All news

336Kr · AI17:08Research

AI score60

LeCun Team Paper: 109M-Parameter Text JEPA Outperforms 72B-Parameter Qwen2.5 in Cross-Modal Alignment

The paper by LeCun and Randall Balestriero uses the JEPA framework to provide a mathematical proof of the Platonic Representation Hypothesis, and claims that on the cross-modal representation alignment task, a 109M-parameter text JEPA outperforms the 72B-parameter Qwen2.5. The research shows that Good Views training requires only 21.6 million samples to reach the alignment level that a next-token prediction model achieves with 900 million samples, reducing the sample count to 1/42, or to 1/14 when counted in tokens. The authors also point out that this comparison is limited to specific alignment metrics, and that the mathematical proof relies on idealized conditions such as a finite world and reversible observations.

Same story · 3 reports

#Papers/Research#Multimodal#English text is being prepared.

336Kr · AI17:08

AI score23

Epi2Diff uses large model reasoning traces to quantify question difficulty, achieving 73.0% classification accuracy on SAT Math

A team from Virginia Tech, the University of Maryland, MBZUAI, and the University of Pittsburgh proposed Epi2Diff, which splits large model reasoning traces into eight types of cognitive segments and extracts 83-dimensional process features, combining them with question semantics to predict human response difficulty. On four datasets—SAT Math, SAT Reading and Writing, Cambridge, and USMLE—the method ranked first on all metrics in a single main experiment, achieving 73.0% difficulty classification accuracy on SAT Math, 3.2 percentage points higher than the strongest baseline. The paper has been accepted by TMLR and can be used for pre-assessment of new question difficulty and to assist human question review.

#Papers/Research#Reasoning#English text is being prepared.

336Kr · AI17:08Opinion

AI score72

Embodied intelligence funding cools: doubled valuations coexist with discounted old shares

Tencent Technology reported that in September 2026, 46 financing deals occurred in China's embodied intelligence primary market, with a reference amount of approximately RMB 11.624 billion, down 4.2% and 40.1% respectively from August. Several investors said that transfer quotes for old shares of some leading robotics companies are concentrated at 65% to 80% of the previous round's valuation, while AgiBot is advancing a new financing of RMB 1 billion to 1.5 billion at a valuation of about RMB 50 billion. The article also mentioned that after Unitree's listing, its stock price has retraced about 41% from its intraday high on the first day, and the Hong Kong listing processes of some robotics companies have fallen short of expectations.

#Industry news#Embodied AI#English text is being prepared.

336Kr · AI17:08

AI score42

Alibaba Supports Dakar Youth Olympic Games with Wan3.0 and Olympic Dedicated Cloud

At the Dakar Youth Olympic Games, Alibaba used its self-developed video generation large model Wan3.0 to let 100 volunteers from around the world drum together on the same virtual stage, and built a 75-square-meter AIGC interactive space called "The Teranga Beat" in the Olympic Plaza, generating personalized videos in real time. Behind the scenes, Alibaba Cloud, based on the Apsara Enterprise Edition, deployed the first Olympic event dedicated cloud locally in Dakar, hosting the paid ticketing system for 7 major venues and over 1 million tickets; after the Games, it will be handed over in its entirety to the Senegalese government and converted into a "national sovereign cloud."

#Industry news#Video#Multimodal

336Kr · AI17:08

AI score42

Today Yixiu's Wang Teng launches the ISHO Ring smart ring and iFlow temperature control system, taking aim at Oura and Eight Sleep

On October 29, Today Yixiu held a launch event in Beijing for the ISHO Ring smart ring and the ISHO iFlow smart temperature control system, and also introduced ISHO Intelligence, a personal sleep health AI technology system. The ISHO Ring debuts the Apollo 330B Plus chip, offers 6 to 10 days of battery life, and its basic features are free; the iFlow starts at 9,999 yuan, which Wang Teng said is roughly 60% to 70% of the pricing of comparable products from competitors. Wang Teng's timetable for building out the system is 5 to 10 years.

#Product updates#English text is being prepared.#English text is being prepared.

336Kr · AI17:08

AI score40

Doctors Using Doubao in the Clinic

A survey of 647 medical professionals by DXY shows that 98.5% of doctors have used AI in the past three months, 81.4% have used Doubao, and Doubao's average daily health-related consultation volume is estimated at the level of 20 million person-times. Doctors use Doubao to check drug interactions and verify prescriptions, while patients use Doubao's results to question doctors' judgments. Dr. Luo once spent nearly half an hour explaining why a 6mm pure ground-glass nodule should first be managed with follow-up. The 2022 "Implementation Rules for the Supervision of Internet Diagnosis and Treatment" has explicitly prohibited AI from automatically generating prescriptions.

#English text is being prepared.#Industry news

336Kr · AI17:08

AI score33

Tencent Maps' LBSwork Reuses WorkBuddy's Same-Source Harness Framework—Can It Replicate the Success?

Tencent Maps has launched LBSwork, a desktop AI product that adopts the same-source Harness framework as WorkBuddy and offers seven map-expert capabilities, including retail site selection, map visualization, and ride-hailing liability determination. In September 2026, WorkBuddy reached 85.34 million monthly desktop visits, ranking first among the 20 AI-native office agents tracked by Analysys. However, LBSwork targets professional clients, which places higher demands on data security and analytical accuracy, and Amap's Wendian and Baidu Huiyan have already entered the market ahead of it.

Same story · 4 reports

#Industry news#Agents#English text is being prepared.

336Kr · AI17:08

AI score30

OpenAI wants to rename "data centers" as "compute campuses"

OpenAI plans to rename AI data centers as "compute campuses," and a company spokesperson has confirmed the idea, citing that it better reflects its long-term investment in computing, energy infrastructure, and technical jobs. A Pew Research Center survey shows that the share of U.S. adults who believe data centers are generally harmful to the environment rose from 39% in January of this year to 54% in the summer, and 60% of respondents do not want data centers built near their homes. OpenAI has previously committed through the "Stargate" community program to bear its own energy costs and not drive up local electricity prices.

Same story · 9 reports

#Industry news#OpenAI

336Kr · AI17:08

AI score66

Xiaohongshu Participates in DeepSeek's Second Round of 105.5 Billion Yuan Financing

The Information disclosed that DeepSeek's second financing round amounted to 105.5 billion yuan, with a valuation of 615.5 billion yuan, and nearly 30 investors participating, with Xiaohongshu appearing on the list. In this round, the proportion of state-owned capital investment rose from 2% in the first round to 27%; Tencent invested 10 billion yuan, unchanged from the first round; CATL and Zeng Yuqun together invested 10 billion yuan. Previously, Xiaohongshu had already participated in Kimi's financing of over 1 billion US dollars and appeared among the strategic investors ahead of MiniMax's listing.

#Industry news#English text is being prepared.#DeepSeek

336Kr · AI17:08

AI score34

Between Zhengzhou and Hengdian: The Cost and Survival Arithmetic of Live-Action Short Drama Factories Under the AI Short Drama Shock

In 2026, revenue sharing on live-action short drama platforms shrank sharply, and AI short dramas poured in at less than half the cost of live-action, forcing production companies in Hengdian and Zhengzhou to pivot to AI production. Chengfeng Polang Film's live-action short drama orders fell from more than 20 per month to around 10, and in April about 90% of its workforce shifted to AI production, with monthly output rising to 5,000 to 8,000 minutes from May to August. AI video generation is billed by the second at about 1 yuan; DataEye statistics show that in the first half of 2026, 221,900 new AI dramas and comic dramas were added on Douyin, and only 1.3% reached the break-even line of 50 million plays.

#English text is being prepared.#Video#Industry news

JJiwei.com17:07

AI score46

First batch of 17 integrated circuit projects under Optics Valley's "Plane Tree Plan" signed, with total contracted investment of over 6 billion yuan

The first batch of 17 integrated circuit projects under Optics Valley's "Plane Tree Plan" were signed in a centralized ceremony, with total contracted investment of over 6 billion yuan, covering the entire chain from lithography, etching, deposition, polishing, and metrology/inspection to design. Within just over two months, 14 companies including AMEC have completed registration of their Wuhan subsidiaries, 10 have confirmed sites and are about to begin construction, and they are expected to be operational in about a year. The plan, centered on leading storage-chain enterprises with a radius of 4.5 kilometers, aims to build a 5-minute industrial collaboration circle and a 30-minute supply circle within 60 square kilometers.

#Industry news#English text is being prepared.

JJiwei.com17:07

AI score40

Xpeedic releases EDA 2027 version, 2026 User Ecosystem Conference focuses on STCO and AI for EDA

At the 2026 User Ecosystem Conference, Xpeedic released the EDA 2027 version, adding a 3DIC PDK and a co-simulation platform for optics and electronics, and launched the Agentic AI super-intelligent agent platform AgentEvo and the trustworthy physical intelligence foundation XAI, driving AI for EDA from single-point functions toward platformization. A live demonstration of EDA Agent 2.0, jointly built by Lenovo and Xpeedic, showed that design iteration efficiency improved by about 5 times and manual operations were reduced by about 90%. The conference also held discussions around STCO (chip-to-system co-optimization), emphasizing that chips, packaging, boards and racks need to collaboratively address multi-physics issues such as electrical, thermal, mechanical and optical problems in the early design stage.

Same story · 2 reports

#Product updates#English text is being prepared.#Agents

JJiwei.com17:07

AI score20

Actions Semiconductor's ATW6095 on-device AI SoC enters mass production, empowering high-end AI Note

The new-generation on-device AI SoC chip ATW6095 series from Actions Technology has entered mass production, with the first batch of end devices being the Yidoor Intelligent YoooClaw C·ONE Pro and Xinren Intelligent N.B.BOSS P2 Plus. The chip is equipped with a self-developed in-memory computing NPU, featuring 100GOPS of single-core computing power and an INT8 energy efficiency ratio of 6.4TOPS/W, and has a built-in 420MHz HiFi5 DSP, supporting far-field pickup at a distance of up to 10 meters. When equipped with a 4+1 microphone array running on-device AI algorithms, the device's average operating current for recording can be kept within 10mA, saving more than 20% power consumption compared with other solutions.

#Product updates#English text is being prepared.#Speech

JJiwei.com17:07

AI score62

Anthropic plans to raise up to $3.5 billion from Japanese investors in its US IPO

Anthropic plans to raise up to $3.5 billion from Japanese investors in its US IPO, with Mizuho Securities responsible for the Japanese portion of the offering, and institutions such as Goldman Sachs and Mitsubishi UFJ Morgan Stanley handling investor orders. Japan's IPO market has dried up to a 14-year low, and investors in Australia, Canada and parts of Europe can also subscribe. Anthropic filed for its IPO in October 2026, and this offering will test investors' appetite for AI-related new stocks.

#Industry news#Anthropic

AAYi AI Notes17:01

AI score25

Altman: An 8-year-old kid makes a 3D game in 10 minutes

In a conversation with Stripe co-founder John Collison, Sam Altman used the example of an 8-to-10-year-old child building a complete 3D game on the spot in ten minutes to illustrate that the way this AI wave most resembles the early internet lies with the next generation. He also mentioned that when OpenAI's internal Codex briefly went down, engineers fell into a panic for a while, realizing they could hardly imagine going back to the days of handwriting code line by line.

Quoted postAYi · @AYi_AInotes

English text is being prepared.

#English text is being prepared.#OpenAI

AAYi AI Notes17:01Product release

AI score73

DeepSeek Open-Sources the dsh-superagent Multi-Agent Architecture

DeepSeek has open-sourced the dsh-superagent multi-agent architecture, turning code repositories into automated pipelines: each repository is assigned a resident overall dispatcher (Router Session), which only reviews GitHub Issues, PRs, or CI errors and does not directly modify business code; after a task passes whitelist evaluation, the system spins up a separate Worker for it, assigning an independent session, an independent directory, and a git worktree branch for parallel execution. The system opens 3 work slots by default, supports a maximum of 1 to 16 concurrent tasks, and always reserves 1 slot for humans to jump the queue; subsequent events for the same task are subscribed directly to the corresponding workstation, bypassing the overall dispatcher to avoid repeatedly consuming Tokens. The author says it takes just 5 steps to get started.

Quoted postAYi · @AYi_AInotes

English text is being prepared.

Same story · 2 reports

#English text is being prepared.#Agents#DeepSeek

SShao Meng17:01Research

AI score52

Zhejiang University team releases open-source personal Agent system report and open-sources nanoMuse

The Zhejiang University team has published an open-source personal Agent system report, using Meta Muse as a reference to define what a "personal Agent" should look like, and has open-sourced their nanoMuse. The tweet provides links to both the paper and the project.

#Papers/Research#Agents#Open-source ecosystem

KKling AI17:01First-party

AI score10

Kling reposts ArtArch Christmas creator event

ArtArch.Ai announced a joint Christmas creator event with Kling, inviting users to create unique Christmas stories with AI; selected creators will each receive 1,000 credits; to apply, send a direct message or join their Discord. The official Kling account reposted this event information.

Quoted postArtArch.Ai · @ArtArchAI

🎄 ArtArch × Kling Christmas Creator Event! Turn your imagination into a one-of-a-kind Christmas story with AI. ✨ Selected creators receive 1,000 credits each to bring their ideas to life. DM us or join our Discord for details & to apply ：https://t.co/sk0bm7nHLe https://t.co/lUMuygD1Ts

#English text is being prepared.

RRohan Paul17:01Research

AI score54

A paper from Tsinghua and Microsoft proposes HarnessMM; DeepSeek Harness on HumanEval…

A new paper from Tsinghua and Microsoft proposes HarnessMM, which lets the model pre-write subsequent commands and their corresponding check conditions; when tool results pass the checks, the commands are executed directly, skipping LLM calls. On 98 HumanEval tasks, DeepSeek Harness used 87.93% fewer tokens than the original version, and its success rate rose from 71.43% to 85.71%; Codex maintained a 90.82% success rate while using 29.54% fewer tokens than standard Codex.

#Papers/Research#Agents#English text is being prepared.

BBaidu17:01First-party

AI score25

Baidu Apollo Go's Progress in London and Hong Kong Right-Hand-Drive Markets

Baidu AI Pulse introduced Apollo Go's progress in two right-hand-drive markets: it completed its first media test ride in London and held its first large-scale public test ride event in Hong Kong. The article also previewed Baidu World 2026 on November 12.

Quoted postBaidu Inc. · @Baidu_Inc

https://t.co/hvqf0fQBUx

#Industry news#Embodied AI