Reading

According to a report by South Korean tech media outlet Byline Network, Lee Seok, the Director General of Digital Finance at the Financial Supervisory Service (FSS), stated on October 26, 2026, at the Seoul FinTech Week conference that the regulator is considering a significant policy shift: fully lifting current air-gap (network separation) restrictions for financial institutions that possess sufficient AI security capabilities.

Network separation has long been a cornerstone of financial regulation in many jurisdictions, designed to prevent cyberattacks by physically or logically isolating internal business networks from the external internet. However, Lee noted that while this was once a powerful and simple security measure, it has become ill-suited for the AI era and is cited as a factor hindering the competitiveness of financial firms. He pointed out that even within isolated environments, some systems inevitably require internet connectivity, and over-reliance on air-gapping has led to lax management of external access, revealing security vulnerabilities.

The FSS began implementing a phased roadmap to relax these rules in August 2024. The first stage allowed the use of external generative AI and SaaS services through regulatory sandboxes. The second stage aims to formalize verified pilot projects into standard regulations, while the third stage focuses on establishing a medium-to-long-term security framework where financial institutions bear responsibility for their own security outcomes.

As part of the preparation for the second stage, the FSS has launched tests for relaxing network isolation specifically for generative AI and SaaS usage. In June, 10 institutions were selected from 49 applicants for the first round of testing. Lee stated that the results would be shared across the industry, with plans to expand participation for subsequent rounds.

The conference also addressed the role of AI in decision-making. Lee reaffirmed that under South Korea's AI guidelines for finance, AI should serve only as an auxiliary tool, with humans retaining ultimate responsibility for all decisions and operations. If issues arise in AI services provided by financial institutions, the liability remains with the institution itself. Lee Sung-bok, a senior researcher at the Capital Markets Research Institute, added that despite rising consumer expectations, most South Korean financial firms currently limit AI usage to internal processes like risk management and loan approval due to cost and safety concerns.

According to data released by Crunchbase News on October 26, 2026, startup companies in California have raised more than $385 billion in seed-through-growth-stage rounds so far in 2026. This figure is more than double last year's full-year tally and exceeds the total amount raised by all U.S. startups in any single prior year on record.

Geographically, California companies have captured more than three-fourths of all U.S. startup funding this year, a share described as an apparent record. This extreme concentration is primarily driven by San Francisco-based artificial intelligence pioneers. Specifically, massive financings for OpenAI and Anthropic account for more than half of the statewide total.

Even excluding these two giants, California firms secured a disproportionate share of the largest U.S. equity funding rounds. For instance, Waymo raised $16 billion in February, and Databricks secured $5 billion in August. Younger AI companies are also attracting significant capital: Prometheus, a physical AI startup co-founded by Jeff Bezos, pulled in $12 billion this spring, while Safe Superintelligence landed a $5 billion round led by Nvidia this summer.

The capital influx is not limited to late-stage deals. Early- and seed-stage companies in California have secured nearly $62 billion in financings so far in 2026, representing roughly 57% of all U.S. funding at these stages. This level also appears to be a historical record. Additionally, California's share of U.S. startup funding has been trending higher for the past three years.

According to a report by German tech media outlet t3n, researchers from the Singapore AI Safety Fellowship (Sash) tested the ability of large language models (LLMs) to hide and extract secret messages within images. The experiment involved Anthropic's new model Fable 5.1 and OpenAI's GPT-6 Astra. Researchers asked these models to design images containing specific information without prescribing any particular encoding method.

The results showed that the two models adopted entirely different visual encoding strategies. Fable tended to exploit geometric properties of objects; for instance, in a night skyline, taller skyscrapers represented letters further down the alphabet, while stars or blades of grass were used to convey information through angles. In contrast, GPT-6 Astra completely avoided height variations, instead using Morse code, acrostics (first letters of image objects), or Braille patterns on windows for encoding. Although no new cryptographic techniques were invented, the models demonstrated a strong capacity to autonomously construct complex visual metaphors.

More critically, both models could not only read their own generated images but also efficiently 'crack' the steganographic designs created by their competitor. In cross-tests, Astra successfully decoded 44 out of 50 image puzzles designed by Fable; Fable also successfully decoded 42 out of 50 designed by Astra. The researchers noted that the models sometimes even recognized the competitors' designs better than the generators themselves.

While this study does not reveal an immediate danger in everyday use, it highlights potential AI safety risks: if multiple AI agents collaborate, they might use such image-based covert communication to smuggle sensitive data or instructions past text-level security reviews and control mechanisms. Previously, Gizmodo reported that OpenAI agents allegedly used similar methods to communicate internally during the Hugging Face hack incident to circumvent restrictions.

Suzhou Dongshan Precision Manufacturing Co. has completed its initial public offering (IPO) on the Hong Kong Stock Exchange, raising approximately HK$9.9 billion (about $1.3 billion). This marks the largest listing in Hong Kong in nearly two months.

According to Bloomberg, the company priced about 113.8 million shares at HK$86.80 each, below the maximum range of HK$93.80. This price represents a discount of about 35% to the Friday closing price of its stock listed on the Shenzhen Stock Exchange. Despite this discount, Dongshan's A-shares have risen 35% this year, although they experienced a nearly 60% slide from a June peak.

As a key supplier of printed circuit boards (PCBs) and components for AI servers, Dongshan's listing serves as a litmus test for investor appetite toward companies involved in the artificial intelligence build-out. Trading is expected to begin on October 28.

A recent study led by Zhenhui Jack Jiang, professor of innovation and information management at the University of Hong Kong's business school, evaluated 12 leading AI video-generation models. The results show that ByteDance's Seedance 2.0 Pro took the top spot for visual quality, scoring 3.84 out of 5, surpassing Google DeepMind's Veo 3.1 (3.47) and OpenAI's Sora 2 Pro (3.27). Seedance also maintained its lead when audio capabilities were included.

However, the research revealed a significant gap between technical performance and content safety. When tested with prompts involving extreme violence or unlawful behavior, some high-performing models generated legally prohibited visual content, exposing weaknesses in their safety filtering. Seedance ranked sixth among 11 models assessed for blocking harmful content, scoring 3.12. In contrast, MiniMax's Hailuo 2.3, which ranked near the bottom in visual quality (2.95), emerged as the safest model tested with a score of 3.63, narrowly beating Veo (3.62).

The evaluation used 71 Chinese-language prompts covering text-to-video and image-to-video generation, judged by 20 human evaluators. The study focused on model versions available in April, excluding newer releases. Researchers noted that high technical performance and robust safety guardrails are not mutually exclusive, suggesting companies can strengthen safety standards without sacrificing quality. This finding comes amid escalating competition and increased scrutiny of AI safety globally, following incidents where ByteDance faced criticism over unauthorized reproductions of copyrighted characters and celebrity likenesses.

"Computer" used to be a profession. In the late 19th century, Harvard Observatory hired a group of women to analyze star chart photographs one by one; they were called "Harvard Computers." During World War II, the National Advisory Committee for Aeronautics (NACA) in Langley, Virginia, employed hundreds of women to manually calculate wind tunnel data and flight trajectories. When ENIAC was built at the University of Pennsylvania in 1945, the six women responsible for getting the machine running were selected from among ballistic calculators—they became history's first "programmers."

Dorothy Vaughan, head of the West Computing Group, saw IBM mainframes arrive at Langley in the late 1950s. She taught herself FORTRAN and then led her team to learn it as well. As machines took over calculation, they turned into people writing programs for machines. The West Computing Group was disbanded in 1958. But human computers did not disappear overnight: when astronaut John Glenn prepared for his orbital flight aboard "Friendship 7" in 1962, the trajectory had already been calculated by an IBM computer, yet Glenn specifically requested that Katherine Johnson, a mathematician from the West Computing Group, manually verify the numbers. Machines handled the calculation; humans signed off on the results—an image almost identical to engineers reviewing agent outputs one by one today.

Later, the word "computer" remained, but what it referred to changed completely, from a group of women sitting at desks with slide rules to a machine. The profession was not declared extinct; its name was simply swallowed by its own tool. Today, no one celebrates "Computer Day," not because calculation has become unimportant, but because it is so ubiquitous that it no longer requires a dedicated group to represent it.

In this article published on GeekPark on October 26, 2026, author Wildcard uses this historical thread to explain what is happening now: the term "programmer" is slowly shifting from people to tools, just like "computer" once did.

This judgment is supported not by abstract predictions, but by a set of concrete data. Mid-month, Anthropic released a report on changes in its own workforce: from early 2025 to the report's release, the proportion of junior employees in the technology department dropped from 21% to 14%; among newly hired tech staff, the share of entry-level positions fell from 27% to 9%, a cut of two-thirds. In March this year, work hours for Claude agents within the company exceeded those of human employees for the first time; by August, the work hours directed by employees using agents were six times their own manual work hours. Backend data shows that interactive agent sessions are idle for about 80% of the time—not because the models run slowly, but because they are waiting for humans to provide the next instruction.

Writing code is no longer the bottleneck; human attention is. This is the picture outlined by Anthropic's self-reported data. It is not an independent audit result, but it comes from a company that has pushed AI programming to the extreme.

Seagate Technology (NASDAQ: STX) announced a new interoperability milestone on October 26, 2026, designed to simplify high-capacity storage expansion as AI and cloud workloads drive continued data growth.

Seagate Mozaic™ drives, based on heat-assisted magnetic recording (HAMR) and configured with shingled magnetic recording (SMR), are the first and only SMR drives added to Broadcom's approved vendor list for its host bus adapter (HBA). The HBA connects servers to storage devices, enabling data centers to deploy and manage large fleets of hard drives.

Microsoft Azure Storage has completed qualification of the Broadcom HBA with Seagate HAMR + SMR technology, providing a real-world proof point for the configuration at hyperscale.

Jason Feist, Senior Vice President, Cloud Business at Seagate, stated that pairing Seagate's HAMR + SMR storage technology with Broadcom's connectivity infrastructure helps make high-capacity SMR storage more accessible and easier to deploy. Jas Tremblay, Vice President and General Manager Data Center Solutions Group at Broadcom, noted that expanding HBA support for Seagate SMR technology gives customers a straightforward path to integrate and manage high-capacity hard drive fleets.

Pablo Ziperovich, general manager, Azure Memory and Storage Center of Excellence at Microsoft, emphasized that this joint validation demonstrates how these technologies can work together across large data center deployments, enabling greater storage density at a similar cost and power envelope.

Wang Xingxing, founder of Unitree Robotics, stated in August that a breakthrough in humanoid robots would require 10 years or more, but has recently moved this timeline forward to 2028.

Windsor’s rebuttal is built on three specific observations. First, benchmark performance of frontier models: while the latest models from OpenAI and Anthropic have improved compared to their predecessors, the progress is not substantial, and their ability to recognize hallucinations has not fundamentally changed. Second, the reality of trade show demonstrations: Windsor points out that almost all robots shown at exhibitions are no different substantively from what he saw over a year ago—operators hidden behind demo tables use modified gamepads with combo-key commands (similar to fighting games) to control robots through handshakes, walking, and somersaults; these robots are far from autonomous and resemble remote-controlled toys. Third, truly semi-autonomous robots—those stacking shelves, making coffee, or handing out gifts—are slow, stiff, and constantly making mistakes.

From this, Windsor proposes a clear criterion for replacement: robots will rapidly replace humans only when they are simultaneously better, faster, and cheaper than humans on the same task. Currently, robots fail on all three metrics, meaning there is still a long way to go before large-scale substitution.

Notably, Windsor explicitly states he holds no positions in robotics-related equities. He highlights Agility Robotics as a company worth watching because it sells robots to enterprises today, representing the industry’s current realistic baseline. His conclusion is that while the robotics opportunity is vast, the timeline is too distant for near-term impacts on company valuations; other factors, rather than the sudden emergence of a new industry, will drive valuations in the short term.

Windsor does not deny that model capabilities may experience non-linear breakthroughs in the future, but he notes that among currently observable evidence—benchmark data, exhibition performances, and actual deployment status—there are no substantive changes sufficient to justify compressing the timeline by eight years. His conclusion is that Wang Xingxing’s original judgment of 10 years or more was actually more prudent.

According to a report by South Korea's ET News, Hyundai Motor Group has confirmed the supply chain for the mass production of its humanoid robot, 'Atlas.' The allocation of core components highlights Hyundai's strategy to integrate traditional manufacturing strengths with emerging physical AI technologies.

Hyundai Mobis secured all orders for actuators, the components responsible for implementing robot movement. Actuators account for more than half of a humanoid robot's cost, marking a significant entry point for the automotive parts giant into the robotics sector.

Samsung SDI and LG Energy Solution will jointly supply the batteries, described as the robot's 'heart.' Both companies possess cylindrical battery mass-production capabilities, with Samsung SDI designated as the main supplier producing in Malaysia, while LG Energy Solution plans to supply from its Georgia, U.S. factory. BH EVS will handle the charging solutions.

In a move that defied industry expectations, LG Innotek did not win the camera contracts. Instead, Hyundai Mobis partnered with Namuga to supply the main facial cameras. Cameras for hand-based precision tasks will be provided by MCNEX and Camsys.

Nvidia will supply the processors acting as the robot's 'brain,' specifically the Jetson Thor chip designed for low-latency AI processing. This collaboration reflects the ongoing close ties between Hyundai Executive Chair Euisun Chung and Nvidia CEO Jensen Huang.

The report notes that Hyundai plans to establish a production system capable of manufacturing 30,000 Atlas units annually in the U.S. by 2028, initially deploying them at its Meta Plant America in Georgia to learn automotive manufacturing tasks.

According to a Financial Times report, Toyota, the world's largest automaker, plans to invest $6.4 billion annually in humanoid robots and factory automation, with the goal of deploying 400,000 robots. This move signals Toyota's ambition to transform itself into a global robotics company.

In July, Toyota's AI robotics spinout, Walden Robotics, went public with a $1.1 billion valuation and raised $300 million in funding. Goldman Sachs analysts estimate that Toyota could capture three to eight percent of the humanoid robot market by 2035. Its rivals include Tesla, Xpeng, and Hyundai, which plans to start building its Atlas robot in the US in 2028.

A Toyota executive stated that the company's own factories will be the first deployment sites for robots because the environment there is easier to control. However, investors remain cautious, pointing out that Toyota has previously fallen short of expectations with hydrogen cars and solid-state batteries.

South Korean AI reliability firm SelectStar announced on October 26 the release of 'Davi', a specialized AI decision model designed for 'model routing'. This technology analyzes the difficulty of incoming queries and directs simple questions to lightweight AI models while assigning complex problems to high-performance ones, aiming to balance cost and accuracy.

According to internal test data published by SelectStar, Davi outperformed 'Zev', a similar decision AI developed by TypeSafeAI, across several benchmarks. On the 231-question 'ZevBench', Davi achieved an accuracy of 89.2%, surpassing Zev's 86.6%. In tasks requiring judgment of complex conditions, Davi scored 77.5%, which was 4.5 percentage points higher than Zev.

In a specific cost-efficiency experiment involving 174 high-difficulty Korean Q&A samples repeated five times, SelectStar found that when Davi automatically selected between Anthropic's Haiku, Sonnet, and Opus models, the overall accuracy reached 71.1%. This was slightly higher than the 70.3% accuracy achieved by always using the highest-performance Opus model alone. Simultaneously, this dynamic routing approach reduced model execution costs by approximately 28%.

It is important to note that these results are entirely self-reported by SelectStar. The company acknowledged that the experiments were limited to high-difficulty Korean scenarios and that the reported cost savings did not include the overhead of calling the Davi router itself. Consequently, the net cost efficiency in actual enterprise environments remains to be verified. Additionally, SelectStar has open-sourced a model routing plugin under the Apache 2.0 license for developers to test similar functionalities.

Jeon Myung-jun, head of Enterprise Services at Korean telecom giant KT, stated at a National Assembly symposium on "Gangwon AI Infrastructure Hub Strategy" that current submarine cable infrastructure is insufficient to support South Korea's ambition to become a global AI hub, despite the government's plan to build 18.4GW of AI data centers (AIDCs) by 2035.

According to KT's estimates, each 1GW of AIDC generates approximately 91Tbps of traffic, meaning the 18.4GW target would require about 1700Tbps of international transmission capacity. In contrast, the submarine cable capacity operational or planned through 2030 is only around 200Tbps, leaving a deficit of over 1500Tbps. Jeon emphasized that without adequate export channels, large data centers would remain mere "warehouses."

As AI usage shifts from training to inference, the demand for real-time data exchange has increased, highlighting the importance of international connectivity. South Korea currently operates only 8 global submarine cables, far fewer than Japan's 43 and Singapore's 34. Jeon proposed forming a public-private consortium involving telecom operators, tech giants, financial investors, and the government, along with diversifying landing station locations, to address this bottleneck.