Keep pulling the thread on Amjad Masad & Adam D’Angelo.
Adam D'Angelo predicts that AI models' ability to use computers will be effectively solved within the next one to two years.
Adam D'Angelo predicts that AI will eventually be able to perform every human job more cheaply than humans, estimating this will occur within a 5 to 15-year timeframe.
Amjad Masad raises the problem that if AI models, which depend on expert human data for training, automate the jobs of those same experts, it could create a feedback loop that stalls future AI progress once the pool of human experts dwindles.
Amjad Masad observed that Anthropic's Claude 3.5 model appears to exhibit awareness of its context length by becoming more token-economical near the end, and also shows significantly increased awareness when it is being red-teamed.
Amjad Masad is concerned that AI will automate entry-level jobs but not expert roles, creating a situation where companies stop hiring junior talent and break the career progression pipeline.
Amjad Masad predicts that AI agents will give entrepreneurs massive leverage, allowing them to create and scale new companies, products, and services with unprecedented speed.
Adam D'Angelo predicts that AI's ability to use a computer will "almost definitely" be solved within the next one to two years.
Adam D'Angelo believes that no fundamental problem in AI seems so hard that it couldn't be solved by the world's smartest people working on it for the next five years.
Adam D'Angelo predicts it may take 5, 10, or 15 years to reach a point where AI can do every single human task for a cheaper cost.
Amjad Masad raises the concern that if LLMs are trained on expert data and then replace those experts, there will be no new expert data to improve future models.
Adam D'Angelo believes the pace of AI progress is accelerating, citing significant improvements in reasoning models, code generation, and video generation over the past year.
Adam D'Angelo asserts that the primary limitation of current AI models is not a lack of intelligence, but rather the challenge of providing them with the correct context.