TechnoMagicians Blog

Everyone in tech has used the phrase “Garbage In, Garbage Out” at one time or another. It’s the notion that if your programs use garbage input as their source of truth, you are going to get garbage out. Meaning incorrect results that users are expecting or think are correct.

I see an even bigger problem with AI. It’s not just “Garbage In, Garbage Out”, we are heading towards “Nothing In, More Garbage Out or ALL Garbage Out”. In the age of AI, its training on existing data and the folks putting up “GOOD data” on the Internet is becoming less and less. Not to mention more incorrect data and false data is being posted.

Take a look Stack Overflow as an example where AI is for the large part getting its training data from for coding. Stack Overflow and very GOOD data, or at least it did in the early days because the site was curated by HUMANS which led to very high quality correct answers. But back in 2022, when ChatGPT was launched, traffic to Stack Overflow dropped by 20-30% and was down more like 40-50% by 2024. In 2023 Stack Overflow was allowing AI answers but mostly the answers ended up being wrong code, plausible sounding nonsense, and hard to detect hallucinations. The moderators were overwhelmed by low quality incorrect AI answers and wanted less AI content or AI banned altogether. Unfortunately Stack Exchange leadership made the opposite decision and wanted more AI content. That led to pretty much the death of Stack Overflow and Stack Exchange altogether. There likely hasn’t been much in the form of GOOD data there in the past couple of years. So one can forget AI training on any recent GOOD code as Stack Overflow was probably the main place or only place that sort of code was found.

This isn’t just technology and coding. The same erosion to GOOD data is happening everywhere. Take a look at reliable, fact based, investigative news journalism. It hardly exists anymore. Replaced 90% by mostly lies, BS, intentionally misleading made up nonsense all with zero facts or truth. This is what AI is training on now.

The list goes on and on.

Don’t get me wrong, AI is very useful in many cases especially if the model is based on GOOD data. There are already 100’s of specialized models that train on very specific topics / data so I expect those to fair well. But any sort of general AI model pulling data from the Internet is heading down a dark deep hole.

I lifted the following image from this post which reminded me to write this in the first place https://lnkd.in/g7tCqYrF