Post #3486301

I think large language models and the cost assigned to them is the reckoning of our terrible closed data mindset when it comes to providing useful APIs. Why? Most of the times they "fix" a system that is broken from the beginning. Someone recently showed me a system he set up. Truck drivers send a picture of a bill. The bill is then analyzed and the data inside fed into a database and further processed. He did this with around 1k€ investment and less than 100€ power bill. In contrast, an "AI Company" offered to do this for 1.5k€ a month. First, I thought, neat. Nice use case. He told me it is not perfect, but good enough, it saves some time. Then I though more about it. The data is there in a machine readable context. It was converted into a bill, then a picture, then back to machine readable data for further processing. What would it take to make the original data available. E.g. with a qr code containing a JSON string. Then it hit me. I put myself into a pre-LLM hype mind. Very simple features like this QR code are incredible valuable. They save so much time and they would help everyone involved in the whole process. But the problem is: they are also very easy to implement. Turning billing data, quantity of items into a json string is a 1 hour job. An you don't want to sell this as a one hour feature, or god forbid, offer it for free. Instead it would become an expensive feature with a subscription or a per-bill fee. Something prices based on the *usefulness* and not based on the actual cost for implementation. And now back to the present with LLMs. And here comes the most ridiculous part. We are fixing an issue that would be very easy to fix by embracing an open data mindset by wasting kilowatt hours of energy every month. Think about this, a feature that would take about one hour to implement ONCE has been blown up to take a couple of extra Watt hours every single time a bill is being processed. And in the end, as I was told, the LLM isn't even perfect and keeps confusing 6's and 8's.