Keep pulling the thread on Eric Zelikman.
Eric Zeltman's new company, Human's End, is focused on developing the "EQ" or emotional and interactive capabilities of AI models, rather than just their "IQ."
Eric Zeltman, formerly of Stanford University and XAI, has founded a new company named Human's End.
Eric Zeltman argues that the most fundamental missing capability in current AI models is the understanding of the long-term implications of their actions and statements.
A key reason for AI model failures in coding is the product trade-off between performance and latency, as models could often perform better if given more time than interactive applications allow.
Eric Zeltman believes that AI models designed to empower people can grow the overall economic pie, whereas models designed for replacement will only reallocate existing segments of GDP.
Eric Zeltman believes that for most industry labs, the human is viewed as an intermediate step to be automated away, not as a permanent collaborator for AI systems.
Current language models are capable of solving some Humanities Last Exam (HLE) questions that are non-trivial for human PhD researchers.
In an early experiment with the Star algorithm, as the model was trained for more iterations, the number of digits it could handle in arithmetic problems continuously increased without an obvious plateau.
A key metric of progress for AI labs is the length of time a model can work autonomously without human intervention, as tracked by benchmarks like METR.
According to sources at Google, task-centric benchmarks persist in large organizations because they are useful for credit assignment and comparing team performance.
The Star algorithm improves model reasoning by iteratively generating solutions and selectively learning from those that produce correct answers, allowing the model to solve progressively harder problems.
The QuietStar paper demonstrated that reasoning algorithms like Star could be scaled up to pre-training scale by using pre-training style data.