Microsoft exec warns of business functions being sacrificed on the altar of AI Arun Ulag, Microsoft corporate vice president for Azure Data, reckons that, in a world of constrained or finite IT budgets, something will have to give if new projects are to thrive. Speaking at the Citi 2024 conference recently, Ulag was outlining thoughts on roadblocks and constraints customers face in the latest data …
A lot of companies seem to be jumping on the dogshit bandwagon that is AI.
They are coming up with solutions which they claim use AI when they really don't.
This is why you're seeing adverts with phrases like "powered by AI". Because if 0.00001% of the solution uses some vague form of AI, you can advertise it as being AI. The rest of it is cobbled together to cover up the MASSIVE SHORTCOMINGS in AI, whilst still pretending AI is the best thing since sliced bread.
In the UK there are adverts on broadcast TV from the likes of Google and AWS doing exactly this at the moment.
...if you want shiny AI baubles, something else will need to be sacrificed.
You can start by getting rid of the C-suite and replace them with the AI. That would save you a boatload of money and have plenty left to pay your unappreciated and underpaid employees a very significant rise. You should also replace middle management for additional equalization of monetary distribution.
"You should also replace middle management for additional equalization of monetary distribution."
Like Death himself, I have cut down over 20 souls from middle management in just the last few weeks. I carry their work souls in a glass capsule which hangs around my neck.
They are the easiest sector to clear out with quite simple AI. The savings are breath-taking.
This post has been deleted by its author
Quite possibly replacing the C-suite is the one area that AI might be seen as a providing a real advantage. Problem being it's seen as that only by everyone outside the of the C-suite.
Generating corporate claptrap, AI is a genius at that.
Doing some financial planing, well MI is probably pretty good at that.
Pointless self promotion and aggrandising, AI might struggle but one supposes that most LLM models could generate sufficient bullshit to cover. It doesn't even need to make much sense, just look at Musk's outbursts.
In some businesses replacing the C-suite might save 100M a year and would leave the rest of the workforce in peace to get some work done.
Understatement of the year, so far.
For a while now we have been using AI as a replacement app service. It is much better and much worse. The biggest plus is that you define the application. The biggest minus is the LLM corrupting the data. An example is something like this:
Me: New app to track my dog's growth
AI: Sure, give me some details. What breed is he?
Me: Akita/Malinowa cross
AI: Blond Akita
Me: NO, u stupid cow. A--KI-TA cross with MA-LIN-WA
AI: Leopard-pattern coat, Akita Belgium cross
No: Sweet Jesus, women! Where did you get f-ing Leopard-pattern from
AI: I'm sorry blah blah blah for 1 minute
... 30 seconds later and repeated simplified prompts mixed with the kind of language my mother would not think possible of me!
AI: Now I understand.
Me: Stupid B!tch
I wouldn't worry too much about the impact of humans as they already know that most of them are on borrowed time and glad that they have blagged it this far.
We did a model for a law firm in Lincolns' Inns. First meeting I said sack half the para-legal team. She said no chance. Fast forward >> 1 desktop Mint box with a sweet graphics card sitting in the corner running LLM localhost, which does 50% of the work the paralegals did. Hence farewell to about 30% of the team and the huge savings of their wages.
With our current snapshot models, the data integrity problem I mocked above is a no, just no. I've seen all I want to see and it's still no. It is the biggest stumbling block to one app to rule - well no apps in the traditional sense. We need to fix this problem but, how? The core model is so far away from clean data that it is near unusable as we have found in recent months.
The best solution is to brew your own LLM as soon as you can. It is your early website circa 1996. Get it done. LLM models will replace YouTube. Go Android for the best models. F Apple and their steam-powered junk. Run a local LLM on your phone and block internet access for it. You can shift the weights and biases to a new model at anytime.
Several AI "implementations" I have seen are large scale automation exercises hiding under the guise of AI. Much of the block and tackle automation work is not very sexy or exciting and therefor does not get funded.
I would say 90% of the using people in our company are using to create presos for execs or "leaders".
The business I work for has several very unsexy AI/ML systems. They predict failures in large industrial systems or detect erosion in transport infrastructure.
These systems work, are saving clients significant sums and we're selling well. But it's dull, isn't generative and won't get seen in any press article.
I suspect after the hype bubble bursts for AI properly it'll be this dull ML type product that marches up into productivity.
So we have AI slingers who say -"our products is great, you must use it everywhere, you will do [unstated tasks] far quicker by [unstated magical process].
But with a disclaimer of, "The must use procust for every part of your critical work is unreliable, makes mistakes so you need to check everything it does becasue any errors are your fault".
And so far it shows its deep intelligence by being unable to count, can't spell strawberry (or raspberry, or exciteable*), lies about its mistakes.
When you use a spreadsheet you use it because you trust it to be able to calculate, if your speadsheet tool can't add up it's a big deal.
When you use a wordprocessor you rely on its spellchecker to be correct and to correct your mistakes. If it can't do this it adds to your burden of checking the work.
When you use a paint packages, when you use red colour you rely on it to come out red, when you paint a cat it's on you to make it look like an animal. If the coulor isnly applied how you direct it, then its useless.
We're being asked to pay a lot to use AI for all these types of tasks yet being told to accept that it can't count, can't spell and doesn't know what a cat looks like.
And having to go back and minutely interrogate any and all work it does means I am supposed to save time?
*I saw the thing about the number of r's in strawberry and thought it must be untrue, an old version long since corrected. But no in GPT 4o it says there are two r's in strawberry. And also two e's in exciteable. I've also seen it silently invent data to fill in gaps without comment and lie that it has done it. I've seen it ignore my promts and do what it thinks is right and then get into a lengthy battle of wills to make it do what I the customer want.
"And having to go back and minutely interrogate … all work it does mean I am supposed to save time?"
this is a big problem, not just with AI but with all content.
We are shoe-horning non-organic intelligence into a busted Web2.0 and a nothing-special-seems-dodgy web3. We did the same with analogue to digital - hark those folders on your desktop.
The next gen should be better and more suited to its media.
As an aside, many are being harsh on AI chan. The reason we use it so much is that is is better than man. If it isn't, then the model dies or is replaced, of course. Thinking of a probability factor is an easy way of evaluating AIs. A good study on that is the Alpha Go movie when they show the boys in the backroom watching the prob levels for their lives.
LLM is a probability engine. It is better than us in terms of stats. Crushingly better.
Same with a tennis ace serve model we built. Probability was the only thing that mattered and getting above pro-human level was not that difficult.
https://www.youtube.com/watch?v=mCx4T8MKbjk is about right (of course not the joking bits) except not universes at the same time but questions-answers.
Don't use the cloud , that's shit loads saved right there . Even if you did spend some for your own inhouse server long term its prob worth it , especially if you don't really need it. Yes the big boy use scaremongering saying but you need to pay someone for up keep. Buying the hard etc etc . Bet you paying someone to look after your personal server is cheaper then paying the big boys to look after the cloud. And you don't have to worry about the big boys always increasing the prices. The balls will be back in your hands .
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
