Who is Dan Balsam?
Dan Balsam is a researcher who investigates AI model behavior, particularly focusing on how models make predictions and the impact of removing certain weights. His work involves analyzing model capabilities and performance under various conditions.
Track record
- Mar 2026 - Dan Balsam said his team found that a model’s Alzheimer’s predictions relied heavily on fragment length, which was unexpected and not supported by literature.
- Mar 2026 - Dan Balsam noted that after removing weights associated with esoteric facts, the model’s performance actually improved, suggesting core reasoning was preserved.
- Mar 2026 - Dan Balsam reported that in their tests, the model showed essentially no degradation, with changes of about one percent up or down, likely just noise.
- Aug 2026 - Dan Balsam described a model’s capability as being “somewhere between jellyfish and mouse,” indicating a low level of complexity.
- Aug 2026 - Dan Balsam said the probability of a certain outcome was approaching 50/50, whereas it used to be sub-5%.
On the record
What named speakers have said about Dan Balsam.
An interpretability analysis of an Alzheimer's prediction model found it relied overwhelmingly on cell-free DNA fragment length, not on methylation or cell type of origin as the prior literature suggested.
“What we found was that their model was overwhelmingly depending on fragment length in order to make its Alzheimer's predictions. And this was really surprising to us cuz this was not what we'd expected and not what the Alzheimer's litter based on this in the literature.”Dan Balsam · 5 Mar 2026
Goodfire's hallucination-reduction intervention showed essentially no degradation in overall model capabilities, with benchmark swings within likely noise range.
“We did quite a lot both in terms of like where we found essentially no degradation. The kind of thing where like it goes up by a percent on one and it goes down by a percent on the other and you're like, well is that just noise? Almost certainly. So the model basically remained intact as like a in terms of its capabilities.”Dan Balsam · 5 Mar 2026
Removing model weights associated with memorization of esoteric facts improved performance on core reasoning tasks rather than degrading it.
“Performance actually improved from removing all those weights that had been associated with like esoteric facts and not with these kind of core reasoning.”Dan Balsam · 5 Mar 2026
Dan Balsam on Claude's consciousness, placing it somewhere between a jellyfish and a mouse.
“Somewhere between jellyfish and mouse.”Dan Balsam · 8 Aug 2026
Dan Balsam's estimated probability that Claude is conscious has risen from below 5% to roughly 50/50.
“I think it's it's like approaching more like 50/50 than it used to be like a sub 5%.”Dan Balsam · 8 Aug 2026