What will the Anthropic SAE paper contain?
6
Ṁ250Ṁ106resolved May 22
Resolved
YESEye-test experiments
Resolved
YESSome cherry-picked proof of concept for a useful *type* of task
Resolved
YESStreetlight edits
Resolved
NODoing PEFT by training sparse weights and biases for SAE embeddings in a way that beats baselines like LORA
Resolved
NOPassive scoping
Resolved
NOFinding and manually fixing a harmful behavior that WAS represented in the SAE training data
Resolved
NOUsing an SAE as a zero-shot anomaly detector
Resolved
NOLatent adversarial training under perturbations to an SAE's embeddings
Resolved
NOExperiments to do arbitrary manual model edits
Resolved
NOFinding and manually fixing a novel bug in the model that WASN'T represented in the SAE training data
This will resolve according to Stephen Casper's judgments.
This question is managed and resolved by Manifold.
Market context
Get 
1,000 to start trading!
1,000🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ26 | |
| 2 | Ṁ12 | |
| 3 | Ṁ3 |
Sort by:
Thus, my assessment is that Anthropic did 1-3 but not 4-10.
That is, YES eye-test experiments, YES streetlight edits, YES cherry-picked proof of concept, NO everything else.
People are also trading
Related questions
When will Anthropic IPO?
12/31/26
What will happen between Anthropic and the Pentagon?
Will Anthropic open-source the training code of their SAE interpretability effort?
What will be the next major event for Anthropic?
Who will join Anthropic by end of 2026?
Will there be a publicly available AI with Anthropic Mythos-level capabilities before 2027?
96% chance
What will Anthropic's initial share price be? (split-anchored Feb 27 2026)
547
Will Anthropic, before 2035, completely halt development of AI and attempt to persuade other organizations to do so?
18% chance
What will be Anthropic first major acquisition?
Anthropic make bigger revisions to RSP by EOY 2028?
76% chance