AI / Tech Predictions for 2026

I lived through the dot-com crash, and what I’m seeing in AI right now is giving me serious deja vu.

You see, I’m not just a tech person. Sure, I write code and teach technology. Still, I’ve also founded three businesses, have extensive domain experience in retail, finance, and insurance, and I provide strategic advisory services to startups, schools, and enterprises.

So, I have a rather unique perspective on what’s happening at the intersection of tech, education, and business. And, despite my AI hype fatigue, I’m going to push forward and predict what I see coming in 2026.

Prediction 1: The Hype Train Slows Down

One of the things that continues to frustrate me as a tech professional is the conflation between LLMs and the term AI. LLMs are only one type of AI. As I teach in my AI literacy workshops for executives and managers (reach out if your company needs some realistic grounding), there are a lot of really useful tools and techniques in the machine learning and AI space that aren’t LLMs.

Many people, especially those who follow me on LinkedIn, think of me as an “anti”, which is a newish derogatory term for someone who doesn’t drink deeply of the LLM Kool-Aid. And yes, I think there are serious ethical and societal issues that are not being adequately addressed, but I do use various AI tools, including LLMs, in my own business.

However, from the beginning, the hype has been out of control, and I often find myself overcorrecting in the other direction. My prediction for 2026 is that the AI/LLM hype train will run into the headwinds of the 90/10 rule.

The 90/10 rule isn’t about exact percentages; it’s a concept that states that the last bits of a complex project are harder and take longer than the early parts.

This is exactly what I predicted before GPT-4 launched, and what we've seen play out. Using OpenAI as an example, GPT-3 to GPT-4 was a significant leap, but 4 to 5? Not so much. Remember that Sam Altman, CEO of OpenAI, said at the beginning of 2025 that they knew fundamentally how to create AGI. They don't! Infinite scale does not create AGI, it never will, and that's a fundamental constraint of how the tech works.

“But look at these benchmarks!” the hypers claim, each time a new version is released. Yes, and if you understand data and machine learning, you also know that overfitting is a thing.

For my novice readers, if you've ever wondered why new model versions are released with much fanfare about benchmark improvements but your experience hasn't changed all that much, it's because those benchmark questions are leaking into the training data.

Humans overfit as well. Let me give you an example: when I'm interviewing people for a coding position, there is always a subset of people who can only give textbook definitions of concepts. They have memorized the textbook's answer wording, but because they don't actually understand the concept, they fail when you change the wording or context.

This gap between hype, benchmarks, and reality is starting to work its way into the C-suite.

Agentic AI will Fail

2025 was supposed to be the “Year of AI Agents.” I poo-poo’d this when the hype started, and I was correct.

The issue isn’t with the concept of automation itself. After all, I have automated countless processes over my career. The root cause is that people have drastically overestimated the value of a non-deterministic process in a deterministic world. Much of this stems from Big Tech’s origins in social media and advertising, where “good enough” is acceptable. If a Netflix recommendation is wrong, you scroll past it. If a supply chain order is wrong, a factory shuts down.

First, let’s define Deterministic. A deterministic process produces the same output for a given input, every single time. Your phone’s calculator is deterministic: 2 + 2 will always be 4. Most enterprise workflows must be deterministic:

• A bank transaction must debit and credit the exact penny.
• A pacemaker must trigger a heartbeat at the exact millisecond.
• A customer survey must trigger after every closed ticket.

Non-deterministic tools, like traditional machine learning and LLMs, deal in probabilities, not facts. They are excellent for sentiment analysis (”Is this tweet angry?”) or prediction (”Who might cancel their subscription?”). In those cases, being 85% correct is still valuable. The output is trends, not facts or truths.

The failure happens when leaders try to apply non-deterministic tools to deterministic problems. Many executives do not fundamentally understand this distinction. Most of my strategic consulting in late 2025 and into 2026 is centered on AI literacy for non-technical leaders. 2026 is looking busy because companies are finally asking the hard questions:

• How do we measure ROI on a system that makes mistakes?
• How do we handle privacy, ethics, and explainability in a “black box” model?
• Where can AI actually be applied safely?

Many executives have been running on FOMO. As the hype cycle slows and the bubble deflates, we will see a return to reason and actual engineering later in 2026. It will take time (and some very public failures) for the message to stick, though I was pleased to see Salesforce executives recently admit they overestimated AI’s current capabilities.

In short, in the social media/ad-tech world, a 1% error rate is fine. Hell, a 40% error rate is often fine. In the world of industrial engineering or finance, a 1% error rate is catastrophic.

The Correction is Coming

The amount of money being dumped into data centers and AI infrastructure is insane. But look closer, and you see the cracks: deliberate opacity in reporting usage numbers, disturbing accounting games with asset depreciation (e.g., claiming that servers and GPUs last 6 years instead of 3), and “megadeals” that turn out to be non-binding press releases.

The bottom line is that if these tools delivered on the hype, there would be a ton of success stories, white papers, demos, and more from big tech. They would give us the playbook for exactly how to 10x your workers’ productivity, backed by data. And they’d do it because it would make them massive amounts of money. The lack of details is the smoking gun.

I am not a financial advisor, and you shouldn’t try to time the market, but when you overstate your revenues by playing accounting games with depreciations and such, the chickens will come home to roost at some point. Right now, a lot of the exposure is in private equity. When publicly traded companies like Oracle, Microsoft, Amazon, Google, and NVIDIA play these games, the general public, including pension funds and 401(k)s, is exposed to the risk.

Now, I avoid sharing my political views publicly for the most part, but there is massive corruption and political turmoil in the U.S. right now. We haven’t addressed how AI companies and products use surveillance, we haven’t addressed copyright issues, and we have a presidential administration that refuses to regulate AI at all. If the midterm elections this November result in a change of parties in Congress, we will see a shift toward stronger consumer protections and greater regulatory pressure. If the bubble hasn’t popped by then, this will surely pop it.

Personally, I think the US needs to get on board. Big tech has way too much power, and the way they have flaunted copyright law, ethics, and regulations is genuinely concerning. We’re seeing monthly reports on issues with inappropriate AI usage, from hallucinations in police reports and legal briefs to deep fakes and non-consensual pornographic image generations, including minors. We need to put in appropriate regulations and penalties to expose and punish both the offenders and the companies that enable them.

The European Commission states that pornographic material accounts for about 98 % of deepfakes. Think about that.

In the United States, there is an argument that we must let these companies do whatever they want, no matter what, or we’ll “fall behind”. This is complete bullshit. As someone with boots on the ground, things are changing so rapidly that there is no falling behind.

AI and Coding

Now let’s talk specifically about software development work. AI can be a genuinely useful tool for coding. But there’s a big difference between using these tools as assistants and outsourcing all the work to an LLM.

When the term “Vibe Coding” showed up, I rolled my eyes and mostly ignored it. For 2026, I stand by that reaction. Conceptually, I see LLM-based building tools as little different than the low and no-code builders that have been around for decades.

Can you vibe code small apps for personal or specific use cases? Sure, sometimes. But the same issues that pop up with no-code tools pop up with Vibe-coded projects.

Simple, small, greenfield projects? They work pretty ok. Complex, layered workflows? Not so much. Eventually, you need to apply actual engineering processes to get things done right.

Many beginners and people outside the field don’t realize that most of a professional developer’s job is research and analysis, not writing code. Code is an expression of a solved problem. You must first analyze the problem, consider trade-offs such as time, budget, security, and scale, and then implement that solution in code.

Writing code was never the bottleneck. If it were, we’d be asking for your typing words per minute on your resume! Also, reviewing code you didn’t write is much harder than writing the code yourself. So, what we’re seeing in the productivity data is that we’re not eliminating much time; we’re just moving it from the initial generation to the review and maintenance side.

I work with many skilled developers who use LLMs daily. They succeed when they use it from an engineering perspective. Small context, clearly defined specifications, proper test coverage. These are things that people without fundamental skills cannot do well. Vibe coding isn’t a real job; it is unlikely ever to be one, but it’s great if small businesses and individuals can get easy stuff done cheaply and quickly. I often point out to people that I started my career with basic HTML, CSS, and JavaScript, as well as drag-and-drop Windows Forms apps. No one will pay my rates for that type of work today, and that’s fine; my skills have moved upstream.

If you’re a beginner trying to break into the field, it is tough right now, but it’s because the bar is going up. You will be expected to bring more process and engineering concepts to interviews, in addition to basic coding skills, to be successful. Though, because of AI interview assistance, leetcode is going to start dying, and I’m glad for that, because I thought it was performative nonsense all along. I won’t miss it. The best way to learn and break into the field is to get to a point where you can build full-stack applications without AI assistance, and then start learning to use them in your day-to-day workflow selectively, where they make sense. From my experiences working with companies and hiring managers, if you use AI as a coding crutch, you will be unemployable.

Wrapping Things Up

2025 was the year of hype. 2026 is when we will see reality start to creep in. The hype cycle has peaked; easy money is drying up, and the magic of LLMs is fading. Accounting games and shiny benchmarks will not substitute for actual engineering.

There’s a reason OpenAI is looking at advertising as a revenue stream instead of, you know, getting more revenue from enterprises: their tools don’t actually work as marketed, and LLMs are a commodity. I can’t stress this enough. If the tools were capable of replacing workers at scale, the revenue would be sufficient that advertisements would be an afterthought.

There’s a reason why all of these companies are still hiring human engineers. And there’s a reason why companies like Klarna and Salesforce are backing away from their bold claims of mass worker replacements.

In 2026, I will continue to promote real, rigorous learning backed by software engineering principles. Open your editor, get solid on the fundamentals, and get to work.

Happy Coding!