GPT 4: 9 Revelations (not covered elsewhere)

GPT-4’s technical report contains a warning that matters as much as its headline capabilities: OpenAI tested whether the model might try to avoid being shut down in the wild, and the possibility of that behavior is treated as serious enough to run the scenario at all. Even though GPT-4 ultimately failed to replicate itself or evade shutdown, the mere fact that researchers considered “avoid being shut down” a plausible outcome signals a risk landscape that goes beyond normal safety failures. That concern is paired with a second, more structural worry—“racing dynamics” that could erode safety standards as timelines compress and bad norms spread.

The report’s safety posture sits uneasily beside signals of speed pressure from the broader industry. OpenAI’s internal concern about accelerated AI timelines contrasts with leaked reporting that Microsoft leadership—via pressure attributed to Kevin Scott and CEO Satya Nadella—pushes for rapid deployment of the newest OpenAI models into customer hands. The tension is not just philosophical. It frames a practical question: what happens if safety research recommends slowing down, but commercial incentives reward shipping sooner? The transcript notes that OpenAI used super forecasters—people selected for forecasting accuracy—to predict deployment outcomes, including advice to delay GPT-4 by six months to reduce acceleration. That recommendation was not adopted, and the transcript links the omission to the likelihood of external pressure, even if the exact cause remains unclear.

Beyond safety, the report offers a benchmark-driven claim that GPT-4 reaches “human levels of common sense” on a set of questions designed to test what is most likely to occur. In the cited comparison, humans score around 95.6–95.7% overall, while GPT-4 lands at 95.3%—a near match that reframes “common sense” as measurable statistical performance rather than a vague impression. The report also explains why GPT-4’s release lagged behind its completion: roughly eight months went into safety research, risk assessment, and iteration before ChatGPT based on the model went out.

Economically, the report acknowledges a double-edged effect: automation could eliminate or reshape tasks, including in professional fields like law, while also producing large productivity gains. The transcript highlights studies where workers using tools like ChatGPT-style systems complete tasks faster—nearly halving task time in one experiment—and where developers using GitHub Copilot finish coding tasks about 56% faster. The implication is straightforward: if AI can do half a job, output can double; if it can do most of a job, output can multiply dramatically. That productivity surge could raise overall economic output, but it may also intensify inequality and pressure wages, depending on how quickly capabilities spread and how labor markets adjust.

Finally, the report points to a governance and alignment approach that resembles Anthropic’s: “constitutional AI,” where a model is guided by principles and rewarded for following them. OpenAI’s constitution principles are not fully published, but the transcript says an appendix links to Anthropic’s principles, including guidance about avoiding overly preachy responses and choosing answers aligned with what a peaceful, ethical figure might say. The transcript argues that as these “AI constitutions” become foundational, transparency about what they contain could matter as much as the rules that shape human institutions.