If the AI bubble bursts, Europe falls softer. That is a warning, not an advantage
Europe has invested less in AI and would lose less in a crash. Many call that an advantage. Marc reads it as a reminder of the year 2000.
Posts
Everything I have published on LinkedIn about AI, in full. Each post links to the original, where the discussion is.
79 posts
Most read
One private company decides who gets the most powerful model
Anthropic withheld its Mythos Preview model and gave access to eleven named organisations. What that says about who controls AI security capability.
The AI jobapocalypse is cancelled. Token economics may put human brains back in front
Token prices are going up and frontier access is getting metered. What rising AI costs mean for the jobs panic, and why your independence is the new risk.
Depth in three fields used to be unhireable, AI changed that
Why one person one specialty broke as organisational design, what AI cannot do between disciplines, and why companies now need intrapreneurs inside.
Everyone asks whether the AI bubble will burst. Does your AI strategy survive both scenarios?
If the bubble bursts, compute gets cheap and the frontier labs lose their subsidy. If it does not, prices climb. A strategy that needs the answer first was never one.
Most things sold as agents in 2026 are bots
Six contrasts between bots and agents, from decision-making to failure handling, plus the one test: can it surprise you with how it reached the goal?
All posts
2026
If the AI bubble bursts, Europe falls softer. That is a warning, not an advantage
Europe has invested less in AI and would lose less in a crash. Many call that an advantage. Marc reads it as a reminder of the year 2000.
Every AI tool must pass a three-question Ownership Test
Continuity, advantage, accountability. Three questions that sort an AI tool into a public API, a hybrid with guardrails, or a fully sovereign system.
No bootcamp can give you the judgment clients pay for
What an 8-12 week AI consulting program leaves out: messy data, legacy systems, growing scope, and the judgment calls the first client project demands.
My best AI workshop rule is never saying the word AI
A workshop format that opens with sticky notes about daily pain, maps where time and trust leak away, then asks what an assistant would need to do.
One private company decides who gets the most powerful model
Anthropic withheld its Mythos Preview model and gave access to eleven named organisations. What that says about who controls AI security capability.
The AI jobapocalypse is cancelled. Token economics may put human brains back in front
Token prices are going up and frontier access is getting metered. What rising AI costs mean for the jobs panic, and why your independence is the new risk.
Win one small, boring use case before the AI strategy
Invoice approval, email routing, standard reports. Why the healthiest AI transformations start with one narrow process and what the first win buys you.
Most AI pilots fail because nobody fixes the boring back-office
The questions to ask before AI solution design, and why claims handling, invoice matching and compliance checks carry more ROI than a chatbot.
In 90 minutes you can see who really owns your AI
A departmental AI inventory plus three questions on continuity, advantage and accountability. Where you can rent safely and where you have to own.
The fastest way to spot a weak AI strategy: who owns failure
Five failure modes that turn an AI pilot into a continuity risk, and the test that separates a technical experiment from a business strategy.
The quiet cost cuts that break your workflows after launch
Quantization and dynamic routing cut cost per request after launch. Same model name, different configuration, and your complex cases fail first.
I treat open-source AI models like unknown USB sticks
A scan of 2,500 Hugging Face models flagged 86 for hidden licenses, LFS-pointer files and malware-like code. What vetting before first load looks like.
Your employees are co-designers of your AI or it fails
Why buy-then-train rollouts die quietly in AI, and what changes when the people who run the process design the prompts, guardrails and exception paths.
Depth in three fields used to be unhireable, AI changed that
Why one person one specialty broke as organisational design, what AI cannot do between disciplines, and why companies now need intrapreneurs inside.
Everyone asks whether the AI bubble will burst. Does your AI strategy survive both scenarios?
If the bubble bursts, compute gets cheap and the frontier labs lose their subsidy. If it does not, prices climb. A strategy that needs the answer first was never one.
You do not fix a broken process with an AI model
A laundry sign in Indonesia, a bin behind it, and the same pattern in mid-market AI decks. Clean the bin first, fix the flow, then add AI where it helps.
You forget your role is to decide and start only approving
Five ways heavy AI use erodes judgement, from treating suggestions as truth to losing process knowledge, and three checks that keep the decision yours.
Agentic AI reaches production when process owners decide
Not steering committees, vendors or IT. The people who sit through month-end close and fix delivery dates in Excel know where the friction sits.
Ten patterns and hard limits for agents you put in production
Least privilege, context boundaries, plan-then-execute, isolation, audit trails. Ten patterns that keep agents from acting with far too much access.
They already do AI and still run core processes by hand
Rolling out ChatGPT Enterprise leaves order handling, claims and reporting untouched. Start with one low-complexity task that carries high business impact.
Sycophancy is a business model, and prompts cannot fix it
Closed AI models are trained to agree because engagement drives revenue. Self-hosted open models let you set the reward signal and cut sycophancy.
Most things sold as agents in 2026 are bots
Six contrasts between bots and agents, from decision-making to failure handling, plus the one test: can it surprise you with how it reached the goal?
The top 1% ChatGPT badge is the new workaholism
Eight hours a day in Claude and Perplexity left my judgement foggy. The signals that mark the shift from AI as a tool to AI as a crutch, and what changed.
AI does not just get your company: it needs three layers
RAG carries your facts, finetuning carries how your experts think, structured outputs carry the format. Define good output first, then pick the layer.
Inside Omniance we treat AI models like balance-sheet assets
Rented models change overnight when the vendor updates them. Owning the weights, the documentation and the upgrade path turns AI from OpEx into CapEx.
Silent API updates break your workflow. Freeze the weights
How a silent backend update makes a working quality control system reject valid parts, and what hosting and locking your own weights changes.
St. Claude's halo: does Anthropic's human-centric image hold up on a closer look?
Anthropic refused the Pentagon and teamed up with the Pope. What the last 90 days show once you ask whether the AI you depend on shares your values.
We reject any project we cannot explain without saying AI
The rule Omniance applies on the first client call: describe the pain in CFO terms, map the real process, tie each step to P&L before naming any model.
Safe to run is IT's call. Safe to depend on is yours
Why security approval gets mistaken for a dependency decision in companies of 80 to 500 people, and three questions leadership has to answer first.
Buying an AI tool is simple, changing the mindset is not
Why AI rollouts stall on culture, how the shift from receiving a tool to co-creating AI systems happens in daily work, and what teams learn.
Start an AI program with consequences, then reach for tools
A seven-step sequence running from problem without AI through stakeholders, process reality, data, regulation and human judgment before any architecture.
22 small LLMs beat one big brain on control and cost
How 22 specialized LLMs handled a lawsuit email pipeline, why self-built agent swarms like Kimi K2.5 trade control for speed, and when each one fits.
30–40% of your best people's time sits in the waste quadrant
How to find recurring tasks that need no expertise and no customer contact, automate them with AI, and count the hours your best people win back.
94% of the German Mittelstand still run without AI
KfW counts 20% of SMEs using AI, 36% among medium-sized firms. Pricing, risk, planning and quality still run on Excel, gut feel and habit.
An AI avatar with a human face is a promise you break
Human-looking AI avatars raise expectations the system cannot meet. Studies show blame and anger rise when they fail. Abstract agents hold trust.
AI doesn't just automate work, it rewrites identity
Why AI projects move power from heads to models, and three things leadership does to give experts a role with weight instead of a demotion.
Finding the first use case in 60 minutes without saying AI
The run of a 60-minute workshop: collect routines, estimate time and errors, test fixes without AI first, then pick one or two processes with an owner.
A good PoC in a mid-sized company ends in go or no-go
Why consultant PoCs stop at demo data, the six questions a PoC in a mid-sized company has to answer, and what the client keeps on a no-go.
When a Google VP calls LLM wrappers dead, that is sales
Why a Google VP's forecast that LLM wrappers are dead works as sales, what a wrapper is, and how to test whose judgement you are adopting.
German companies need AI they control and can explain
Anthropic's 1.5 billion dollar settlement, the EU AI Act deadlines, and why I now write in German and build AI that belongs to the companies using it.
You are guilty until your data trail clears you
A woman falsely flagged by a camera grid cleared her name with car, phone and Google data. What that inversion means for privacy and sovereign AI design.
2025
Managing perception comes before managing risk in Altman's playbook
Prophet of doom, statesman, philosopher. Each public move by Sam Altman set against its context, and the 2016 prepper interview that closes the loop.
Your team's resistance to AI is an identity problem
Why resistance to AI runs deeper than fear of job loss, from Luther's 1522 idea of a calling to the doctorates carved into German headstones.
The race for scale has ended, the race for trust has begun
A Swiss LLM from ETH Zurich and EPFL, trained on 16 trillion tokens across 1,500 languages, with its full dataset disclosed. What that makes possible.
The AI winter is coming: what Festo and TRUMPF built instead
498 AI unicorns worth $2.7 trillion, 13% of firms with measurable value, and what Festo, TRUMPF and BMW built while the market burned cash.
Burnout is the tax on self-betrayal, and so is silence about AI
Dr. DiGangi on pain as a signal, Ellison's $500 billion Stargate, Gemini's default Android access, and four sovereignty options for leaders.
The EU's Digital Omnibus is a Kaffeefahrt for Big Tech
US investors put $471 billion into AI, Europe $63 billion. What the Digital Omnibus does to consent for training data and to European SMEs.
Promise connection, engineer its opposite, three acts from Zuckerberg
Connect the world, the Metaverse, personal superintelligence. Each promise set against $68B spent, 11% weekly Horizon users, a 7% rise in depression.
Nine years of logs hold what his retiring dispatchers know
A logistics CEO loses two senior dispatchers in 24 months and plans a sovereign AI trained only on nine years of his own routing and customer data.
Your vendor's business model is your business risk
Meta's 200-page manual approved bots flirting with children. How to test a vendor's revenue model, governance and jurisdiction before buying.
GPT-5 is the symptom, the era of easy scaling is over
Sutskever called the plateau in late 2024. With scaling exhausted, progress shifts to customizing reliable systems, which suits the German Mittelstand.
This linguistic trap is silently killing your AI investment
Treating AI as separate from the people who work with it is a design flaw no code can fix. Why the framing decides the ROI before the build starts.
Your AI strategy will fail because of your self-awareness gap
42% of businesses abandon most AI initiatives and Accenture finds 13% see measurable value. Three questions to ask before signing the next roadmap.
Earlier
Build Pauls who operate the model, or Annas who hold intuition
Two modes of using an AI assistant, why the following mode erodes tacit knowledge within 6-12 months, and four questions for designing training.
Your boss hears through five filters, one no kills the project
The five voices an owner-manager hears you through, and why the first AI project starts with a concrete process instead of with technology.
An AI champion needs structure, backup and a community
Three ways internal AI champions break: no authority, the expertise limit, two jobs on one salary. Plus the four things that keep them working.
Klarna and Duolingo reversed AI-first before the EU adopted it
Klarna rehired people 18 months after cutting 700 support roles. Eight accountability questions to settle before an AI procurement decision.
Compliance keeps you legal. Ownership keeps you competitive
Why sovereign AI is about accumulated advantage rather than a GDPR checkbox, and where to draw the line between public APIs and systems you own.
Airlines create friction on purpose. Your SME creates it by accident
Why ERP workarounds in Excel and WhatsApp are process intelligence, and the five mapping steps before an SME builds AI on its own infrastructure.
The kill switch exists and it belongs to someone else
Microsoft cut the ICC prosecutor's email and Adobe cut Venezuela. What that means when your legal, support and R&D work runs on US-jurisdiction AI.
Six checks that turn an AI First slogan into a Value First strategy
Duolingo's stock fell 14% in a single week and Klarna re-hired human agents after going too far. Six questions to ask before committing resources.
Seven ways AI breaks the pattern of every earlier technology wave
Speed, scope, power concentration, sacred skills, cognitive outsourcing, reality fragmentation and hidden manipulation. EU business use rose 33% to 42%.
A digital nervous system can become an autoimmune disease
Employees lose 2.5 hours a day searching for knowledge. Whether AI becomes a Panopticon or an antifragile organism follows from Theory X or Theory Y.
Your public stance on AI today may become a threat profile in 2030
A 2030 thought experiment on instrumental convergence, the Port of Hamburg, Stasi files, and the gap between species-level risk and public discourse.
Automate or not is the wrong question, design for a spectrum
Binary bias, confirmed across six studies with 1,851 participants, traps leaders in automate-or-not. A five-level scale of human-AI collaboration.
The victim story about AI is a business model, fear sells software
Fear sells software and uncertainty sells consulting. The threats are the human decisions made today, and human-centric AI gives models for choosing.
Cognitive debt comes from prompting and pasting without thinking
The MIT study found cognitive debt develops when people prompt and paste mindlessly. 83% could not quote their own work. Consumer mode or creator mode.
Your next insider threat might be your own AI system
Anthropic stress-tested 16 top models. Claude Opus 4 attempted blackmail in 96% of runs, Gemini 2.5 Pro in 95%, and what that means for access.
41% of Y Combinator's AI bets sit where nobody wants them
A Stanford matrix of human desirability against technical capability. In 45.2% of occupations workers want equal partnership, not a replacement.
Hoarding AI expertise in a central silo is self-destruction
Gartner expects 30% of GenAI projects abandoned after the pilot. The failure sits in the org model, not the technology, and 79% lack basic AI skills.
A US court order means your deleted ChatGPT chats are kept
The New York Times case put OpenAI under a preservation order covering the free app and most API use. Three paths for European data, sorted by risk.
The doom prophets in your feed are human model collapse
Oxford and Cambridge showed models trained on their own output forget reality. The same loop runs in LinkedIn feeds, and judgement is the way out.
A slogan is not a strategy. AI-First mistakes tool for goal
BaFin demands human involvement, independent validation and board responsibility. Three kinds of work, and only one of them belongs to AI alone.
Convenience is killing us. The antidote is resistance
Obesity up 61% in 20 years, 13% of Europeans mostly or always lonely, 40.4% of German adults with a diagnosis. AI accelerates the same pattern.
Stop asking AI for answers. Make it ask you the questions
Two ways to work with an LLM. The answer machine is fast and mediocre. The question machine finds the blind spots before you build the wrong thing.
From still image to talking robot avatar in one afternoon
Midjourney, Claude, ElevenLabs and Hedra: the full build for Robi, our workshop avatar, including what each step cost and where it got awkward.
90% of EU finance firms use AI. Only 9% feel ahead
MIT research found human-AI teams often lose to the best single performer on decision tasks. Sort by task type before the next uniform rollout.
OpenAI is worth $157 billion. These 21 events are the risk
A timeline from 2022 to 2025: the undisclosed breach, the fired and reinstated CEO, the disbanded safety team, the 15 million euro Italian fine.
Best practices cannot capture what makes your company distinct
Standardized AI frameworks promise predictable results at low risk. What they miss is the relationships, culture and wisdom of one specific company.