AI Daily Brief — 02 January 2025
The first US workday of the year started cautiously. The only verifiable corporate move came from Beijing — Baidu disclosing its Kunlunxin AI-chip arm had filed for a Hong Kong listing — while arXiv produced three notable papers including Qwen’s CodeElo benchmark. Western frontier labs and Wall Street were still quiet.
Top stories
- Baidu reveals its Kunlunxin AI-chip unit has filed for a Hong Kong IPO. A stock-exchange filing disclosed that Baidu’s chip subsidiary submitted a confidential Hong Kong listing application the previous day. Baidu retains roughly 59% of Kunlunxin, recently valued around CNY 21B (~$3B). Baidu shares rose 8.1% on the news — a marker of Chinese capital markets rewarding the domestic AI-chip self-sufficiency narrative. via Yahoo Finance
- Qwen team publishes CodeElo — a CodeForces-style LLM coding benchmark with human-comparable Elo ratings. Tests 33 models with private test cases and special judges. OpenAI’s o1-mini scores 1578 Elo, QwQ-32B-Preview 1261; most open-source models still land below the bottom 25% of human contest participants. via arXiv
- GPT4Scene (HKU / Shanghai AI Lab) gives VLMs 3D-scene understanding from video. A visual-prompting pipeline feeding Bird’s-Eye-View images plus per-frame object IDs to vision-language models; fine-tuned on 165K annotated videos, it reports zero-shot wins over GPT-4o on 3D scene benchmarks. via arXiv
- G2PT — Graph Generative Pre-trained Transformer. Tufts / Cornell / Northeastern propose treating graphs as token sequences and learning them autoregressively rather than via adjacency-matrix diffusion, with strong results on molecular and generic graph generation. via arXiv
Who shipped
Frontier labs again sat out. OpenAI, Anthropic, Google DeepMind, and Meta AI made no dated announcements; xAI‘s Grok logo refresh happened on this Thursday but was a cosmetic note, not a product release. Sam Altman’s “Reflections” blog post would land three days later on Jan 5, not Jan 2.
Open-source pulse
No fresh open-weights releases. DeepSeek-V3 remained the most-downloaded item on Hugging Face’s DeepSeek org; community benchmark replications and cost-efficiency analyses continued circulating on X and HF. Chinese labs held cards for the late-January cluster.
Quiet corners
Funding rounds, M&A, lawsuits, executive transitions: nothing dated Jan 2. Markets were thin after the New Year’s Day close. The FBI continued investigating the Jan 1 Las Vegas Cybertruck explosion, but its eventual AI angle — the perpetrator’s documented ChatGPT use for planning — would not be disclosed by LVMPD until the Jan 7 press conference.
Research & papers
A moderate arXiv day with three solid submissions above. CodeElo from Qwen was the highlight, picked up by MarkTechPost the next morning; many papers on the day’s Hugging Face Daily Papers panel were actually late-December submissions appearing under January arXiv IDs.
By the numbers
- 3 notable arXiv papers shipped (CodeElo, GPT4Scene, G2PT)
- 8.1% Baidu’s HK-listed gain on the Kunlunxin IPO disclosure
- Most-mentioned lab: Alibaba/Qwen (CodeElo authorship)
- Notable absences: OpenAI, Anthropic, Google, Meta
Compiled by AI Feed’s editor from verified web sources for 2 January 2025.