2023 AI Landscape: Public Perceptions, Emerging Trends, and the Road to AGI

2023 was a year of explosive growth for large language models (LLMs). Starting with ChatGPT’s debut in November 2022, the release of GPT‑4 in March 2023 marked the first public encounter with a system that resembles a “world model”. The author notes that OpenAI’s strategy of scaling data and compute aims to compress all human knowledge into a single model, a view echoed by former OpenAI chief scientist Ilya Sutskever.

Short‑term hype vs. long‑term reality. The public overestimates immediate applications—only a few apps like ChatGPT and Character.ai achieve user breakthroughs—while most “star‑apps” fade quickly and cannot cover costs. In the long term, technologies such as video generation, audio synthesis, agents, memory, and model mini‑aturization have made significant progress, though commercial readiness remains uncertain.

World model, emergence, and self‑evolution. Emergence allows models to acquire abilities they were never explicitly trained for (e.g., GPT‑3.5’s spatial reasoning). The author speculates that if model parameters approach the number of human neuronal connections, AGI could appear within 2‑4 years. Synthetic data generation (e.g., self‑play “SPIN”) is presented as a way to overcome data bottlenecks, with many models already training on ChatGPT‑generated data.

Startup ecosystem and routes to market. Three pathways are described:

API + Prompt + product shell – suitable for companies with existing user bases but lacking a moat (e.g., Jasper.ai’s collapse after GPT‑4).

Open‑source model + fine‑tuning – viable for data‑rich firms, yet risky because upstream models may not stay open and upgrades can break fine‑tuned performance.

Building a proprietary base model – feasible only for large corporations with billions of dollars and massive GPU clusters.

The author warns that AI‑focused startups often lack technical moats and that “speed over perfection” is the only viable strategy for small teams.

Model compression and small‑model era. Miniaturized models (6‑7 B parameters) such as LLaMA‑7B, Mistral‑7B, and domestic equivalents dramatically reduce training cost (e.g., LLaMA‑7B costs ~9.3 years of A100 time vs. 100 years for GPT‑3). However, compression sacrifices stability and memory, and small models typically lag a generation behind their larger counterparts.

Multimodal explosion. Late‑2023 saw a surge in video and audio generation. Tools like Runway, Pika, and OpenAI’s Sora demonstrated high‑quality text‑to‑video generation, while audio cloning (e.g., ElevenLabs) achieved near‑realistic speech. The author argues that these advances reaffirm the generational superiority of AGI‑scale models over narrow AI.

AI agents and memory bottlenecks. Agents require goal setting, task decomposition, tool use, and decision making. Current agents suffer from limited memory, token length, and efficient recall. Newer models such as Gemini 1.5 (10 M tokens) and domestic long‑token efforts (e.g., “Moon of Darkness”) aim to address these issues.

US‑China gap. Despite a flurry of domestic model releases, the author observes that the technology gap has not narrowed in 2023. Chinese models lag behind GPT‑4, and hardware constraints (lack of domestic training‑grade chips) persist, though inference‑grade alternatives from Nvidia and Huawei are emerging.

Hardware bottlenecks. Training large models still requires billions of dollars and thousands of GPUs. While inference on smaller chips (e.g., Huawei Ascend) shows promise, stability and software ecosystem gaps remain, making hardware‑centric startups a more reliable opportunity.

2023 AGI timeline. A concise chronology lists key milestones: ChatGPT launch, ControlNet, LLaMA open‑source, GPT‑4, Nvidia H100, Stanford “AI Agent” paper, Mistral‑7B, GPT‑s + Assistants API, OpenAI internal power struggle, Google Gemini release, Nvidia Chat with RTX, and OpenAI’s Sora video model.

Overall, the piece serves as a snapshot of the “infant” stage of AGI, highlighting the need for rapid iteration, realistic expectations, and strategic positioning for 2024.