AI Roadmap LLM Learning Roadmap
✨ Awesome-Anything
Awesome-AI-Tutorials
Awesome-AI-Researchs
Awesome-GitHub-Repos
Awesome-AI-Examples🖼️ Digital Image Processing
DIP I Image Demosaicing and Histogram Manipulation
DIP II Image Denoising
DIP III Edge Detection
DIP IV Digital Half-toning
DIP V Geometric Image Modification🍃 LLM Components
1.1 Tokenization
1.2 Embedding
1.3 Attention
1.4 FFN, Residual Addition, and LN
1.5 Positional Encoding
1.6 Architecture & Decoding Policy🌱 LLM Pre-training
2.1 Pre-training Definition
2.2 Pre-training Data
2.3 Pre-training Workflow
2.4 Continued Pre-training (CPT)
2.5 Pre-training Evaluation☘️ LLM Post-Training
3.1 Supervised Fine-Tuning (SFT)
3.2 Reinforcement Learning
3.3 RLHF: Reinforcement Learning from Human Feedback
3.4 Optimizing PPO-Based Algorithms
3.5 Direct Preference Optimization (DPO)🍀 LLM Popular Models
4.1 BERT & Variants
4.2 PaLM Series
4.3 GPT Series
4.4 LLaMA Series
4.5 GLM Series
4.6 Qwen Series
4.7 DeepSeek Series
4.8 MOE Series
4.9 Other🪴 LLM Applications
5.1 Prompt Engineering
5.2 LLM-based Agents
5.3 Retrieval-Augmented Generation (RAG)
5.4 Enterprise-level RAG Practice (I)
5.5 Enterprise-level RAG Practice (II)🌿 LLM Optimization
6.1 GPU Usage Analysis
6.2 Latency Analysis for Inference
6.3 FlashAttention
6.4 PagedAttention
6.5 Packing for Efficient Training
6.6 PEFT🌾 LLM Compression
7.1 Quantization
7.2 Pruning
7.3 Knowledge Distillation🌵 LLM Hands-on Practice
8.1 nanoGPT: Training a GPT from Scratch
8.2 Instruction Fine-Tuning Qwen3-0.6B: A Minimal Working Example
8.3 OpenRLHF: SFT Qwen3-0.6B and Code Analysis
8.4 OpenRLHF: Reward Model Training and Code Analysis
8.5 OpenRLHF: PPO Tutorial and Code Analysis🌴 LLM Must-read Papers
9.1 Technical Reports Explained
9.2 MedLLM
9.3 o1 Techinical Paper
9.4 Prompting
9.5 LLM-as-Judge🌳 LLM Q&A
10.1 Why GRPO training loss starts at 0
10.2 VLM-R1 Reproduction Results
10.3 KL(P||Q) or KL(Q||P)?
10.4 Left Padding or Right Padding?
10.5 What is QK-Norm?
10.6 Why LLM training rarely adopts Dropout?
10.7 Dropout and DropPath🐝 VLM Image Encoders
11.0 Introduction
11.1 Vision Transformer
11.2 CLIP
11.3 BLIP
11.4 EVA
11.6 BLIP-2
11.7 SigLIP📝 MISC.
OpenAI ChatGPT Progression🤔Large Language Models Pass the Turing Test.
Turing Test: A recent paper by a UCSD team on arXiv mentions that LLMs have now been able to pass a true Turing Test. The experiment recreated the original three-party Turing Test, in which an interrogator engages in text-based conversations—without knowing their identities—with both a randomly selected human and an AI (e.g., GPT-4.5, LLaMa, etc.). These two participants are referred to as witnesses. After 5 minutes, the interrogator must decide which participant is the human, providing both a written justification and a confidence score.
During the test, the two witnesses communicate with the interrogator via separate windows on the left and right, but they cannot see each other’s conversations. This is also known as a triple-blind experiment. The specific chat interface is as follows:
Experiment Setup: The experiment recruited a total of 126 undergraduate psychology students from UCSD and 158 participants from the Prolific platform. Each human participant took part in 8 rounds of testing, randomly serving as either the interrogator or the witness in each round. This means that on average, each person interacted with an AI in about 4 rounds.
Additionally, the researchers designed two types of prompts for the AI models:
- Basic Prompt (NO-PERSONA): The model was simply instructed to try to convince the human that it was a real person.
- Persona Prompt (PERSONA): The model was given a specific personality to portray—that of an introverted young person who is familiar with internet culture and frequently uses slang in casual conversation.
Key Findings:
- When GPT-4.5 was prompted with the persona mode, it was judged to be human 73% of the time—significantly higher than the actual human participants.
- LLaMa-3.1-405B, using the same persona prompt, had a "win rate" of 56%, which was not statistically different from humans. (this is creepy enough though)
- In contrast, more basic models like ELIZA and GPT-4o performed well below random guess level, with only 23% and 21% of judges identifying them as human.
ELIZA is an early chatbot program developed by Joseph Weizenbaum in 1966, and is considered a classic in the history of AI. Unlike modern LLMs, ELIZA relied on pattern matching and template substitution to generate replies. For example, it would extract keywords from a user's input and match them with predefined response templates. In this study, ELIZA and GPT-4o were used as baseline comparisons.
From the above results, it’s clear that persona prompts have a statistically significant impact: both GPT-4.5 and LLaMA performed noticeably better than real humans when operating under the persona condition (ironically, the imitated human seemed more real than the real one...). However, the same models without a persona performed much worse, which really underscores the importance of prompt engineering.
Of course, passing a Turing Test doesn’t necessarily mean the model has human-level intelligence or consciousness. It just shows that today’s LLMs have become extremely good at mimicking humans—or maybe more precisely, that they’ve been well-trained in role-playing 🎭😂.
Participant Biases:
- Among the general population recruited from Prolific, several factors were significantly correlated with accuracy in identifying the AI:
- Gender differences played a notable role: accuracy varied significantly between genders.
- The more rounds a participant played, the worse they got at correctly identifying the AI.
- The more confident participants felt, the less accurate they were in reality (which is... kinda hilarious).
- However, none of these trends held in the undergraduate student group.
The only significant correlation found among UCSD students was:
The more confident they were, the more accurate they actually turned out to be — at least showing a positive trend overall.
Interrogators’ Strategies for Spotting the AI:
- In most conversations, interrogators tried to determine if they were talking to an AI by asking about daily life, personal opinions, or specific life details. Some even just bluntly asked, “Are you a human or a bot?”
- But here's the twist: the people who were actually good at identifying AIs didn’t rely on those everyday questions. Instead, they used:
- Weird or unusual talking approach (out-of-distribution data actually), designed to trip the model up
- LLM jailbreak-style tricks, trying to “trap” the AI into revealing itself
- Interestingly, the strategy of asking normal, casual stuff ended up performing even worse than random guessing — a strong hint that LLMs have gotten very good at sounding human nowadays, although still struggle with the unexpected from professionals.
Prev
11.7 SigLIP
Next
OpenAI ChatGPT Progression
Loading...