Google Engineer Robert O'Callahan Resigns Due to Rapid AI Progress
Robert O'Callahan, a former prominent engineer at Mozilla and a member of Google DeepMind's New Zealand team, publicly announced his resignation, stating that his team's goal was to make the next generation of AI chips cheaper and with lower latency. He believes that the current pace of AI progress is too fast and has doubts about the end goals, which led him to leave Google.
He clarified that he was not directly involved in model capabilities but worked on chip design tools; initially, he considered himself relatively harmless, but later assessed that if these tools succeeded, they would enable the next generation of AI chips to be faster and cheaper, making AI more widespread and stronger due to increased inference tokens. He mentioned that being in New Zealand and Google not having many other engineering projects locally made it nearly impossible to transfer to internal projects that do not accelerate AI; he had previously expressed to a manager that he was not pleased with the project's success. He stated that while millions are pressing the accelerator, his withdrawal has a small but non-zero impact, as some skills are scarce; attempts to positively influence from within were insufficient, and external influence could also be exerted.
In his letter, he stated that the probability of superintelligent destruction is not 100%, the risks are real but uncertain, and pursuing ASI in the near future is irresponsible in his view. He also listed concerns such as cognitive abandonment, AI-induced mental issues and loneliness, concentration of power, economic disruption, cybersecurity, and lack of accountability. He emphasized that his decision is not directly related to other recent warning-style resignations and does not wish to ignore the consequences of his work. His next steps include continuing to maintain the recording and debugging tool rr and the commercialization platform Pernosco, and he plans to research how AI can debug code and whether debugging tools can intervene, requiring future work to be clearly pro-human.
Google has not publicly responded to this resignation. Earlier in the year, there had been public resignations in DeepMind's safety direction, and Alphabet has also experienced departures of core researchers like Jeff Dean and management adjustments; TPU and self-developed accelerators remain its vertical integration pillars against Nvidia and to lower inference costs. The exit from the chip tools position does not change the schedule for tape-out and cluster expansion but turns "cost reduction equals acceleration capability" into a public statement.
Who is buying cheaper inference and who is selling design tools: cloud and consumer products require low-latency tokens, while chip tools trade electronic design automation time for mass-producible accelerators. Funds continue to flow into wafers, packaging, and data centers, not into the personal debugging products of the resigning employee. The beneficiaries are the cloud businesses that continue to lower inference unit prices; the pressured parties are the employer brand that leaves safety concerns within the laboratory and the geographical employment structure that cannot transfer roles in New Zealand. The event driver is the open letter, not the cancellation of tape-out.
Source: Public Information
ABAB AI Insight
O'Callahan's background is in browsers and deterministic debugging, not alignment papers. During his time at Mozilla, he worked on Gecko, rr records and replays Linux processes, and Pernosco sells the same capabilities to engineers. His move to DeepMind was to bring scarce tool capabilities into chip design, rather than writing loss functions. The logic behind his resignation parallels that of Geoffrey Hinton leaving Google and some safety personnel at Anthropic and OpenAI publicly breaking away: individuals are part of an acceleration chain, and once they determine that the success of that chain's product equals the premature arrival of capabilities, internal transfers are geographically and organizationally non-existent.
The capital path is very clear: Google uses TPU to integrate training and inference costs into its own cloud and search, and every time the toolchain shortens a design cycle, there is one more round of deployable cheap computing power. The work available to the New Zealand office has almost been reduced to this line, meaning that location locks employees into only accelerating AI. After his withdrawal, he continues to sell debugging tools, shifting skills from "making chips land faster" to "helping people see what the program has done," with money transitioning from salary and stock to small products and uncertain new projects.
In comparison to Nvidia's CUDA ecosystem engineers, TSMC's advanced packaging capacity, and the inference clusters bound to OpenAI and Microsoft: the real accelerator is in wafers and interconnections, not in a resignation letter. The industry phase is a competition among model laboratories to see who can bring inference down to consumer prices first, with chip tools serving as a hidden lever. The safety faction uses resignations to voice concerns, while laboratories maintain delivery through restructuring and poaching.
Structural changes belong to the supporting layer of technological substitution. The substitute is not a specific chatbot, but rather "how long it takes to design an AI chip and how much it costs to infer a thousand tokens." As prices drop, capabilities increase with usage, and regulation and accountability follow unit prices. The mechanism is: tool positions consider themselves far from models, yet after product success, they become multipliers for capability expansion; once the multiplier is exposed, employers can no longer package hardware tools as neutral infrastructure.
ABAB News · Cognitive Laws
Tools that reduce latency are levers for increasing capability.
Locations that cannot transfer roles mean only pressing the accelerator is allowed.
Those who do not touch models can also bring the endpoint closer.