AI does have its uses

[Data Swami]

Before I get linched, I am not saying I will do this...but...

As @ctrlbrk has brought up AI and forum integration, and you're a bit of an AI guru @Data Swami , how could AI add value to UKBF, if at all?
I'm a bit curious, as the University of Northampton a few years back spoken to me about running a project with their AI students and UKBF, which never saw the light of day.

So i definitely think an "AI Bot" for a forum can work. Its just that it cant be some generic advice replier. Anything and all of its replies should have a very strict setup as to how it replies to someones post. Ive done this with Reddit at the moment. It doesnt automate the responses it just grabs someones post/question and researches uses my knowledge base of what i know of Data and AI and then drafts a response for me. I can then edit it get it to rewrite bits and then get it to post it. That I could see as something for a forum to use. It could be sending responses straight away but I would only do that after its proven itself to source the right info and also write in the tone of voice you want it to.

We also talked about something in relation to younger users too and where there could be a mash up of discord and a forum utilising AI to help get the conversation going around pieces and even tagging those we know have specific expertise when a question is posted.

Problem is as you know how my brain works there are far too many possibilities when it comes to this sort of thing. But the most important part is that its tailored to the UKBF audience. You dont want it talking like some douche canoe tech bro when someone asks a question or gets far too technical quoting some random specs

 

[Data Swami]

I just want to take this opportunity to clarify:

I, for one, am not against AI, per se.

Image or movie generation? Love it. AI models used to spot patterns in diseases, DNA or science in general? Big fan.

But the case I alluded to in my previous post is a classic example of how grossly misusing a LLM can quickly turn off the very people the technology is meant to serve (an update: that forum's AI model has declared that it's been directed to remove any criticism of itself from the forum).

Its turned into a reddit mod then XD

 

[Ozzy]

You dont want it talking like some douche canoe tech bro when someone asks a question or gets far too technical quoting some random specs

See, I wouldn't want it posting at all. I would however think something that helps you search intelligently through member past posts to find if a question has already been asked, then provides responses quoting what other members of said, giving you a summary of a thread, and then taking you to read the full thread.
A bit like a librarian helping you find a good book, and being able to quote you parts from books to help you decide which one is right for you...if anyone else remembers libraries?

I would not want it interacting in the threads.

 

[Data Swami]

See, I wouldn't want it posting at all. I would however think something that helps you search intelligently through member past posts to find if a question has already been asked, then provides responses quoting what other members of said, giving you a summary of a thread, and then taking you to read the full thread.
A bit like a librarian helping you find a good book, and being able to quote you parts from books to help you decide which one is right for you...if anyone else remembers libraries?

I would not want it interacting in the threads.

ahhh so more a seperate little widget to ask questions then ye that can work too. The real work there is more getting all the data and content from the forum setup nicely for the AI to be able to read it properly and then give the right summary and links to cite

 

[Paul Kelly ICHYB]

ind if a question has already been asked, then provides responses quoting what other members of said, giving you a summary of a thread

I think that is a great idea, however, if it starts quoting bad advice, it could backfire.

Maybe the first step is to list a few 'similar topics' maybe driven by keywords/tags?

Another idea might be to get AI to actually create an answer, posted by the 'UKBF AI Bot', but highlighting that this is a sample of what AI can deliver, good or bad. The grumpy contributors (whoever we they may be) could then point out the errors in the AI response, if any. This could also be used to build a UKBF LLM?

 

[Ozzy]

I think that is a great idea, however, if it starts quoting bad advice, it could backfire.

Here's a real world type example of how I feel a useful amalgamation of AI and UKBF could look. I'm currently using Claude to help research some more articles to publish on here, and here is Claude giving me some member contributions to quote within the article.

The majority view in the thread was clear. @JEREMY HAWKE , who runs a courier operation and knows this world from both sides, put it simply: you don't have the time to chase compensation, and most experienced sellers factor losses into their pricing from the start.

@Marantzdigital described an approach I think is genuinely useful: treating a small portion of every postage charge as a contribution to an imaginary loss pot. If you need to dip into it during the year, it's there. If you don't, it becomes profit. It's not glamorous, but it's honest accounting.

@Paul Norman laid out the standard costs of doing business online as a clean list: lost parcels, frivolous returns, and chargebacks. His position was that chasing each case individually would cost more in time than you'd recover in compensation, and that the right approach is to manage the aggregate rather than fight every individual claim.

@pentel drew the comparison to shoplifting in a bricks-and-mortar business. It happens. You build it into your model and you get on with it.

That's all reasonable. But it's also incomplete.

So here Claude has read a thread in full, and provided me with a summary of member contributions to that thread, which I could quote within the article, quoting their advice and tagging the members. It also helps journalists who read the forums, as Jeremy has experienced.
I have a bespoke Project and rulebase configures in Claude to do this, as well as various brand guidelines, a tone and voice and personal profile document on me, and guidance rules on how to read UKBF and what to ignore and not to ignore. So it isn't perfect yet, but has taken a considerable amount of time to configure to where it is so far.

A variation on this that asks as a knowledgebase search and summary tool, that someone can query, ask questions of, and deep dive into, and it always quotes actual comments and not make stuff up, would be useful. That I feel would be quite a good addition to this site, especially if it also includes the published articles too.

 

[Ozzy]

(an update: that forum's AI model has declared that it's been directed to remove any criticism of itself from the forum)

That is just...bad.
Criticism is where you where to make your best improvements; of course so long as it is constructive and comes from a place of wanting improvement. Criticising just for the same of trolling, that's different

 

[Paul Kelly ICHYB]

@Ozzy that looks good, as a post creation tool.

Re the knowledgebase, if people can search for answers, could this reduce asking questions/posting? The combination of the 2 (topic creation with a knowledge base post) could be good.

 

[Data Swami]

Here's a real world type example of how I feel a useful amalgamation of AI and UKBF could look. I'm currently using Claude to help research some more articles to publish on here, and here is Claude giving me some member contributions to quote within the article.



So here Claude has read a thread in full, and provided me with a summary of member contributions to that thread, which I could quote within the article, quoting their advice and tagging the members. It also helps journalists who read the forums, as Jeremy has experienced.
I have a bespoke Project and rulebase configures in Claude to do this, as well as various brand guidelines, a tone and voice and personal profile document on me, and guidance rules on how to read UKBF and what to ignore and not to ignore. So it isn't perfect yet, but has taken a considerable amount of time to configure to where it is so far.

A variation on this that asks as a knowledgebase search and summary tool, that someone can query, ask questions of, and deep dive into, and it always quotes actual comments and not make stuff up, would be useful. That I feel would be quite a good addition to this site, especially if it also includes the published articles too.

Claude is defo good at reading 1 thread on its own. Its the numerous threads that is where it would be better for it to read a proper data source than browsing the site. And the rules need to be easy to update too (learnt that the hard way accidently baking them into a process once XD) A scoring system is usually good with different types of categories to score on in trying to work out quality of content/contributions. And then its making sure it speaks how you want it to in the summary rather than the generic AI summary replies

 

[Data Swami]

@Ozzy that looks good, as a post creation tool.

Re the knowledgebase, if people can search for answers, could this reduce asking questions/posting? The combination of the 2 (topic creation with a knowledge base post) could be good.

i think if you track the data and types of queries it could help to identify common questions that keep getting answered for then to create a bigger post on that specific topic so it covers all aspects of it in one place and keep the conversation going still

 

[ctrlbrk]

AI Models Can’t Agree on Basic Facts Most of the Time, Study Shows

Ask five of the world's most advanced AI systems whether a statement is true, and two-thirds of the time, at least one will give you a different answer. That's the finding of a new study published this month by researcher Kosta Jordanov at Lenz Research.

The study gave GPT-5.4, Claude Opus 4.7, Gemini 3 Pro, Gemini 3 Pro with Search, and Sonar Pro the same 1,000 real-world fact-check claims submitted by actual users. The models had to pick one of four labels: true, mostly true, misleading, or false.

[...]

This matters because people are increasingly turning to AI systems for fact-checking. If you paste a claim from a news article into ChatGPT, Claude, or Gemini, you might get three different answers. Which one do you trust?

AI companies love to tell you their models are getting more accurate. They publish benchmark scores showing steady improvement. But the Lenz study tested these models on the kind of jagged, ambiguous claims that real humans actually argue about—and found that the models argue too.


Article here.

 

[DontAsk]

LLMs believe false statements even after explicit warnings that they’re false​

https://arstechnica.com/ai/2026/05/...en-after-explicit-warnings-that-theyre-false/

 

[NickGrogan]

LLMs believe false statements even after explicit warnings that they’re false​

https://arstechnica.com/ai/2026/05/...en-after-explicit-warnings-that-theyre-false/

Just like people. Dont forget most AI is trained on Reddit posts.

 

[DontAsk]

Never use it for anything you might need to rely on.

Asked Google "can pension tax relief be claimed against dividend tax" and it said "... total relevant UK earnings (which do include dividends)."

Asked it "do relevant earnings for pension contributions include dividends" and it replied "No, relevant earnings for pension contributions do not include dividends."

Unfortunately, AI is trained on good and bad data and the answer you get will always depend on how you phrase the question.

Always do your own research.

 

[ctrlbrk]

Never use it for anything you might need to rely on.

This is my experience too, which brings me back to:

What's the point of these LLM if we can't rely on them?​

 

[fisicx]

What's the point of these LLM if we can't rely on them?

Data analysis. Coding. Trends. Reworking prose.

What’s it’s not very good at is answering questions. Traditional search is still the better option (if you ignore the AI summary).

 

[ctrlbrk]

Data analysis. Coding. Trends. Reworking prose.

I take your point but even in these tasks it needs a lot of hand-holding doesn't it?

Reworking prose is probably the most reliable answer, but I wouldn't trust current LLMs doing any data analysis or trends. It's always a case of the LLM telling me what it thinks I want to hear, rather than serving as an objective tool.

I've tried coding and, while some output is okay, in most cases the results leave a lot to be desired.

 

[fisicx]

I take your point but even in these tasks it needs a lot of hand-holding doesn't it?

Yes, but once you get the hang of writing prompts it's can save a load of time.

We have two spreadsheets (thousands of rows, hundreds of columns) with a lot of data. ChatGPT did a really good comparison analysis and found some outliers we could then address. Would have taken hours to do manually.

I've tried coding and, while some output is okay, in most cases the results leave a lot to be desired.

But claude is brilliant at regex. It's also very good at debugging. And I use it a lot for sanitizing and escaping. It means I can focus on the fun bits of coding and let the AI tools do all the drudge work.

The people who are getting the most benefit are crims. They have discovered LLM tools are brilliant at phishing and social engineering and can get people to hand over passwords and keys far more efficiently and effectively than the old iffy email campaigns.

 

[gpietersz]

I've tried coding and, while some output is okay, in most cases the results leave a lot to be desired.

its a matter of using it appropriately. It is good at some things such as writing unit tests, but of course you review what it does. It can help with debugging and code review. It can search through documentation very effectively and find multiple sources. It can do small well specified tasks very well/

I was quite amused a few days ago when Claude used one of my own blog posts in response to a question though.

It means I can focus on the fun bits of coding and let the AI tools do all the drudge work.

Agreeing, and fun bits wider programming, which is a lot more than coding.

Its quite interesting to look at what really good programmers say about how they use it. This is one I read recently: https://antirez.com/news/164

 

[NickGrogan]

Just remember its a text generation engine, not actually thinking and its fine.

I use chatgpt/gemini with DuckDB - it writes the queries, but the database holds the actual data. So the data is always real and AI is focused on writing SQL which it is very good at.

Its will forget field names now and then, but give it the DB describe output and its sorts itself out.

Much easier to combine 7/8 separate databases if someone else is writing the SQL.

Just have to remember to tell it what the data is and how it should combine.

 

[fisicx]

The other thing to remember is these tools can’t do maths. Ask it to do sums and it often goes awry. Ask it to look at trends and you get far better results.

 

[Paul Kelly ICHYB]

I think it boils down to, at the moment, AI is best for automation and repetitive tasks and, maybe, simple decisions/activities.

Using it to manage your accounts, SM research & publishing, video/graphics or closing sales, probably, for most, should be where is starts and stays for a while.

 

[gpietersz]

@Paul Kelly ICHYB I think it is subtly different from that. For automation I would use it to write scripts that do the automation, rather than having it do it directly. What it is good at are tasks that can be clearly defined, and for which success and quality can be verified.

@fisicx but they will do a good job of writing code to do maths. You do not ask it to do a multiplication, instead you ask it to write a calculator.

 

[louis292]

Nice! That's the thing about AI - don't expect a perfect finished solution. For me it is never perfect, but can provide a great base to push on from. One of the key things is having a decent prompt - ask it for slop and you will get slop! The more specific the prompt the better result.

Spot on. It’s all about the input you give it. Garbage in, garbage out.

 

[fisicx]

@fisicx but they will do a good job of writing code to do maths. You do not ask it to do a multiplication, instead you ask it to write a calculator.

Exactly that. Can’t do numbers but know how to do the calculations.

I’m working on an energy project for a client in Finland (not microgrids!). I gave it a set of datapoints and told it what I wanted and it delivered the code. Dropped it into the function and it all worked. Needed a bit of tweaking as there are some weird energy outliers in Helsinki but that was just an additional filter.

 

[DontAsk]

As per an amusing Instagram post "AI didn't make us smarter. It just gave idiots a superpower".

You can still do all the bad stuff, but faster.

 

[Ozzy]

AI didn't make us smarter. It just gave idiots a superpower

I should have used this as the subject line of today's newsletter!

 

[seabro]

I seems it's more cool to criticise AI these days but it has allowed me to do years of work in months, I am definitely a fan.

Initially I asked it if it could extract data from emails and add it to a website templates, to create new web pages from form submissions. It could of course. Then I got it to create a stats/click tracker report for the business profiles (it's a directory website) and auto mail it out to businesses monthly. Then I asked it perform SEO analysis and implement the recommended changes. (Which ultimately lead to a full web stack migration - which it also implemented).

Then I asked it to perform outreach for the same web directory, contact niche businesses and ask them if they would be happy to be added, then check the email and process the affirmative replies, gather data about the company, create a profile for them, then send them a link to their profile in a welcome message. All in about 10 mins of them agreeing, multiple customers expressed amazement how fast 'I' did it.

Then I setup a review request and chase system for the work I do. It very politely asks for a review, periodically checks if one has been left and if it has, it makes a public reply and sends them a thank you email. If one hasn't been left, it sends a gentle reminder. In my experience, people rarely leave a review after a single ask and manual checking and chasing is easy to overlook and time-consuming.

I also update Facebook business pages every day for multiple businesses, and Google Business Profile several times a week.

Once of my favourite projects is a for client with a very high ticket business (6, 7 or 8 figure transactions). It performs around 60 searches per day, identifies leads all over Europe, enriches the data and sends a very detailed 'market intelligence' report to the sales director. His sales process is long but he is over the moon with the data he is getting. I have been working on further automations for the same co. I have also successfully duplicated this approach to uncover and present leads for service businesses in the UK and the data looks good. It's nothing the right person couldn't do given the time, but I think the level of research would be a full time job.

This is just a selection of the automations I am doing, the capabilities continue to blow my mind.

It has cost me a lot of time and money on API's and tools over the months but it's definitely been worth it.

I agree, AI definitely has it's uses!

 

[Data Swami]

I seems it's more cool to criticise AI these days but it has allowed me to do years of work in months, I am definitely a fan.

Initially I asked it if it could extract data from emails and add it to a website templates, to create new web pages from form submissions. It could of course. Then I got it to create a stats/click tracker report for the business profiles (it's a directory website) and auto mail it out to businesses monthly. Then I asked it perform SEO analysis and implement the recommended changes. (Which ultimately lead to a full web stack migration - which it also implemented).

Then I asked it to perform outreach for the same web directory, contact niche businesses and ask them if they would be happy to be added, then check the email and process the affirmative replies, gather data about the company, create a profile for them, then send them a link to their profile in a welcome message. All in about 10 mins of them agreeing, multiple customers expressed amazement how fast 'I' did it.

Then I setup a review request and chase system for the work I do. It very politely asks for a review, periodically checks if one has been left and if it has, it makes a public reply and sends them a thank you email. If one hasn't been left, it sends a gentle reminder. In my experience, people rarely leave a review after a single ask and manual checking and chasing is easy to overlook and time-consuming.

I also update Facebook business pages every day for multiple businesses, and Google Business Profile several times a week.

Once of my favourite projects is a for client with a very high ticket business (6, 7 or 8 figure transactions). It performs around 60 searches per day, identifies leads all over Europe, enriches the data and sends a very detailed 'market intelligence' report to the sales director. His sales process is long but he is over the moon with the data he is getting. I have been working on further automations for the same co. I have also successfully duplicated this approach to uncover and present leads for service businesses in the UK and the data looks good. It's nothing the right person couldn't do given the time, but I think the level of research would be a full time job.

This is just a selection of the automations I am doing, the capabilities continue to blow my mind.

It has cost me a lot of time and money on API's and tools over the months but it's definitely been worth it.

I agree, AI definitely has it's uses!

Having been embedded in the world of Data and AI i can see it from both sides.

AI side of things we have the charlatans and the massive amount of focus its getting and just like digital transformations, data etc its being pushed hard in really bad ways. There are also some bad actors in respect of expecting AI to solve their employment problems and be able to bin off their staff which the likes of Microsoft etc have found its bitten them on the arse.

But then I also see how damn powerful it can be. I dont know if many here have been using Hermes Agent or the likes of openclaw etc. They have been an absolute revelation for my use of AI. Its like Cursor on steroids with what I am able to do and create. Plus making it sensible automations too not just having claude code do everything and burn through tokens I can make it sensible with python scripts for easy automation and only add the LLM into it when I really need it.

However i know that something like Openclaw or Hermes is well above the knowledge and know how of most SMEs and I would not advise them to try and set it up themselves. Even with my experience i still find ways of burning myself especially at the moment where I have found a bug with token leakage while i use Discord as the chat interface

 

[DontAsk]

Meanwhile large companies are rowing back on AI use as it's cheaper to employ humans.

 

[Data Swami]

Meanwhile large companies are rowing back on AI use as it's cheaper to employ humans.

Well thats all in how they tried to implement it. Just like all the failed digital transformations and also in part the way that wage growth has been suppressed too. Throw bodies at it is also another way of failure especially that old Capita model