Key Takeaways

  • OpenAI plans to bring its in-house custom chip program online in the first half of 2027 to manage surging inference demand.
  • Technology diffusion follows a jagged exponential path rather than clean linear progression across workplace workflows.
  • Altman wants to eliminate the manual model picker in ChatGPT to remove cognitive friction for users.
  • Expanding raw bullet points with AI only for colleagues to summarize them back into bullets creates useless communication debt.
  • Alignment of superintelligent systems remains an open scientific question rather than an ordinary engineering checklist.

The Jagged Exponential and the 2027 Silicon Timeline

At OpenAI DevDay 2026, Sam Altman described the current pace of AI development with a blunt observation: “It's all just one jagged exponential that keeps going and probably the next six months will be even crazier.” The expansion is not neat. Some weeks bring quiet iterations, while others bring jumps like autonomous computer control and dedicated agents.

To sustain this curve, OpenAI is moving directly into silicon. Compute bottlenecks dictate how fast developers can run agentic workflows. Altman highlighted a project that often runs under the radar: “One thing that OpenAI has that doesn't get as much attention is our chip program. Which will start to come online in the first half of 2027 and will scale a lot in future years.”

This internal silicon pipeline aims directly at inference scaling. Running continuous autonomous loops across millions of users burns through tokens at rates that general cloud infrastructure struggles to support affordably. By controlling hardware design for their own architecture, OpenAI seeks to slash unit costs and unlock longer execution loops for agents.

Stop Inflating Text and Kill the Model Picker

Altman also targeted bad software habits that emerged around generative models. The first is communication bloat. Employees draft three bullet points, ask ChatGPT to inflate them into a five-paragraph email, and send it to a colleague who runs it through ChatGPT to extract the original three bullets.

Altman called for clear cultural boundaries around this loop: “The right answer is to just send the bullet points.” Adding words for politeness or corporate theater wastes compute and human attention.

The second friction point is user interface clutter. The era of manual model selection is ending. “I think people are really tired of the model picker,” Altman said. Forcing users to choose between speed, reasoning depth, and context size creates hesitation. The interface should route queries automatically behind the scenes based on task complexity.

Alignment Is an Unsolved Science

The most serious warning from Altman concerned safety. Treating alignment as a solved technical routine is a severe mistake. “Anyone who says we have solved the science of alignment I believe is wrong in a very dangerous way,” Altman said. “We need to make more research progress. I assume we will.”

Scaling compute and training bigger models does not automatically make them obedient or safe. As systems take proactive actions across computer interfaces, the risk moves from bad text generation to unauthorized system actions. Alignment remains an open research problem that requires dedicated scientific discovery, not just standard software testing.

What to Do With This

Audit your team's internal communication this week. If employees use AI to turn quick updates into multi-paragraph memos, establish a clear rule: send raw bullet points. Then, check your product interface and remove user-facing model pickers in favor of automated backend routing.