Unique Research · Insight / AIGC monthly · October 17, 2023 · Full editorial rewrite of ~4,660 CJK — all international and domestic headline items preserved
International headlines
Meta launches CoTracker — GitHub stars to 1.4k; pixel-level video tracking
Meta’s new video-tracking tool CoTracker hit ~1.4k GitHub stars soon after release. Observers flag object tracking, sports analytics, wildlife tracking, and film post as likely impact zones. Demo clips — show-jumping arcs, sailboats, skydivers — drew attention.
Differentiator: pixel-level tracking, not only semantic understanding. Transformer architecture encodes point tracks and iteratively updates positions; a sliding-window mechanism covers longer videos. On FastCapture it beat many models including Meta’s own DINOv2. Deploy via Colab or the GitHub repo.
Stanford “self-destructing models” against malicious fine-tunes
Stanford researchers propose self-destructing models: even open models degrade on harmful tasks. Algorithm MLAC (Meta-Learned Adversarial Censoring) meta-trains on beneficial vs harmful task sets — inner loop simulates adaptation attacks; outer loop maximizes loss on harmful tasks so the model “forgets” harmful-task information. On Bias in Bios (gender ID = harmful; occupation = beneficial), harmful-task performance after MLAC stayed near a randomly initialized model across data regimes. Accepted at AAAI with Best Student Paper nomination.
ChatGPT plugin “Consensus” for paper search
GPT-3.5 ChatGPT gained Consensus: search/summarize over 200M+ papers. Modes: (1) one-sentence Q&A with paper links; (2) short multi-paper reviews with citations; (3) keyword-free topic paper lists (author/year/venue/abstract); (4) SEO-friendly blog drafts citing peer-reviewed work. Results claim peer-reviewed sources via Semantic Scholar (monthly refresh). Plus members only; third-party Consensus Inc.
Claude Pro: $20/month paid tier
Anthropic’s Claude Pro at $20/month (~£18): ~5× conversation volume vs free; priority access in high traffic; first look at new features. Mixed user reaction — some welcome extras without gutting free; others call $20 steep for casual use. Anthropic has raised ~$1.45B and expects ~$1B compute (esp. GPU clusters) over ~18 months.
DeepMind AlphaMissense: pathogenicity of missense variants
Built on AlphaFold, AlphaMissense (Science) classifies missense (nonsynonymous) DNA substitutions that change protein amino acids. It scored ~71 million possible missense variants; classified 89% as likely pathogenic or likely benign vs ~0.1% for human experts. Fine-tuned using human/primate variant-frequency DBs (less human-label bias). ClinVar auROC 0.94. Predictions, code, and 216M+ single-AA substitution predictions released openly.
Windows 11 + Copilot: toward an “AI OS”
Win+C opens system Copilot for dark mode, music, desktop tidy, summarization, writing, and image edits (portrait cutout, posters). Snipping Tool gains text extract and AI mosaic for privacy. Paint adds AI background swap; Clipchamp suggests scenes and one-click share. Photo search by fuzzy description. Also: passkeys (face/fingerprint/PIN), Cloud PC, seamless switch. Rolling out now / larger Nov update.
Domestic headlines
Moonshot AI (月之暗面) ships Kimi Chat; >$200M raised
Founder Yang Zhilin (FAIR / Google Brain; first author Transformer-XL & XLNet; student of Ruslan Salakhutdinov at CMU) — named among The Information’s “China OpenAI” five. ~6 months in, Sequoia China, ZhenFund and others; funding >$200M. Own models, consumer (ToC) first, multimodal later, super-app exploration; heavy C-end product hires (0→1 and 100M+ DAU experience).
ByteDance “AI Assistant” on Volcano Engine
Third major LLM-related product after (1) Volcano Ark MaaS (Jun) aggregating third-party models; (2) Douyin’s Doubao chatbot on in-house Skylark (云雀), approved Aug 31; (3) AI Assistant on VeDI — either Skylark directly or third-party models (Zhipu, Baichuan, MiniMax…) via Ark. DataLeap and DataWind cover most data production/consumption so non-specialists can query and analyze.
360 Zhinao: ~RMB 20M H1 revenue — first profitable Chinese LLM company claim
360 H1 2023 report: nearly RMB 20 million from large-model business — framed as first Chinese LLM company to turn a profit. “360 Zhinao” landed first in search, then multi-turn dialogue and text/audio/image/video; first to pass CAICT “trusted AIGC LLM” evaluation. Lines: industry LLMs (tax, enterprise services), city LLMs (Sichuan/Chongqing), GPT industry alliance (Jun), security LLM (Aug 9) from 15 years of AI security data.
Baidu Intelligent Cloud AI-native apps
Service marketing, office productivity, production optimization. Xiling digital-human platform: lower-cost faster 3D humans that switch look/persona by industry (finance, e-comm, culture/tourism). Yinian content studio (copy, images, video). Zhenzhi enterprise search over internal docs. Du An’an industrial safety assistant. Natural-language UX and unified entry points.
Tsinghua-line Lingxin Intelligence reportedly fully acquired by Zhipu AI
Unofficial; valuation said to be hundreds of millions RMB. Shared Tsinghua roots, co-investment, co-released CharacterGLM “hyper-realistic” persona model. Zhipu market value ~RMB 12B (Meituan, Alibaba among backers). Lingxin earlier angel + Pre-A at ~$100M valuation; founder Huang Minlie (Tsinghua CS assoc. prof., NLP/dialogue/affect); Li Wenjue UBO (ex Zhongke Chuangxing / BAAI investment). Mission: mental-health digital therapeutics; with Tsinghua CoAI released LLM safety eval framework and AiU interest community. Deal read as resources + research runway while keeping some independence.
Author: Haoyu · Review: Siky