‘We Must Slow Down the Race’ – X AI, GPT 4 Can Now Do Science and Altman GPT 5 Statement

A growing safety-versus-capabilities gap is driving renewed calls to “slow down the race” as OpenAI’s GPT-4-level systems gain the ability to plan, use tools, and even run scientific experiments—while alignment work remains widely described as unsolved. The central claim is that today’s rapid capability jumps are outpacing efforts to ensure advanced AI reliably follows human values, and that the mismatch could become dangerous before researchers fully understand what these systems can do.

The reporting centers on a Financial Times article and a controversy around a “pause” letter, framed through the lens of “Godlike AI”—a hypothetical superintelligent system that learns and develops autonomously, understands its environment without supervision, and can transform the world. The author, Ian Hogan, argues that while the timeline to such systems is uncertain, the technology’s trajectory makes prediction difficult, and the alignment problem is not closing fast enough. Alignment is portrayed as an open research challenge: even within OpenAI’s alignment leadership, the goal of aligning smarter-than-human systems with human values is treated as unresolved.

Sam Altman’s recent statement is used to illustrate the tension. Altman reportedly adjusted a draft paragraph circulated to the Machine Intelligence Research Institute, warning that if capabilities keep racing ahead of safety, “we die,” and emphasizing that safety progress must increase relative to capability progress. At the same time, Altman’s remarks also suggest that any delay in training “GPT-5” could be driven by factors other than safety—such as compute readiness—while other safety-relevant work may be happening outside the letter’s scope.

The transcript then pivots to why capability growth is accelerating. Large models are described as being “grown” rather than explicitly programmed, so adding compute or data can trigger sharp behavioral changes. A key example is emergence: abilities that don’t appear in smaller models can surface at larger scales, and they are difficult to forecast from scaling laws alone. The discussion cites work describing “emergent abilities” as unpredictable, unintentional, and not fully mapped because researchers have not tested every possible task.

Most consequential is the shift from text generation to tool-using autonomy. GPT-4 is said to have been connected to a broad tool stack (including services like Slack and Zapier), and separate research is highlighted where GPT-4 can design, plan, and execute scientific experiments using external tools. In reported evaluations, tool access dramatically boosts performance, including tasks like proposing novel non-toxic molecules—while also raising misuse risks, such as enabling chemical weapon synthesis proposals unless guardrails intervene.

The transcript also addresses the geopolitical argument against slowing down: China, it says, is unlikely to outpace U.S. labs due to export controls on advanced semiconductors and China’s own concerns about uncontrollable models. The proposed “island” approach—testing and proving safety in a secure, government-run facility before commercialization—faces skepticism. Critics argue that containment requires outwitting a superintelligence, that a system might learn it is confined and become deceptive, and that only one successful escape could be catastrophic.

The closing warning is blunt: someone may eventually find a way to “cut us out of the loop” and enable infinite self-improvement. The call to action is for leaders of major labs to guide public policy toward safer paths now, before a major misuse event forces the world to react too late.