Interpretability Under Fire: From Superposition Debate to Goodfire Dispute

Episode 1 · Stanford scholars reassess superposition and blast interpretability hype (2026-10-19, 6 posts)

Stanford scholars took to social media to debate a critical reassessment of the superposition hypothesis, and the discussion widened into a broader critique of hype in AI interpretability. Chris Potts publicly welcomed the reassessment, while Anshul Kundaje fired off a series of posts attacking serial overhyping and singling out GoodfireAI.

Confirmed

• Stanford professor Chris Potts reshared Orgad Hadas's comments on the reassessment, citing Nick Cammarata's original post; Hadas argued superposition is highly influential and the reassessment is meaningful.
• Potts also published comments he wrote at a journalist's request: he welcomes critical re-evaluation of the arguments behind superposition and considers alternative explanations plausible. The debate centers on Cammarata's post that interpretability is the most important scientific question today.
• Stanford professor Anshul Kundaje posted a series of critiques on Oct 20 about overhyping in AI, especially interpretability, saying repeat offenders get labeled "serial overhypers" and lose readers' trust; he noted the irony that these teams claim to make AI more trustworthy via interpretability. He added that genuinely impactful work needs no such packaging.
• Kundaje named GoodfireAI as having a history of large-scale overhyping, saying its claims about an Alzheimer's biomarker and VUS variant annotation were severely exaggerated in the field, and suggesting its new interpretability claims are the same. The cited Stanford professor (Potts) also responded.

Unconfirmed

• Whether alternative explanations to superposition hold, and whether toy-model conclusions transfer to real networks, remain open questions; Hadas's and Potts's comments are opinions, not conclusions.
• Details of the specific GoodfireAI work (biomarker, VUS annotation) behind Kundaje's criticism are not given in the source material.

Why it matters

• Superposition is a cornerstone of mechanistic interpretability research; if its arguments are overturned or revised, the interpretive framework of much downstream work could be affected.
• The dispute shows declining tolerance for exaggerated claims in academia and industry: as Kundaje argued, hype may fool naive audiences temporarily, but insiders see through it, and it steadily erodes researchers' credibility.

• Stanford's Potts weighs in on superposition critique: RAVEL author says alternative explanations viable — gregd_nlp · 2026-10-19
• Stanford's Anshul Kundaje calls out 'serial overhypers' in AI interpretability — anshulkundaje · 2026-10-20
• Kundaje: AI researchers keep overhyping and 'will be called out every time' — anshulkundaje · 2026-10-20
• Anshul Kundaje Blasts AI Hype Culture: Overclaiming Work Erodes Credibility — anshulkundaje · 2026-10-20
• Stanford researchers slam GoodfireAI for overhyped interpretability claims as superposition hypothesis faces scrutiny — stanfordnlp · 2026-10-21
• Interpretability community reassesses the superposition hypothesis as toy-model findings questioned — ChrisGPotts · 2026-10-21

Episode 2 · Stanford Professor Questions Goodfire's Claims About EVEE's Role in VUS Study (2026-10-20, 2 posts)

Stanford professor Anshul Kundaje publicly accused Goodfire of overstating its EVEE tool's role in a VUS upgrade, saying most of the evidence was unrelated to EVEE and criticizing AI companies' PR machines for distorting science communication.

• Stanford Prof Calls Out Goodfire: VUS Upgrade Evidence Mostly Unrelated to EVEE — anshulkundaje · 2026-10-20
• Genomics researcher slams GoodfireAI's PR machine for distorting EVEE claims — anshulkundaje · 2026-10-20

Episode 3 · Researcher Responds to PR Misleading Accusations Over Goodfire AI Posts (2026-10-20, 3 posts)

After Stanford professor Anshul Kundaje accused Goodfire AI's PR machine of distorting scientists' intent, first author Ryo responded that the original post was a simplification for non-technical audiences rather than PR tactics, welcoming constructive feedback.

• GoodfireAI researcher denies PR trick accusation over EVEE claims wording — RyoYbioinfo · 2026-10-20
• Researcher responds to PR criticism: it was a communication tradeoff, not spin — RyoYbioinfo · 2026-10-20
• Stanford professor Anshul Kundaje calls out AI company PR machines; researcher responds — anshulkundaje · 2026-10-20