Eric Bigelow
A researcher at interpretability AI startup Goodfire, holding a PhD in psychology from Harvard University. His work focuses on the internal mechanisms and interpretability of large language models, and he has proposed a forking-paths theory of neural text generation.
Timeline (last 90 days)
The experimental method in Eric Bigelow's paper "Forking Paths in Neural Text Generation" involves resampling more than 30 continuations at every token position during inference and recording the distribution of final answers corresponding to each continuation.
In that study, changing the left-parenthesis token when writing "kWh" altered the model's final output answer, indicating that fate-turning points are often not the semantically salient parts.
Eric Bigelow argues that AI decision-making occurs largely at the sampling stage rather than in pre-planned computation inside the model.
Eric Bigelow's broad definition of in-context learning is: any process that does not occur inside the model's weights but dynamically changes the model's behavior at runtime.
Eric Bigelow considers the 7–8B parameter range to be the golden sweet spot for interpretability research, as it is large enough to exhibit rich zero-shot emergent behaviors while remaining computationally tractable.
Much frontier interpretability research relies heavily on Chinese open-source weight models, because US open-source models have a gap at the highest capability tier.
Eric Bigelow believes that many capabilities already exist in pre-trained base models, and that subsequent SFT, RL, and agent fine-tuning are more about sharpening the distribution than endowing entirely new capabilities.
All Facts (7)
The experimental method in Eric Bigelow's paper "Forking Paths in Neural Text Generation" involves resampling more than 30 continuations at every token position during inference and recording the distribution of final answers corresponding to each continuation.
50%UnverifiedIn that study, changing the left-parenthesis token when writing "kWh" altered the model's final output answer, indicating that fate-turning points are often not the semantically salient parts.
50%UnverifiedEric Bigelow argues that AI decision-making occurs largely at the sampling stage rather than in pre-planned computation inside the model.
50%UnverifiedEric Bigelow's broad definition of in-context learning is: any process that does not occur inside the model's weights but dynamically changes the model's behavior at runtime.
50%UnverifiedEric Bigelow considers the 7–8B parameter range to be the golden sweet spot for interpretability research, as it is large enough to exhibit rich zero-shot emergent behaviors while remaining computationally tractable.
50%UnverifiedMuch frontier interpretability research relies heavily on Chinese open-source weight models, because US open-source models have a gap at the highest capability tier.
50%UnverifiedEric Bigelow believes that many capabilities already exist in pre-trained base models, and that subsequent SFT, RL, and agent fine-tuning are more about sharpening the distribution than endowing entirely new capabilities.
50%