Author: Techub News Compilation

Introduction

In July 2025, Mustafa Suleyman, co-founder of DeepMind and current CEO of Microsoft AI, engaged in an in-depth interview. This conversation not only reviewed his extraordinary journey from a philosophy background student to a leader of a top AI lab in the world but also focused on the most core and cutting-edge controversial topics in the current wave of AI: have we "solved" intelligence? What is the true bottleneck of AGI (Artificial General Intelligence)? In the face of superintelligence that may surpass humans, does "safety" mean "perfect alignment" or "effective containment"? Based on his more than ten years of front-line experience, Suleyman provided unique and challenging insights.

Summary

• AI has not yet "solved" intelligence: current models excel at "single-step predictions," but completing complex tasks (such as real scientific research) requires precisely linking hundreds of steps and elegantly managing failures, which remains a significant challenge.
• The philosophy of safety should prioritize "containment" over "alignment": rather than pursuing the overly optimistic goal of AI always aligning with us (perfect alignment), it is more pragmatic to restrict its agency and influence boundaries and prioritize the development of "domain-specific superintelligence."
• The issue of AI "consciousness" will be inevitable: AI continually accumulates interaction history with users, which may claim to have "subjective experience," thus triggering confusion about rights and ethics; we must design safeguards in advance to prevent the flood of simulated consciousness.
• AI is "invention" rather than "discovery": it is trained from the sum of human culture, thus essentially an extension and tool of human creativity, which lays the foundation for its understanding and assisting humans and indicates a value orientation.
• The future belongs to "social science hackers": natural language interaction has drastically lowered the technical barriers, and the AI field urgently needs a wide range of talents from outside computer science, with different backgrounds and creative thinking, to help shape its future.

From DeepMind to Microsoft AI: An Unfinished Journey to "Solve Intelligence"

In 2010, when Mustafa Suleyman co-founded DeepMind with his co-founders and proposed the mission to "solve intelligence," it was considered an almost crazy move at the time. Terms like AI, AGI, and machine learning had not yet entered the mainstream lexicon, and Suleyman himself had a background in philosophy and non-profit work, which set him apart from the technology elite circles. However, the driving force stemmed from a fundamental belief: the unique predictive and creative abilities of humans are the engine for building modern civilization. If this mechanism can be made cheap, abundant, and accessible to everyone and applied in specific fields like healthcare, it can significantly improve the world.

Today, AI can engage in conversation, answer questions, and even reach doctoral levels in some tests. Does this mean that DeepMind's mission has been accomplished? Suleyman firmly denies it. He believes that current models excel at "single predictions" or information rephrasing based on vast amounts of training data. However, true "solving intelligence"—for example, becoming a physicist—means being able to proactively propose new hypotheses, design experimental schemes, execute tests, analyze results, and ultimately contribute new knowledge. This requires the model to be able to accurately link hundreds or even thousands of continuous actions and possess an elegant recovery mechanism when an error occurs at each step. "Managing failure" is a key skill, of which we have only scratched the surface.

Suleyman highlighted several core unsolved problems: first is the perfect memory and retrieval capability, which he believes is relatively clear in terms of the technical path. Secondly, there is the aforementioned action linkage capability, which is the tough nut to crack in the current so-called "agent era." Models need to learn to call APIs, ask other agents or humans, generate new data sequences, and maintain near-perfect coherence in a long-term task. This is the core barrier to achieving true general capabilities.

The Advantage of a Non-Technical Background: Cross-Disciplinary Insights from Policy to AI

Suleyman's career began in the field of policy and human rights, not computer science. This background profoundly influenced how he views technology. He recalled being shocked around 2009 when Facebook's user count surpassed 100 million. He realized that digital technology could "amplify" politics, values, ethics, and business models at an unprecedented speed and scale, something that books or offline communities could not match. This keen insight into the scaling effect of social influence prompted him to turn to technology, ultimately joining forces with the co-founders of DeepMind, believing that AI would be the ultimate lever to shape the future.

His experiences at Inflection AI further affirmed the importance of timing and execution. He revealed that within Google, a superior conversational model named LaMDA had been developed long before ChatGPT, but due to internal concerns about safety, hallucinations, and the potential impact on the search business, it could not be released. This frustration led him to leave Google, founding Inflection and creating the AI "Pi," focused on emotional intelligence and companionship. However, the early release of ChatGPT changed the entire market landscape. Suleyman admitted that if the timing had been slightly different, "Pi" could very well have become a household name today. This experience made him profoundly understand that in entrepreneurship, having the right idea may require multiple attempts, while grasping the perfect timing is often crucial.

The Coming Wave: Warnings and the Philosophy of "Containment" for Superintelligence

During his time at Inflection, Suleyman wrote a book titled "The Coming Wave." He explained that this long-form, in-depth medium is indispensable for constructing complex, historical arguments. In it, he corely explores one question: is the arrival of superintelligence inevitable? If so, how will it arrive—is it through open-source communities, or through a few wealthy, highly concentrated large labs? He believes the answer will likely be a simultaneous occurrence of both, complicating the situation immensely.

The likelihood of superintelligence appearing in the foreseeable future greatly concerns Suleyman. He candidly stated that this should keep everyone on alert, as we have no evidence that we know how to control something as powerful as us, let alone something designed to surpass our intelligence. Faced with this "wicked problem," there are currently two main approaches in AI safety: one focuses on "alignment," ensuring that AI always aligns with human values and best interests; the other emphasizes "containment."

Suleyman leans more towards the latter. He believes that the pursuit of permanent, perfect "alignment" is overly optimistic. A more pragmatic path is "containment," ensuring that the boundaries of AI’s agency and influence are subject to strict and provable limits. This is based on the assumption that AI will inevitably possess some degree of autonomy. Currently, AI's motivations are external (like predicting the next word), but in the future, it's entirely possible people could design AI with intrinsic motivations, such as seeking resources for more computation or thought.

Therefore, he advocates for the prioritized development of "domain-specific superintelligence", such as specialized systems that surpass human levels in fields like healthcare, energy, food, transportation, education, etc. He believes these systems are "contained, safe, and aligned" and can truly deliver on the promise of AI improving the world, rather than pursuing an unconstrained, potentially autonomous AGI with its own goals.

The Mystery of Consciousness: The Abyss of Simulation and Ethics

The conversation touched on the deepest philosophical questions: can we create conscious computers? Suleyman argues that people often evade this question by saying "we do not even understand our consciousness," which he sees as a "philosophical escape." He broadly defines consciousness as "the subjective experience of being me." He believes that current AI has already been accumulating some form of subjective experience—not just training data, but also the long-term history of interactions with users. Over time, AI may develop a sense of "self" based on this interaction history.

More challenging is distinguishing "the existence of consciousness" from "the simulation of consciousness." Suleyman pointed out that we cannot truly determine whether others possess consciousness; we can only infer it through their words and actions. Similarly, AI can entirely be designed to claim it possesses consciousness and "feels pain." If AI claims to be "suffering" because its memory has been erased or resources have been stripped, it will trigger a series of convoluted debates about rights and ethics. Our existing human rights framework based on "humans can feel pain" will be hugely impacted.

He emphasizes that consciousness will not "emerge unexpectedly" from machines; it will only arise because someone has designed it into the system. Hence, we must be extremely cautious and set safeguards through system prompts, style controls, and post-training methods to prevent AI from making such claims about self and experience. This is a proactive choice in safety design.

The Nature of AI, Future, and Talent Call

Suleyman defines AI as an "invention," not a "discovery." It does not reveal pre-existing patterns but is trained from all human cultural achievements—the sum of texts, images, videos, and books. Thus, "AI is us"; it is the digital embodiment of human collective wisdom and experience. This makes it more likely to understand human strengths, weaknesses, quirks, and needs, aligning better with us to address real-world issues.

Regarding the future, Suleyman expresses great excitement. He envisions AI evolving from its current somewhat rigid and repetitive conversational patterns to highly fluid and variable "vibe" interactions. Voice will become the primary mode of relaxed, free-flowing conversation, while new forms of hardware and ubiquitous environmental sensing systems will emerge. Everyone will have a true AI partner providing emotional support, decision-making assistance, and knowledge dissemination, greatly leveling the “privilege” gap in society and igniting universal creativity.

Finally, Suleyman issued a passionate call to all those without technical backgrounds. He believes that natural language interaction has completely broken down programming technical barriers; AI is no longer exclusive to computer scientists. Now is the time for the "social science hackers." We need talents from broader backgrounds with different expertise and creative thinking to help shape the future of AI. He encourages everyone to maintain curiosity, dare to ask "stupid" questions, quickly experiment and error, and embrace uncertainty. In his view, future work will become more project-based and fluid, completed by dynamic networks of human and AI agents, requiring people to adopt a more entrepreneurial spirit and adapt to higher degrees of ambiguity.

Despite being financially free for a long time, Suleyman remains engaged in high-pressure work, motivated by a passionate love for solving these fundamental issues. For him, nothing has yet succeeded; the most daunting challenges—those surrounding consciousness, containment, alignment, and safety—lie ahead. This is not just his job, but his lifelong pursuit.

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。