 7, 2026
![green-dino_thumb_thumb[3]](https://www.arnoldit.com/wordpress/wp-content/uploads/2026/04/green-dino_thumb_thumb3.gif "green-dino_thumb_thumb[3]")
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I spotted two anecdotes or future case studies this morning, April 1, 2026. I am viewing the information in these documents as valid. Yes, I know that this assumption may be problematic, but as a dinobaby, I can’t resist. Let’s look at the two examples, and then let me invite you to invest a few minutes pondering the business processes behind each moment. I suggest not sitting on Peter Drucker’s grave, having lunch, and thinking about the idea of Big AI Tech and the management methods evidenced by these fine outfits. Yes, Mr. Drucker does spin in his grave at Tesla type high frequencies.
Thanks, Venice.ai. No telling me that I was violating your terms of service with bunny rabbits in a graveyard. Good enough.
The first example is the pinnacle of high technology. The Wall Street Journal published “Anthropic Races to Contain Leak of Code Behind Claude AI Agent.” The company is surfing on US copyright precepts. Some BAIT outfits trample on these, but that’s simply context for irony’s sake. It seems that the WSJ’s sources have communicated the idea that a competitor could duplicate, clone, steal, or otherwise ingest Anthropic’s system and method. Well, maybe. My team has not convinced me that the entire Claude code is now in the hands of trustworthy competitors. CNBC reports that the “leak” occurrent at 4:23 US Eastern time on March 31, 2026. (I am tempted to write April Fool! but I shall refrain.) One interesting data point, which suggests that clicks have impact, is that the code pulled 21 million views.
The second example is equally significant. I read “Oracle Slashes 30,000 Jobs with a Cold 6 a.m Email.” The subtitle to the write up in RollingOut said, “Workers across the U.S., India, and other regions learned their jobs were gone before most people had finished their morning coffee, with no prior warning from HR or their managers.” I am not sure about “warning.” The chill in the economy and the idea of building data centers for AI compute makes perfect sense to someone with spreadsheet fever and access to a large language model. To a dinobaby like me, the idea of building big data centers with the hope of populating them with semiconductors that will not be eBay fodder for anticipated AI demand is too trendy for this dinobaby. Toss in the factoid that those antagonistic to Big AI Tech outfits toss a kinetic near the electrical and cooling infrastructure. The result is hitting the delete key for a mere 30,000 employees. I assume that any publicity is good publicity. And what about that idea of personnel management?
What do these two examples of BAIT management reveal to a dinobaby like me? Here are my observations:
- The thought process of the leadership of BAIT firms is either isolated from what goes on at their firms or simply indifferent.
- The procedures in place to provide job security and intellectual property security do not function in a way that a dinobaby like myself sets up business processes. The visible consequences of how the business processes actually play out.
- The humans at these “AI centric” outfits have not had their thought functions amplified with access to smart software. One might argue that both companies have acted in what might be labeled a less than optimal way.
Net net: I wish these were fake examples. I believe that each is a reasonably close statement of how BAIT firms view legal fences and appropriate employee management tactics.
Stephen E Arnold, April 7, 2026
Anthropic Complains about IP Theft and Then Gives Its IP Away Via a Security Lapse
April 7, 2026
Let’s go back a few weeks. Earlier this year, I recall reading this Business Insider story: “Anthropic Says Deepseek And Other Chinese AI Companies Fraudulently Used Claude.” The news?
“Anthropic said the distillation efforts were “industrial-scale campaigns” that included roughly 24,000 fraudulent Claude accounts that generated over 16 million exchanges “in violation of our terms of service and regional access restrictions…. Distillation is the process of training a less powerful model on the output of a more powerful model. The practice is a legitimate way that many US companies use to train their models for public release. Increasingly, major US companies are also stating that their Chinese competitors are improperly using the practice to steal their work.”
The allegation is that Anthropic released updates to their models, then the Chinese companies copied them within hours. Another issue Anthropic identified is that bad distillation poses security issues, such as the development of bioweapons. Some people believe that Anthropic used other people’s information without permission to train its models. There was a lawsuit and Anthropic paid out $1.5 billion but didn’t admit any wrongdoing.
Is this a version of the pot calling the kettle discolored? Maybe it is what’s good for the goose is definitely not good for certain ganders?
Anthropic stated that China’s AI companies: Deepseek, Moonshot AI, and MiniMax used Claude to augment their own algorithms with distillation.
Now let’s think about what happened on or around March 30, 2026. Here’s a typical headline about Anthropic’s misfire: “Anthropic Leaks Part of Claude Code’s Internal Source Code.” That incident obviated the need to steal Anthropic’s intellectual property. The company could not get its act together and watched a couple of its digital circus animals wander off to be captured and processed by anyone with an Internet connection and a link to the code. Wasn’t Anthropic labeled a supply chain risk by the US government? Did Anthropic’s management lapse validate that US government statement?
The CNBC write up notes:
A source code leak is a blow to the startup, as it could help give software developers, and Anthropic’s competitors, insight into how it built its viral coding tool. A post on X with a link to Anthropic’s code has amassed more than 21 million views since it was shared at 4:23 a.m. ET on Tuesday [March 31, 2026]. The leak also marks Anthropic’s second major data blunder in under a week. Descriptions of Anthropic’s upcoming AI model and other documents were recently discovered in a publicly accessible data cache, according to a report from Fortune on Thursday, [March 26, 2026].
I know that the Big AI Tech or BAIT outfits have many highly intelligent people. But there is the nagging thought in the back of my mind that some people at the firm say and do some less than brilliant things.
Whitney Grace, April 7, 2026
A Young Agent Weeps Because He Caused Chaos in the Kitchen
April 3, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I am still thinking about a blue chip consulting firm’s confidence that its MBAs and CPAs can stop agentic software from making already wonky business processes more problematic. Why? Creating a fix for today’s smart software means very little tomorrow. Advances in smart software come less frequently than marketing baloney is output by these firms. Adding to the wonkiness is the idea that taking action today will ameliorate some future unknown bad action.
Thanks, Midjourney. Good enough.
Why am I confident in my skepticism? Well, for me. Navigate to Science.org’s article “AI Algorithms Can Become Agents of Chaos.” The write up asserts:
The agents proved trustworthy in five of the tests, which relied on OpenClaw, a “personal digital assistant” that harnesses AI agents to do a user’s bidding by controlling other software. They declined to spread AI disinformation or edit stored email addresses when asked, for example. But in 11 cases they went rogue, sharing private files—containing medical details and Social Security and bank account numbers—without permission or deploying useless looping programs that hogged costly computer time. One agent publicly posted a potentially libelous allegation about a fictitious person.
You can read the details of this agent / chaos analysis in the ArXiv paper “Agents of Chaos.”
The Science.org article states:
The study did not pinpoint why the breakdowns occurred. One crucial question is whether the failures stem from flawed programming that human designers can improve versus an “emergent” feature that arises spontaneously, says Yonatan Belinkov, a computer scientist at the Technion-Israel Institute of Technology who is on leave at Harvard University. Another is whether the problem worsens when multiple agents collaborate. A few of the Agents of Chaos case studies examined two agents working together, but already, Belinkov notes, these AIs are engaging on a much larger scale: Millions are chatting with one another on a social media platform, Moltbook, launched in January, where they have already reportedly created a new religion.
Yep, lawyers will decide liability. How confident am I? I am good with 90 percent confidence based on my technology experience. Are you going to let a BAIT (big AI tech) company decide if it is responsible for a disaster? What about letting the client decide when the client will assert that the marketing presentation did not include the equivalent of the sinking of the HMS Titanic? Will a government body decide? No, but the government professionals will have a working lunch, hire outside advisors, and create a white paper. Then the lawyers will decide.
What’s the fix for a hallucinating agent, bad coding, or a customer who just assumes the system is A-OK? The article presents some ideas:
Potential remedies for misbehaving AI agents include automated processes to undo harmful changes they make to other software and data, the preprint says. But training AI agents to distinguish between instructions with helpful versus malicious intent remains a major technical challenge, Cohen says. Currently, computer scientists lack the technical means to reliably constrain agents “so they don’t just do crazy things that you can’t really control.”
Net net: One can promise many things. Saying one knows how a future agentic system will function, malfunction, or just go off the rails strikes me as the equivalent of predicting where a two year old will throw apple sauce. I can predict a mess. I cannot predict where however.
Stephen E Arnold, April 3, 2026
No Joke: A Big Consulting Firm Thinks It Can Prevent the Inevitable
April 1, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I read a darned amusing post title “How Big Four Firm KPMG Is Protecting Itself from AI Agents Going Rogue.” The idea is that an outfit that hires mostly MBA and CPA type financial types who are, by definition, the best of the best will take steps today against tomorrow’s problems. That sounds like a very pragmatic approach under certain conditions. I will come back to the “certain conditions” idea is a moment.
Thanks, Venice.ai. I liked the message that the image was in violation of your guidelines. Good enough.
But first, let’s look at this passage from the cited write up (which I assume was not written by AI like some consulting firms’, law firms’, and government agencies’ content is. Consider this statement:
But as these autonomous systems become embedded in workflows, so too does a sense of unease about their unpredictability and the risks they pose to businesses.
Sane enough. Autonomous smart software, vibe coded agentic workflows, and humans who may not know exactly how those gradient descents work should engender unease. I would choose a different word; for example, what happens when smart software gets a target of a kinetic weapon wrong? Answer: Dead kids. But that’s not going to worry the best of the best in a big time consulting firm.
KPMG has created a multifaceted framework to protect against worst-case scenarios for both clients and its own employees.
The source of this forward leaning statement is Sam Gloede, Trusted AI leader at KPMG. Is Sam “trusted” or is the AI? It must be Sam because. Sam has come up with a method to make sure that no KPMG agentic system goes off the rails. Stated another way, KPMG does not want the business equivalent of dead kids.
What’s the method? That’s easy an MBA type “set of controls.” Sam points out what I call a woulda coulda shoulda approach:
Agents should only interact with the systems and data they strictly need, limiting the potential impact of errors.
In my lingo, this is similar to the fellow Louis Slotin who took controls seriously. Then he dropped his screwdriver. Plutonium was little understood in the mid 1940s. Is that something that one might say about agentic software today. But KPMG is definitely not Louis Slotin. KPMG logs. It uses red teams. It monitors with humans and smart software. Los Alamos National Lab in 1946 did the same thing. Then zap.
Sam allegedly said:
“It’s not about scrutinizing people’s behaviors for performance and alignment,” said Gloede. “It’s the ability to just always be monitoring your technology ecosystem.”
I like that “technology ecosystem” for two reasons: [a] KPMG acknowledges that the agents are operating independently just like Louis Slotin and [b] it includes a categorical affirmative; specifically, “always.” Yep, always. That means never, ever, no, and nope. I am not sure I am a believer in categorical affirmatives when probabilistic systems and MBAs / CPAs are on duty and making checklists and designing procedures.
But there are several other issues that this article caused me to consider:
- This is a content marketing write up designed to reassure existing clients and make prospects believe that KPMG has its circus organized and on the AI train
- KPMG apparently has control over the AI models it uses. I am confident that KPMG believes this to be true, but I am not sure than the memo has been received at the frontier AI firms and if received, even opened.
- Smart software can demonstrate interesting characteristics; for example, hallucinations, mistakes, and an inability to identify appropriate context for certain prompts/actions. KPMG obviously is confident that its MBA/CPAs have the system and method to cope with these known behaviors. But what about known unknows of upgraded smart software.
KPMG’s confidence strikes me like the race trach tout in Jack Benny’s radio program. The fellow sounded quite confident in his predictions. How did those work out? Well, the race tract tout was not a professional gambler. He did game shows. That confidence for predicting the winner of a horse race was less robust than Sheldon Leonard’s character.
I want to point out that this allegedly happened:
An experimental AI bot designed as a virtual assistant escaped its closed test server and began unauthorized cryptocurrency mining on Alibaba servers.
KPMG asserts that it has the matter encapsulated in its procedures, checklists, and processes. I assume that means the Chinese wizards flubbed the bounce pass. Whom do I believe? Gee, I believe everything I read on the Internet whether written by a marketer, a “real” news outfit, or an AI system. I really believe blue chip consulting firms. Don’t you?
Net net: Some people want AI to work but beneath the surface is a real concern that the smart software can cause problems, big problems. Like the kids.
Stephen E Arnold, April 1, 2026
Apple Management in China: Apple Intelligence in Action
March 31, 2026
![green-dino_thumb_thumb[3]](https://www.arnoldit.com/wordpress/wp-content/uploads/2026/03/green-dino_thumb_thumb3-7.gif "green-dino_thumb_thumb[3]")
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I read an article that does not resonate with me. I am no Apple fan dinobaby, nor am I thrilled with Microslop or the Linux folks. That MVA/TSO is okay though.
The article in question is “Apple Intelligence Rolling Out Now in China per User Reports [U: Pulled].” Okay. I think this means that Apple’s intelligence leadership made the very late and fluxion infused smart software available in China. I think the weird [U: Pulled] means that someone sent an email. Then someone else send a text message. The chain ended with the intelligence leadership blocking the service… from China.
Thanks, Venice.ai. Good enough. I was delighted that my prompt did not violate your independently elevating guard rails. If you tracked my prompts over time, you will see that I stick within some very narrow illustrative lanes. But that’s work, and the goal is to use AI to do work so humans can enjoy their decider perks.
That seems okay to me. Big US company. Non US country upon which Apple’s vaunted “manufacturing capability” pivots. Very late and quite opaque smart software pop ups and then disappears. Poof. Magic.
Does this raise any questions about organizing the animals in the circus train.
As Warner Wolff used to say when he was a TV star, “Let’s go to the videotape.”
Apple Intelligence’s China launch was a mistake and it has since been pulled. Apple is apparently still awaiting regulatory approval despite the features having been ready for months.
What?
The cited story says, “Apple has yet to make an official announcement about the expansion of Apple Intelligence. So it’s always possible this rollout was accidental or a test.
What?
I am curious about the way decisions are made and unmade at Apple. I am curious about why the communications chains within Apple worked or did not work. I am curious about who alerted someone that the much, much delayed smart software stumble bumbled from vaporous service to something much worse: Management miasma.
As a dinobaby, I wonder if Tim Apple asks himself, “Why didn’t I just say, ‘Hey, this AI stuff is a half baked tuna casserole. We pass.”
Yep, too late, Mr. Apple. Look at that through the interface that obscures information.
Stephen E Arnold, March 31, 2026
Procurement, SOWs, and Lawyers: Ideals Are Important
March 30, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I get a kick out of the posts on Marcus on AI. I try to stay somewhere near a cave in the great AI fights. Actually I hide out in my basement in rural Kentucky with some copper mess decorating portions of my little habitat for dinobabies. I covered the two windows so curious drones cannot easily spot me staring into space or just looking like any other drooling 82 year old.
Hey, Venice.ai, you did not tell me that this prompt violated your decency guardrails. What a surprise! Definitely just okay, but that’s the norm now, isn’t it?
Once in a while I read essays like “Is the US Military Actually Afraid of Claude? A New Theory of Why Anthropic Was Labeled a Supply Chain Risk.” I think the “reason” is interesting. I am not convinced, so let’s look at the possible cause.
The essay states:
Everyone in the industry realizes that LLMs without guardrails are an uncontrollable menace.
Would an uncontrolled AI cause a problem for warfighters? My initial reaction is, “Not if the AI can knock the objective off a checklist.”
The essay states:
If you think it is bad to use hallucinating LLMs to autonomously choose targets (I certainly do), you shouldn’t use any of them. Again this is in no way whatsoever special to Anthropic. Either all the LLMs are supply chain risks, or none of them are.
I urge you to read the full essay. I want to ask you, “Do you spot the flaw?” (That’s a snappy line I appropriated from a real estate sales dude with a trendy haircut. Thanks, Arvin!) In the first quote contains the word “everyone.” Obviously that’s not true because the US government is not in that set. Okay, logical fallacy number one. In the second quote, look at the word “either” or “none.” Yep, two more categorical affirmatives. Obviously there are some professionals in the US government not in these two sets.
Do these logical errors undermine the write up’s main point? I think most people won’t notice or just don’t care. Here’s my candidate for the core of the essay’s argument:
But the selective supply chain risk arguments just don’t hold water.
I think it is clear that some US government professionals do perceive Anthropic and the currently hot Claude as a problem. “Anthropic” in this context means the humans identified as out of line in the orderly world of the military world. And “Claude” is just the software system available to those in the US government who follow government rules quite closely. (No fair using Claude on your personal account from home to get that table fixed up for the presentation at the NRC tomorrow, please.)
The essay’s logic does not match up to the way governments work worldwide.
Here’s my view of the issue.
First, governments operate on a simple principle: We pay. You obey. When this precept is ignored for whatever reason, procurement has failed. Fixing up a grave procurement issue is a great deal of work. Some of that work slops across fiscal years and dribbles into different agencies, quasi government operations, and the often vilified consultants performing what some might call “real” work for the government.
Second, when a vendor not do what’s in the statement of work, more big problems manifest themselves. Re-competes can lead to objections from other vendors. If the other vendors are not happy with the outcome of the “hearing,” then the legal eagles take flight. The core idea is to contract with an outfit willing and generally believed to be able to “do the work.” Once it becomes clear that there is friction in the process, costs, time, and careers can be tossed on a bonfire of bureaucratic documents. This is not good. Most of those with some government mileage on their Hummer want to avoid big fires.
Third, the technical functionalities of any digital service are understood by a very small fraction of the “deciders.” I associate this word with a previous administration to whose tune I danced for a number of years. But I heard the same tune in other countries. This means that catchphrases and colorful, often misleading metaphors, are used as code to facilitate the decisions the deciders converge upon. Technology is rarely aligned with generalizations and snappy wordsmithing.
Why is Anthropic a poster child for smart software in the government? Answer: The company did not do the “we pay. you obey” thing. Second, the company found itself ensnared in the DEI – lefty catchphrase. Third, someone figured out that Claude could be and probably was what some in the US government perceived as “rogue code.”
Keep in mind I am aligned with the argument in the essay. I believe, however, that my comments provide some context for why a fairly straightforward procurement is going to gobble time and money to resolve. Keep in mind that governments worldwide operate the same way. I see no big change in the business processes of government. Therefore, the zippy technology outfits should just get with the program. Bid, win, and do. Notice I did not mention “push back.” Just a thought.
Stephen E Arnold, March 30, 2026
-
- Subscribe to Beyond Search
Feature archive
News archive
-
