Why Large Language Models Still Matter: Insights into Industry Trends and Research Directions

As September 2024 approaches, the author observes that what can be added to a résumé has changed little over the past two years, while personal mental states have shifted from hopeful optimism to confusion, despair, and a search for new meaning within the AI landscape.

The capital market’s frenzy for AI applications has long cooled; investors no longer place high expectations on AI products. Even as star startups are acquired, skepticism resurfaces, exemplified by Nvidia’s stock dropping during product announcements and the free release of GLM’s flash version, suggesting monetization challenges for large models.

Personal experiences with Claude highlight how conversational agents can precisely address knowledge gaps and emotional needs, yet accessible, always‑on products remain elusive. The author also notes a growing admiration for DeepSeek, which rose from obscurity to the top tier, attributing its success to strong infrastructure and talent rather than mere scale.

The piece critiques the myth that larger parameter counts guarantee better performance, recalling how the trillion‑parameter model from Zhiyuan failed to deliver proportional impact, while a 3.5 B InstructGPT proved more valuable, underscoring that talent and data quality often outweigh raw compute.

Programming language analogies illustrate how Rust and Python won due to efficiency and simplicity, and how natural language may become the ultimate programming interface. Tools like Cursor and early ChatGPT API integrations faced post‑processing frustrations, but improving model capabilities are gradually reducing such overhead.

Future directions are identified in multimodal models that accept text and images and produce both modalities, as well as advanced optimization techniques such as MCTS and reinforcement learning from prover feedback. Communities around formal verification tools like Lean and Coq are thriving, indicating a broader research ecosystem.

The author stresses that scientific research—scaling laws, data provenance, and model generalization—is essential to navigate the fog of rapid AI development. While engineering can cut costs, rigorous theory guides which variables truly matter.

However, the high cost of research and limited willingness to fund it, especially during economic downturns, pose significant barriers. Infrastructure costs are decreasing, yet many fundamental scientific questions about LLMs remain unsolved.

Despite these challenges, the explosive popularity of ChatGPT has attracted more people and capital into the field, offering hope that the technology will eventually benefit a broader audience. Patience, rather than haste, is advised as the ecosystem matures.