Actual Idiots
It was only a matter of time until I wrote an article on AI. I really wish I didn't have to as anyone who has spent any significant time around me would be aware, I really dislike AI.
I dislike AI for many of the reasons some of you might also dislike AI. I dislike it's potential job displacement, I dislike it's theft of art, I dislike how dumb it's starting make people, I dislike that it's called "AI" (It's not intelligent and it doesn't "feel" or "think") and I dislike that it's being rammed into every aspect of our lives whether we like it or not. Seriously, your rice cooker does not need ArTiFiCiAl InTeLlIgEnCe, it's a heating element and a scale...
All of those things and others are great reasons to dislike AI, but what I really HATE about AI is the fucking hype.
Sure, use the power of the sun to make a crappy glossy picture of a cat with a party hat on it's head and 4.75 legs!
Why not ask ChatGPT to help you write a poem whilst it drinks more water than a blue whale would see in it's lifetime?!
Get Claude to tell you everything is ok and you're actually the best person in the world and it's totally your friend!
If you're comfortable with unplugging your brain, stealing other people's artwork and having a non-sentient robot dictionary sycophant for your best friend then you go for it champ! Just don't talk to me about how great your new HAL 9000 replacement is and that it's the second coming of Jesus.
Bubbling Up
The current AI situation is no different to the dot com bubble of the early 2000s.
- You have a new technology that seems like it might hold some promise.
- Everyone starts to use or build on it.
- Venture capitalists are all too eager to fund every person and their dog's random projects no matter how crappy.
- Crappy projects are crappy and fail.
- Technology is seen as useless and people stop using it.
- People lose money and funding dries up.
As can be seen from the now widespread use of the web, these technologies don't tend to go away, they just get refined and all the hype boils off so they can actually be useful.
This has happened multiple times before and it will happen multiple times again, AI is just the current flavour of the decade. The cat is out of the bag and we are stuck with AI for better or worse, we just need the silicon valley tech bros to find something new and shiny to chase so we can get on with our lives.
You'd think that since I work in IT I would be all for the latest and greatest tech advancements but you'd be very wrong. A new technology in the form of a new standard or protocol that stands to change the way systems work behind the scenes is exciting, "Billy Bobs Synergy Scaler Gold Edition" is pure hype and nonsense. When people take a solid piece of technology, wrap it in shiny bits and oversell it we end up with the current situation. Remember, 95% of current AI projects fail to return a profit but that doesn't stop people from throwing shit at a wall to see what sticks.
I've seen what happens first hand when people fall for the hype despite my advice.
Another Waste of Silicon
At {org} we use Microsoft Azure as our major cloud platform provider, this makes sense as we are very much embedded in the Microsoft ecosystem.
One day I get a message from the boss: "Hey do you want to help with an AWS deployment?"
"Weird" I thought, I was under the impression we were pretty opposed to the multi-cloud design. I said yes regardless as there were only a handful of people in our team with any AWS experience and I came from a workplace that was 100% AWS.
Turns out it was one of those "throwing shit at a wall to see what sticks" projects.
Essentially {org} had decided it wanted to be on the forefront of AI development and utilisation and so had put aside a pot of cash to spin up multiple AI projects across various departments. Some in Azure, some from other orgs and this one in AWS. Fine, whatever, I'll help.
The AWS AI service is called Bedrock and provides access to a fair few LLMs. The plan was to deploy a sample application into AWS that would utilise Bedrock and provide a web app to create RAG bots. A RAG is essentially an AI technique that allows a bot to retrieve extra information beyond the dataset it's been trained on.
For example you could have a generic bot that uses Claude Sonnet 4.5 and can answer questions for you about anything it's been trained on. If the bot was given RAG powers it could be fed an additional dataset such as a company knowledge base or other non-public info and it would use that dataset on the fly to answer domain specific questions.
In {org}'s case the plan was to use the Bedrock Chat sample app and allow various teams around the {org} to create bots and provide knowledge bases to said bots to tailor them to specific use cases. All sounds fine to me if a little bit gimmicky.
Over Promise and Under Deliver
{org} brought in an external consultant who was extremely well versed in this new world of sentient toasters to help guide us through the various projects that were being initiated. To work in tandem with this consultant, {org} found the most AI glazing individual they could who was immediately enamoured with the idea of staff everywhere building bots to offload their cognitive functions on to. We will refer to them as {glazer} from now on.
The consultant presented the Bedrock Chat sample app almost as a finished product and promised that it could change the course of history, well not quite but it was very much over hyped, as with all things AI.
The problem with this was in the name: Bedrock Chat SAMPLE.
AWS had designed a bunch of applications and code snippets for organisations to dissect and integrate into other projects that when iterated on would provide a finished product. They weren't designed to be in production and they were very rough around the edges.
That did not stop {glazer} from thinking this particular deployment was tantamount to nuclear fusion however.
So on we plodded working with AWS Solution Architects directly (who I must say were fantastic) to get this deployment into something resembling a useful application. During the course of the initial deployment we even managed to help the AWS team with some gotchas that we had figured out but they were struggling with while working with other clients:
- We integrated Entra ID SSO into the application.
- We added banners to various pages to denote what kind of data was acceptable to be fed into the application.
- We removed models that we deemed risky or not useful (Deepseek, Amazon Nova, Mistral).
- We managed to make it use alternative regions for inference when the model requested wasn't available in ap-southeast-2 but still store the data and results locally.
- We hacked together some scripts that could pull data out of on-premises systems and feed it into S3 for the bots to reference.
All of these changes were way above and beyond the "just deploy the sample app" brief and at some point I had to stop the requests and say "This is a sample, it's supposed to have a team of developers work on it and create a finished product. I am not a developer", but the requests kept coming.
Most of the drive to make this something useful came from {glazer}. The consultant knew full well that it wasn't a production ready system and once the initial hype and excitement had died down it would be a fairly useless crappier version of ChatGPT. Of course they didn't let on, consultants laugh all the way to the bank after all.
Meeting after meeting ensued which only seemed to serve as a platform for {glazer} and the consultant to hype up more AI slop and promise techno-nirvana to the proof of concept trial users, who of course lapped it up.
Minimising
The final nail in the coffin for me was when the consultant hosted a "lunch and learn" on the project. My distaste for "lunch and learns" is palpable. Sure let me use the only time I get to collect my thoughts and be away from work for a short period of time during the day to listen to someone talk about MORE WORK.
It's bad enough that my downtime was stolen, however they are optional and you don't have to attend if you don't feel you would get anything out of it. This was a project I was actively involved in though so I thought there might be some mention of the insane amount of work we'd put in and people might be able to try the "finished" application...
Not only did the consultant actively lie about it's capabilities to the people asking questions as if to make this crappy AI RAG seem omnipotent (seriously it cannot read data from arbitrary databases, why even say it can?!) but they also said something that made me leave the call and go for a long long walk.
If you want to have a go with it and test all it's capabilities, someone in IT can add you to the necessary groups and give you access.
SOMEONE IN IT?!
Apparently I'm just some drone in IT. I helped rewrite this stinking thing into something useful for {org} and spent countless late nights making sure it was ready for the test users the next day.
The gloves were off, I felt no need to pander to this grifter after that statement.
I physically went up to the person that asked about the apps capabilities and was told "sure it can read from a database", I informed them in front of a full office: "He's lying, it 100% cannot do that".
The next useless "could have been an email" meeting was tense, it was obvious to everyone on the call that I was not happy. I'd wasted the best part of a year fucking about with this stupid thing only to be minimised to "someone in IT".
I believe the first thing out of my mouth was something along the lines of "This is a toy at best and I do not see it being anything other than that" to a shocked audience. {org} does not like it when you minimise the magical AI overlords as I've come to realise on multiple occasions, but I don't like having my work minimised either and I bite back unlike the ultimate sycophant we seem to have created.
The consultant seemed sheepish from here on out and would randomly insert hollow praise for the AWS staff and myself into sentences. It was as transparent as this entire project's end goal, to make them money.
This Costs Money?!
AWS had granted us a large block of credits to get started so we could deploy the proof of concept, get some users involved and see what kind of use cases they could come up with. Time and time again I mentioned that AWS can get really really expensive if you don't design things in a cloud native way. Time and time again I was told "it's fine we have loads of credits".
Well, loads of credits turned into a handful of credits and then "What does our spend look like?" At this point in the project barely anyone was using the application anymore, probably because the next shiny thing had come along and promised the earth whilst delivering nothing.
My reply?
About $200 a day, and it has no users and is doing nothing.
...shut it down
So I did, I turned it off and closed the AWS sub-account.
The next day {glazer} called me and said something I never thought I'd hear:
This AI stuff doesn't really do what they say it does
No shit
AI can be useful if used in moderation and to supplement people's existing knowledge. If it's used to replace people it falls over and it certainly isn't the deus ex machina people think it is.
That being said.
