non-LLM’s and Overview
2026-04-23
Outline
- non-LLM AI [complete]
- image generation experiment
- on consciousness with Amanda Askell.
Stochastic Parrots are Random
A discussion item: is creativity really just randomness?
The Examples of Firms
1. Palantir Foundry Analytics Platform
Paradigm: Knowledge graphs + traditional ML + deterministic rule engines
Why not an LLM: Palantir explicitly designs Foundry to avoid LLM dependency for core data integration and predictive modeling. It uses symbolic AI, graph traversal, and gradient-boosted trees to ensure auditable, deterministic outcomes for defense, healthcare, and logistics.
Market capitalization: 338 Billion dollars (2026-04-23)
2. DeepMind AlphaFold (Protein Structure Prediction)
Paradigm: Evoformer architecture (biological-sequence-specific transformer) + diffusion-based 3D folding
Why not an LLM: Though it uses transformer-like attention, it processes amino acid sequences and physical constraints, not natural language. Its outputs are 3D protein coordinates, not text.
3. Spotify´s Recommendation Engine (Discover Weekly, Release Radar)
Paradigm: Collaborative filtering + matrix factorization + CNNs analyzing audio spectrograms
Why not an LLM: Core ranking and content matching rely on user-item interaction graphs and acoustic feature extraction. No language modeling or text generation is involved.
Market capitalization: 106 Billion dollars (2026-04-23)
4. Netflix Content Recommendation System
Paradigm: Deep learning for metadata tagging + reinforcement learning for ranking + collaborative filtering
Why not an LLM: Optimizes for watch-time and engagement using behavioral signals and multimedia features. Its architecture is built for temporal sequence modeling and content-based filtering, not linguistic prediction.
Market capitalization: 390 Billion dollars (2026-04-23)
5. Stripe Fraud Detection AI
Paradigm: Gradient boosting (XGBoost/LightGBM) + graph neural networks + anomaly detection
Why not an LLM: Analyzes transactional metadata, device fingerprints, and network graphs in real-time. Operates on numerical/time-series data with strict latency and interpretability requirements unsuited to LLMs.
Market capitalization: ca 150 Billion dollars (2026-04-23)
6. DeepMind AlphaGo & AlphaZero (Game AI)
Paradigm: Deep reinforcement learning + Monte Carlo Tree Search (MCTS) + CNNs for board/state evaluation
Why not an LLM: Learns through self-play in rule-based environments. Uses value/policy networks for spatial-state evaluation, not textual token prediction.
7. Apple FaceID & Core ML (On-Device AI)
Paradigm: Specialized neural networks for 3D facial mapping, health metrics, and on-device sensor fusion
Why not an LLM: Optimized for low-power, real-time vision and physiological signal processing. Runs on Apple´s Neural Engine with architectures designed for pixel/depth analysis, not language.
8. Waymo Autonomous Driving Perception & Control Stack
Paradigm: Computer vision (CNNs/Vision Transformers) + LiDAR/radar sensor fusion + control theory
Why not an LLM: Core deployment focuses on real-time spatial reasoning, object tracking, and trajectory planning. While some competitors experiment with LLMs for high-level reasoning, Waymo´s commercial stack remains vision-control driven.
Market capitalization: ca 126 Billion dollars (2026-04-23)
9. DeepMind AlphaStar (StarCraft II AI)
Paradigm: Deep reinforcement learning + CNNs for minimap and unit-state processing
Why not an LLM: Operates in a discrete action space with raw game-state inputs. Uses hierarchical policy networks and attention over game entities, not linguistic corpora.
10. Amazon Rekognition (Enterprise Computer Vision)
Paradigm: CNNs + object detection models + facial embedding networks
Why not an LLM: Processes images/video for face recognition, scene classification, and OCR. Its OCR component uses vision-based text detection (e.g., CRNN architectures), not language modeling.
11. Insilico Medicine / Generative Chemistry Platforms
Paradigm: Diffusion models + Graph Neural Networks (GNNs) for molecular design
Why not an LLM: Works with chemical graphs, bond topology, and physicochemical properties. Generates novel drug candidates by optimizing molecular fitness functions, not linguistic tokens.
Market capitalization: 41 Billion dollars (2026-04-23)
12. Boston Dynamics Spot & Advanced Robotics Control
Paradigm: Reinforcement learning + inverse kinematics + sensor fusion + control theory
Why not an LLM: Core AI handles locomotion, balance, and manipulation through physics-based simulation and real-time feedback loops. Relies on deterministic control algorithms, not probabilistic text generation.
13. Google Photos Object/Scene Detection
Paradigm: Computer vision models (Inception variants, Vision Transformers) for image indexing
Why not an LLM: Extracts pixel-level features to classify scenes, faces, and objects. Uses embedding spaces for similarity search, fundamentally different from semantic language modeling.
14. DeepMind AlphaGeometry (Mathematical Reasoning)
Paradigm: Hybrid symbolic AI + neural networks + automated theorem provers
Why not an LLM: Explicitly designed to avoid LLMs to ensure formal, verifiable proofs. Combines neural pattern recognition with symbolic logic engines for geometry problem-solving.
15. Google Cloud Speech-to-Text / Microsoft Azure AI Speech
Paradigm: Acoustic models (Conformers, RNNs) + phonetic mapping + language-independent signal processing
Why not an LLM: Core deployment converts audio waveforms to text using deep acoustic modeling and phoneme-level alignment. While newer versions may add post-processing LLMs, the foundational speech recognition stack relies on signal processing and acoustic neural networks, not language generation.
Key Takeaway
The AI landscape is highly modular. While LLMs dominate text-centric applications, prominent real-world AI systems in enterprise analytics, biotech, robotics, finance, autonomous systems, and multimedia rely on computer vision, reinforcement learning, traditional ML, symbolic reasoning, acoustic modeling, or specialized generative architectures (e.g., diffusion/GNNs). These paradigms are chosen for determinism, latency, domain specificity, or hardware efficiency that LLMs do not natively provide.
LLMs are Random; when do we and do we not want that?
On Consciousness
BUS 1301-SP26