evacideJun 3, 2023
It's always the most mediocre people who are worried about AI replacing them or taking their jobs or becoming sentient and/or murderous.
Show threadTHIS ACCOUNT HAS MOVED
@evacide an interesting observation. LLMs are by far the best bullshitters on the planet.
Jun 3, 2023 at 12:39amWeb
Show threadfemavenJun 3, 2023
@noplasticshower @evacide reminds me about this article http://www.onthelambda.com/2014/02/20/how-to-fake-a-sophisticated-knowledge-of-wine-with-markov-chains/?utm_source=pocket_reader
Show threadTHIS ACCOUNT HAS MOVEDJun 3, 2023
@femaven @evacide I am getting a 404 on the link
Show threadTHIS ACCOUNT HAS MOVEDJun 3, 2023
@femaven @evacide here's what I think about LLMs https://www.darkreading.com/vulnerabilities-threats/tech-insight-dangers-of-using-large-language-models-before-they-are-baked
Expert Insight: Dangers of Using Large Language Models Before They Are Baked

Today's LLMs pose too many trust and security risks.

Dark Reading
Show threadfemavenJun 3, 2023
@noplasticshower @evacide it’s about using Markov chains to fake knowledge about wine http://markallenthornton.com/blog/wine-markov/
Automatically generating wine tasting notes with Markov chains

Creating pseudorandom wine back labels customized by price, rating, type, or region using data from Wine.com

Mark Allen Thornton
Show threadTHIS ACCOUNT HAS MOVEDJun 3, 2023
@femaven @evacide fun and games. LLMs are much better bullshitters than Markov chains in my experience. Would be fun to prompt chatGPT into pretending to pontificate about wine.
Show threadDaniel LakelandJun 3, 2023

@noplasticshower @femaven @evacide

LLMs are literally markov chains. They just have really really complicated state and transition functions.

The way they work is... you give them a sequence of N tokens (that's the state) and they predict the most likely next token (that's the transition function). Then you've got N+1 tokens so lather-rinse-repeat... Once the number of tokens exceeds some value M drop the tokens off the front of the text... So basically it's Markov chains but really big.

Show threadTHIS ACCOUNT HAS MOVEDJun 3, 2023
@dlakelan @femaven @evacide I see your perspective but disagree with that characterization
Show threadfemavenJun 3, 2023
@noplasticshower @evacide the summary of the Markov chain for wine comes to a similar conclusion per the 2nd article - Overall, the results of the Markov generation process were mixed - some are nonsensical, self-contradictory, or ungrammatical, but many are quite coherent and convincingly human. Ultimately, perhaps they're best considered as the tasting notes a critic might write after finishing the bottle.