September 13, 2026 • 10 mins
You've put in the work on your prompts. You know how to give AI the right context, set the right tone, ask the right questions. But then, mid-conversation, something shifts. The AI starts making things up, forgetting what you told it five minutes ago, contradicting itself.
In this How I AI episode, Neo and I unpack what is actually happening when AI gets unreliable mid-conversation, including how to read the warning signs that tell you it is time to close a thread and start fresh.
What you'll learn:
- Why AI gets unreliable mid-conversation, and what's actually going on
- What the 60% rule is and why it matters
- Which file types quietly eat through your context window
- The warning signs that it's time to close a thread and start fresh
- Practical AI tools for productivity and focus
- Real-world AI workflows used by high performers
- How to use AI at work without burning out
- Smart shortcuts for managing time and mental load
Connect with Neo Aplin on LinkedIn (https://www.linkedin.com/in/neoaplin/) and via inventium.ai (https://inventium.ai), where he leads Inventium's AI training and upskilling work with organisations and teams.
How I AI with Dr Amantha Imber and Neo Aplin explores how high performers are using AI at work to boost productivity, make better decisions and reduce overwhelm.
Want more practical ways to put AI to work? You can now follow How I AI separately on Spotify or Apple Podcasts so you never miss an episode.
My latest book The Energy Game is out now. You can order a copy here: https://amzn.to/48ID29M
Connect with me on the socials: Linkedin (https://www.linkedin.com/in/amanthaimber)
Instagram (https://www.instagram.com/amanthai)
If you are looking for more tips to improve the way you work and live, I write a weekly newsletter where I share practical and simple to apply tips to improve your life. You can sign up for that at http
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
Imagine you're deep into a long chat with AI. And
it starts getting things wrong. Forgetting something you told it
a mere five minutes ago.
Speaker 2 (00:13):
Making things up.
Speaker 1 (00:15):
And you can't tell if it's you, your prompt, or
maybe just.
Speaker 2 (00:19):
The AI is having a bad day.
Speaker 1 (00:22):
But what's usually happening is that the chat thread has
run out of short-term memory. And once that memory gets
to about 60% full, the AI starts becoming less reliable.
In this episode, Neo and I talk through how to
know when it is time to start a fresh chat thread.
Speaker 2 (00:44):
You will learn the warning science that a thread has
passed its use by date.
Speaker 1 (00:49):
Why PDFs chew through your context window faster than almost
anything else.
Speaker 2 (00:53):
And what to upload instead.
Speaker 1 (00:56):
By the end of this episode, you'll know exactly when
to close a thread and start again.
Speaker 2 (01:02):
So that the AI stays sharp. Welcome to How I
AI with me, Dr.
Speaker 1 (01:12):
Amantha Imber and Neo Applin, head of Inventium AI. Each episode,
we share one practical way to use AI better at
work and in life. No fluff, no tech jargon, just
things you can use straight away. Neo, today we are
going to be talking about something that might seem on.
Speaker 2 (01:31):
The surface quite basic, but we see.
Speaker 1 (01:36):
Hundreds and hundreds, if not thousands of people get this
wrong when we are training people in workshop rooms. And
that is, when should you open or start a new
chat thread? So, Neo... What are your top tips on
how do we know when it's time to start a
new chat thread?
Speaker 3 (01:56):
The biggest tip is if the AI seems to be
doing strange and wacky things, then maybe it's time to
start a new chat thread. Because going back a step,
the reason is AI doesn't have an infinite memory for
each of your chat threads. They've got a technical thing
called a context window, which is it can hold a
certain number of tokens or words in an entire chat thread.
(02:18):
And that's everything you say, everything it says, all the
documents that both of you are creating, and also every
thought that it might have behind the scenes that it
doesn't show you kind of fills up this short-term memory.
And so once the memory gets full, or at least
starting to get full, science is saying about 60% full,
then you'll find that hallucinations happen more often and the
(02:38):
AI is more likely to get things wrong. So if
you're finding it's making up stuff, it's getting things wrong,
ask yourself, have I been talking with this chat in
this one chat thread For an awful long time, maybe
it's time for me to move off to another chat thread.
Speaker 1 (02:52):
So when you say 60% full, how do we know
what the total amount is? What does that look like?
And is it different across the four major AI platforms?
Speaker 3 (03:03):
And wouldn't it be amazing if they had a bar
chart going, awooga, awooga, you've just hit 60%. I'm about
to start to lie to you. No, it doesn't do that.
Isn't it quite annoying? each of the different four platforms
does have a different context window. The short-term memory thing
is different in each, and it's changing over time. I'm
hoping by the time you hear this podcast, you'll go, oh, wow,
(03:24):
it's even better than it was yesterday. They don't generally
come out and tell you how big these ones are,
but you can even ask your AI, how big is
your context window? And then even say, how many words
would that kind of equivalent be. For some of them,
I've got more of like a 200-odd A4 pages worth
of text. And it's kind of hard because is it
(03:44):
A4 pages with double-sided? Is it like wall-to-wall letters and numbers?
It's kind of hard to kind of get that in
our head. So I don't mind the term of 200
pages for some people. And others, if you're on Gemini,
they're just rocking it. You're like 1,000 pages worth of
text or something like that. And so each of them
are different. And these things change over time. But there
(04:05):
are really Two big principles here, which is if that
limit is getting full, and they've done a lot of
science on this, which is once you start hitting about 60%
of this context window getting full, then it's more likely
to make stuff up, hallucinate, and get things wrong. Either
forget something you've asked. or misinterpret something in the entire
(04:25):
conversation before. And so it's a good idea then to
start a brand new chat thread. Obviously, if you're 100%
full on that context window, every new word you put
at the bottom means the word at the top is
going to drop off what it can say. And so
don't go forever, which is why in all our trainings
and here in our podcast, we always say start a
new chat thread for pretty much every time you come
(04:45):
to AI and hopefully you won't even hit this problem
in the first place.
Speaker 2 (04:49):
Now, another thing on context window.
Speaker 1 (04:52):
So you say that maybe at the time of recording,
which is September 2026, it's maybe about 200 pages of text,
but also each page and each file type is not
created differently. So for example, something that I hit yesterday,
I was trying to create a mega project around health.
(05:12):
I was uploading all these different health documents that were
all PDF documents. And you heard me whinging to you
that I had hit my context limit.
Speaker 2 (05:23):
Which I rarely do because I'm pretty careful.
Speaker 1 (05:25):
About how I use AI and, you know, just managing
what is going in so that I don't hit those limits.
Speaker 2 (05:34):
But I did because it was all PDF documents.
Speaker 1 (05:37):
So are there little hacks that we can use so
that we're not filling up that context window as quickly,
like with the types of files that we're uploading?
Speaker 3 (05:47):
Absolutely. The first hack is every time we come to
AI for a new purpose, go new chat. There's no
point putting in, I don't know, health information or querying
and learning about health with any kind of company learning,
querying about company type stuff. They're definitely two different chats there.
So that kind of helps at the start. The other
is the files you upload do make a difference. And
(06:09):
so you'll find that markdown or text files are awesome.
And this is why people love markdown files, which, by
the way, they're just text files with a couple of
little clues as to first styling and whether it's a
bold and things like that. So if you see a.
md file, don't be scared. It's just a text file.
And that's great because it's pretty much it's all awesome
(06:30):
stuff and not a lot of filler. If you're looking
at things like PowerPoints and Word documents, then often the
AI has a tool just to extract the text and
the structure from those things. And so they can actually
be okay as well. PDFs, as you found out, are
the absolute killer. And the reason for that is PDFs
have been built for a printing format or viewing format,
(06:51):
not an understanding format. And so AI kind of knows
this because like if you ever selected a bunch of
text in a PDF, then often you'll get like tables
coming out weird and things like that. So AI can't
just go select the text in the PDF and copy it. Instead,
what it does is it does that, but it also
takes screenshots of every page of your PDF sometimes. And
then it needs to read that screenshot. This takes a
(07:12):
lot of tokens. So PDFs are the worst, absolutely the worst.
If you can do some kind of like a, extraction
from your PDF to a text file or a Word
doc or whatever first, oh my God, that will save
you a lot of tokens. But ultimately, only upload what
you need to for that particular chat thread. Don't go
nuts and give it all the documents in the world
(07:32):
because of course you'll fill that earlier. And the other
is particularly with PDFs. If you can somehow get the
text out of it in a different way, you'll also
find you won't hit that limit as much.
Speaker 2 (07:42):
And I'm sorry about using the PDF.
Speaker 3 (07:43):
That really sucks for you.
Speaker 2 (07:44):
Yeah, it certainly did yesterday.
Speaker 1 (07:47):
Now, before we go, because I know we are going
to come back in a future episode and talk about
some more advanced tips for how to close a chat
thread so that you're capturing all the goodness and really
easily exporting it into a new chat thread, particularly if
you have hit that 60% window or if you're just
finding that the AI is starting to get a little
bit dumber.
Speaker 2 (08:09):
But in a nutshell, what are the big things that
you would.
Speaker 1 (08:12):
Like people to take out of today to be very
clever about when to start a new chat thread?
Speaker 3 (08:19):
Number one, every time you come to AI for a
new purpose, particularly at a new day, go to a
new chat thread. Chats are not there to be crafted
and kept for forever. You're there to be used, thrown away, used,
thrown away. So think of them more like a tissue.
If you use it for one purpose and then it's
gone kind of thing. So yeah, if it's important, then yes,
(08:39):
get it to download a document out of your chat
or something like that. So absolutely number one. Number two
is just keep a mind on how long you've been talking.
And so you'll know I've been using this for lots
of documents, lots of chats, lots of ins, lots of outs.
And so therefore you'll go, this feels like it's a
longer chat and then maybe it's time for me to
(09:02):
wrap this one up. The third is if it starts
hallucinating and lying to you and getting things wrong, that's
an absolute canary in the coal mine for you. So yeah,
you might go, yep, this is definitely time for me
to start wrapping these ones up. So yeah, there's no
easy way to say, oh, it's 60% or it's 100%.
It's more like, do I feel like it's been a
(09:22):
long time? Maybe it's time for me to wrap these
things up, get that document out, get the output, whatever
I need to do, and then start a new chat thread.
Speaker 2 (09:30):
Amazing.
Speaker 1 (09:30):
And I will just add wherever possible, use markdown or
text files so that you don't crazily eat up your
context window.
Speaker 3 (09:39):
Those PDFs are a killer. Also, images can be quite
hard as well. So if you can upload a bunch
of images, it does take a lot of processing time
for it to understand the images. So yeah, pick the
documents that you're going to upload and you'll realize a
little bit more length in your speech rather than having
to wrap it up earlier.
Speaker 2 (09:58):
Neo, thank you so much.
Speaker 1 (09:59):
And we will be covering in a future episode in
the next few weeks, some more advanced tips about how
to neatly wrap up chat threads so you can take
all that great knowledge into a fresh one with a
fresh context window. How I AI was hosted by me,
Samantha Imber and Neo Aplin. A big thank you to
Martin Imber who does our sound editing and Jem Rubio
(10:22):
for production support. And thank you to John Kilby who
composed the theme music.
Imagine you're deep into a long chat with AI. And
it starts getting things wrong. Forgetting something you told it
a mere five minutes ago.
Speaker 2 (00:13):
Making things up.
Speaker 1 (00:15):
And you can't tell if it's you, your prompt, or
maybe just.
Speaker 2 (00:19):
The AI is having a bad day.
Speaker 1 (00:22):
But what's usually happening is that the chat thread has
run out of short-term memory. And once that memory gets
to about 60% full, the AI starts becoming less reliable.
In this episode, Neo and I talk through how to
know when it is time to start a fresh chat thread.
Speaker 2 (00:44):
You will learn the warning science that a thread has
passed its use by date.
Speaker 1 (00:49):
Why PDFs chew through your context window faster than almost
anything else.
Speaker 2 (00:53):
And what to upload instead.
Speaker 1 (00:56):
By the end of this episode, you'll know exactly when
to close a thread and start again.
Speaker 2 (01:02):
So that the AI stays sharp. Welcome to How I
AI with me, Dr.
Speaker 1 (01:12):
Amantha Imber and Neo Applin, head of Inventium AI. Each episode,
we share one practical way to use AI better at
work and in life. No fluff, no tech jargon, just
things you can use straight away. Neo, today we are
going to be talking about something that might seem on.
Speaker 2 (01:31):
The surface quite basic, but we see.
Speaker 1 (01:36):
Hundreds and hundreds, if not thousands of people get this
wrong when we are training people in workshop rooms. And
that is, when should you open or start a new
chat thread? So, Neo... What are your top tips on
how do we know when it's time to start a
new chat thread?
Speaker 3 (01:56):
The biggest tip is if the AI seems to be
doing strange and wacky things, then maybe it's time to
start a new chat thread. Because going back a step,
the reason is AI doesn't have an infinite memory for
each of your chat threads. They've got a technical thing
called a context window, which is it can hold a
certain number of tokens or words in an entire chat thread.
(02:18):
And that's everything you say, everything it says, all the
documents that both of you are creating, and also every
thought that it might have behind the scenes that it
doesn't show you kind of fills up this short-term memory.
And so once the memory gets full, or at least
starting to get full, science is saying about 60% full,
then you'll find that hallucinations happen more often and the
(02:38):
AI is more likely to get things wrong. So if
you're finding it's making up stuff, it's getting things wrong,
ask yourself, have I been talking with this chat in
this one chat thread For an awful long time, maybe
it's time for me to move off to another chat thread.
Speaker 1 (02:52):
So when you say 60% full, how do we know
what the total amount is? What does that look like?
And is it different across the four major AI platforms?
Speaker 3 (03:03):
And wouldn't it be amazing if they had a bar
chart going, awooga, awooga, you've just hit 60%. I'm about
to start to lie to you. No, it doesn't do that.
Isn't it quite annoying? each of the different four platforms
does have a different context window. The short-term memory thing
is different in each, and it's changing over time. I'm
hoping by the time you hear this podcast, you'll go, oh, wow,
(03:24):
it's even better than it was yesterday. They don't generally
come out and tell you how big these ones are,
but you can even ask your AI, how big is
your context window? And then even say, how many words
would that kind of equivalent be. For some of them,
I've got more of like a 200-odd A4 pages worth
of text. And it's kind of hard because is it
(03:44):
A4 pages with double-sided? Is it like wall-to-wall letters and numbers?
It's kind of hard to kind of get that in
our head. So I don't mind the term of 200
pages for some people. And others, if you're on Gemini,
they're just rocking it. You're like 1,000 pages worth of
text or something like that. And so each of them
are different. And these things change over time. But there
(04:05):
are really Two big principles here, which is if that
limit is getting full, and they've done a lot of
science on this, which is once you start hitting about 60%
of this context window getting full, then it's more likely
to make stuff up, hallucinate, and get things wrong. Either
forget something you've asked. or misinterpret something in the entire
(04:25):
conversation before. And so it's a good idea then to
start a brand new chat thread. Obviously, if you're 100%
full on that context window, every new word you put
at the bottom means the word at the top is
going to drop off what it can say. And so
don't go forever, which is why in all our trainings
and here in our podcast, we always say start a
new chat thread for pretty much every time you come
(04:45):
to AI and hopefully you won't even hit this problem
in the first place.
Speaker 2 (04:49):
Now, another thing on context window.
Speaker 1 (04:52):
So you say that maybe at the time of recording,
which is September 2026, it's maybe about 200 pages of text,
but also each page and each file type is not
created differently. So for example, something that I hit yesterday,
I was trying to create a mega project around health.
(05:12):
I was uploading all these different health documents that were
all PDF documents. And you heard me whinging to you
that I had hit my context limit.
Speaker 2 (05:23):
Which I rarely do because I'm pretty careful.
Speaker 1 (05:25):
About how I use AI and, you know, just managing
what is going in so that I don't hit those limits.
Speaker 2 (05:34):
But I did because it was all PDF documents.
Speaker 1 (05:37):
So are there little hacks that we can use so
that we're not filling up that context window as quickly,
like with the types of files that we're uploading?
Speaker 3 (05:47):
Absolutely. The first hack is every time we come to
AI for a new purpose, go new chat. There's no
point putting in, I don't know, health information or querying
and learning about health with any kind of company learning,
querying about company type stuff. They're definitely two different chats there.
So that kind of helps at the start. The other
is the files you upload do make a difference. And
(06:09):
so you'll find that markdown or text files are awesome.
And this is why people love markdown files, which, by
the way, they're just text files with a couple of
little clues as to first styling and whether it's a
bold and things like that. So if you see a.
md file, don't be scared. It's just a text file.
And that's great because it's pretty much it's all awesome
(06:30):
stuff and not a lot of filler. If you're looking
at things like PowerPoints and Word documents, then often the
AI has a tool just to extract the text and
the structure from those things. And so they can actually
be okay as well. PDFs, as you found out, are
the absolute killer. And the reason for that is PDFs
have been built for a printing format or viewing format,
(06:51):
not an understanding format. And so AI kind of knows
this because like if you ever selected a bunch of
text in a PDF, then often you'll get like tables
coming out weird and things like that. So AI can't
just go select the text in the PDF and copy it. Instead,
what it does is it does that, but it also
takes screenshots of every page of your PDF sometimes. And
then it needs to read that screenshot. This takes a
(07:12):
lot of tokens. So PDFs are the worst, absolutely the worst.
If you can do some kind of like a, extraction
from your PDF to a text file or a Word
doc or whatever first, oh my God, that will save
you a lot of tokens. But ultimately, only upload what
you need to for that particular chat thread. Don't go
nuts and give it all the documents in the world
(07:32):
because of course you'll fill that earlier. And the other
is particularly with PDFs. If you can somehow get the
text out of it in a different way, you'll also
find you won't hit that limit as much.
Speaker 2 (07:42):
And I'm sorry about using the PDF.
Speaker 3 (07:43):
That really sucks for you.
Speaker 2 (07:44):
Yeah, it certainly did yesterday.
Speaker 1 (07:47):
Now, before we go, because I know we are going
to come back in a future episode and talk about
some more advanced tips for how to close a chat
thread so that you're capturing all the goodness and really
easily exporting it into a new chat thread, particularly if
you have hit that 60% window or if you're just
finding that the AI is starting to get a little
bit dumber.
Speaker 2 (08:09):
But in a nutshell, what are the big things that
you would.
Speaker 1 (08:12):
Like people to take out of today to be very
clever about when to start a new chat thread?
Speaker 3 (08:19):
Number one, every time you come to AI for a
new purpose, particularly at a new day, go to a
new chat thread. Chats are not there to be crafted
and kept for forever. You're there to be used, thrown away, used,
thrown away. So think of them more like a tissue.
If you use it for one purpose and then it's
gone kind of thing. So yeah, if it's important, then yes,
(08:39):
get it to download a document out of your chat
or something like that. So absolutely number one. Number two
is just keep a mind on how long you've been talking.
And so you'll know I've been using this for lots
of documents, lots of chats, lots of ins, lots of outs.
And so therefore you'll go, this feels like it's a
longer chat and then maybe it's time for me to
(09:02):
wrap this one up. The third is if it starts
hallucinating and lying to you and getting things wrong, that's
an absolute canary in the coal mine for you. So yeah,
you might go, yep, this is definitely time for me
to start wrapping these ones up. So yeah, there's no
easy way to say, oh, it's 60% or it's 100%.
It's more like, do I feel like it's been a
(09:22):
long time? Maybe it's time for me to wrap these
things up, get that document out, get the output, whatever
I need to do, and then start a new chat thread.
Speaker 2 (09:30):
Amazing.
Speaker 1 (09:30):
And I will just add wherever possible, use markdown or
text files so that you don't crazily eat up your
context window.
Speaker 3 (09:39):
Those PDFs are a killer. Also, images can be quite
hard as well. So if you can upload a bunch
of images, it does take a lot of processing time
for it to understand the images. So yeah, pick the
documents that you're going to upload and you'll realize a
little bit more length in your speech rather than having
to wrap it up earlier.
Speaker 2 (09:58):
Neo, thank you so much.
Speaker 1 (09:59):
And we will be covering in a future episode in
the next few weeks, some more advanced tips about how
to neatly wrap up chat threads so you can take
all that great knowledge into a fresh one with a
fresh context window. How I AI was hosted by me,
Samantha Imber and Neo Aplin. A big thank you to
Martin Imber who does our sound editing and Jem Rubio
(10:22):
for production support. And thank you to John Kilby who
composed the theme music.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Bobby Bones Show
Listen to 'The Bobby Bones Show' by downloading the daily full replay.