Weekly AI Roundup Issue 9: OpenAI Vision, LeCun Interview, ByteDance HLLM, and DeepSeek‑V3 Highlights

Weekly AI Roundup Issue 9

Editor: Wang Jidi

Market and Voices

Sam Altman releases 2025 technology vision – what should OpenAI do?

On December 24, 2024, Altman asked the public which new features and improvements they would like OpenAI to prioritize. The post attracted over ten thousand comments and 3.6 million views. Selected user suggestions include adding a family account with parental controls, enabling multi‑chat drag‑and‑drop, and releasing an image generation capability for GPT‑4o. Altman responded positively to each request and later summarized ten focus areas such as AGI, agents, upgraded 4o, better memory storage, longer context windows, and stronger personalization.

Opinion: Altman's feedback session was a success, showing strong anticipation for OpenAI’s upcoming updates, especially in the agents space where major tech players are competing.

LeCun says goodbye to the heated interview – deep dive into AI’s future

Meta’s chief AI scientist and Turing Award winner Yann LeCun gave an interview to UC San Diego professor Brian Keating, focusing on technical content rather than controversy. He reiterated his 2024 viewpoints and offered a roadmap for AI in 2025, emphasizing that reinforcement learning is too inefficient to explain human‑level learning, that human‑level AGI is unlikely within 5‑6 years even with architectures like JEPA, and that future challenges will include finding home robots, fully autonomous cars, and advanced personal assistants. He also highlighted AI’s potential societal impact comparable to the 15th‑century printing press.

Opinion: LeCun’s insights provide valuable guidance for researchers, policymakers, and the public, urging a balanced view of AI’s possibilities and risks.

Industry Solutions

ByteDance HLLM – a new paradigm combining large language models with recommendation systems

Cold‑start problems in recommendation systems arise when new users or items lack sufficient interaction data. Traditional ID‑based embeddings struggle in these scenarios. ByteDance introduced a Hierarchical Large Language Model (HLLM) that separates item and user modeling. The Item LLM extracts detailed features from item descriptions, converting them into embeddings, while the User LLM processes these embeddings to predict user behavior. This architecture reduces computational complexity and improves cold‑start performance.

Opinion: The work demonstrates that using LLM‑derived embeddings for items can solve tokenizer issues for non‑textual sequences, offering broader applicability to recommendation and behavior modeling.

Valuable Technologies

DeepSeek‑V3 – a domestically developed, open‑source large model

DeepSeek‑V3 is a 671 B parameter MoE model with 37 B active parameters, trained on 14.8 T high‑quality tokens using less than 2.8 million GPU hours—about one‑twentieth the cost of GPT‑4 (≈ 5.6 M USD). Benchmarks show it surpasses open‑source rivals like Qwen2.5‑72B and Llama‑3.1‑405B, and competes with closed‑source models such as GPT‑4o and Claude‑3.5‑Sonnet. In multilingual programming tests, DeepSeek‑V3 outperformed Claude‑3.5‑Sonnet, ranking just after o1.

Users have reported impressive results, including generating SVG images of a pelican riding a bicycle and integrating the model with AI video editors and the Cursor programming assistant, dramatically reducing reliance on FFmpeg commands.

Opinion: DeepSeek‑V3 delivers strong performance while saving memory and compute resources, embodying a cost‑effective AI solution comparable to a “Pinduoduo” of the AI world.

END