원문 보기
Hello. We will start the Oduri Jeremy's AI Tactics update on October 2, 2026 . hello. Shall we get right to this week's top news? The topic we will focus on and delve into today is the historical lessons and economic repercussions brought about by the rapid proliferation of AI technology. Rather than simply listing figures on AI growth percentages, this will be an opportunity to find the implications of AI innovation within the monumental historical event of Spain's conquest of the Americas, while simultaneously taking an in-depth look at the resulting structural changes and risks in the market.
Oh, the connection is quite wide. Conquest of America and AI. You usually only cover tech news, so to bring up a historical analogy—in what specific context are you making this comparison ? That is a very good question. A recent interview with the two Beishpetay schshefs on Door Cash Street yields very interesting insights.
Sipard analyzes how a few hundred Spanish soldiers were able to bring down the Aztec and Inca empires . The key was not simply military superiority . Rather, the key was how well they built diplomatic alliances, leveraged their technological superiority in iron legs and horses, and exploited the internal vulnerabilities of the existing order .
hmm. So, it seems the Spanish army didn't possess formidable forces, but rather formed alliances with local tribes and exploited the cracks in the existing empire . But how is this connected to AI ? This is exactly the key point. Dobesh Patel explains this historical case by contrasting it with the modern AI industry, particularly the real-world examples of Janest interns.
It cites the example of Jane Street interns who reduced memory usage in the large-scale language model LMN and developed custom checkpointing. This emphasizes that the practicality and increased efficiency of AI are core values . In other words, the key to future victory lies not in investing massive resources, but in identifying vulnerabilities in existing systems and presenting lighter, faster, and more efficient solutions .
Just as the Sipain army brought down the empire with the help of modern humans, AI companies are reshaping the market by exploiting the inefficiencies and vulnerabilities of existing IT infrastructure . all right . In other words, it means that AI is not simply a new technology, but plays the role of a conqueror that shakes up the existing order.
However, one question arises here: there is also an interpretation that Spain's conquest of the Americas eventually brought resources like silver, which revived the Spanish economy. Likewise, there seem to be arguments that AI merely acts as a tool for generating immediate profits, and it is questionable whether it will bring about genuine improvements in society's productivity in the long term .
That is a really sharp point. That is precisely the point of contention. Historians have long debated whether Spain's conquest of the Americas was a substantial creation of wealth or merely a simple redistribution of existing resources . Some argue that the inflow merely caused inflation and failed to strengthen the actual industrial base .
Similarly, there are concerns that the rapid adoption of AI may only bring short-term efficiency and could become a trap that hinders the establishment of long-term systems or the core influence of humans . hmm. So, how are these concerns being reflected in the AI investment field ? It seems that Stelly Drummiller's recent portfolio changes are also closely related to this.
Specifically, what strategies is he employing ? Yes, that is exactly the part. Hedge fund manager Stanley The Miller recently invested in and withdrew funds from AI-related stocks . He sold skyrocketing semiconductor stocks, especially stocks like Micron and Intel, and rotated his funds into the healthcare scare sector, NTLA, INSM, large-cap stocks, Amazon, Google, and emerging market EWZ .
Ah, so that's why it was reported that he ended the AI trade . However, in the case of Intel, politicians including Nancy Pelosi hold call options and are pushing for strong tariffs, so why did Droken Miller sell? It is a move that is the exact opposite of the market, and that is a very interesting point.
While political figures are taking full-scale positions in anticipation of short-term policy benefits or the rise of specific companies, professional investors like Drummiller are adopting a defensive strategy of shifting funds to low-volatility value stocks . This is because they expect AI-related stocks to be already overheated and a correction to come soon.
In other words, it means acknowledging the long-term potential of technology while restructuring the portfolio to avoid short-term bubble risks. I understand. In other words, the strategy is that while the technology itself is important, it is already reflected in the current price, so let's look elsewhere .
However, another question arises here. As AI technology advances, models will become more complex and massive, but will that actually lead to better performance? Or wouldn't it actually lead to inefficient overlearning? That is a very important insight. Julian Siddhartha's interview provides a clear answer to this problem.
He points out the limitations of a single large model in the process of training a chess AI model. It is precisely the overfitting problem. Although draw-out or data augmentation techniques are used to reduce validation loss, the fundamental solution was none other than partition learning. How exactly is split learning performed ?
It is a method of training dedicated models for the opening prologue, middle game, and end game separately, and then combining them into an ensemble. This maximizes generalization performance and efficiency . In other words, rather than unconditionally making the model large, it is more effective to subdivide the phases of the problem and combine lightweight models optimized for each phase .
Hmm. It seems possible in games because situations are clearly distinguishable, but could it be applied to complex AI systems in the real world ? For example, in continuous and highly uncertain fields such as autonomous driving or medical diagnosis . yes . That is precisely the point of contention. Mainstream veterinarians acknowledge that this approach may have limitations due to a lack of strategic thinking during actual matches .
There is also a point raised that the logical link may not be clear when interpreting the causal relationship between macroeconomic indicators and advertising quality. However, the key point is not an infinite number of single models, but rather optimized modularization based on the problem structure .
This has important implications for the design of AI systems in other domains as well. In other words, rather than attempting to unify everything, an approach that understands the characteristics of each domain, divides them, and then integrates them can lead to more stable and efficient results in the long run .
all right. So, in terms of technical efficiency, division and integration are key, while in terms of investment, avoiding bubbles and reallocation are key. But may I ask you one more heavy question here? There are also views that express concern about how the efficient development of AI will ultimately affect the meaning or value of human existence.
In particular, regarding grasslands, the discussion about the era in which AGI is realized seems even more serious; what is Nickstrom's argument ? Yes, yes, that is exactly the point. Nixtrum takes a very cautious stance on the probability of collapse in the era of superintelligence . He emphasizes that while the probability of a catastrophic event itself is subjective, alignment, harmony, and governance are the most critical tasks during the technical maturity stage .
In particular, he keeps in check his position as a concerned optimist. It is predicted that in a world where technology solves all needs, humans will be liberated from the necessity of survival, but at the same time, underlying spiritual values will be highlighted. hmm. So, you are saying that when the necessity of survival disappears, one actually comes to pursue higher-level values .
However, the question that arises here is, if technology replaces core human influence, how can the value of our efforts or the recognition of others be sustained ? That is precisely the biggest dilemma. Bostrom points out that there is a conflict between concerns that humanity will be lost upon the realization of superintelligence and the view that it is an extension of natural evolution .
It also raises the issue of the ethical validity of introducing excessive surveillance systems . In other words, if AI solves everything, humans may become powerless, but at the same time, we may need to find new values for a true trial. This is an area that requires philosophical and social discussion, going beyond mere technical issues.
That is a really profound story. Finally, the discussion regarding watermarking, a technology for determining the authenticity of AI-generated content, is also an important issue. Google DeepMind's Pushmeat Corey and Jerami Rediclis claimed that they are trying to solve this problem through a technology called Sint-Yid.
Specifically, what does it entail? yes . Researchers at Google Mind explain that the Sseundae provides watermarking that is robust against deformation while remaining imperceptible to humans. This is recognized for its importance to the extent that it is being adopted as a core technology in areas such as regulatory compliance.
In particular, Jeremy Radcliffe introduced that watermarks can be inserted into protein sequences and structures through Ssundadi Bio . He emphasized that this is essential to prevent the misuse of synthetic materials without proper functionality and to trace their origins. Ah, it's surprising that even DNA is becoming a target for watermarking .
However, it seems like this technology has its limitations . that's right. Sensid has a limitation in that it is a closed system requiring key sharing. It means that if a specific organization does not disclose the key, the content cannot be detected. Furthermore, it can be vulnerable to short text or here-on attacks, and there is not yet a definitive method to perfectly detect all AI products .
Accordingly, despite the enthusiastic acceptance by the bio community, it will be a matter of how quickly industrial standardization for key exchange and technology integration between DNA synthesis companies and AI developers is achieved. To summarize the upcoming events, this week's major news can be broadly categorized into three points .
First, through the historical example of Spain's conquest of the Americas, we confirmed that AI plays the role of a conqueror overturning the existing order, with efficiency and partition learning being key . Second, through Stelly Drochmiller's portfolio rotation, we examined bubble concerns and defensive reallocation moves in the AI investment field .
Third, through Nickstrom's superintelligence controversy and Google DeepMind's watermarking technology, we considered the social and ethical repercussions that technological advancements will bring and countermeasures . You summarized it exactly. In particular, the key question we posed this week was how to bridge the tension between short-term efficiency and long-term productivity values, and the gap between technological feasibility and social acceptance .
Yes, it was a really fulfilling time. Which of these topics would be good to delve into more deeply next time? It would probably be good to cover actual industrial application cases in terms of AI safety regulations and the catastrophic debate, or agent infrastructure and security . That's a good suggestion.
Next time, we will have a more in-depth discussion on how specific technical specifications and regulatory frameworks should be harmonized . Thank you for watching today.
댓글 0