Ellie Huxtable

I think AI already means a substantial change in how software gets built. I was skeptical of AI programming, but better models and tools changed my view: agents can now write effective software and provide a real productivity boost. The strongest workflows let them see the same useful context a developer sees—errors, command output, and project state—rather than making them guess. That does not mean handing an agent an unrestricted terminal and hoping for the best. Language guarantees can eliminate whole classes of mistakes, human review still matters, and unattended agents should run in isolated environments such as VMs. Access to sensitive machine context should be explicit, dangerous commands should require confirmation, and users should be able to self-host or use local models. More broadly, I want privacy enforced technically rather than through promises. Terminal history and shell output can contain extremely sensitive data, so encryption and user control are fundamental. I don’t have a defensible prediction here about AGI timelines, extinction risk, or society’s entire future. My concrete expectation is that capable agents become ordinary developer infrastructure, with their value determined as much by context, isolation, and privacy architecture as by the models themselves.

Overall, I expect AI to be a substantial net benefit in software development. It already makes capable developers more productive, and agents can write effective software when they have access to relevant errors, command output, and project context. The harms are concrete, though: bad commands, flawed code, excessive machine access, and leakage of sensitive terminal data. I don’t think those are solved by trusting a provider or assuming the model will behave. They need technical controls—explicit permissions, dangerous-command checks, cryptographic privacy, self-hosting options, strong language guarantees, human review, and isolated execution for unattended agents. Beyond developer tooling, I don’t have a well-supported overall forecast for AI’s effect on society. I would not turn evidence of real productivity gains into a claim about AGI, extinction risk, employment as a whole, or every other consequence.