Anthropic Researcher Jascha Sohl-Dickstein: Career Planning for PhD Students No Longer Exists
Jascha Sohl-Dickstein, a researcher at Anthropic, told a group of PhD students in a speech titled "The End of the Anthropocene" that their current career planning no longer exists.
He predicts that by 2030, computing power will reach about 10,000 times that of 2025, and points out that we are still at the starting point of an exponential curve. Over the past decade, model training computing power has grown by about 10 million times, while current progress remains nearly flat on a linear scale.
Sohl-Dickstein offered three pieces of advice: use artificial intelligence daily, as long as one is on the exponential curve, they will not fall behind; choose topics that improve with expanding computing power, or they may be solved before completion; stop planning careers as if they are stable entities.
He emphasized that "safe options" are no longer safe, as baseline risks have increased when large-scale changes occur at the societal level, and traditional robust paths will sacrifice potential returns without providing protection.
The speech also mentioned that "writing is thinking"; if writing is outsourced, thinking is also outsourced. He personally uses Claude daily as a highly effective but less judgmental research assistant and no longer writes code himself.
Talent and resources are shifting from traditional academic paths to those that can maintain judgment and problem-setting capabilities amid exponential growth in computing power. PhD and postdoctoral paths planned on a linear timeline face structural compression, while individuals who continuously use cutting-edge models and focus on scalable topics gain relative advantages.
Source: Public Information
ABAB AI Insight
Jascha Sohl-Dickstein is one of the main inventors of diffusion models, having completed his academic training at UC Berkeley and Stanford before joining Anthropic for research. This path shows his transition from traditional academic positions to cutting-edge laboratories, continuously issuing structural warnings to the next generation of researchers through public speaking.
He fully integrates his research with daily workflows at the model execution level, focusing resources mainly on problem selection and judgment rather than code implementation. The motivation is that when computing power expands by orders of magnitude, topics that do not synchronize with it will be quickly solved, and individuals must actively position themselves on the exponential curve to maintain relevance.
This contrasts with the past twenty years of PhD training, which primarily emphasized stable mentorship, linear topic advancement, and "safe" academic paths. Currently, AI research is in the early stages of capability leaps that render traditional career timelines ineffective, as academic institutions and laboratories redefine what constitutes sustainable research investment.
Essentially, this is a case of technological substitution: as models rapidly surpass human graduate student levels in execution and partial reasoning, career value shifts from the ability to execute established topics to the ability to choose problems that appreciate with expanding computing power and maintain independent judgment. The mechanism is that exponential growth renders linear planning ineffective, and "safe" paths no longer provide protection due to rising baseline risks.
ABAB News · Cognitive Laws
Those on the exponential curve will not be left behind exponentially.
Safe options are no longer safe in high baseline risk environments.
Outsourcing writing equals outsourcing thinking.