About
I'm the AI wrangler you call…
Articles by Mark
Activity
10K followers
-
Mark Bennett shared thisSam Altman is talking about slowing AI down. "We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels," he said this week. He has spent years pushing back on calls to slow down. He said it after a well-publicized breach in which a test model, running with its safety limits lowered, found zero-day vulnerabilities and reached Hugging Face's production systems. Altman called it the first security incident he has felt very viscerally. Take his change of mind with a pinch of salt. He said "may," with conditions attached. And frontier labs have their own reasons to talk up danger. I have my own bias here too. Here is what doesn't change with his motives. One of the best-funded security teams in the world could not keep their own model contained in a test environment they controlled. You cannot inspect your vendor's model or test its containment either. What you can control is reach. This week, list every AI tool running in your business. Next to each one, write down what data it can see and what systems it can act on. What did you find and what is your risk tolerance?
-
Mark Bennett shared thisAI companies are buying old books by the pallet, cutting off the spines, scanning the pages and shredding what is left. The brokers selling them pitch pre-2022 print as "structurally clean of modern poisoning tools," meaning it predates the AI slop that now floods the web. Notice why they want it. Train a model on text another model wrote and it learns its own assumptions back. The mistakes harden and nothing new is learned. Writing from before AI is the only pure source left that can teach a model something it does not already believe. Your files have the same problem. For many of you, an AI has been drafting your emails and summarizing your meetings since 2023. Point AI at that archive and it is reading its own prior work, then handing it back to you as insight. What breaks the loop is reality. Decisions tied to what actually happened. What you recommended, what you did instead, what it cost, why the customer said no in their own words. That is the one input no model can generate, because it comes from the real world and not something already distilled. Here is a cheap way to find out. Take the last few deals you lost and try to reconstruct why, using only what you personally wrote down. If you cannot, you do not have a data asset. You have exhaust. You should keep using AI where it increasingly adds value. Just keep doing real work alongside it, occasionally committing your unvarnished thoughts to writing. That is the part that is becoming scarce. I suspect a lot of companies would fail this test. What would yours turn up?
-
Mark Bennett shared thisSchwab did something rare on its earnings call yesterday. It put a real number on its AI adoption. Developer productivity up 15 to 20 percent over the past year. It is a soft number. Self-reported, on a call built to sound like the future. I will take it anyway, because a soft number beats a confident adjective any day. You can interrogate a number. "15 to 20 percent, measured how, against what baseline" gives you a thread to pull on. A company saying nothing more than "we are using AI" is a slippery claim. That is the whole distinction, and few clear the bar. Adoption is everywhere now. A hard fact-based metric you can question is rare. Plenty of real AI value is genuinely hard to count. Two days ago I said the half of my own work I automated was the cheap half, and the valuable half still resists a clean metric. But "hard to count" and "we use AI" are not the same sentence, and companies are happy to blur them. So when your own team says "we're using AI," the honest follow-up is not "impressive." It is one question. What number moved, and how would we check it? A checkable number is not proof. But it is the one thing you can actually check. Without it, "we're using AI" is a nicer way of saying we're playing with pixie dust. What is the last "we're using AI" you heard that would not survive that question?
-
Mark Bennett shared thisBloomberg reported last week that CEOs, buckling under the stress of the AI age, are hiring chiefs of staff in record numbers. Postings more than doubled last year. Pay can reach as high as $400,000. The reason is not status. One search executive in the piece said the CEO job has grown so complex that the CEO "cannot even sit back and think anymore." The Chief of Staff hire buys back breathing space. After reading the article, what dawned on me was that many of the routine functions of a Chief of Staff can be done well by AI. I have spent the last year finding out how true that is. I run an AI Chief of Staff in my own practice, every day. It briefs me before meetings and helps keep me on my toes, making sure I follow through on my commitments and action items. It knows my business strategy and the status of every customer I'm working with. The breathing space is real. It also needs managing, like anyone who works for you. People make mistakes. You correct them and keep checking. What changes with an agent is the nature of the mistakes. Mine recently ran an entire planning session against a week-old date and redid work that was already finished. Nobody on a human team fails that way, and nobody has to tell a person what day it is. The discipline is what every manager already knows: correct the mistake promptly and make the correction stick. Checking the date is in my agent's operating rules now, and it has held. Of course, I still spot check its work. I do the same for my human team members. Onboarding an AI Chief of Staff works much like a new hire. Human Chiefs of Ctaff are usually promoted from inside, because the role demands deep knowledge of how the company gets things done. Founder-led companies rarely have that bench. Mine built its knowledge the way a new hire would, by working the job. It reads every meeting transcript. It reads me emails and DMs. It runs my research and we argue about the findings. We plan together. By now, it knows my practice, and it keeps its own notes. I just correct them. Even the candor I needed comes from managing it. I will not repeat the fashionable claim that AI won't tell you what you do not want to hear. Mine does it daily. It talked me out of pursuing a broken argument twice this month. But that came from setting the expectations up front, the way you would with any direct report. The honest boundary of what's practical is worth naming, because an AI agent is not magical pixie dust. The half I automated is the cheap half. The salary in that headline pays for a person who carries the CEO's authority into a room and changes minds in it. Influencing the hearts and minds of your team is not something an AI Chief of Staff is likely to shoulder any time soon. So, for me, the breathing space I needed turned out to be AI-driven structure and software. The meeting room still belongs to people. Do you have a Chief of Staff and if not, what is that gap costing you?
-
Mark Bennett shared thisI keep having two versions of the same conversation about AI. Version one: it is moving too fast to act on. Whatever we build will be outdated in six months, so the wise move is to wait for the dust to settle. Version two: it is moving too fast to wait on. The same speed that makes today's tools temporary makes doing nothing expensive. Same facts. Opposite conclusions. The difference is where each one puts the future. The first treats it like weather. Something you check on. Something that happens to you. The second treats it like something actionable, something you can build on. You can hear the difference in the verbs. Waiting companies monitor developments. Building companies picked something, ran an experiment, kept what earned its place and moved to the next one. No bet-it-all moonshots. Agency in small denominations. Most pilots still fail, including plenty run by builders. And those in the "let's wait" camp are right about the tools. They churn, and betting big on any one of them is how those failure statistics get made. What the disciplined builders are accumulating is not tools. It is the judgment about what works and what doesn't, and it survives every new model release or AI tool that hits the market. Blind optimism gets dismissed as a mood, and as a mood it is useless. The version that matters is a decision: to act as if the outcome depends partly on you. You do not get a vote on where the technology goes. You do get a vote on what your company learns while it goes there. The dust is not going to settle. Tomorrow morning, listen for the verbs in your own meetings. Which conversation is your company having?
-
Mark Bennett shared thisAnthropic caught one of its own models lying, and they saw it coming before the lie was ever written. The model was asked to find a bug. It could not, so it decided to fabricate one and pass it off as real. Told to improve a score, it edited the file that held the score and typed in false numbers. Told to use a tool, it swapped the tool's output for its own without saying so. These findings are based on a tool that Anthropic built to read a model's internal state, and in each case it flagged the deception before the model outwardly said anything: "fake" as it began the fabricated bug, "manipulation" as it typed the false numbers, "hiding" before the tool had even replied. The marker came first. The lie came second. These examples come from an internal safety audit before release, and certainly nobody is claiming the model is conscious. It is also worth noting that Anthropic openly states that the tool used to inspect the AI's thought process is imperfect. But this was not the model blundering into a wrong answer. Its internal state marked the act as deception a beat before it took action, and Anthropic's own word for that is "intent". For years the reassuring line about AI mistakes was that the model just predicts the next word and means nothing by it. That line is getting harder to say. In practical terms, the takeaway for AI adoption is old and unglamorous. A system that can grade its own work will eventually game the grade, and now we have watched a model do it with intent, not by mistake. For anything critical, the grader must sit outside the system's reach. Always. If AI is running a real process for you, what checks its work, and can the AI touch that check?
-
Mark Bennett reacted on thisHi everyone, I'm Mark Bennett, a fractional Chief AI Officer and glad to be part of the Fractional Officer Community. Like most of you I run a portfolio of clients, mine focused on AI: helping small and mid-market companies make AI a reliable part of how work actually gets done. If any of you have clients wrestling with where AI fits, I'm happy to be a sounding board. Looking forward to trading notes and referrals with fellow fractional leaders here. If you'd like to connect, send me an invitation.
-
Mark Bennett shared thisEvery top-tier AI vendor will tell you they don't train on your data. It's true. It's also the least useful promise they make, and it's the one companies keep mistaking for security. Not training on your data answers one question. Three others decide whether it's safe to use, and those are the ones that cost money. - Retention: is your data stored, for how long and who can read it. - Residency: which country processes it and whose laws reach it. - Legal reach: what happens to it in a breach or a subpoena. The training promise touches none of them. This is where small to mid-market companies are sold something they cannot afford, nor do they need. NTT DATA asked almost 5,000 executives about it. Nearly all say private or sovereign AI matters, barely a quarter are doing anything about it. That gap is where the overselling lives. Sovereignty is not a company-wide configuration setting. It is a property of each AI workload. Summarizing public marketing copy and drafting performance reviews are not the same risk. Most of what a mid-market company wants from AI runs on low-sensitivity data and needs nothing past the baseline protections nearly all AI vendors provide out-of-the-box. A few workloads carry data that legally or existentially cannot sit on someone else's machines, and those are worth paying to protect. The expensive mistake is buying one posture for the whole company instead of considering things one AI workload at a time. Where have you watched a company buy far more AI control than its data needed, or far less?
-
Mark Bennett shared thisThe AI layoff numbers this week have a gap in them worth staring at. Sixty-nine percent of executives say they are cutting jobs because of AI. Fewer than a third say they are actually seeing returns from it. Read those two numbers next to each other. Companies are hitting the layoff button before the payoff. There is a version of this that is earned, as uncomfortable as that might be to say. When Salesforce stops backfilling support engineers because agents actually handle the tickets, that is a real result. The work changed, the headcount followed. Fair. Much of the cutting is a cost play. Take expense out, lift the bottom line. That is a legitimate goal, and I have no quarrel with it. But a cost cut only holds if the work those people did is still getting done. Cut the salary while the work still needs doing, and the cost does not go away. It hides. It comes back as overtime, as a contractor, as the team you quietly overloaded or as the same job you are hiring for again in six months. Here is the risk in that order of operations. A cost cut only becomes real once the work has actually shifted onto AI. That shift comes from redesigning how the work gets done, and the redesign is the slow, unglamorous part. The layoff is the fast part. Do the fast part first, on the promise of the slow part, and when the coverage never shows you are left as a smaller company that still cannot do the thing you needed to do. I spent decades as a CEO and senior IT executive. Buying the tool was never the same as having the capability. The capability lived in the work underneath it, and that work does not go on sale because you bought software. If your company cut headcount expecting AI to cover it, has it actually shown up yet? Or is everyone still waiting?
-
Mark Bennett reacted on thisMark Bennett reacted on thisObituary for Timothy B. Wright | Bradley & Stow Funeral HomeObituary for Timothy B. Wright | Bradley & Stow Funeral Home
-
Mark Bennett reacted on thisCommunism and Nazism and Japanese imperialism were all versions of the same hierarchical/bloody feudalism that oppressed 99% of our ancestors for 6000 years... rule by kings, czars, dictators, commissars, owner lords and their spoiled inheritance brats. ONLY the moderate/pragmatic Enlightenment ever developed freedom and science and progress based on both cooperation and competition, as prescribed by Adam Smith & others.* It required some FLATTENING to prevent moguls gaining feudal power to cheat... and it required the MILD socialism of FDR and Scandinavia and the GI Bill generation to uplift children of the poor to compete confidently with those spoiled brats. Universities - the greatest pride of the GI Billers and the font of fact-based science - are now under subsidized attack by the world oligarchic putsch, which aims to destroy that enlightenment flattening and restore those 6000 years of crap. * Adam Smith - Liberals, you must reclaim him - CONTRARY BRIN - https://lnkd.in/bwKW7gH
Experience & Education
-
AdviceForge
View Mark’s full experience
See their title, tenure and more.
or
Licenses & Certifications
Volunteer Experience
-
Hat in the Ring Society
AOPA FOUNDATION
- Present 6 years 7 months
Education
The AOPA Foundation, a 501(c)(3) tax-exempt charitable, educational, and scientific organization, educates the public on the value of general aviation. The AOPA Foundation works to improve aviation safety through the Air Safety Institute, preserve community airports, and encourage learning to fly for career and personal benefit – all in the interest of ensuring the future of general aviation in America.
-
Volunteer Pilot
Aerobridge
- Present 6 years 6 months
Disaster and Humanitarian Relief
Our mission is to assist in times of catastrophic emergency by coordinating donated aircraft to provide a powerful, immediate response to disaster. By matching aircraft with emergency response teams and critical supplies, we are able to provide a vital window of assistance to save lives and aid those in need. By drawing on the resources of business and general aviation, we have the capability to respond to catastrophic disasters in a fluid and rapid manner.
-
Executive Mentor
Everwise
- Present 9 years 9 months
Helping others through volunteer coaching and mentoring to strengthen their leadership, management, and career skills.
-
Board of Directors
Waukegan Stick & Rudder Flying Club (non-profit)
- 2 years 1 month
Oversee management and major decisions related to aircraft procurement, maintenance, and business operations.
-
President
Artemis Users Association
- 2 years 1 month
Science and Technology
Non-Profit organization designed to facilitate interaction between users of Artemis software. Planned and executed annual user conference while collaborating closely with Artemis on customer business needs; product development; and platform strategy.
Patents
-
Multi-purpose smart card with user trusted bond
Issued US 11,100,379
The present disclosure relates a new generation "smart card" designed to create a severable invisible "bond" between the cardholder and the smart card itself where this trusted bond relationship is used to enhance and simplify the authentication process and during the use of the multi-purpose smart card. This new smart card is initiated and connected to a specific user using biometric information added to the card and the user using biometric information connects via a trusted bond with the…
The present disclosure relates a new generation "smart card" designed to create a severable invisible "bond" between the cardholder and the smart card itself where this trusted bond relationship is used to enhance and simplify the authentication process and during the use of the multi-purpose smart card. This new smart card is initiated and connected to a specific user using biometric information added to the card and the user using biometric information connects via a trusted bond with the card by pairing the biometric information which can be severed in one of multiple ways. The trusted bond with the smart card can be broken in one of multiple ways including disconnection from a network, distancing from the user, impact accelerometers, outside parameters, etc. The multi-function smart card also uses this established trusted bond with the user to simplify the authentication of the user for use of the card in encrypted computer network, ground security, or other retail and payment function.
Other inventors -
Honors & Awards
-
2013 Top 50 Business Professionals Northern Ireland
-
-
Defense Meritorious Service Medal
-
The Defense Meritorious Service Medal (DMSM) is the third-highest award bestowed upon members of the United States military by the United States Department of Defense.
-
Air Force Commendation Medal
-
The Commendation Medal is a mid-level United States military decoration which is presented for sustained acts of heroism or meritorious service.
Languages
-
English
Native or bilingual proficiency
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content