NTrending

AI trending

Trending

Last 48 hours, most discussed AI stories, top 20 stories

Nov 2 13:20 Updated
NO.01

The DeepSeek Harness team shares its Super-agent design exploration

On the evening of October 31, 2026, the DeepSeek Harness team released "Super-agent Design Exploration: Toward More Autonomous Agent Systems," proposing the Super Agent concept and open-sourcing the prototype dsh-superagent-prototype on GitHub. According to Sina Tech, the architecture consists of five modules: a two-layer Router/Worker, event-driven operation, asynchronous multi-channel interaction, storage-compute separation, and an external Watchdog. The goal is to enable Agents to stay continuously online, run multiple tasks in parallel, and proactively respond to environmental events; in preliminary experiments, the p50 task completion time was 12.4 minutes and the p95 was 15.7 minutes, and Router batching reduced the inference time allocated to a single decision from 6.4 seconds to 1.6 seconds. TMTPost·AGI reported that the article splits tasks using two roles, Router and Worker, triggers work from activity such as group chats, emails, and code checks through two filters—rules plus models—and emphasizes notifying only when results are usable or a human decision is needed. The team stated that the prototype is only a research prototype and not an official DeepSeek product; it currently mainly targets a GitHub repository and can also connect to Feishu, and some designs have not yet been implemented.

NO.02

Karpathy builds HackerTube with Claude Opus 5.5 in two hours

Over the weekend, Andrej Karpathy used Claude Code together with Opus 5.5 to filter 43 million Hacker News comments spanning twenty years down to 6,454 frequently recommended long videos, and built HackerTube, a video site that requires no login, in about two hours. According to a report by A Yi AI Notes, the site's cover images are taken directly from the middle frame of each video rather than the optimized thumbnails uploaded by creators, and browsing state is saved in the local browser session. Karpathy only manually applied in the browser for a Google Cloud YouTube API key, rented a VPS and bought a domain, after which he handed the IP to Claude and let it complete deployment, security hardening and code optimization. indigo's report adds that the comment analysis data was packaged into a zip file, allowing Claude to generate its own version based on it. The project is currently still in the post-release showcase stage, with no further development progress yet.

5 sources · 10 participants
79
NO.03

OpenAI unveils its AI policy agenda after the midterm elections

After the U.S. midterm elections, OpenAI plans to launch a public relations campaign that includes renaming data centers as "AI Compute Campuses." OpenAI Chief Global Affairs Officer Chris Lehane said the company is also considering setting "speed limits" for AI development, helping the public more intuitively understand how AI improves daily life, and pushing for U.S. participation in setting global AI standards. The report also noted that since the start of this year, multiple data center projects with a total value of up to $100 billion have been forced to delay or stall. The plan is currently at the disclosure stage, with no further progress on implementation.

7 sources · 7 participants
61

More stories No.04–20

Participants · 24-hour trend · Heat index

  1. 04

    OpenAI announces that ChatGPT and Codex share a single token allowance

    OpenAI announced that ChatGPT and Codex for the Plus, Pro, and Business plans will share a single token allowance, with the change taking effect in December. According to a report by Rohan Paul, OpenAI stated on its official documentation page that it will then merge the agent features of ChatGPT Work, an agent built on Codex in July, into the main ChatGPT experience; currently Work already shares limits with Codex, while regular ChatGPT conversations are metered separately, and after the switch each request will be priced by token. Phoenix Technology reported that OpenAI sent a notice to users explaining that allowance consumption depends on the selected model and task complexity, that using Instant mode chat does not consume the shared allowance, and that users can view their combined remaining allowance and reset time on the usage page in settings; once the allowance is exhausted, users can wait for the reset, upgrade their plan, or purchase additional allowance. OpenAI also plans to integrate the chat and work tabs, testing first with a small group of users, and Codex will still retain a separate experience in the desktop app.

    7 sources56
  2. 05

    Meta collaborates with Fourier to develop humanoid robot prototype Max

    Meta has partnered with Chinese robotics company Fourier to develop a humanoid robot prototype named Max. The two sides completed the first-generation model in about three months and in September of this year launched development of the second generation, Max 2. According to Jiwei.com, Max is about 1.5 meters tall and weighs about 48 kilograms, uses Nvidia's Jetson Thor chip, and its hardware foundation comes from Fourier's previous three generations of robot products; Max 2 will redesign the torso, improve the mechanical design of the hands, and add a projector and sensors that can climb up and down stairs. 36Kr AI reported that Max is customized based on Fourier's GR series hardware and is used to test the performance of Meta's agent Muse in real-world environments. A Meta spokesperson said that the collaboration is currently limited to internal research and experimentation, and there are no sales plans at this stage; the company has recruited for positions such as supplier development and product design in Shanghai and Shenzhen.

    4 sourcesInsufficient trend data56
  3. 06

    Sam Altman discusses the DRAM utilization issue in local inference

    On October 30, 2026, OpenAI CEO Sam Altman put forward a view on the podcast "Cheeky Pint": having everyone leave globally scarce DRAM memory idle in personal computers at only 5% utilization is quite bad from the perspective of global resource allocation; memory should be concentrated in cloud data centers to achieve maximum utilization. He also called the progress in resolving the memory shortage "terrible," arguing that memory manufacturers are conservative in capital expenditure due to historical loss cycles, and that the "invisible hand" is failing. There are currently no subsequent developments.

    11 sources41
  4. 07

    poteto used Grok Bot and Cursor to build a web game in four days

    lauren (@poteto) bought the domain for $5.99 and, in four days and with 1,150 PRs, used Grok Bot and Cursor to build an embeddable web game for X called "Doge Squad vs. the Endless Swarm." When meng shao reposted it, they called it "Day 28/ Tibo Reset" and remarked that it was "maximum mockery." The game is now complete and can be embedded in X.

    1 sources34
  6. 09

    Peter Steinberger recommends using your own harness

    On October 30, 2026, Peter Steinberger commented on the new Claude Code rules, advising developers to use their own harness rather than a harness whose permission scope the company can change at will. Earlier posts said that around October 27, Claude Code distributed a rule named "tengu_heron_brook" through Anthropic's servers, and it cannot be turned off: original games can be built, but specific existing games must not be replicated. Subsequent reports said that Anthropic uses this server-side switch to prohibit replicating specific existing games, even if the name is changed or the work is claimed to be original, and there is no option to turn it off; the rules still allow original games, games of the same genre, and projects based on users' own designs and assets. Other reports said that Anthropic confirmed adding the relevant rules to the Opus 5.5 system prompt, and after the ban took effect, some developers creating original games reported in GitHub issues that they had been mistakenly judged and rejected.

    7 sources26
  7. 10

    English text is being prepared.

    Gabriel Ocana-Santero described that the Exin Therapeutics team had the Claude agent autonomously select a gene therapy and perform brain surgery on epileptic mice. According to Phoenix Technology, the team modified a surgical platform with Arduino, and what ultimately ran was Sonnet 5; the agent selected one of 19 gene therapy vectors, set the injection point at the lateral entorhinal cortex, took coordinates from the average of three papers, injected 250 nanoliters of solution within five minutes, and took screenshots to observe 32 times throughout. The team's initial choice, Opus 5.5, refused to run as soon as it detected the task involved living organisms, and Opus 5 stopped at the in vivo step after simulating the full procedure; the craniotomy was still performed by humans, with manual safeguards such as a veterinarian on site and a physical emergency stop switch. A subsequent IT Home · AI report likewise stated that the craniotomy and suturing were performed by human doctors.

    4 sources26
  9. 12

    OpenAI lowers its latest funding round target by $10 billion

    OpenAI has lowered its latest funding round target from $30 billion to $20 billion, a cut of $10 billion; at the same time, it has lowered the valuation target for this round from $1.4 trillion to $1.25 trillion. According to a 36Kr·AI report, as of the Friday commitment deadline, actual investment commitments stood at about $11 billion, less than half of the original target. This funding round is led by a UAE consortium, with UAE state-owned technology funds such as MGX as core investors, and with participation from Founders Fund, Andreessen Horowitz, SoftBank and others; OpenAI plans to expand its fundraising scope to family offices and high-net-worth individual investors. East Money reported that, because some investors have reduced their subscription amounts, the actual funds raised at the first closing are expected to be less than $20 billion. There is currently no further progress.

    7 sources24
  10. 13

    Apple announced that the iOS 27.3 developer beta will be released next week

    Apple announced that the first iOS 27.3 developer beta will be released next week. Phoenix Technology reported on October 31, 2026, citing an announcement on Apple's official website, that Qwen will be integrated into Siri, Writing Tools, and Visual Intelligence, and will be rolled out with iOS 27.3 in December; this update does not include Siri AI or the next-generation features announced at WWDC. X.PIN reported on November 1, 2026, that the first batch of Apple Intelligence features will enter China in December, with Alibaba's Qwen providing answers in Siri, Writing Tools, and Visual Intelligence; Qwen is an optional opt-in, and Alibaba says it will not use requests and responses to train models, and will seek consent before sending any question, document, or image. There is currently no further follow-up progress.

    10 sources24
  11. 14

    The U.S. Department of Justice announces the results of Operation Ghostbusters

    The U.S. Department of Justice announced the results of "Operation Ghostbusters," an investigation into student loan fraud in which stolen identities and artificial intelligence were used to complete applications and pose as students enrolled in online courses. Phoenix Technology reported on November 1, 2026, that the U.S. government is estimated to have lost more than $200 million as a result, that five people have been charged so far, and that more than 40 open cases involve 26 FBI field offices and 20 U.S. Attorney's Offices. cnBeta reported on November 2 that the FBI has confirmed related losses exceeding $32 million, that one defendant obtained about $1.44 million through more than 700 fraudulent refunds, and that the Department of Education has identified more than 830,000 suspected fraudulent aid accounts. Earlier reports said losses exceeded $200 million, while later reports said the FBI confirmed losses exceeding $32 million.

    3 sources23
  12. 15

    lauren shares a guide to creating Grok bots and managing teams

    lauren (@poteto) released a Grok bot usage guide, explaining how to create high-quality bots and manage a team. She recommends first adding her own bot designer, dr eggbot, keeping the number of bots under control, letting bots build their own skills during tasks and store them in a private GitHub repository, connecting tools through connectors and vaults, having bots register independent email addresses and accounts to isolate permissions, and using voice mode to dump context and then having the bot repeat it back to confirm understanding. Another paraphrase also mentions that the guide argues bots should not be numerous—a few good ones are easier to manage than a whole bunch—and that when teaching a bot to do something, you should have it save the method as a skill to keep it stable in long conversations, and most recommends speaking your ideas all at once by voice and then having the bot restate them in its own words to confirm; commenting @bot under the original post gets you a personalized optimization guide. A Yi AI Notes and indigo each paraphrased this guide.

    3 sources21
  13. 16

    Juqiao Industry announces completion of 100-million-yuan Series A financing

    On October 29, 2026, Juqiao Industry announced the completion of a 100-million-yuan Series A financing round. According to a November 1 report by Zhidongxi, this round was led by Ant Group, with Lingyi Venture Capital, Weihai Gangrong, and Qingdao Hexin participating as follow-on investors, and the funds will be used for the large-scale mass production and delivery of fabric electronic skin. The report stated that Juqiao Industry's tactile gloves are each equipped with hundreds of sensing points per hand, with a minimum point spacing of 2mm and a minimum trigger force of 0.01N, and have achieved delivery at a scale of nearly 10,000 units, serving more than 400 customers, including more than 200 embodied intelligence customers; its own production line in Weihai has a production yield rate exceeding 95% and can rapidly produce samples within one day. A November 2 report by 36Kr·AI likewise mentioned that Ant Group led the round and stated that this is Ant's second move in the tactile sensing track. Currently, there is no public information on further progress following the financing.

    3 sources21
  14. 17

    Hidden mention of Claude Mythos 5.5 appears on Anthropic's official website

    On November 1, 2026, TestingCatalog reported that a hidden mention of Claude Mythos 5.5 was discovered on Anthropic's official website, with the related description being "the strongest model category for cybersecurity and biology research." The post cited Chubby's claim that Mythos 5.5 had already appeared in the website's HTML source code; based on this, TestingCatalog speculated whether a Mythos 5.5 aimed at cyber defense would be launched next week, and surmised that Fable 5.5 might also arrive alongside it. Currently there is only this one report; Anthropic has not responded, and there have been no further developments.

    1 sources20
  16. 19

    UBTECH Releases the Thinker-Sim Embodied Intelligence Simulation Platform

    On November 2, 2026, UBTECH released Thinker-Sim, an embodied intelligence simulation platform centered on Real2Sim2Real, covering the digitization of real-world scenarios, large-scale synthetic simulation data generation, and robot capability evaluation. According to a Sina Tech report, the platform's rendering efficiency has improved by 67.3%, a single node (8 GPUs) can synthesize an average of 390 hours of simulation data per day, and the efficiency of the self-developed communication module has increased by more than 400%; the platform comes with built-in standardized task templates, and its visual workflow can adapt to more than 90% of target simulation scenarios with zero code. The Beijing News reported the same day that the platform has built-in standardized simulation task templates, through which users can launch simulations simply by selecting templates and configuring parameters, and that it supports real-time video streaming and task status monitoring; based on the self-developed Ulab-Agent, the platform uses natural language as its unified interaction entry point and can adapt to more than 90% of target simulation scenarios. No further developments have been disclosed at present.

    2 sourcesInsufficient trend data18
How heat is calculated

Heat is a time-decayed index over a 48-hour window, counting each independent participant's best contribution once. It is not a count of articles, views, likes or participants. Compare values within the same ranking only. Missing intervals between real observations are interpolated; observed zeros remain on the baseline.