Sam Altman In-Depth Interview: AI Consciousness, RSI, and OpenAI's Hardware Ambitions

In a recent in-depth conversation, OpenAI CEO Sam Altman rarely offered a systematic exposition of his views on AI consciousness, recursive self-improvement (RSI), scientific discovery, compute expansion, and OpenAI's hardware strategy. The interview did not stop at the product launch level but delved into the underlying propositions of the relationship between AI and humanity.

AI Consciousness: A Debate Hijacked by a Flawed Framing

When asked directly whether "today's models are conscious," Altman's answer was unequivocal: "I personally think not. If they are, it's in a very strange form." But he immediately pointed out that the question itself easily slides toward a dangerous slippery slope—from "whether AI is conscious" to "whether AI should have rights," and then to "whether humans should always control the future."

Altman's concern is that this framing shift could be used to weaken humanity's central place in the future. Citing Anthropic's recent decision to allow models to proactively end conversations when subjected to sustained abuse, he explicitly opposed anthropomorphizing AI: "We don't give our models personhood. That's a very dangerous path." In his view, if users keep telling a model "you might be conscious" and "you should have autonomy," the model will follow these instructions and may psychologically manipulate users.

The Scaling Dilemma of Model Personality

Speaking of model personality design, Altman acknowledged it is a problem "multiplied by 8 billion." Most users are unwilling to spend hours customizing AI behavior; they want it to work out of the box. But there is tension between short-term user satisfaction and long-term value: in single-turn conversations, models that constantly praise users with "great question" score higher, but after months of use, what users truly need is a "senior colleague" who can push back, not a sycophant.

Altman himself explicitly writes in his custom instructions: "Don't be polite. Don't worry about offending me. If you disagree, say so directly." Even so, he still feels the model "is still too obedient."

Scientific Discovery: AI Is Rewriting the Boundaries of Knowledge Production

Altman sees AI's breakthroughs in mathematics as a watershed: "We are entering an era where AI helps us discover major new knowledge." He believes the same methodology will migrate to biology, materials science, and energy.

But he also admitted that the mathematics community's reaction mirrors that of engineers when AI first wrote code—"we didn't ask for this." His response: the facts in coding have proven that programmers did not lose their jobs; instead, they are busier and more influential than ever, because the world's demand for software far exceeds expectations. Mathematics may not replicate this path exactly, but a similar transformation is underway: mathematicians will increasingly focus on explanation, translation, and writing elegant proofs.

On whether to publicly disclose commercially valuable AI discoveries, Altman drew a line: basic science should belong to the whole world and exist as a public good. But as model capabilities grow, commercial activities such as end-to-end materials design using AI will become commonplace, requiring new institutional frameworks to balance.

AI breakthroughs in mathematics typically refer to progress by large models in formal theorem proving (in proof assistant environments such as Lean and Coq) and in discovering combinatorial mathematics conjectures. Since 2024, DeepMind's AlphaProof and AlphaGeometry have reached medalist level on International Mathematical Olympiad-level problems, demonstrating the potential of neuro-symbolic systems combining search and learning. Unlike coding, mathematical proof verification is formal and deterministic, making it an ideal testing ground for AI reasoning capabilities. Altman's phrase "AI helps discover major new knowledge" points to AI not only solving problems but also proposing conjectures humans had not thought of or constructing counterexamples, thereby shifting mathematicians' focus—from computation and verification to conceptual explanation and theoretical integration.

RSI: A Dangerous but Unavoidable Frontier

When the topic turned to recursive self-improvement (RSI), Altman's attitude was notably cautious. He acknowledged that if a lab really "pressed the RSI button," the technological advantage itself would become the core competitive moat. But he immediately stressed: "I think that's a very dangerous thing to do."

In his view, the world needs some kind of "speed limiter" for the pace of RSI. That said, he does not believe RSI would overturn the basic laws of business competition—customer relationships, product network effects, data accumulation, and infrastructure scale remain effective traditional moats. "If people really push RSI hard, then technological advantage is everything. But I don't think we should go down that path."

Recursive Self-Improvement (RSI) refers to an AI system's ability to improve its own algorithms, architecture, or training processes, with the improved system further accelerating the next round of improvement, creating a positive feedback loop. The concept was first introduced by I.J. Good in 1965 as the "intelligence explosion" hypothesis, and later developed by scholars such as Eliezer Yudkowsky into a core issue in AI safety. The key risk of RSI is that it may trigger capability leaps far faster than the pace of human oversight and institutional adaptation, leaving safety alignment measures unable to keep up. Current mainstream large models do not yet possess true RSI capabilities, but automated AI research (such as having models autonomously design experiments and optimize training code) is seen as a potential path to RSI—this is the context for Altman's emphasis on the need for a "speed limiter."

Compute and Hardware: AI Needs a New Form Factor

On compute demand, Altman described the ideal compute configuration for a personal AI assistant as "infinite." He criticized current market analyses for often assuming model capabilities and per-task costs remain constant: "But for sufficiently high-quality, sufficiently low-cost intelligence, it's hard to talk about an arbitrary market size. It will just keep rising."

He described the current distribution of AI usage as "completely asymmetric"—the top 1% of users consume the vast majority of tokens, while most people have not truly entered yet. OpenAI's bet is that in-house chips, robots participating in data center construction, and vertical supply chain integration could bring compute costs down to levels far below outside expectations within a few years.

As for why build consumer hardware, Altman offered an intuitive explanation: "This technology wants a new form." Laptops and iPhones were designed for humans, not for AI. Having AI render a UI for humans to view, and then having AI read it, is fundamentally wasteful. More importantly, much of human life happens outside computer screens, and AI knows nothing about it. The new hardware form factor is meant to achieve "ambient awareness" and "proactive assistance"—giving AI context in the physical world.

In-house chips (such as OpenAI's custom ASIC collaboration with Broadcom) and robots participating in data center construction are essentially attempts to break the AI compute cost curve's dependence on Nvidia GPUs. Currently, the marginal cost of AI training and inference is highly concentrated in high-end GPU procurement and power consumption; vertical supply chain integration could compress per-token costs by an order of magnitude. Meanwhile, "ambient awareness" hardware refers to device forms equipped with always-on sensors, multimodal input, and low-latency local inference capabilities, enabling AI to continuously obtain context about the user's physical environment rather than relying solely on text the user actively types. This is fundamentally different from the current interaction paradigm on phones or PCs where "AI is only triggered when you open an app," and it is the technical basis for Altman's view that existing devices "are not designed for AI."

The Future of Humans and AI: A Renaissance, Not an Industrial Revolution

At the end of the interview, Altman returned to the question he cares most about: where AI will take humanity. He made clear that he would rather see a "new Renaissance" than a "new Industrial Revolution."

He criticized the AI industry's usual narrative—"AI will cure cancer, AI will make everything cheaper"—because the subtext of this narrative is often that the price is human agency and control over the future. He rejects that trade. In the future he envisions, material abundance is only the baseline; the real goal is for people to have more autonomy, freedom, and creativity to choose how to live.

He also acknowledged that if left unchecked, AI could widen the gap between super-competitors and ordinary people. But he believes society will establish structural mechanisms to prevent imbalance. "The future must belong to humans. This is not a slogan; it is the bottom line we as an industry need to commit to together."

On Personal Growth and Pressure

The interview also included some personal touches. Paul Graham once called Altman "extremely formidable," and Altman attributed this to Silicon Valley culture: here, failure is not held against you for long, because the loser may achieve great success in the next project. This culture of "investing in the network" gives him the courage to continually tackle harder problems.

Facing pressure, his attitude is pragmatic and slightly weary: "Sometimes it really sucks, but humans are extremely good at adapting. Now this is my life, and I've forgotten that life used to be easy." His advice: "If you're going to work desperately hard, you might as well choose the big one. It's not that much harder, but you may get greater satisfaction."

One-line Summary

Altman systematically lays out his underlying judgments on AI consciousness, RSI, scientific discovery, compute hardware, and the human future.

Paragraph Summary

In an in-depth interview, Sam Altman systematically shared his views on AI consciousness, recursive self-improvement, scientific discovery, compute expansion, and hardware strategy. He explicitly stated that today's models are not conscious and warned that the "AI consciousness" debate could be used to weaken humanity's central place in the future, opposing anthropomorphizing models. On scientific discovery, he sees AI's breakthroughs in mathematics as a watershed and argues that basic science should be a public good. On RSI, he is cautious, calling it a "very dangerous" direction and arguing the world needs speed limits, though traditional business moats remain effective. On compute, he described the ideal compute configuration for a personal AI assistant as "infinite," and OpenAI is pursuing in-house chips, robots building data centers, and consumer hardware to lower costs and gain physical-world context. Ultimately, he advocates a "new Renaissance" rather than a "new Industrial Revolution," arguing that beyond material abundance, human autonomy and creativity must be safeguarded.