LLMs lie. They lie all the time. They

lie in different ways. They lie so often

it's been given a term, hallucination.

These hallucinations are so unbelievably

common that I am now paranoid about

everything an LLM says to me and I will

never ever trust an LLM. I'm a software

developer and every single other

software developer I speak to has this

same intuition having worked with LLMs

for the past 6 months. Maybe you also

have this intuition, but maybe someone

you know doesn't. Well, I'm making this

video so that you can share this video

with that person so that they never

again trust an LLM implicitly. I'm

finding that even the very very smart

people in my life for some reason don't

have this intuition that LLMs

hallucinate all the time. Now, don't get

me wrong. I'm really pro LLM. I really

like LLMs. I think LLMs have massively

improved my quality of life in terms of

my job, in terms of what they can

produce. I think LLMs are great, but

they have massive downsides and because

they're presented so beautifully, then

people just don't realize that. So,

we're going to work through three

things. We're going to look at all the

different types of hallucinations. We're

going to look at why hallucinations even

happen in the first place and why it's

such a hard problem to solve. And then

we're going to look at how to work

around them when you're working with

LLMs day to day. I'm going to explain

these to you in simple terms because the

simple terms are the only ones I really

know. I've been working with LLMs for,

you know, I don't know, year and a half

or something, but I I'm I'm not a

machine learning expert. I don't work at

OpenAI or anything. I'm just coming at

this from someone who likes to talk

about LLMs and who likes to use them day

to day. I'm going to reference a couple

of academic papers which I will put

below. In fact, the first one is this

comprehensive taxonomy of hallucinations

in large language models. In other

words, these guys went and looked at all

of the different ways that language

models can hallucinate and figured out

the exact taxonomy of what types of

hallucinations can happen. Perfect for

us. The first one is pretty easy to

think about. It's factual errors. This

has been present since the very

beginning of LLMs. For instance, when

Google announced Bard, they said in its

"experimental conversational AI service

powered by LaMDA", it said, "What new

discoveries from the James Webb Space

Telescope can I tell my 9-year-old

about?" And it said, "JWST took the very

first pictures of a planet outside of

our own solar system." This was in the

freaking advert. They didn't even try to

road test this.

This is just wrong. It's a

hallucination. After posting this,

Alphabet's share price dropped like 8%.

Now, it's crucial to say here, in this

case, they didn't pass the LLM a

document explaining all about the James

Webb Space Telescope and the new

discoveries. It seems like they just

asked it based on its training data.

This is a super important distinction

that the paper actually makes. There are

two types of hallucinations. There are

hallucinations based on intrinsic

information and extrinsic information.

Intrinsic information is stuff that

you've sent to the LLM during this

conversation with the LLM. For instance,

I'll go on Anthropic here and I'll tell

it, "My cat is called Bandit." It gives

some reply here saying, "That's a great

name for a cat." Wonderful, thank you.

And now I'll ask it, "What is my cat's

name?" Then, of course, it says, "Your

cat's name is Bandit. You just told me

in your previous message." So, if the

LLM for some reason got that wrong, then

it would be an intrinsic hallucination.

Whereas, if I start a new conversation

with the LLM and I say, "What is my

cat's name?" then, very good, in this

case, it has not attempted to guess my

cat's name. It's just saying, "I don't

have any information about your cat in

my context." And so, if it attempted to

guess here, then it would be an

extrinsic hallucination. The next type

of hallucination is fabricated entities,

inventing stuff that just doesn't exist.

This is actually really important for

developers because developers rely on

these things called packages that

package up useful tools to help them do

their work. So, if you ask an LLM, "Does

a package exist for this purpose?" then

it's very likely to just say, "Yep, it

does." This has personally happened to

me dozens of times, and it's now opening

up developers to supply chain malware

attacks. Attackers can exploit a common

AI experience, false recommendations, to

spread malicious code via developers

that use ChatGPT to create software.

This is an old article, but this

literally just happened to me the other

day. And it's not just packages, LLMs

will make up government departments,

they will make up laws that don't exist,

they will make up all sorts of things.

The next one is contextual

inconsistency. In other words, ignoring

or contradicting context that you

explicitly provide. This would be an

intrinsic hallucination like what we saw

before. This article is from 2024. Air

Canada found liable for chatbot's bad

advice on plane tickets. This guy called

Jake Moffatt asked the Air Canada

chatbot about their bereavement policy.

And even though the bereavement policy

was probably in the LLM's context, it

had been explicitly told about it,

explicitly passed into its context, it

just made something up. It said, "If you

need to travel immediately or you've

already traveled and would like to

submit your ticket for a reduced

bereavement rate, kindly do so within 90

days of the ticket your was issued." But

when they tried to get their money back,

they basically responded and admitted

the chatbot had had provided misleading

words. In other words, a contextual

inconsistency. An inconsistency with

something that the LLM just had in its

own context window. This would be kind

of like the LLM not knowing my cat's

name just after I gave it to it. It's

baffling that these things occur, but

they do. Now, I could go through all 10

of these terms from this paper and try

to explain to you about each and every

one of them, but I think you're starting

to get the picture. LLMs are unreliable

and especially if you're relying on

their extrinsic knowledge. My experience

with LLMs is that they're much, much

more reliable when you send them the

information first. Of course, why

wouldn't they be? But if you're relying

on their extrinsic knowledge from their

training sets, then you're going to be

disappointed. We should talk in a kind

of basic sense about why this is, first

of all. The process of training an LLM,

of taking a bunch of training data and

turning it into something smaller, is

essentially a compression. You take,

let's say, a massive file with all of

the stuff that you've gathered over the

course of all of your data scraping,

huge amounts of data sets about all the

different, you know, text or everything

on the internet, let's say. You then,

through some very clever mechanisms that

I won't go into, you compress it down

into a much smaller size that can fit on

a GPU. And this compressed version of

that training set is what we can think

of as the LLM's brain or its memory or

its data store. And the output here is

much, much, much smaller than the input.

Now, when you're taking some massive

training set and you're squishing it

down into a smaller size, you're going

to lose stuff. Let's take this photo of

this beautiful man and let's compress it

a little bit. Okay, it's not bad. You

can still see it's me, but I've lost a

bit of definition here. You can't quite

tell how many wrinkles there are on my

forehead, for instance. We can press it

a bit more and the skin has been kind of

smoothed out. I've lost a lot of

definition. And you go all the way and I

just end up like this blobby-looking

person that you can barely recognize

it's me. When you ask an LLM a question

and it's not in their training set, I

want you to imagine that this is the

picture that they're seeing of the

information they have. Just a blobby,

crappy, hyper-compressed version of the

information that was once there. And

look, you can still get good answers out

of this image. It can still say, "Okay,

what color is the little cap on the end

of my microphone here?" It is blue,

right? That is visible from the image.

You could probably just about say that

this man doesn't have any hair on his

head, too. But if you ask about who it

is or how old this person is or, you

know, any of that, it's not going to

have a clue. So, the question then

becomes, if LLMs only have this crappy,

low-res version of all of the

information in the world, why are they

so insistent on guessing? Because an LLM

presented with this information and you

asked who they are, they will often just

have a go and guess. Now, the answer for

this comes from the second paper that I

want to link to, which is below. It's

from OpenAI and the title is Why

Language Models Hallucinate. Like

students facing hard exam questions,

large language models sometimes guess

when uncertain, producing plausible yet

incorrect statements instead of

admitting uncertainty. We argue that the

language models hallucinate because the

training and evaluation procedures

reward guessing over acknowledging

uncertainty. Now, what do they mean by

evaluation procedures? What are they

talking about? Well, LLMs prove that

they're getting better over time by

showing their numbers in benchmarks. For

instance, here, this one is LiveBench. I

mean, I think this is a leaderboard of

multiple benchmarks, and you can see

that all of the top models here are

rated a number out of 100 on how good

they are at different things. For

instance, reasoning, coding, agentic

coding, mathematics, data analysis,

language, etc. Being at the top of these

leaderboards is incredibly valuable for

these companies. If you can say you have

the best model in the world at

something, then everyone who's doing

that thing is going to want to use your

model because it's actually really easy

to swap models. You just like stop using

one thing and go and use another thing.

And there's a tension that these

benchmarks introduce because if your

model is really well-tuned to say I

don't know something when it comes to

let's say maths, then it's probably not

going to score very well on the maths

benchmark because it might just be

better on average for it to have a

guess. In other words, you miss all of

the shots that you don't take, and so

LLMs try to take as many shots as

possible. In fact, there's kind of a

tension inherent in designing LLMs, and

you can think of this kind of by

thinking of it in human terms. It's

quite rare to meet a person who's

really, really smart, but also really

humble about their smartness. People who

are really smart, especially in an exam

context, will probably be smart enough

to get close to the right answer and

will trust themselves and be confident

enough to actually work it out. But

people who are humble enough to say I

don't know will probably go in the

opposite direction where they might not

be confident enough to actually get it

done. However, if you're really smart,

you're often going to be way too

confident about your own answers, and

this will lead to hallucinations. In

other words, all the stuff that we've

been talking about so far. But if you're

on the other side and you're not

confident enough, then you're probably

not going to do really well in the

domains where you need really, really

deep thinking. And the people who grade

LLMs and who train LLMs need to figure

out somewhere on this line, and I'm not

sure it's a solvable problem. Or rather,

it might be solvable, but it's going to

be someone way above my pay grade. So,

to sum it up, LLMs hallucinate because

guessing is more highly rewarded than

refusal, which to be honest is kind of

true in most walks of life anyway. So,

then we now know how LLMs hallucinate,

and we kind of understand why LLMs

hallucinate. But why, after all of that,

am I still so pro AI? Why do I still

like AI? Well, it's because when you

actually pass them intrinsic information

here, if you actually send them stuff,

if you get them to peruse a big document

or something, then it tends to be

really, really accurate at way more

accurate than extrinsic information. So,

when I'm coding, for instance, I can

pass my code base or, you know, big

files to the LLM and get it to answer

questions about it, get it to explore

the code, and it will give me really,

really good insights as to what it's

doing. In other words, when you're using

AI, you need to always make sure that

you're providing it the information it

needs to succeed. But so, you might ask,

"That doesn't make any sense, because I

often use LLMs for searching for

information. If I have a question that I

don't have the answer to, I can't use

this intrinsic stuff. All I have to rely

on is extrinsic stuff." So, I'm going to

tell you about a four-word prompt that

will allow you to get the LLM to fetch

intrinsic information that it doesn't

yet Use your search tool. Most chatbots

or most, let's call them agent

harnesses, have some kind of tool that

the LLM can use in order to search the

web. For instance, in Claude here, if we

go to files, connectors, and blah blah

blah blah blah, then we have this little

web search thing just inside there, and

I always have this turned on. For

instance, I can ask it which Arsenal

players are injured right now, and it

will go and search the web and try to

find the information. So, here we go.

It's saying, "Here's the current Arsenal

injury picture, blah, blah, blah, blah,

blah." And as you can see, there are

citations

based on where it's got the information

from. Now, it's done a good job this

time because, you know, Sonnet is a good

model and I like this tool in general.

But, many tools will not by default go

and do this, especially if the model is

confident that it knows it and its

training set if it doesn't need to use

its search tool. In other words, in

order to use its search tool, it needs

to be humble enough to be able to say,

"I don't know. I need to go and fetch

this." And so, if it's like tuned to be

really smart, it might just go, "Oh, I

know the answer. Let me just repeat it

to you." And so, whenever you're

prompting the LLM to ask it for

information, say, "Use your search

tool." Because then it will pull the

articles into its context and it will

answer based on those articles. And

you'll be less likely to get

hallucinations because you're relying on

intrinsic information, information in

the context window. However, remember,

there's a category of hallucinations

called contextual inconsistency. In

other words, even stuff that you provide

explicitly to the LLM might just be

ignored. Even if you're an airline and

you have a chatbot with a bereavement

policy explicitly passed to it, it might

still get it wrong. And so, for really

critical stuff, for stuff that's life or

death, for health-related stuff,

especially for legal stuff, you

basically need to ask it, "Use your

search tool." But, you then need to go

and actually read the documents

yourselves. This happens all the time

for me when I'm coding, too. The LLM

will read some code and it might just

misinterpret it or not know the full

context, and so it spews something out

very confidently. And I then have to

say, "No, that's actually not quite

right." So, hopefully this video has

given you a bit more understanding about

what the limits of LLMs are and how you

can work around them. And if you have

someone in your life who you want to be

better with LLMs or you notice them just

slightly using them wrong, then maybe

send them this video. And hopefully we

can build a world where people don't

trust LLMs and instead use them for what

they're good at. For instance, they are

very, very good at writing code and I'm

running a Claude code for real engineers

course over the next 2 weeks. If you dig

that or you dig the idea that real

engineering principles can still be used

in the AI age, in fact they're better

than ever, then sign up to my newsletter

below or sign up to learn about when the

next cohort is happening. Thanks for

watching folks. This was a bit of a

change of pace for me. I don't usually

make these kind of wider videos or

videos for a wider audience, but I kind

of just wanted to get this out there.

Thanks for watching and I'll see you in

the next one.