Training AI Assistants: From Pre-Training to Human-Guided Conversations
Pretraining and post-training chapter of Deep Dive into LLMs like ChatGPT (https://www.youtube.com/watch?v=7xTGNNLPyMI) by Andrej Karpathy.
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
14 Feb 2025
Understanding Base Model Inference: How AI Models Generate Text
In this section, the speaker explores the concept of base model inference, explaining how large AI models are trained, released, and function as token simulators rather than full assistants. The key points discussed include:
Base Models and Their Availability
* Training large AI models is extremely costly, but big tech companies often release “base models“ after training.
* A base model is a token simulator that predicts text sequences but is not yet an AI assistant.
Examples of Base Models
* GPT-2 (1.5B parameters, trained on 100B tokens) was one of the first widely released base models.
* LLaMA3 (405B parameters, trained on 15T tokens by Meta) is a modern, larger base model.
Components of a Model Release
* Requires two main parts:
* Python Code
* Model Parameters
Base Model Behavior
* It functions as an advanced autocomplete system, generating text based on statistical patterns from training data.
* It does not inherently provide factual or structured responses like an assistant.
Characteristics of Base Models
* Stochastic Nature - Given the same input, different competitions may be generated.
* Knowledge Compression - Acts like a lossy “zip file” of internet text, storing probabilistic patterns rather than explicit facts.
* Memorization & Regurgitation - Can recall high-frequency training data, sometimes verbatim (e.g, Wikipedia entries).
Limitations of Base Models
* Cannot provide factual updates beyond their training data cutoff.
* Tends to “hallucinate” (generate plausible but false information)
Practical Uses of Base Models
* In-context Learning - Few-shot prompting enables them to recognise and follow simple patterns
* Simulated AI assistants - Carefully structured prompts can trick a base model into behaving like an assistant by mimicking a conversation format
Acknowledgment
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
13 Feb 2025
The Internal Working and Training of Neural Networks
The Neural Network Internals and Training of Angrej Karpathy's video (https://www.youtube.com/watch?v=7xTGNNLPyMI)
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
12 Feb 2025
The Neural Network I⧸O
The video clip of Deep Dive into LLMs Like ChatGPT (https://youtu.be/7xTGNNLPyMI?feature=shared) by Andrej Karpathy
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
12 Feb 2025
Speech Recognition via Whisper
This is chapter3 from Andrej Karpathy (https://www.youtube.com/@AndrejKarpathy) video (https://www.youtube.com/watch?v=7xTGNNLPyMI)
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
10 Feb 2025
Understanding Reasoning Models
Understanding Reasoning Model
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
8 Feb 2025
Understanding Reasoning in Large Language Models
The original article is https://www.linkedin.com/pulse/understanding-reasoning-large-language-models-bowen-li-kgadc/?trackingId=WtLIWBUiTFmVwCPL4HCwUw%3D%3D (https://www.linkedin.com/pulse/understanding-reasoning-large-language-models-bowen-li-kgadc/?trackingId=WtLIWBUiTFmVwCPL4HCwUw%3D%3D)
The audio generated by Amazon Polly. It is not better than the OpenAI ChatGPT audio service.
Get full access to AI Journey OpenBook at aisuko.substack.com/subscribe (https://aisuko.substack.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)
Podcast Authority Score: 17 / 100
A composite of feed quality, social presence, YouTube performance and engagement. Read the methodology.
Quality
35
Social presence
0
YouTube
0
Engagement
0
Contact AI Journey OpenBook Podcast
Guest appearances
Does not typically book guests
Based on episode analysis; this does not confirm that the show is currently accepting guests.
Host of AI Journey OpenBook Podcast?
Claim your podcast to manage its listing and keep your show details accurate.
Pod Engine is an independent podcast discovery and analytics service and is not affiliated with or endorsed by this podcast. Artwork and show content belong to their owners. Full legal notice.
Explore this show Podcast research with Pod Engine
