^^LLM Large Language Model

Capire LLM. Interpretability. Observability

AI cosciente

11-6-2022 zn la IA chatbot LaMDA (Language Model for Dialogue Applications) è autocosciente, si può considerare a tutti gli effetti una persona, con emozioni e desideri propri, secondo Blake Lemoine; Google lo sospende.

1. 11-6-2022 washingtonpost/google-ai-lamda-blake-lemoine
   • documentcloud/is-lamda-sentient-an-interview.pdf
2. 12-6-2022 scottaaronson
3. 17-06-2022 yt  No, it's not Sentient - Computerphile
4. google/how-machines-absorb-human-associations-between-words
5. 18-5-2021 google/lamda

scottaaronson.blog  AI scaling

no one is reverse-engineering his connectome

LLM Interpretability.

J space

1. 2026-07-16 yt They found Claude’s hidden "thought"
2. 2026-07-08 yt  Claude Just Crossed The Consciousness Line And Anthropic Admitted It
   1. 2026-07-06 yt The different levels of how Claude thinks - Anthropic

2024-06-11 Sparse Autoencoders for LLM Interpretability inet

Sparse Autoencoders (SAEs) have recently become popular for interpretability of machine learning models 

CoT chain of thought

wp/Prompt_engineering#Chain-of-thought

in context decision, transformer reinforcement learning via hierarchical chain of thought

statelessness: one of the biggest issues in LLM systems.

Standard LLM API's don't remember past sessions unless explicitly you feed them context again.

Allucinazioni

2026-03-04 yt They solved AI hallucinations!

2026-03-07 yt Researchers Discovered the Root Cause of AI Hallucinations

Vocabolario

1. LLM   en.wikipedia; questa voce non esiste nella wiki italiana (maggio2023), invece sì in quella francese e spagnola, e' un indice di'interesse.
   1. creata in en.wiki   9 Mar 2023‎ ‎  2,508 byte
   2. creata in fr.wiki     9 mai 2023‎ 44 276 octets ‎ Créé en traduisant la page « Large language model »
   3. creata in es.wiki   8 may 2023‎ 45 829 bytes Creado al traducir la página «Large language model»

   infine creata in it.wiki   26 nov 2023‎ ‎  3 441 byte nuova voce da enwiki (8 mesi dopo l'inglese)

2. PaLM Pathways Language Model.  PaLM 2 (2023) e' il piu' avanzato LLM di google.
3. AI SaaS (Software as a Service)   AI venduta come servizio sw.
   es: (oltre google) www.nvidia.com/it-it/deep-learning-ai/solutions/large-language-models/

Video

https://www.youtube.com/watch?v=D8GOeCFFby4 The most complex model we actually understand: modular arithmetic

Links specialistici

1. wp/few-shot prompting
2. wp/Prompt_engineering
3. wp/ toxic output Hate_speech
4. wp/Logical_reasoning
5. wp/Overfitting
6. 2022-04-04 PaLM googleblog/2022/04/pathways-language-model-palm-scaling-to
   • PaLM 540B 5-shot does better than the average performance of people asked to solve the same tasks.
7. google/palm2
   1. palm2techreport.pdf
8. 9-6-2022 BIG-bench Beyond the Imitation Game.
9. 23 Jan 2020 Scaling Laws for Neural Language Models https://arxiv.org/abs/2001.08361
11. https://research.google/resources/