# LLMs, hallucinations and claims that do not add up

Canonical page: https://thejasonsamuel.com/llm-hallucinations

[Home](https://thejasonsamuel.com/) > [Essays](https://thejasonsamuel.com/essays)

# LLMs, hallucinations and claims that do not add up

Restored from Instagram stories · 16 August 2026

so i am going to write this as soft as possible since i don’t want to discourage anyone from building in tech, especially ai - in a field where everyone hates ai (the film space)

but i have to do this to remain honest to tech.

1. the claim about randomness. they say - ask the exact same question yesterday, today, or tomorrow and it will always produce exactly the same answer. there is no randomness.

this is a bit funny. more let’s see how llm’s work.

llms fundamentally produce a probability distribution over possible next tokens when they generate a word or a sentence for an answer. randomness usually enters when you sample from that distribution.

for example:

p(next token) = { "pondy": 0.94, "chennai": 0.02, "india": 0.01, ... }

so, when they say their answers are exactly the same every single time,

1. it’s not an llm at all. which is impossible since we know that it’s an llm

2. did they mean they do a rag (hope you know what this is or it’s homework) over a finite text and they give the same answer and not probably different answers?

3. or they made the llm’s deterministic by decoding it. (there are dirty ways to decode it)

you saw how the llm’s work by probability of tokens last time - but a typical deterministic decider to produce the same answer works by choosing the highest probability every single time.

this doesn’t mean the underlying model somehow stopped being probabilistic. they’ve just chosen a deterministic method of decoding its predictions.

now this is the part where it gets really interesting. let’s assume they did get the same answer without randomness by a deterministic decoder.

(i’d be more surprised if they built a deterministic inference end to end but that’s hard as hell and the talent who could do that - i know they’re not here working at that startup)

let’s get back to the topic again

if you’ve a work/workflow where you need a classifier/a classification agent, extraction, routing, structured outputs, compliance checks and any form of output that would require a deterministic agent - this deterministic decoding to avoid randomness is insanely useful.

(although i’d still do tracing of every workflow run and do automated evals (short for evaluation: a form of testing for ai outputs and workflow) and compare it with benchmarks + auto adjust regularly - my style)

but but but

for anything that’s related to brainstorming and creative writing - this is so so so so bad.

you literally want to NOT avoid randomness here.

idk the exact username with which they’re doing, but if the decades are around brainstorming and creative writing - it’s totally bad and the worse way to implement ai.

2. now let’s come to the stupid ass claim that they trained llms to read every single word and read it word by word. i laughed so hard that i nearly farted.

(ok maybe i didn’t laugh and used laughing to cleverly hide why i farted) but let’s get to the point.

every llm whenever you input a text document to it - literally tokenized every text and it processes every single token through it’s attention mechanism. every single token.

if something is missing - it’s mainly the fault in the document - you can’t parse the text and so on. you just need to build few things before the ingestion of the document to ensure everything is tokenizable for the model. easy peasy.

3. so let’s come to the third thing. they literally blamed a false reason from point 2 to say that’s how they solved hallucination.

top tier regarded tbh.

we need to understand why llms hallucinate.

you remember where i explain how llms work through probabilities earlier in this essay?

good. llms hallucinate because their fundamental objective and how they work is to predict the next plausible text to construct an answer - not to determine what’s true. (the fact that you get more truth than false stuff proves how good human engineering is)

and hallucination is a solved problem. all you need to do is to build a system where the llm  has fewer opportunities to invent information and where unsupported claims are caught before reaching the user. boom - solved.

how do you build a system like that? you use a combination of a good agentic rag/even simple rag, architecture a better retrieval, structured extraction, (use your memory architecture for a good combination of the above), citation verification and so on.

that’s all.

exactly why i get pissed when non tech larpers do tech startups and then larp complete nonsense.

[← Previous essay](https://thejasonsamuel.com/a-filter-for-long-term-trust)[All essays](https://thejasonsamuel.com/essays)[Next essay →](https://thejasonsamuel.com/learning-history-through-rome)All rights reserved. Copyright © 2026 Jason Samuel