Karpathy’s skepticism, as reported by ITPro, isn’t just an off-the-cuff remark; it’s a pinprick to a balloon of what we must call AI capability inflation. He suggests it could take “about a decade” to work through the cognitive issues plaguing these agents before they’re truly viable. This isn’t some Luddite complaining about a new machine; this is one of the architects of the revolution telling us the blueprints are still very much a work in progress. And it forces us to ask a difficult question: are we hurtling towards a future of intelligent automation, or are we simply caught in one of the most aggressive tech hype cycles we’ve ever seen?
The Great AI Expectation Game
So, what exactly is AI capability inflation? It’s the ever-widening canyon between what AI is advertised to do and what it can actually deliver reliably in a real-world business context. It’s the shiny demo that works perfectly under controlled conditions but falls apart when faced with the messy, unpredictable nature of your company’s actual data. Think of it as a form of marketing-driven wishful thinking, where the promise of future capability is sold as a present-day reality.
The implications are far from trivial. Businesses are pouring billions into AI, driven by a fear of missing out. According to a prediction from Microsoft, we could see an astounding 1.3 billion AI agents in use by 2028. This creates immense pressure to adopt, to integrate, and to show results. But when the technology isn’t quite there, you end up with wasted investment, frustrated teams, and a growing cynicism that could poison a genuinely useful technology. The danger is that we inflate expectations to a point where the inevitable disappointment triggers a backlash, potentially leading to premature predictions of an AI winter.
The Ghost in the Machine is Still Just a Glitch
To understand why Karpathy used a word like “slop,” we have to look soberly at current machine learning limitations. An “agentic” AI is supposed to be more than a chatbot. The idea is that it can understand a complex goal, break it down into steps, use various tools (like browsing the web, accessing a database, or sending an email), and execute that plan autonomously. The problem is, they’re not very good at it yet.
Imagine you’ve hired a very keen, very fast, but utterly inexperienced intern. You ask them to compile a report on your main competitor’s quarterly performance. The intern races off, but instead of using the official financial filings, they pull numbers from a random blog post, misunderstand the CEO’s comments on an earnings call, and then “creatively” fill in the gaps with plausible-sounding but entirely fabricated data. The final report looks vaguely correct, but it’s a liability. This is the state of many AI agents today. They suffer from a list of cognitive limitations:
– Poor long-term planning: They can handle a two or three-step task but often get lost or an go in circles when faced with a more complex, multi-stage project.
– Lack of robust reasoning: They struggle to “think” critically about the information they find, often failing to weigh the credibility of different sources or spot contradictions.
– Brittleness: They can fail spectacularly if they encounter a situation or a tool that is even slightly different from what they were trained on. One unexpected pop-up on a website can derail an entire process.
This is the gap between expectation and reality. We want a seasoned senior analyst; we’re getting an over-confident intern who requires constant supervision. The “slop” Karpathy refers to is the mess a human employee has to clean up after the agent has had its go. This is a critical point that the marketing materials from big tech firms often conveniently omit.
The Industry’s Double-Edged Response
So, how are the big players responding to these very real limitations? On the one hand, you have companies like Microsoft and Salesforce doubling down on the agent narrative. Microsoft’s ambitious forecast and Salesforce’s push with its “Agentforce” framework are classic examples of trying to will a market into existence. They are building the platforms, betting that the capabilities will eventually catch up. It’s a high-stakes, high-reward strategy.
On the other hand, there’s a more insidious trend analysts are calling “agent washing.” This is where vendors take existing, relatively simple automation tools—like an API connector or a basic rules-based workflow—and rebrand them as a sophisticated “AI agent.” It’s a cynical marketing ploy that further fuels AI capability inflation and erodes trust. It tricks customers into believing they are buying a piece of the autonomous future, when in reality, they’re just getting a repackaged version of yesterday’s tech. It’s this exact behaviour that has analysts at Gartner sounding the alarm bells.
Are We Staring Down an AI Winter?
Gartner’s analysis, highlighted in the same ITPro article, is a bucket of cold water for the entire industry. They predict that a staggering 40% of enterprise projects building agentic AI systems could be abandoned or cancelled within the next two years. That’s a massive rate of failure. It’s not because AI is useless, but because projects are being launched with wildly unrealistic expectations about what the current state of technology can achieve. When a company invests millions in an “AI agent” that ultimately creates more work for its human employees, the project sponsor’s enthusiasm—and budget—dries up very quickly.
This leads to whispers of AI winter predictions. For those who haven’t been around the block, an AI winter is a period of reduced funding and interest in artificial intelligence research, typically following a period of intense hype. We saw them in the 1970s and late 1980s when promises of machine translation and expert systems failed to materialise.
The lesson from history is not that the technology was a dead end, but that the hype got way ahead of the science. The same risk exists today. If too many businesses get burned by over-investing in half-baked “agent” solutions, we could see a broader pullback. To be clear, this wouldn’t mean a stop to the progress at OpenAI or Google DeepMind. It would mean that mainstream enterprise adoption stalls, and the transformative potential of AI gets delayed by a decade of cynicism and distrust. How can companies avoid this? By being brutally honest about ROI and starting with problems that are solvable with today’s technology, not the technology promised for 2028.
Navigating the AI Hype Cycle
What we are witnessing is a classic example of tech hype cycles. A breakthrough happens (like the transformer architecture that powers modern LLMs), leading to a “Peak of Inflated Expectations.” This is where we are now with agentic AI. The demos are amazing, the promises are boundless, and everyone is a true believer. This peak is inevitably followed by the “Trough of Disillusionment,” where the technology’s failures and limitations become apparent, and the backlash begins. This is where Gartner’s 40% failure rate comes in.
Look at past examples. The dot-com bubble was a hype cycle built on the promise of e-commerce, but many of the companies that crashed and burned had no viable business model. The underlying technology (the internet) was revolutionary, but the application was premature. Similarly, virtual reality has gone through multiple hype cycles, with the promise of the metaverse always seeming to be just five years away. Genuine innovation is a slow, iterative process. Hype is a marketing phenomenon. The trick is telling the two apart.
Microsoft’s grand prediction of 1.3 billion agents is the sound of the peak. Salesforce’s approach seems a bit more grounded, focusing on giving developers the tools to build agents, implicitly acknowledging that this isn’t an out-of-the-box solution. The most successful businesses will be those that treat agentic AI not as a magic black box, but as a low-level capability that requires significant engineering effort, careful process design, and, most importantly, a human in the loop.
Where Do We Go From Here?
The path forward requires a pragmatic recalibration. The gap between the hype and the reality of AI is not a sign that AI is a failure; it’s a sign that it’s a maturing technology. For business leaders, the message should be clear: resist the hype. Don’t buy an “AI Agent”; buy a solution to a specific business problem.
Here are a few recommendations for navigating the landscape:
– Start Small and Prove Value: Instead of trying to automate an entire department, identify a single, high-friction, repetitive task. Automate that with a human-supervised AI tool and measure the return on investment.
– Keep a Human in the Loop: Acknowledge the “slop.” Design your workflows with the assumption that the AI’s output will need to be reviewed, corrected, and approved by a person. Think of the AI as a productivity multiplier for your existing team, not a replacement for them.
– Question Everything: When a vendor sells you an “AI agent,” ask tough questions. What are its specific machine learning limitations? What happens when it fails? Can you show me a demo using my company’s messy, real-world data, not your perfectly curated sample set?
Karpathy’s dose of realism is a gift. It’s a call to be more rigorous, more critical, and more honest about what we are building. The future of AI is not one of flawless autonomous agents magically appearing to solve all our problems. It’s one where we painstakingly, iteratively, and intelligently integrate these powerful but flawed tools into our existing human processes. The companies that understand this distinction are the ones that will truly succeed.
What do you think? Is the industry’s focus on “agentic AI” a dangerous distraction, or is it a necessary step to push the boundaries of what’s possible? Have you seen examples of AI capability inflation in your own work? Let’s discuss it in the comments.
