Geordin Hill-Lewis won't grant Cape Town departmental management additional budget for salaries unless they implement AI

[RedViking]

The big problem behind AI is that the policies and processes in using AI are weak. This is a management problem, starting at the top.

Also, nobody seems to know how to solve issues created by AI other than throwing more AI at the problem.

People aren't trained. Neither do they know how to prompt or vet the output. AI users assume that AI is the truth bearer. I have seen these issues prop up in reporting and all other documentation. I can't speak to all applications. Frameworks suck, and most employees are given prompt templates or that what is limited to the chatbot build into whatever suite they are using.

We have even seen legal cases being thrown out due to unskilled AI use.

Sure, blame hallucinations, but it is a human problem that lacks due diligence.

AI is such a common word in business now. So much garbage is automated.

I do pity the city, but then again, perhaps AI will hallucinate less than the politicians.

Seeing as GHL is named in the headline I do expect him to respond in this all being misleading as per usual.

I don't disagree , but the people running this country is far more useless.

 

[Sideshow Bob]

I don't disagree , but the people running this country is far more useless.

GHL is not running the country though. Well, not yet.

 

[Sapphiron]

unless they implemented artificial intelligence (AI) in their operations.

It should read:

Unless they have showed efficiency improvements in their operations using automation tools.

 

[vigras rojara]

I don't disagree , but the people running this country is far more useless.

AI just enables uselessness at scale.

 

[PlanesTrainsAutomobiles]

Pushing LLM slop for the sake of having LLM slop with no clear directive or adoption plan - that has clear, beneficial, deliverables - is a time-waste of note and any management twerp doing so needs to be smacked until they stop.

There is nothing magical in LLM land. Forcing its use - or as the saying now goes, "tokenmaxxing" - is destructive noise for no good reason.

Why do you *not* have an issue with this?

How does this not improve the speed and quality of certain jobs?

 

[Sideshow Bob]

I read an article that mentioned that only around 25% of enterprises currently track token usage in real time, and that cost shocks are happening all the time.

Although the article was probably AI generated and accordingly wildly inaccurate.

If I have to guess, token cost is going to decide future of AI. I hope I will be out of workforce by then. Some of the generated code is very obscure but then I am old school.

 

[rustypup]

How does this not improve the speed and quality of certain jobs?

Please describe how you would achieve this, what model/s or "agents" would be used, what metrics would serve to measure outcomes as succcessful......

AI - as in "artificial intelligence" - as it is being currently punted is BS. We don't have "AI". We have raw math. LLMs don't reason. They are incapable of doing so. They echo the input they were trained on in interesting ways, based on the prompt provided, the vector store/s that would see said prompt play the drunk-walk game through and various other bits and bobs, including other LLMs trained on more focused data sets.

Explain how any of this is going to "speed up" anything, especially in service delivery functions.

Successful LLM implementations are focused at solving specific problems in specific ways. Simply throwing LLMs at something and expecting something magical in outcome is .... delusional thinking.

Without a detailed problem statement and a fair stab at scoping use cases, the only guaranteed outcome is wasteful expenditure and failure.

 

[PlanesTrainsAutomobiles]

Please describe how you would achieve this, what model/s or "agents" would be used, what metrics would serve to measure outcomes as succcessful......

AI - as in "artificial intelligence" - as it is being currently punted is BS. We don't have "AI". We have raw math. LLMs don't reason. They are incapable of doing so. They echo the input they were trained on in interesting ways, based on the prompt provided, the vector store/s that would see said prompt play the drunk-walk game through and various other bits and bobs, including other LLMs trained on more focused data sets.

Explain how any of this is going to "speed up" anything, especially in service delivery functions.

Successful LLM implementations are focused at solving specific problems in specific ways. Simply throwing LLMs at something and expecting something magical in outcome is .... delusional thinking.

Without a detailed problem statement and a fair stab at scoping use cases, the only guaranteed outcome is wasteful expenditure and failure.

In local govt, there are tons of uses of even LLMs, let alone proper AI.

Chatbots to handle FAQs, translations, and drafting communications in plain language rather than bureaucratic speak; legal work like policy reviews, drafting bylaws, compiling agendas, minutes, tracking resolutions; HR work like shortlisting against job ads, drafting job profiles; spotting patterns in data.

I can go on and on.

 

[rustypup]

Chatbots to handle FAQs

Poorly and very easy to jump guard rails. If you need something halfway meaningful you would need to spend time and effort training a model aimed at this, only to discover that Google does it better.

translations

Thousands of services already available delivering this with zero issue

drafting communications in plain language rather than bureaucratic speak

Thousands of services available delivering this with zero issue.

legal work like policy reviews

So many points of failure without human input you end up doing the work anyway - We've played with multiple of these and they all suck uniformly.

drafting bylaws

See, you're still buying the delusion these things are capable of reasoning. They are not.

compiling agendas, minutes, tracking resolutions

No argument. They do this no worse than garbage humans. Trusting that they are perfectly accurate is a mistake. But go ahead.

HR work like shortlisting against job ads, drafting job profiles

They suck at this so badly it's gotten certain corps into hot water. Being too lazy to vet job applicants is very much the HR dream... except these LLMs end up being somewhat biased and not all that great at hitting the mark on auto assessment requiring - once again - human oversight to verify.

Sadly humans are lazy shiats who accept LLM output on faith and end up creating endless drama.

spotting patterns in data

Unless specifically trained on data, no. Even when trained, accuracy is closer to 90%. Again, no worse than humans but humans can correct for noise or intuit informance not available in the data set/s

Do yoou understand *anything* about how LLMs function?

 

[EmmanuelGoldstein]

Poorly and very easy to jump guard rails. If you need something halfway meaningful you would need to spend time and effort training a model aimed at this, only to discover that Google does it better.


Thousands of services already available delivering this with zero issue


Thousands of services available delivering this with zero issue.


So many points of failure without human input you end up doing the work anyway - We've played with multiple of these and they all suck uniformly.


See, you're still buying the delusion these things are capable of reasoning. They are not.


No argument. They do this no worse than garbage humans. Trusting that they are perfectly accurate is a mistake. But go ahead.


They suck at this so badly it's gotten certain corps into hot water. Being too lazy to vet job applicants is very much the HR dream... except these LLMs end up being somewhat biased and not all that great at hitting the mark on auto assessment requiring - once again - human oversight to verify.

Sadly humans are lazy shiats who accept LLM output on faith and end up creating endless drama.


Unless specifically trained on data, no. Even when trained, accuracy is closer to 90%. Again, no worse than humans but humans can correct for noise or intuit informance not available in the data set/s

Do yoou understand *anything* about how LLMs function?

With due respect I know what LLMs are, nothing more than a ton of math applied to large datasets.
I use it as a tool, I let it automate things that would take me hours to do. The input is defined properly, the expected output is defined properly and it does the job pretty well- no different to say, an effect used in Photoshop. I have seen evidence in creative industries where it is used to fix images up quickly. I do not use it to create things from scratch because 99% of the time that comes out wrong... However, if I do the basic hard graft and tell it to apply effects such as thicken lines, smooth them, smooth curved lines, etc.. it does that astonishingly well. But then again we can see why... Mathematics- these things have been in software for a long time.

It is also good at testing code, I do not go for the vibe coding scene, because there is an embarrassing incident of that at the day job, which fortunately does NOT have my name on it. I use it to fix things in code, that's about it.

 

[PlanesTrainsAutomobiles]

Poorly and very easy to jump guard rails. If you need something halfway meaningful you would need to spend time and effort training a model aimed at this, only to discover that Google does it better.


Thousands of services already available delivering this with zero issue


Thousands of services available delivering this with zero issue.


So many points of failure without human input you end up doing the work anyway - We've played with multiple of these and they all suck uniformly.


See, you're still buying the delusion these things are capable of reasoning. They are not.


No argument. They do this no worse than garbage humans. Trusting that they are perfectly accurate is a mistake. But go ahead.


They suck at this so badly it's gotten certain corps into hot water. Being too lazy to vet job applicants is very much the HR dream... except these LLMs end up being somewhat biased and not all that great at hitting the mark on auto assessment requiring - once again - human oversight to verify.

Sadly humans are lazy shiats who accept LLM output on faith and end up creating endless drama.


Unless specifically trained on data, no. Even when trained, accuracy is closer to 90%. Again, no worse than humans but humans can correct for noise or intuit informance not available in the data set/s

Do yoou understand *anything* about how LLMs function?

You believe you have an answer for it all. Nobody said LLMs are without fault. Try the human propensity for mistakes. It compares terribly.

Let's do it your way. I mean, it is not as if SA can fall any further behind, right?

 

[konfab]

One of our clients started on this path recently and got a rude awakening when tokens were exhausted in no time. LOL. Only couple of people (architect & some senior devs) have access to consume tokens now and that too after approval from top.

This token thing is going to be costly in my opinion, even if it seems cheap in the beginning.

Nope.

Every single new model that has been released has gotten cheaper for the same task.

If you look at 5.6 Luna, which is the dirty cheap model that you use for things like generating the name of a session, it is on par of intelligence (that is the ability to do stuff), with the previous generation, whilst being a 10th of the cost.



Why people are spending so much on models is they are doing far more difficult things with them.

 

[konfab]

AI just enables uselessness at scale.

AI is a productivity multiplier. For a diligent person who wants to get their work done, they will do more far more work.
If you give it to someone who is already a net drain, they will produce far more slop.

 

[Sideshow Bob]

Nope.

Every single new model that has been released has gotten cheaper for the same task.

If you look at 5.6 Luna, which is the dirty cheap model that you use for things like generating the name of a session, it is on par of intelligence (that is the ability to do stuff), with the previous generation, whilst being a 10th of the cost.

View attachment 1929096

Why people are spending so much on models is they are doing far more difficult things with them.

Sure, I am old school. Enjoy the AI boom.

 

[konfab]

Poorly and very easy to jump guard rails. If you need something halfway meaningful you would need to spend time and effort training a model aimed at this, only to discover that Google does it better.


Thousands of services already available delivering this with zero issue


Thousands of services available delivering this with zero issue.


So many points of failure without human input you end up doing the work anyway - We've played with multiple of these and they all suck uniformly.


See, you're still buying the delusion these things are capable of reasoning. They are not.


No argument. They do this no worse than garbage humans. Trusting that they are perfectly accurate is a mistake. But go ahead.


They suck at this so badly it's gotten certain corps into hot water. Being too lazy to vet job applicants is very much the HR dream... except these LLMs end up being somewhat biased and not all that great at hitting the mark on auto assessment requiring - once again - human oversight to verify.

Sadly humans are lazy shiats who accept LLM output on faith and end up creating endless drama.


Unless specifically trained on data, no. Even when trained, accuracy is closer to 90%. Again, no worse than humans but humans can correct for noise or intuit informance not available in the data set/s

Do yoou understand *anything* about how LLMs function?

Do you think data scientists reads a bunch of numbers and magically comes about a conclusion?

They don't. They write a bunch of code that tests hypothesis against data and checks whether the hypothesis is correct or not. So for example. Is there a correlation between X and Y, is there a periodicity, what does the distribution look like, what do the regression coefficients look like etc.

All of that is easily within what an LLM can do. Even the retarded ones like Gemini.

 

[vigras rojara]

AI is a productivity multiplier. For a diligent person who wants to get their work done, they will do more far more work.
If you give it to someone who is already a net drain, they will produce far more slop.

Exactly, hence my comment.

Sadly, in day to day life we experience much more of the latter, than the former.

 

[dualmeister]

Someone lost the plot.

The DA lost the plot the moment they jumped into bed with the ANC.

 

[ToxicBunny]

Exactly, hence my comment.

Sadly, in day to day life we experience much more of the latter, than the former.

For me, the problem is that Management think its a universal "productivity multiplier" so everyone must do AI!

For those who know how to use it, or where it will be useful it can be a productivity multiplier, but for those who don't the productivity drain would likely outweigh any multiplication in other areass.

 

[konfab]

Exactly, hence my comment.

Sadly, in day to day life we experience much more of the latter, than the former.

Just make producing continuous slop a fireable offence.

 

[vigras rojara]

Just make producing continuous slop a fireable offence.

Amen, brother!