AI Weekly Digest Issue 6: OpenAI’s AI Christmas Season, LeCun’s AGI Forecast, Chinese Text‑to‑Image Breakthrough, and EchoMimic V2

AI Weekly Digest Issue 6

Welcome back, friends! This week we have carefully compiled the most significant AI industry developments, offering a comprehensive look at the latest trends and breakthroughs.

Market and Voices

Sam Altman: OpenAI Launches an “AI Christmas Season” with 12 New Products in 12 Days

On December 4, OpenAI announced a surprise: starting at 02:00 UTC on December 6, the company will release a new product each day for twelve consecutive days. Details remain scarce, but the OpenAI team confirmed the plan with brief replies such as “Correct” and “IYKYK”.

The first half of the launch introduced six items:

Full‑power o1 and ChatGPT Pro: 60 % faster inference and multimodal reasoning.

Reinforcement fine‑tuning: create expert models with minimal training data.

Sora Turbo: the second version of the video‑generation model released after the February preview.

Canvas (AI assistant for creators and programmers) opened to the public.

ChatGPT integration with Apple’s ecosystem.

OpenAI Vision: advanced voice mode with video and screen‑sharing on mobile.

Trend analysis shows that Sora dominates discussion, far outpacing o1, Apple Intelligence, Canvas, and reinforcement‑learning features. Users focus mainly on broadly applicable application products, while Canvas and fine‑tuning are viewed as niche productivity tools.

Despite the hype, Sora’s domestic interest has declined to less than 30 % of its early‑year peak, and reviewers note issues such as unrealistic hand movements, text garbling, and occasional “flying” animals.

LeCun: AGI Is Within 5‑10 Years, LLMs Are a Dead End

In a recent interview, Yann LeCun dramatically shifted his stance, predicting human‑level AI within five to ten years—contrasting with his earlier claim of a 10‑20‑year horizon. He remains skeptical about large language models (LLMs), calling them a dead end and advocating for new architectures like JEPA that learn from the world through goal‑driven, hierarchical planning.

LeCun argues that current LLMs exhibit “System 1” fast, intuitive responses, whereas AGI requires “System 2” deep, rational thinking. He proposes a goal‑driven AI framework that can learn from real‑world interactions and perform layered planning.

LeCun’s reversal offers fresh hope for the AI field, but achieving AGI will still demand sustained scientific effort and innovation.

Industry Solutions

Generating Chinese Text in Images: A Breakthrough from Doubao

Accurately rendering Chinese characters in text‑to‑image models has been a long‑standing challenge due to the complexity and sheer number of characters. Earlier methods resorted to post‑processing with separate font rendering.

ByteDance’s Doubao now supports end‑to‑end Chinese text generation directly in the app. Tests show a dramatic improvement: previously garbled “ancient‑script” results are replaced by clear, correctly positioned Chinese characters, with support for layout specifications such as vertical text, color, and font size.

While most cases are handled well, occasional misspellings or missing characters still occur, especially with rare glyphs. The feature is currently available on Doubao’s mobile app and will soon roll out to the web version.

The shift from image‑text mixing to full image‑text generation marks a significant step forward for multimodal AI.

Valuable Technology

Digital‑Human Project EchoMimic

Ant Group’s open‑source EchoMimic V2 advances from generating static digital faces (V1) to full‑body digital humans. By providing a reference image, an audio clip, and a gesture video, the system produces high‑quality animated videos with synchronized lip‑movement, facial expressions, and body gestures.

Key advantages:

Comprehensive animation covering head to torso, with expressive lip sync and gesture‑driven motion.

Simple workflow requiring only three inputs: reference image, audio, and gesture video.

EchoMimic V2 runs on 16 GB GPU memory (int8 quantization under 12 GB) and can generate 24 fps video in real time. The project is open‑source and continues to receive performance and speed upgrades.

Compared with Alibaba’s AnimateAnyone and Tencent’s MimicMotion, EchoMimic V2 demonstrates higher precision in finger details and lip sync while using relatively modest computational resources.

For further reading, see the linked articles throughout the newsletter.