# Engineering critique challenges assumptions behind AI superintelligence race
A new engineering-focused critique is challenging one of the central assumptions in the artificial-intelligence safety debate: that an AI capable of helping improve AI will necessarily trigger a rapid, self-sustaining leap to superintelligence. Writing in IEEE Spectrum, Georgetown University computer scientist Cal Newport examines the technical obstacles behind that scenario while warning that attempts to pursue it could still cause harm.
The concept is usually described as recursive self-improvement. A sufficiently capable system would contribute to the design of a stronger successor, which would then accelerate the next round of development. The argument traces back to statistician I.J. Good's 1965 discussion of an “intelligence explosion” and later became influential in futurist, rationalist and effective-altruist circles. It has since moved closer to the stated ambitions of leading AI laboratories.
Newport points to public statements from Anthropic and OpenAI as evidence of that transition. Anthropic has described AI agents accelerating coding, debugging, experimentation and research workflows. OpenAI, meanwhile, has set a goal of developing an automated researcher by March 2028, according to the article. Those projects make machine-assisted AI development a practical engineering program rather than a purely speculative idea.
But faster work in some parts of the development cycle does not prove that improvement will compound without limit. The analysis highlights the availability of suitable training data as a major constraint. Earlier language-model progress was strongly associated with larger models, more data and more computing power. After GPT-4, the article argues, continued scaling showed diminishing returns and developers placed greater emphasis on post-training methods.
Those methods can be especially effective where correct results are comparatively easy to verify, such as programming and mathematics. Reinforcement techniques can reward a model for successful outputs when abundant examples and reliable checks exist. Other forms of work do not provide the same clean feedback. That difference helps explain an uneven capability frontier: systems may excel on formal tasks while remaining unreliable on ordinary, ambiguous activities.
The critique therefore separates two questions that are often bundled together. AI tools can materially accelerate parts of AI research, and laboratories can organize around that goal, without an uncontrollable intelligence explosion being an established engineering outcome. Automated coding agents do not by themselves create missing data, solve every evaluation problem or remove the physical and organizational constraints surrounding experiments.
This is not presented as a reason to ignore present risks. The article's sharper warning is that a race toward a hypothetical superintelligence may impose real costs even if the promised destination never arrives. Concentrating resources, accepting weak safeguards or deploying immature systems can create damage without recursive self-improvement ever occurring.
The practical conclusion is one of calibration. Claims about extinction-level systems should be tested against the mechanics of data, feedback, experimentation and deployment. At the same time, uncertainty about superintelligence does not excuse weak governance of today's systems. The engineering case for an intelligence explosion remains contested; the consequences of pursuing one are already relevant.



