The Great Data Devaluation: Why AI Is Eating Junk Food
The Great Data Devaluation: Why AI Is Eating Junk Food
I love the idea of a fully self-aware hammer. You've got this incredible, powerful tool that is constantly learning, talking to you, and offering advice on how to build a better wall. That is the promise of today's LLMs.
But what if you discover that the hammer was trained entirely on plans drawn in crayon and measurements scrawled on the back of a beer mat?
That is the question now looming over the entire field of AI. We're all focused on the spectacular capabilities of the models themselves, but we're slowly realising the data underneath them is getting poisoned. AI is eating junk food, and it's losing its reasoning skills as a result.
The AI is a Filter, Not a Fact Machine
The first problem is that these models live in a separate reality.
When you use Google, you get links to the known web. When you ask an LLM, it is citing information from sources that are often less known and different from what a search engine would pull up. This is a massive shift. We are building a layer of knowledge that doesn't always align with the widely vetted parts of the internet.
Now add bias into the mix. This week, Elon Musk's Grok was shown to use a service called Grokipedia, which is essentially a biased, built-in wiki. Researchers found the content to be biased AI slop that reinforced controversial views.
The Problem with Training on a Rubbish Tip
The internet is becoming a rubbish tip of its own making. Generative AI tools are pumping out oceans of synthetic content that is scraped up and fed back into the next generation of models. We are, quite literally, creating an information echo chamber.
Bill Gates recently called AI the biggest technical thing ever in his lifetime. But he also warned of a bubble. This is that bubble. We are building the most advanced technology on the weakest foundation since someone invented a house made of sand.
The Corporate Escape Route
How are the biggest companies trying to fix this? They're running away from the public internet.
OpenAI, for example, is positioning ChatGPT as a search engine for work data. They want you to feed the LLM your company's internal documents. They want you to pay them to turn their hammer onto your own, clean workbench.
We must use these flawed systems to stay competitive. That is the brutal truth. But we must also approach every output with the same caution a foreman uses when inspecting a questionable foundation. We need to stop trusting the hammer blindly just because it talks nicely. We need to check its blueprints.
News Roundup: The Five Business Imperatives
AI-Driven Manufacturing is Here
Samsung and Nvidia have partnered to build an AI Mega Factory, powered by 50,000 Nvidia GPUs. The goal is to embed AI across the entire manufacturing process, from design to quality control.
NVIDIA and Samsung Build AI Factory
Leaders Aren't Aligning on AI Strategy
One report found that 53% of CEOs struggle to get their C-suite to agree on AI and talent strategy. This lack of unity means a third of leaders still do not have a corporate AI policy.
The Urgent Crisis in Applied AI Skills
A recent UK report found a 51% skills shortage in the digital sector for AI-related roles. The shortfall isn't just in deep technical roles. It's also in the applied AI skills needed for adjacent roles, where workers simply need to know how to use the tools effectively.
No-Code Tools Create New Developers
Gartner projects that over 70% of new applications will be built using low-code or no-code technologies by 2025. No-code platforms can reduce development time by up to 90%.
Low-Code Platforms Reshaping Apps
Marketing Becomes Hyper-Personalised
Salesforce has rebranded its AI offering to Agentforce, using it to auto-generate personalised emails and service responses.
The Internet is now a data rubbish tip. If your AI is eating junk food, its reasoning will fail. Stop trusting the self-aware hammer.