AI, where art thou goin'?

What is covered

Fierce competition for Large Language Models

We are witnessing a fierce competition on the LLMs market (large language models) resembling a veritable "gold rush". These are artificial intelligence systems designed to learn and generate human language trained on vast amounts of unstructured data. The most prominent proprietary models are OpenAI's GPT-4V (supported by Microsoft), Google DeepMind's Gemini AI, Anthropic's Claude 2. In the open source world the leaders are Meta's LLaMA 2 (including many others finetuned from this base model), Mixtral-8x7b, 01.ai's Yi 34B.¹

Scientific effervescence

The global AI scientific community is living an effervescent and fruitful age of innovation. This knowledge is used to train deep learning multimodal models capable of understanding the meaning expressed as human natural language of any data source: text, voice, image, video, audio, code. Notable papers are submitted on a fast pace aiming at improving the capabilities of existing models. "Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation"²) is exploring strategies of self-improvement, enabling a model to edit its own code. "Generative Agents: Interactive Simulacra of Human Behavior" presents how multiple specialized collaborative generative AI agents can "remember and reflect on days past as they plan for next day"³.

Vector databases revolution

Vector databases market is booming to support ever larger AI models. They differ from the relational databases built to store structured data in tables with fields and rows. Vector embeddings are numerical representations of meaning, the attributes of multimodal data sources. These vectors are clustered in the database by similarity, a numerical representation of distance in meaning. Some of the popular vector databases are: Pinecone, Chroma, Weaviate, Milvus, Qdrant, Vespa, Elasticsearch.

AI chip market going up

AI chip market is exploding. It is expected to reach almost 54 billion U.S. dollars by the end of 2023 and more than double by the end of 2027⁴. NVIDIA is the leader featuring the H100 Tensor Core GPU aiming at "supercharging large language model inference"⁵. Other major corporations started the trend of designing their own custom AI chipsets to increase efficiency and profitability: IBM's Telum and NorthPole, Microsoft's Maia and Cobalt, Google Alphabet's TPUv5, Meta's MTIA, AWS' Graviton4 and Trainium2, Qualcomm's Cloud AI 100, Apple’s M3, AMD MI300.

Potential adverse effects

AI is the latest tool humans are enthusiastically developing to improve their life. It comes with a promise of a "golden age". Unfortunately as any other instrument at our disposal may also have adverse effects. Some scientists and business leaders argue that unregulated AI may cause jobs displacement, civil unrest, discriminating groups by model bias, loss of privacy, new tools for a police state, weakening democracy.

Embrace it

My view is we should embrace this new development of human ingenuity as it might bring progress on an unprecedented scale. We should use a systematic approach to building trustworthy, explainable, fair, robust, transparent, privacy safeguarding AI. Regulating AI must be organic, according to necessity, able to favor innovation.

1. Hugging Face Chatbot Arena Leaderboard
2. "Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation" by Eric Zelikman and others
3. "Generative Agents: Interactive Simulacra of Human Behavior" by Joon Sung Park and others
4. "Artificial intelligence (AI) chip market revenue from 2022 to 2027" on Statista.com by Thomas Alsop
5. "NVIDIA H100 Tensor Core GPU Unprecedented performance, scalability, and security for every data center."