Large language models, explained

A plain-English explainer on how tools like ChatGPT work, checked claim by claim.

YouTubePublished Sep 30Original source · youtube.com

claims in 7 statements

7 confirmed
1 needs review

Published by the Pikaan team. This review is read-only, and reading it uses no credits.

The video plays from YouTube. Use a claim’s time to play it in context.

Claims and sources

Every claim in the video, in order, to play from its time. The summary under each claim is written by AI from its sources; open the sources to check it.

Claim 1 of 8Confirmed

Factual claim · Statement 1 of 7 · 99% confidence
AI summary
OpenAI’s help documentation explicitly describes ChatGPT as an artificial intelligence-based service and an AI chatbot. Independent reference sources from IBM and Britannica likewise characterize ChatGPT as a generative AI chatbot/application.
Saved source links15ibm.com · help.openai.com · openai.com
Factual claim · Statement 2 of 7 · 85% confidence
AI summary
Large language models are trained on vast corpora of text assembled through web scraping and web-scale crawling of billions of pages from the internet, matching the description that they scan or consume internet text.
Saved source links9dev.to · linkedin.com · bildungsportal.sachsen.de
Factual claim · 86% confidence
Statement 2 of 7
Large language models scan every word on the internet.
AI summary
The evidence shows that LLMs are trained on or ingest large corpora that include blogs, websites, research papers, books, newspaper articles, and code repositories. The wording 'scan' is imprecise, but the underlying claim that these sources are part of LLM training data is supported.
Saved source links11educative.io · cohorte.co · ibm.com
Factual claim · Statement 3 of 7 · 76% confidence
AI summary
Sources from IBM and PNAS describe LLMs as pretrained on massive text corpora containing billions or trillions of words/tokens, and Stanford notes a 1-trillion-word web-text corpus as a major language-model dataset. That supports the statement’s core idea, but the exact “snapshot” being referenced is not independently identified in the search results.
Saved source links10understandingai.org · ibm.com · web.stanford.edu
Factual claim · Statement 4 of 7 · 95% confidence
AI summary
Extensive documentation on large language models confirms that ChatGPT operates by tokenizing text, analyzing relationships and patterns within a training model, and generating responses through next-token prediction that yield human-like quality.
Saved source links13lesswrong.com · neuroscience.stanford.edu · techtarget.com
Factual claim · Statement 5 of 7 · 97% confidence
AI summary
The evidence says language models and similar AI systems work by taking prior text and predicting the most likely next word or token. That is exactly what the claim says, just in simplified language.
Saved source links11cset.georgetown.edu · science.org · apxml.com
Factual claim · Statement 6 of 7 · 95% confidence
AI summary
Common Craft’s explainer says the model is created by computers that attempt to capture every word on the internet, around 1 trillion words. That aligns with the claim’s training-size and context description, and the statement was made in the 2023 explainer content surfaced on the page.
Saved source links11ibm.com · commoncraft.com · elastic.co
Factual claim · Statement 7 of 7
AI summary
The evidence shows AI chat systems typically generate replies from patterns learned during training on large text datasets, including publicly available internet text, and not by copying a fixed script. That supports the general idea, but it does not establish that every AI response is directly based on what is currently published on the internet.
These links belong to the claim, but no exact sentence was found on their pages. Review the sources before relying on this result.
Saved source links11arxiv.org · grammarly.com · ibm.com

Review your own source

Check your own videos, articles and documents the same way: every claim, its state and its sources. Starter includes a 7-day free trial. Cancel before it ends and you are not charged.

Claims and sources

1ConfirmedAI verdict: Supported · 99% confidenceChatGPT is an artificial intelligence tool.15

AI summary

2ConfirmedAI verdict: Supported · 85% confidenceLarge language models scan every word on the internet.9

AI summary

3ConfirmedAI verdict: Supported · 86% confidenceLarge language models scan blogs, websites, research papers, books, newspapers, and computer programs.11

Statement 2 of 7

AI summary

4ConfirmedAI verdict: Supported · 76% confidenceThe snapshot of over a trillion words creates the foundation for large language models.10

AI summary

5ConfirmedAI verdict: Supported · 95% confidenceBy analyzing the words in the model, ChatGPT can answer our questions and requests with human like quality.13

AI summary

6ConfirmedAI verdict: Supported · 97% confidenceArtificial intelligence tools use artificial intelligence to find what word should come next in a sentence.11

AI summary

7ConfirmedAI verdict: Supported · 95% confidenceThe large language model was trained on a trillion words and their context.11

AI summary

8Needs reviewAI verdict: MixedAI responses are based on what's published on the internet.11

AI summary

Review your own source