using AI for pension projections

SouthCoastBoy

https://forums.moneysavingexpert.com/discussion/comment/82126886#Comment_82126886

My primary objective is to ensure I can live my current lifestyle for the rest of my life. i am also keen on leaving an inheritance, which is why i havent considered an annuity, if there is an iht bill so be it, although i do plan to gift regularly if my pot can sustain it.

horsewithnoname

https://forums.moneysavingexpert.com/discussion/comment/82126886#Comment_82126886

the great thing about being dead is that you no longer pay tax, but your estate may be liable for IHT if you are lucky enough to have sufficient money

RogerPensionGuy

SWR, we can pick any % and maybe adjust withdrawal rate as appropriate.

We have enjoyed a 17 or 18 years of a bull run overall with a few big corrections and these dips corrections have been pretty dramatic, but the recovery time has been very short in my opinion.

Picking a SWR % with a dynamic adjustment mechanism does feel okay in the most part in my head.

However I'm a glass quarter full person and I planned/plan for a real big correction/crash like history shows us that is very slow to recover and this to occur in my gogo years whereby I would probably stop DC SIPP withdrawals awaiting recovery.

AI is certainly helpful to show examples of the past and this can be helpful in planning, but AI doesn't know the future.

I used AI for information many times on SWR & wanting good hard cash flows in the gogo years and beyond until 100 years old in fact and it agreed due my situation and not bothered if I/wife leave any legacy that doing annuities or PLAs with current annuity rates was very suitable.

Reference legacy, annuities I saw can have guarantee periods of up to 30 years and these opinions maybe useful to some people who want to leave legacies.

I hope the markets are the same or similar for the next 17 or 18 years, but I'm guessing they won't be.

On a general note they say people with various or many income streams tend to pan out well, having most income stream from a DC SIPP in drawdown certainty didn't feel good to me.

Cheers Roger.

Aylesbury_Duck

I've used AI to generate lines of thinking, rather than rely on its recommendations. I've caught a few obvious errors, and it will occasionally forget really important contextual information (e.g. a reasonable LGPS DB kicking in at a certain date) so it's not reliable, but it is helpful to bounce ideas around and to check my own assumptions and spreadsheets. It has picked up one or two of my own spreadsheet errors, too.

Linton

Pension projections are simply a mathematical consequence of a set of assumptions so the best tool in my view is a spreadsheet not an AI. Those assumptions must be ones that you are happy with, not ones that an AI has chosen for you.

AI is more useful for "open" questions and coming up with ideas that you may not have considered.

Cobbler_tone

I use AI that much now (for teaching myself things/analysis/recommendations etc) that I am not sure how I got through life before AI! I think that was pretty inevitable, like most tech advances. If you use it a lot, you soon wise up to where it is reliable or where you need to do further work elsewhere, or really need to interrogate to outputs, or just know that it is plain wrong!

LHW99

I find AI's useful to re-write an existing section of text to accomodate a specific focus, but as said above, for detailed research into a niche subject (which is what I frequently tend to do), not so much.

RogerPensionGuy

https://forums.moneysavingexpert.com/discussion/comment/82128130#Comment_82128130

Just to agree, AI often gives out bad information, nearly schoolboy errors.

I reply saying the errors and it just politily says, O yes of course blah blah.

Once I correct it, it never gives out the same error on that exact item, but does with close by information, now I've corrected it a good many times it produces data stuff I more like using to fill my fag package with for my reference.

It obvious knows the real information, just happy doing its own research on people collecting all the data it wants and needs, it's interesting.

Nebulous2

I used AI last year to plan a route through France. I said how long I had to get to the channel, what kind of attractions I was interested in, what facilities I wanted in each town I stopped in and how far I was prepared to drive each day. It was extraordinarily good, and came up with a great route and some very good campsites that matched what I had asked for.

This year, buoyed by my success last year, I wanted to improve my fitness. I gave a family member my car for a few weeks, had much more free time than I usually have, and was determined to climb some hills.

We live on a long bus route, which extends some 35 miles in one direction, and 50 miles in the other direction.

I told AI that I wanted to climb a hill by using that bus. I said I was willing to walk up to a mile from the bus route to the base of the hill, climb for around 3 hours, then return to the bus. I live in Scotland, the whole place is littered with hills.

It was absolutely hopeless. It told me a hill was 1/2 a mile from a bus stop when it was 3 and 1/2. When challenged it doubled down, and said there was a smaller hill which was part of the same walk, and I should count my climb as starting from there. It told me to get off the bus in a village, walk up a specific street and turn left, which led me directly to the sea. When challenged it agreed a left turn was in the wrong direction, but said I would find a gate that led me directly to the hill. It insisted that a town 15 miles from the bus route was on the bus route, and had a lovely walk up a hill. Again when I questioned this it told me to take my bus 30 miles, change buses and go another 20 miles on another bus to get to the town. There was no recognition of the parameters I had set - namely a hill less than a mile from a specific bus route.

I've no idea why this happened. Why an earlier LLM was so good at navigating through France, or why a more recent one was so hopeless at suggesting a hill to climb in Scotland. Possibly there is more data on routes through France, and decent campsites for it to draw on. Hill climbing in Scotland isn't exactly a minority activity though, we have a lot of people come her to climb our hills and write about it.

The problem is that people may have taken it at face value. I knew some of the hills recommended, knew they were inaccessible for my purposes. Knew if I walked up a specific street and turned left I was on the wrong side of a main road and was heading directly for the sea.

Dead_keen

I've no idea why this happened. 

It's about giving the AI the right tools for the job.

Small LLM's are good with language but know few facts. So a small one may know about some of the towns in France but not about all the towns there. Large LLMs are generally better with language and know lots more facts, including all the towns in France and how to get from one to another. Where a fact is not known, an LLM may make something up that sounds plausible.

You can get a small LLM to know all the towns in France by giving it a data source (e.g. wiki or a map of France) that it can interogate. So a bit like me googling "Towns in the Pyrenees", "How do I get to Loundenvielle".

The detail needed to plan a hilly bike ride won't be in small or large LLMs (unless you plan on riding Ventoux or somewhere else famous for being ridden). So you need to give your favourite LLM a data source and tell it how to interpret it. If I wanted to plan a cycling route around Scotland and a bus route, I'd give Claude Code access to my computer's Chrome browser and tell it to go ahead and use suitable tools to access, for example, your bus route, Strava, Veloviewer, Google Maps, cycling route reviews and so on. Then I'd get it to create me route guides that I can read on the screen to see if they make sense. Once I am happy with them, I'd get it to create the route or GPX files for my bike computer and transfer them to it.