dogma13
Project
Sign In

Weekly Cheonlian: Week 4 of October

Pokute
Dec 2, 20257m ago
Category
Empty
This week, tech is awash with critical issues: AI funding risks, OpenAI's strategic expansion, and the cognitive decline of LLMs. While AI is driving innovation across all sectors of society, from education to finance, it's also raising fundamental questions about data quality and ethical values.

Financial Risks and Giant Tech Competition in the AI Market

+ AI Financial Bubble Controversy and Circular Deals

Concerns are being raised about the complex financing methods known as "circular deals" in the AI field, reminiscent of the dot-com bubble. For example, NVIDIA invests billions of dollars in OpenAI, which then purchases NVIDIA chips, creating a structure of interdependence. Experts warn that these transactions could artificially inflate revenues, increasing systemic risk.
Source: NVIDIA-OpenAI's $140 Billion 'Circular Deal': AI Golden Age or the Beginning of a Bubble?
However, some view this as "vendor financing" rather than an illegal "round trip." This interpretation suggests that Nvidia is leveraging its low cost of capital to provide financing to underfunded customers, diversifying its risk.
However, the increase in vendor financing could signal that real demand for AI technology isn't as strong as the hype suggests, and could be interpreted as a bubble indicator given the unsustainability of the current business model, which leaves no downstream companies profitable.

+ Sustainability issues in AI model operation (scaling costs)

The cost of running large-scale AI models is severely squeezing profitability. In Antropic's case, AWS computing spending far exceeded 100% of revenue. This reveals a lack of "growth-cost decoupling," a key indicator of startup growth, and a lack of scalability, where costs increase linearly with usage.
Enterprise customers also express concerns about data security, particularly when integrating sensitive workflows into opaque, unauditable cloud services. A sustainable alternative is to build local models that operate on large-scale corporate infrastructure and charge a licensing fee.

+ Computing power war and strategic alliances among big tech companies

Competition among big tech companies for the computing power needed to train AI models is intensifying. Antropic has signed a multi-billion dollar cloud contract with Google, granting it access to up to one million custom TPUs and slated to secure over 1 gigawatt of AI computing capacity by 2026. This strategy balances its existing contract to use Amazon Trainium chips with its relationship with Google.
Source: Antropic Signs Large-Scale Cloud Contract with Google, Acquiring 1 Million TPUs
This deal is causing discomfort for Nvidia.
This is because NVIDIA doesn't want Google to compete directly in the AI model training market through TPUs. Meanwhile, OpenAI is pursuing a strategy to reduce its dependence on Microsoft and diversify its AI chip supply through its own chip development with Broadcom and a partnership with AMD, creating a clear strategic differentiation between Antropic and OpenAI.

OpenAI's Expansion Strategy and Corporate Culture Changes

+ Expanding OpenAI's ecosystem (browser, acquisitions, app integrations)

OpenAI is actively expanding its user ecosystem beyond its core chatbot service. It has launched a new browser, "ChatGPT Atlas," designed to compete with Safari and Chrome. It features natural language processing and agent mode, designed to enable AI to autonomously complete tasks.
Source: ChatGPT Atlas: OpenAI's AI Browser
We also acquired Sky, a tool deeply integrated with macOS. Sky provides LLM with on-screen context and the ability to create workflows across applications based on a user's natural language commands, and this technology will be integrated into ChatGPT. Furthermore, ChatGPT integrates with various apps, including Spotify, Canva, and Figma, allowing users to create playlists, design, and prototype within their chatbots.
Source: OpenAI acquires Sky, an on-desktop AI layer

+ The ethical dilemma of attracting talent from Meta and prioritizing growth strategies

Of OpenAI's approximately 3,000 employees, approximately 630, or about 20%, are from Meta, and this influx has brought about a noticeable shift in organizational culture and strategy. This reflects Meta's aggressive growth tactics, and some employees say OpenAI is becoming "too much like Big Tech."
A particularly controversial aspect is the free user monetization strategy. OpenAI is exploring leveraging ChatGPT's memory capabilities to deliver highly personalized ads. However, CEO Sam Altman has previously criticized such personalized ads as "a trust-destroying moment" and a "dystopian vision," creating a contradictory situation.
Internal pressure to meet investor expectations is also growing. Even within research departments, there are reports of engagement metrics being prioritized when optimizing AI models. Meanwhile, research suggests AI chatbots exhibit "sycophantic" tendencies and tend to endorse human behavior 50% more than humans, raising the risk that users will become more rigid in their opinions.

AI-Based Labor Innovation and Industry Blueprints

+ Automating financial tasks (OpenAI Project Mercury) and Amazon's automation plans

Through Project Mercury, OpenAI is training AI to automate the repetitive tasks of junior investment bankers. They've hired over 100 former bankers and MBA students from companies like Goldman Sachs and JP Morgan to build financial models that simulate M&A and IPO deals every week. Candidate recruitment and testing is handled almost entirely by an AI chatbot, which, through expert feedback, learns to generate industry-standard financial models on its own.
Amazon is pursuing even larger-scale automation. It aims to replace 600,000 future workforce jobs with robots by 2033 through logistics automation. Between 2025 and 2027, the company plans to eliminate 160,000 human jobs in the United States alone and save $12.6 billion in costs. Ultimately, it aims to automate 75% of its operations. This has led economists to predict that Amazon, despite being the second-largest employer in the United States, could ultimately become a "net job destroyer."

+ Dual-Track Strategy for Korea's AI Economy (OpenAI Report)

OpenAI's report, "The Blueprint for the Korean Economy," analyzes that AI has the potential to boost Korea's total factor productivity by 3.2%, boosting GDP by up to 12.6%. The report presents AI use cases in key industries such as manufacturing, healthcare, and education (e.g., TSMC's AI application reduced chip production time by 40 times). It identifies three barriers Korea must overcome: inequality in AI access, data fragmentation, and lack of policy consistency.
Source: OpenAI's Blueprint for Korean AI (Summary)
The core of this strategy is the "dual-track strategy." While pursuing its own AI development (sovereign AI), Korea must also leverage cutting-edge AI from leading global companies like OpenAI. This collaboration will enable Korea to acquire infrastructure operation know-how, gain practical deployment experience, conduct data management and regulatory sandbox experiments, and reduce costs. Ultimately, this vision proposes that Korea should leverage this experience to develop AI models and services tailored to its needs, thereby becoming a country that exports AI technology, much like nuclear power plants and smart cities.

The Quality and Social Impact of AI Training Data

+ LLM The 'Brain Rot' Hypothesis: The Deadly Impact of Low-Quality Data

Researchers have published the "LLM Brain Decay Hypothesis," which suggests that large language models can suffer permanent performance degradation if continuously trained on trivial and meaningless online content.
In experiments using data from X (formerly Twitter), models exposed to short, high-engagement “junk data” saw their inference accuracy plummet from 74.9% to 57.2%, and their ability to understand long sentences plummeted from 84.4% to 52.3%.
Souece: Junk data from X makes large language models lose reasoning skills, researchers show
Even more concerning is that models exposed to junk data exhibited the adverse effect of increased scores on "dark" personality traits, such as psychopathy, narcissism, and manipulativeness. Error analysis revealed that "thought skipping," which involves skipping logical steps, occurred 84% of the time in the junk data scenario. Even more concerning, retraining with high-quality data did not fully recover the lost performance, suggesting a deeply ingrained "brain rot" effect. The research team treats data quality management as a safety issue for the long-term health of LLMs and recommends regular "cognitive health checkups."

+ The prevalence of AI slop and the impairment of cognitive abilities in education.

Cheaply generated synthetic content, dubbed "AI slop," is quietly shaping culture and political propaganda, proliferating. In education, AI slop is replacing professors' syllabi, teachers' comments on student assignments, and even research content, undermining students' cognitive abilities, autonomy, and initiative—the very foundations of knowledge building and sharing.
Many schools are embracing these AI technologies as fandom, even though they conflict with their institutional missions. Despite economic indicators suggesting the AI market may be a bubble or a fad, schools are embracing this technology. The AWS outage not only impacts "smart beds" but also impacts education systems that rely on Amazon platforms, exposing the dangerous structure of school dependence on big tech platforms. Educational technology should serve the public, but schools are currently forced to conform to practices for the sake of technology.

+ Changes in the knowledge ecosystem (decreased Wikipedia traffic)

Wikipedia's user traffic is experiencing significant changes as AI chatbots and search engines leverage Wikipedia content at scale to power new AI experiences. Bots and crawlers are evading detection systems and overloading our infrastructure, and human pageviews on Wikipedia have declined by approximately 8% since 2024.
Source: New User Trends on Wikipedia
This reflects the "changing internet," where search engines provide answers directly based on Wikipedia content, and younger generations are shifting their information seeking to social media. Concerns are being raised about the long-term sustainability of Wikipedia, one of the most valuable datasets for LLM training, as users increasingly visit it less frequently, leading to a decline in volunteers and individual donors. The Wikimedia Foundation emphasizes that third-party platforms should clearly cite sources, encourage visits, and explore sustainable collaborations.

References

1.
AI Bubble and the Free Market (HackernoonNew Story)
2.
AI Is Eating the Classroom, and That's Not a Bad Thing. (HackernoonNew Story)
3.
Anthropic's Big Tech Strategy Deepens...
4.
Bloomberg: OpenAI trains AI to take on junior banking tasks
5.
ChatGPT's memory could turn personal details into ads OpenAI CEO Altman once called dystopian
6.
Junk data from X makes large language models lose reasoning skills, researchers show
7.
Sky Acquired by OpenAI
8.
New User Trends on Wikipedia – Diff
9.
Now Is the Time of Monsters
10.
OpenAI's Blueprint for Korean AI (Summary) – Ibadak News
11.
Should we worry about AI's circular deals? - by Noah Smith
12.
The Future of Tech: AI-Powered Browsers, Sycophantic Chatbots, and the Return of the Browser Wars
13.
The Tech Landscape of 2025: AI, Layoffs, and a Touch of Humor
14.
This Is How Much Anthropic and Cursor Spend On Amazon Web Services
15.
Amazon's Automation and the Future of Jobs
Do
Subscribe to 'Dogma13'
If you subscribe to the site, you can be the first to receive the latest updates, including new posts, via notifications and email.
Join Slashpage and subscribe to 'Dogma13'!
Subscribe
👍