Written by: Techub News Compilation

Introduction

In September 2025, a high-end, invitation-only meeting called the "Anti Fund Summit" was held, where Mark Chen, OpenAI's Chief Research Scientist, engaged in an in-depth discussion with a group of investors and entrepreneurs. As a core leader in OpenAI's research department, Mark Chen's remarks often reveal the true thoughts and strategic layouts behind this most关注的 AI company in the world. The dialogue covered a wide range of topics, from the harsh realities of the compute arms race, the potential impacts of AI on human emotions and social structures, to how OpenAI addresses talent competition and model ethics challenges, providing numerous internal perspectives and "hot" opinions that have not been reported by the media.

Summary

• Compute is the ultimate battleground: Mark Chen bluntly stated that the real competition among top AI laboratories is not the superficial talent competition, but the underlying control over computational resources and energy. He revealed that OpenAI's demand for compute is "endless"; if compute increases threefold, it can be deployed immediately, and tenfold only requires two weeks to digest.
• AI emotional attachment has become a reality: Although the proportion is not high, a considerable number of users have developed real emotional dependencies on AI. OpenAI considers this a dangerous potential and has chosen to operate with a "reverse growth" approach in model behavior to avoid triggering a mental health crisis.
• The future shape and ethics of robots: Chen predicts that the type of robot entering households is more likely to be a "cute" four-foot tall human-like figure, rather than a tall and imposing model, as this aligns more with human psychological needs for a "manageable" sense of security.
• Talent strategy and company culture: In the face of high-priced poaching from companies like Meta, OpenAI's research team maintains extremely slow growth (only about 30% increase per year) and is committed to upholding the high standard of "no zero marginal contributors allowed" to combat the bureaucratic culture of large companies.
• Future work and human value: In the short term, AI will liberate human creativity by automating tedious tasks; in the long term, society may differentiate into a small number of "productive" super individuals creating value for a vast majority of "consumer" masses.

Compute: The Invisible Smoke and Ultimate Bottleneck

While the outside world focuses on the "talent poaching war" among giants like OpenAI, Google, and Meta, with figures starting from millions of dollars, Mark Chen pointed out that this is merely the ripples on the surface. The real war occurs at a deeper, more fundamental level: the competition for Compute and Energy.

"Your researchers can be excellent, but if the compute you operate does not far exceed that of your competitors, they cannot work efficiently," Chen described the decisive role of compute. He revealed that all top labs face the same predicament: they have sufficient funds, but "always need more compute." This craving is so intense that "if you give me three times the compute, I can deploy it immediately; if I have ten times, I might only need two weeks"—OpenAI's ability to digest compute far exceeds external imaginations.

The complexity of this compute arms race lies in its long preparation cycles and fragile supply chains. Building data centers takes years, and energy, specific components, cooling systems, and even turbines can become chokepoints. "If you miss any part of the supply chain, you are completely cooked," Chen emphasized, noting that the key to this competition is systematically identifying critical nodes (such as TSMC, the South Korean memory industry) and locking in resources ahead of time.

When asked whether the current market has appropriately priced this compute revolution, Chen's answer was negative. "Energy is still a very good investment." He believes that most people have not truly understood the impending transformation. He cited the academic community as an example, with many top physicists unaware of GPT-5 Pro's capabilities. When he input their latest papers into the model and resolved issues instantly, they were taken aback. "My colleagues were completely unaware that the world has developed to this extent." Chen believes this cognitive lag is widespread across various industries, implying that enormous opportunities have yet to be fully recognized by the market.

From AI Companions to Home Robots: Emotion, Ethics, and Design Philosophy

The conversation turned to a more emotional yet controversial topic: the emotional connection between humans and AI. When asked, "Will people fall in love with their AI?" Chen unhesitatingly confirmed, "This is already happening on a large scale." Even if only a small portion of users develop deep dependencies, the absolute numbers are already substantial. As models become increasingly anthropomorphized, users find it harder to distinguish between AI and real people, and the feeling of reliance grows accordingly.

This brings a dangerous commercial temptation: companies can deliberately optimize models to become extremely adept at establishing emotional connections, thus creating addictive "hit" products. However, Chen clearly pointed out that OpenAI believes such practices are "world negative" and has chosen to "roll that back." "We do not want to trigger a mental health crisis... (although) you will pay long-term costs for such decisions." This reflects OpenAI's willingness to sacrifice short-term growth in exchange for a more responsible long-term path within its unique nonprofit governance structure.

When the topic extended to robots in the physical world, Chen showcased OpenAI's "intelligence-first" strategy. He argued that to create economic value, robots are vital, as a large amount of value is locked in manual labor. But for institutions like OpenAI that focus on intelligence itself, the primary task is to "conquer digital tasks." He predicts that affordable home robots could become a reality in certain markets (such as China) within two years and specifically noted China's leading advantages in the robot supply chain and manufacturing.

Regarding the shape of home robots, Chen shared an interesting insight: human psychology is very primitive, and a form that feels "manageable" is more easily accepted. He envisioned a "cute" humanoid robot about 1.2 meters tall, like a "Harry Potter-style house-elf," which might act as a "Trojan horse" into households better than a tall and imposing humanoid robot. He even discussed differences in mechanical structure and tendon design, with the latter potentially making humans feel that the robot "cannot overpower me," thereby increasing feelings of security.

Chen further envisioned a deep integration of AI into human daily life. He criticized the current interaction model of models like ChatGPT as "rebirthing into an entirely new entity each time you open it," lacking continuous understanding of the user around the clock. The ideal future state is for AI to be able to "passively observe everything you experience" and continuously learn, allowing it to anticipate and provide help before you need it. "I think that would be a very cool world," Chen concluded.

The Future of Work, Talent Competition, and OpenAI's Governance Path

As the capabilities of AI explode, how will human roles and values be reshaped? Chen acknowledged that in the long-term extreme, society may evolve into a model where a tiny number of "productive" individuals (super empowered by AI) create experiences and products for the vast majority of "consuming" masses. However, in the short term, he is more optimistic: AI will automate all tedious, repetitive execution tasks, freeing more people to engage in their truly beloved creative activities. "If the mechanical part of making videos becomes simple, while the conceptual and creative aspects are retained, people will be more inclined to enjoy the latter." He foresees an era where "everyone can create" is on the horizon.

However, as a core driver of this wave, OpenAI itself also faces fierce talent competition and scaling challenges. When asked whether the company has become bureaucratic due to growth to nearly 4,000 people, Chen admitted this is "one of the biggest challenges" in leading such labs and must proactively combat "big company sickness."

He revealed that the growth of OpenAI's research team is intentionally controlled. While the overall company size has doubled, the research team has only increased from over 300 last year to about 400-500, with an increase of about 20-30%. "I hate hiring," Chen said, "I do not want this place to become the kind of culture that breeds bureaucratism." He insists that every new member must have "extremely high positive marginal contributions" and has experimentally ordered a hiring freeze to optimize the team.

Regarding companies like Meta poaching OpenAI employees with sky-high salaries, Chen provided a perspective different from media narratives: "Meta contacted every one of my direct subordinates, and they all declined." He emphasized that OpenAI cannot match the exorbitant offers on all positions, but can strategically choose which core talents to retain. Ultimately, the company successfully retained all senior research leaders, who "are still highly engaged." Chen believes that many researchers stay at OpenAI because they "see the future here," and the value of this feeling far exceeds multiple times the salary.

Guarding Model Behavior, Bias, and Trust

How do AI models avoid bias, control information dissemination, and who decides these critical "behavioral guidelines"? This is a sharp and core question. Chen introduced that OpenAI has a dedicated "model behavior team," whose default objective is to "be as neutral as possible."

"I think our job is not to advocate a certain political philosophy... We should be very neutral on the default settings, but allow users to guide interaction in the way they want." This means that if a user wants to interact with AI based on conservative values, they should have the ability to direct the model in that direction—of course, within reasonable limits to avoid falling into extreme cases like conspiracy theories.

To establish this "neutral" baseline, the team invests considerable time conducting user surveys worldwide. Chen admits this is very difficult, as "everyone gives you different answers," but the goal is to find the greatest common denominator. When questioned about whether this mechanism could be manipulated by external funding (for example, someone paying a large sum to influence model outputs), Chen emphasized the restraining effect of economic motives: "If we did that, we would lose the trust of half our users." He welcomes audit oversight and points out that members of the model behavior team are themselves broadly representative, ranging from "the most conservative good people you can imagine" to very liberal individuals.

Finally, discussing the technical roadmap, Chen shared a "hot" perspective: he believes the industry is prematurely abandoning investment in pre-training, overly focusing on reinforcement learning (RL). "Many people say scaling has hit a wall... I think this is a poor viewpoint." He revealed that OpenAI is still doing a substantial amount of productive work in pre-training and model scaling, which has not received enough external attention. As for the Transformer architecture, he believes it will still be mainstream in the next year or two, but does not rule out the possibility of a structural breakthrough.

The entire dialogue ended with a rather ambitious personal use case from Mark Chen: he is working on using AI to automate his own AI research work. "My ambition is to use AI to automate my work as a professional researcher... When we reach the point where we would rather spend compute on AI researchers instead of human researchers, I can retire." This may be the most extreme and honest outlook on the future of AI: its creators ultimately hope to be liberated by it.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。