Anna's DayBreak News
Posts
Deep Dive: AI - Part IV: Inside the AI Black Box - How AI Works

Deep Dive: AI - Part IV: Inside the AI Black Box - How AI Works

Anna Eisenberg
February 22, 2025

Anna's Deep Dives

_{Just facts, you think for yourself}

Inside the AI Black Box: How AI Works

Neural Networks: How They Mimic the Brain

Neural networks process information like the human brain. They use layers of artificial neurons that pass signals through weighted connections. Each neuron activates based on input strength, just as biological neurons fire when stimulated. These networks learn by adjusting weights, enabling pattern recognition and predictions.

Spiking neural networks (SNNs) function more like the brain, transmitting information through discrete pulses, reducing energy use. Processing 70,000 MNIST images consumes only 0.015 kWh with SNNs, compared to 0.64 kWh for traditional models. Intel’s Loihi chip, based on SNNs, powers robotics and autonomous systems efficiently.

Neural networks replicate human perception. AI models interpret optical illusions using quantum tunneling, similar to how human eyes and brains process deceptive images. Georgia Tech’s RTNet integrates uncertainty into decision-making, improving reasoning. These advances help AI perform complex tasks, from navigation to risk assessment.

Hierarchical structures improve neural network performance. fMRI scans of 18 subjects revealed consistent brain-like structures in artificial models. This organization enhances communication within networks, improving accuracy and stability. The Thousand Brains Project aims to replicate the neocortex, using 150,000 cortical columns to advance AI cognition.

Neural networks aid scientific research. Graph Neural Networks (GNNs) analyze relationships, predicting protein interactions and improving fraud detection. DeepFRI, a GNN-based model, maps amino acid interactions to predict protein function. In protein binding, the ProBAN model achieved a Pearson correlation of 0.6, highlighting AI’s ability to model biological systems.

In 2024, John Hopfield and Geoffrey Hinton received the Nobel Prize in Physics for their work in neural networks. Their discoveries transformed AI, influencing deep learning and generative models. Neural networks continue evolving, bringing AI closer to human intelligence.

The Role of Data: Garbage In, Garbage Out

AI relies on data. Without quality data, AI systems fail. Poor data leads to biased results, incorrect predictions, and unreliable recommendations. In finance, bad data distorts credit scores, affecting loans and interest rates.

Organizations struggle with data quality. A study found 59% of companies do not measure data quality, despite investing heavily in AI. A 2020 Gartner report estimated poor data quality costs businesses $12.9 million per year. Without proper governance, AI cannot function effectively.

Cleaning data is critical. Errors, missing values, and inconsistencies must be corrected before training AI models. Preprocessing techniques like normalization and scaling standardize data. If test data leaks into training sets, AI models become overconfident and fail in real-world applications.

Bias is another issue. AI trained on flawed data reflects and amplifies societal inequalities. In healthcare, one study showed an algorithm assigned lower risk scores to Black patients, limiting care access. Recruitment AI has also favored male candidates when trained on biased hiring data.

The best AI systems rely on strong data validation. Automated methods catch errors quickly, while manual reviews add context. AI-assisted cleansing improves data accuracy by 20% and reduces cleanup time by 50%. High-quality data ensures AI delivers reliable and fair outcomes.

How Models Like GPT-4, Gemini, and DeepSeek Generate Content

AI models like GPT-4, Gemini, and DeepSeek process vast text datasets to generate human-like responses. These models use neural networks to identify patterns and relationships in data. They predict the most likely next word based on probabilities learned from training data. Larger datasets and parameters improve accuracy and coherence.

GPT-4, developed by OpenAI, uses over one trillion parameters to generate text. It maintains context over long conversations, improving coherence. DeepSeek R1 employs a Mixture-of-Experts system, activating different parts of the network based on the query. This optimizes efficiency by reducing unnecessary computations. Gemini, created by Google DeepMind, integrates text, images, and audio, making it more versatile than traditional text-based AI.

These models rely on transformers, a neural network architecture using self-attention mechanisms to weigh word importance. This enhances context understanding. With longer context windows, modern AI can process entire documents, improving responses in legal, technical, and academic domains.

Training these models requires massive computational power. Datasets include internet text, books, research papers, and proprietary sources. Engineers filter and preprocess data to remove bias and improve accuracy. Despite advancements, AI models still produce errors, hallucinate facts, and struggle with ambiguous queries.

AI models continue evolving, improving efficiency and reliability. Innovations in neural architectures and training methods push AI closer to human understanding. However, transparency remains a challenge. Researchers develop explainability tools to make AI decision-making more interpretable.

Chain of Thought (CoT) Reasoning: Why Giving AI a "Thinking Process" Matters

AI models produce impressive results but struggle with complex reasoning. Chain of Thought (CoT) reasoning helps AI break down problems into smaller steps. This structured approach improves accuracy, making AI more reliable in math, coding, and decision-making.

Researchers at UC Berkeley demonstrated CoT's impact by training models on 17,000 examples. Their method increased accuracy by up to 40% across various benchmarks. OpenAI’s latest models use multiple reasoning chains, leading to more nuanced responses. DeepSeek’s R1 model, with 671 billion parameters, excels in CoT reasoning, handling multi-step queries with improved clarity.

AI’s reasoning capabilities extend beyond logic-based tasks. In healthcare, CoT reasoning reduces misdiagnoses by clarifying decision paths. Investments in explainable AI (XAI) reached $11 billion in 2024, driven by the need for transparency. AI models like HamiltonianGPT2 achieved 89.5% accuracy in multi-hop question-answering, improving complex information retrieval.

Businesses benefit from CoT reasoning. Forrester research shows CoT prompting improves AI accuracy, reducing operational errors. Predictive analytics powered by CoT has cut logistics costs by 15%. In architecture and engineering, 76% of firms plan to expand AI use, expecting a 44% productivity boost.

AI's ability to reason transparently is becoming a competitive advantage. DeepSeek’s Deep-Think feature enhances interpretability, while OpenAI’s o1 model outperformed PhD students in multiple fields. As AI evolves, structured reasoning methods like CoT will play a key role in making AI more effective, reliable, and trustworthy across industries.

^{(Click on any section to start reading it)}

1. Introduction: The Dawn of the Age of Intelligence

Why AI is the defining technology of our time
AI hype vs. reality: Cutting through the noise
Why are we at an inflection point?
The impact of AI on society, economy, and human cognition

2. What is Intelligence? Human vs. Machine

Defining intelligence: Biological vs. artificial intelligence
The different types of AI: Narrow AI, General AI, Superintelligence
How AI "learns": Supervised, unsupervised, and reinforcement learning

3. The Evolution of Artificial Intelligence

Early AI: Symbolic reasoning and expert systems
The Machine Learning revolution
The Deep Learning era and the rise of neural networks
The Transformer revolution: How GPT-3 changed everything
Breakthroughs in generative AI and multimodal models (images, video, speech, code)