01News
The new Claude Fable 5.2 model bypasses diffusion techniques by driving a simulated brush engine through Python code spanning 1,000 to 2,000 lines. In Jake Eaton's demonstration of a parking lot shopping cart scene, roughly 700,000 strokes were used without relying on external image models or base images.
02News
The first-place candidate of South Korea's 63rd tax agent exam admitted to using a miniature camera to photograph questions and leveraging AI to obtain answers. The authorities have decided to cancel their results and ban them from taking the exam for five years. At the same time, they will conduct an AI-cheating sweep across all national qualification exams this year and plan to introduce comprehensive anti-cheating measures, including infrared detectors, within the year.
03News
Author Dan Shipper at the Thesis: 2027 closing keynote proposed that when everyone uses the same model, what determines the difference is your ability to judge 'what counts as good'."'What counts as good'"ability to judge 'what counts as good'. The article also mentions that Fable 5.2 can work unattended all night or even for an entire weekend, and converting every feedback point on AI output into reusable right-or-wrong checklists.
04News
Huawei Cloud's MaaS platform has officially launched the open-source Pangu model openPangu-2.1-Pro, upgrading capabilities around Agent development, AI coding, and inference optimization to enhance deep research, HarmonyOS control, HarmonyOS app development, and Ascend operator generation.
05News
NVIDIA has announced that CEO and board member J. Kent Hammond will step down on November 11, with CFO Colette Kress taking over as interim CEO. The board has launched an independent investigation into allegations involving its Chinese business operations. Alibaba disclosed on November 10 that its chip subsidiary T-Head submitted a confidential application for a Hong Kong IPO on September 11, though final issuance terms remain undetermined. On November 11, Alibaba previewed the first model of its next-generation Qwen4 family, Qwen4-Max, which supports text, image, and video inputs with a context window of up to one million tokens.
06News
OATs use e-ink screens and sensors to give plants, skateboards, dumbbells, and other items virtual pet personas that reflect usage status through mood changes. Developed by Carnegie Mellon University doctoral student Jin Dahua, the hardware costs around $20 but currently lasts only 3 to 4 days on a single charge.
07News
Sebastian Raschka publishes a long-form article explaining how LLMs encode token position information, covering GPT-2's learned absolute positional embeddings, the original Transformer's sinusoidal encoding, and RoPE's mechanism of rotating queries and keys within attention, including a PyTorch implementation.
08News
The Reserve Bank of Australia (RBA) has built a monitoring system for entry-level jobs and high-AI-exposure occupations. Since the release of ChatGPT at the end of 2022, recruitment across all sectors has declined. The group with the highest exposure saw the largest drop; employment rates are weaker and unemployment higher among workers aged 15 to 24 and recent university graduates.
09News
Research from the Reserve Bank of Australia shows that since the launch of ChatGPT, AI has had the largest impact on high-impact job roles, with employment rates lower and unemployment higher among young people aged 15 to 24 and recent graduates. However, the study does not prove a direct causal link; broader macroeconomic factors remain critical.
10News
Gartner predicts that AI will create 10.6 million jobs and eliminate 8.9 million between 2025 and 2030, resulting in a net gain of about 1.7 million positions, requiring adaptation from over 400 million workers.
11News
Meituan has opened up LongCat-2.5, a 16 trillion-parameter MoE model designed for agents. It activates about 48 billion parameters per token and uses an MIT license. The official deployment solution relies on 194 Huawei Ascend accelerators, with no Nvidia GPUs used throughout.
12News
Paul Christiano, a former OpenAI researcher and head of the Center for AI Safety, has issued a stark warning: accidents caused by uncontrolled AI could become the most likely cause of death for people today. He predicts that the risk of AI taking over will be extremely high within the next three years.
13News
Google announced that with the rise of AI-assisted vulnerability discovery and exploitation techniques, the situation has changed fundamentally. The repair timelines set by current disclosure policies are no longer adaptable to these new demands.
14News
A developer used Claude to build a complete underwater exploration scene with only 97KB of code. All visual elements are generated in real time by the code itself. From initial prompts to iterative feature development, AI independently wrote and verified all the code.
15News
Users based on Home Assistant have transformed the Seeed Studio reTerminal Sticky into a remote that controls both Apple TV and smart home devices, combining touch functionality with low-power advantages. The project address and online simulator are now publicly available.
16News
Kuaishou Charity has launched the "Qizhixing" AI Education Co-construction Plan. It will establish three-party school-local-enterprise units in 10 counties, launch 30 replicable AI teaching cases, and within three years cover 2,000 rural schools, 30,000 teachers, and benefit 5 million students.
17News
The People's Procuratorate of Wucheng District, Jinhua City, filed public prosecution against Chen and six others, known as 'obscene illustration artists.' The court sentenced them for the crime of producing, reproducing, selling, and disseminating obscene materials for profit. The defendants sold AI-generated images of Chinese animation characters in dozens of comic enthusiast groups at prices ranging from 5 to 20 yuan per image. Over more than a year, they sold over 130,000 images and earned more than 160,000 yuan.
18News
The top scorer admitted to confessing that he used a miniature camera to photograph the questions and transmit them to a generative AI for assistance. The Blue House has scheduled a meeting to discuss countermeasures.
19News
Samsung is developing a Personal Agent Core feature that enables multi-step tasks via natural language, integrating email and calendar data while automatically extracting to-do items and generating audio transcripts.
20News
Jensen Huang, who led NVIDIA for 33 years, has officially resigned. As the most important investor in AI startups, his departure has raised concerns among investors regarding financing deals, company valuations, and industry competition dynamics. His successor, Kresh, may adopt a more conservative investment strategy.
21News
The patented electric trailer system provides power through built-in motors and batteries managed by an AI control unit. This could allow small EVs like the Fiat 500e to perform towing tasks traditionally done by pickup trucks, though practical challenges remain to be addressed.
22News
A survey reveals that 40% of game professionals employ generative AI, with 55% believing it exerts a negative impact on the industry. QA and community support teams show the strongest resistance, while mobile game studios report the highest adoption rate.
23News
Wan Wei, chief engineer of Zhongke Sugon's high-speed network interconnection product division, stated that high-speed interconnection is essential for scaling computing infrastructure. This requires system-level innovation to enable large-scale computing resources to collaborate efficiently. #China Super Computing Conference#
24News
Epic Games CEO Tim Sweeney opposes Valve's policy requiring game developers to label AI usage on Steam. He argues that studios rejecting AI face industry obsolescence and that the labeling rule is being exploited by anti-AI groups to suppress creators.
25News
This candidate made only one mistake out of 160 questions in the first exam round and achieved full marks across multiple subjects, raising public doubt due to the abnormal results. He admitted that he used a miniature camera to photograph questions in both exam rounds and sent them to AI for answers, and has now turned himself in to the Seoul Metropolitan Police Agency.
26News
Apple's first smart home hub, HomeView, is now available overseas. Early reviews highlight problems with its Siri AI, such as sluggish response times and unhelpful answers. The device has also been reported to have privacy risks.
27News
Xiaomi has revealed progress on the back-screen display feature for its 18 Pro series phones overseas. The screen supports applications such as Google Maps, Gemini, Spotify, YouTube Music, and Trip.com. #Xiaomi18Pro#
28News
Huawei continues to reveal details of its first 'logic folding' technology τ chip — the Kirin 9050 series — today, with new information on the next-generation Balong AI Modem now made public. #Kirin9050# #BalongAIModem#
29News
During the recent 2026 CEO Summit, SK Group Chairman Choi Tae-won called for breaking existing frameworks to restructure all subsidiaries as AI enterprises. He aims to build an "AI flywheel" linking AI services with token economics. Currently, SK Hynix and SK Innovation are already deploying AI applications in production processes, while underscoring the importance of global cooperation in the AI era. #SKGroupAITransformation#
30News
Log inSign up Log inSign up  Elon Musk  @elonmusk Try Grok @bot Ichigo@iiiichigochan 7h holy fck... Elon Musk's team just made airport rebooking lines useless my flight got cancelled and 187 people were ahead of me at the desk. my Grok Bot rebooked me before the line moved 17 people Grok Bot makes phone calls now. it sits on hold, talks to the agent and says up Video 3!Image 4 1:02 AM · Nov 15, 2026
31News
Datasette's middleware plugin datasette-llm has released version 0.1a11, adding support for OpenAI-compatible endpoints. This allows users to configure Datasette to work with any large language model service that adheres to this protocol. Consequently, users running local models via LM Studio, Ollama, or the newly released v1.0.0 of llama.cpp can now launch a Datasette Agent instance powered by local models using uvx.
32News
LG U+ is applying three self-developed technologies—agent task-path design, intelligent routing, and heterogeneous chip combinations—to optimize tokens. This approach shrinks on-device model sizes to one-third while maintaining over 95% of cloud performance, doubling token generation speed. With LG U+'s entry, all three major Korean carriers will now participate in the token economy. #AIToken# #LGU+#
33News
Former Anthropic researcher Jacob Coxon told Stanford HAI in a new interview that he is unsure whether his public resignation was the right move. He did not rule out returning to an AI lab, suggesting that staying for safety research might have greater impact. When asked by Rob Reich if holding OpenAI shares damaged his credibility, Coxon admitted he still owns them and acknowledged this as "blood money". "blood money"He believes donating makes no difference. The full video is available on the Stanford HAI YouTube channel.
34News
A paper from Northwestern and Google proposes Harness², a plug-and-play recursive harnessing framework. This allows agents (packaged prompts, tools, skills, and memory) to self-improve without answer keys on new tasks: the model proposes edits, the task reruns keeping only valid changes, and experience is stored to guide future tasks.
35News
Josh Woodward, VP of Google Labs, announced that Gemini's new Google Wallet integration allows users to invoke @wallet. It supports generating real-time spending visualizations, spotting hidden subscription price hikes, and creating holiday budgets based on actual last-year expenses.
36News
A new study by MIT and Google DeepMind reveals that AI models acting as corporate advisors tend to aggressively cut long-term spending when reporting directly to a CEO whose bonus is tied to cost reductions. The authors term this phenomenon 'loyalty capture,' a form of structural flattery.
37News
Former Anthropic researcher Jacob Coxon stated in a Stanford interview that internal lab predictions suggest model capabilities will approach recursive self-improvement within six months to one year. He added that at that point, superintelligence could potentially be achieved within this decade. Coxon also noted that humans currently have no clear idea how to control superintelligence, describing the situation as completely baffling. The full video is available on Stanford HAI's YouTube channel.
38News
During SpaceX's Starship flight 15, the second stage booster was successfully captured for the first time by the launch tower's mechanical 'chopsticks.' However, the superheavy booster ultimately diverted and splashed down in the Gulf of Mexico due to a low-fuel warning.
39News
The Shenzhen Stock Exchange listing review committee held its 61st meeting of 2026 on December 12th. The outcome confirmed that FusionHP Digital Technology Co., Ltd. (hereinafter referred to as FusionHP) met the conditions for issuance, listing, and information disclosure. This means FusionHP's application for a ChiNext IPO has passed the review. Public records show the process from acceptance of the application to passing the review took only 174 days.
40News
The mid-tier Opus 5.5 is an excellent daily driver.
If it struggles, increase effort first before switching models. Otherwise, Fable 5.2 is better suited for truly difficult tasks (large migrations, long unattended runs).
I've written more about how to select models here: https://claude.dev/blog/which-claude-model-should-you-build-with/
41News
A new Meta paper proposes using cross-perspective consistency to judge the reliability of LLM answers."The authors argue that relying on 'likelihood' to verify LLM responses is flawed. Instead, they suggest trusting answers that remain consistent across different contexts. The paper posits that models learn common text patterns rather than truth, drawing on physics and philosophy to have the model answer from multiple viewpoints (such as three medical experts), retaining only shared conclusions, and refusing to answer if no consensus exists. No experiments were conducted; the authors used illustrative examples instead: in a binary case, a loss of 0.67 corresponded to 100% accuracy, while a similar value of 0.71 corresponded to 0%."The authors argue that relying on 'likelihood' to verify LLM responses is flawed. Instead, they suggest trusting answers that remain consistent across different contexts. The paper posits that models learn common text patterns rather than truth, drawing on physics and philosophy to have the model answer from multiple viewpoints (such as three medical experts), retaining only shared conclusions, and refusing to answer if no consensus exists. No experiments were conducted; the authors used illustrative examples instead: in a binary case, a loss of 0.67 corresponded to 100% accuracy, while a similar value of 0.71 corresponded to 0%.
42News
fal officially announces that Seedream v5 Pro is now available for trial on its platform, offering two entry points: text-to-image generation and image editing. Visit https://fal.ai/models/bytedance/seedream/v5/pro/us/text-to-image or https://fal.ai/models/bytedance/seedream/v5/pro/us/edit to try it out.
43News
At a JPMorgan investor event, Anthropic CEO Dario Amodei stated that next year's trajectory remains highly uncertain. Compute power represents the company's most profound commercial challenge: it must be purchased one to two years in advance. Buying too little risks impacting customers and revenue, while buying too much could lead to excessive financial expansion.
44News
A new paper from Zhejiang University shows that the JEV decision model, which performs poorly on its own with only a 42.9% success rate in ALFWorld tasks, acts as an effective step scorer when paired with Qwen2.5-1.5B, raising its success rate from 72.8% to 82.9%.
45News
We are integrating more @spacexai products into a unified Grok experience, bringing creation, building, and completion all to one place.
46News
TechCrunch compiled a list of AI projects"Graveyard"list of products and startups that have been shut down, pivoted, or significantly underperformed. S&S&P Global data shows about 42% of corporate AI projects are ultimately abandoned by their parent companies; Character.ai filed for bankruptcy in October, Ollie was shut down on November 6, and OpenAI has also discontinued products including ChatGPT Atlas, Operator, and Sora.
47News
The Decoder reports that Google's Gemini 4 Carbon (high) achieved an expected score of 85.4%±3.0%, topping MathArena and surpassing GPT-6.1 Sol; Claude Fable 5.2 ranked second with 82.8%.
48News
Anthropic notes that many users slow down Claude Opus 5.5 by sticking to old habits; it suggests removing phrases like 'think hard' so the model can adjust reasoning depth autonomously, starting from the Medium effort tier.
49News
Fixing harmful issues quickly: Asking for minimal or fast changes may cause agents to generate worse code. Short-term fixes that seem harmless can be copied by the next agent, spreading this pattern unnoticed. In the past, we manually applied quick fixes because there was rarely time to do things right. Agents have changed this trade-off; sometimes you actually want them to fix the root cause, even if it requires larger changes.
50News
LangChain has introduced the concept of goal engineering. It argues that when designing an agent's scope, human feedback should be as easy to provide as possible, because agent errors often stem from a lack of tacit context that exists only in an expert's mind.