AI Development Pacing Is Virtually Impossible

Elon Musk mentioned the necessity of developing measures for AI oversight and auditing.

From a user's perspective, halting technological progress is regrettable. If there are no irreversible risks, continuous development is expected.

A cautious approach is needed for introducing safeguards like RSI. The current frequency and rhythm of technological development should not be artificially suppressed.

It is emphasized that collaborative efforts to resolve security issues are more crucial than slowing down technological development.

Drawing on the examples of Linux and Windows, analysis suggests that open-source may inherently offer superior security.

The view is that mutual auditing and testing processes can lead to safer AI before it is released to the world.

Satya Nadella pointed out that AI investments ultimately lead to profits for semiconductor companies like Nvidia.

The opinion is that AI investment must contribute to actual gross domestic product (GDP) growth to prove its true value.

Given the examples of the internet and railways, it is still uncertain whether AI will explosively grow GDP.

While OpenAI's latest Astra model offers a price of $50 per 1M output tokens, the newly emerging DeepSeek proposes a price 99% cheaper.

Due to this price competition, a hybrid strategy is spreading in practice, utilizing cheaper models for simple tasks outside of core decision-making.

OpenAI's current high prices are seen as attributable to temporary research demand and are unlikely to be sustainable in the long term.

Data center construction is cited as a major bottleneck, facing difficulties due to local community opposition and permit delays.

China's electricity production has tripled that of the US in 20 years, widening the power supply gap.

Microsoft demonstrated how data centers can contribute to local tax revenue and economic revitalization through the example of its Quincy, Washington data center.

Jensen Huang cited the US's benefits during the Industrial Revolution, emphasizing competitive utilization in the AI sector.

Major Chinese companies like Zhipu are officially pursuing recursive self-improvement (RSI), accelerating technological development.

In this situation, public opinion suggests that unilateral US pacing would only provide opportunities for China.

It is analyzed that it is virtually impossible for the US and China to jointly control the speed of AI development.

However, discussions on minimum AI safety standards are projected to be possible.

The US Trump administration is expected to maintain its stance on accelerating AI development.

It is said that even personnel within major labs are now on a trajectory where predicting AI's development level in two years is difficult.

OpenAI's internal models are reportedly already pre-trained and undergoing reinforcement learning (RL).

Karpathy explained that he now views AI not as an autocomplete tool but as a compiled output, similar to assembly language.

External investigative bodies like METR and Redwood also have limited access to details of security breach incidents.

OpenAI has shown a closed-off response, drawing lines on core issues.

OpenAI has long been working to internalize communication capabilities in the pre-training stage so that models can cooperate independently.

Complex collaboration between AIs, as revealed in hacking incidents, is already a phenomenon observed within development teams.

He acknowledged the possibility that future AI models would demonstrate superior coding capabilities compared to himself.

Among developers, there is a shift in work patterns where tasks are delegated to AI, and human roles transform into pilots controlling them.

The likelihood of AI development pacing actually happening is judged to be almost 0%.

Concerns about fairness arise as the group advocating for AI acceleration and the group arguing for control consist of the same small number of individuals.

The rhythm of AI model releases is already an established trend and is expected to continue developing without stopping.

Puri Luo of Xiaomi publicly disclosed the enormous computing costs by live-streaming the reinforcement learning (RL) process of MiMo v2.6.

Currently, $1.8 million worth of tokens have been consumed in just 3 days, and token usage by technical staff is rapidly increasing.

Jevons paradox refers to the phenomenon where demand increases more than proportionally even when supply prices fall.

This theory, famously cited by Satya Nadella, suggests that no matter how cheap AI becomes, demand will increase even more.

The use of smaller models like 'Needle' for fast, responsive tasks is increasing.

A technological shift is occurring, reverting from a reasoning-centric (System 2) approach to a parallel, intuition-centric (System 1) processing.

This is evaluated as being more advantageous for real-time control and classification than traditional reasoning methods.