Srreyansh Sethi

documenting journey building.

Purdue 2029

[Srreyansh Sethi @SSrrey97428]

@jxmnop EARL beats state-of-the-art image editing with 5x less training data. Are we finally moving past the 'just throw more data at it' era, or is this just autoregressive models being more sample efficient?

[Srreyansh Sethi @SSrrey97428]

@dmizrahi_ I didn't realize that by just matching we see a 2x boost! Though, I was curious about decontamination. If you n‑gram prune out benchmark overlaps, does BETR still pick the same docs, or does performance dip?

[Srreyansh Sethi @SSrrey97428]

@jiqizhixin When you make many on‑demand actors, does that blow out your central Progress Module’s memory footprint, or do you have smart eviction/compaction in place? IMO, memory is the biggest problem for anything AI rn.

[Srreyansh Sethi @SSrrey97428]

@liuziwei7 @_akhaliq When tuning the search-penalty factor in your reward (you set it to 0.1), did you perform a sensitivity analysis to observe how the policy’s search ratio vs. accuracy Pareto frontier shifts?

[Srreyansh Sethi @SSrrey97428]

@natolambert Have you profiled how emoji insertion alters the layer-wise cosine similarity of token embeddings—specifically, do you see larger distortions in early layers (e.g. layer 1–4) versus later layers (e.g. 8–12)?

[Srreyansh Sethi @SSrrey97428]

@htihle The cost vs. accuracy trend is so clear. At what point do you see diminishing returns? Is there a ‘sweet spot’ cost for the models?

[Srreyansh Sethi @SSrrey97428]

@gm8xx8 Zero outliers sounds great, but does a 2% training overhead really stay negligible when you push to a trillion tokens?

[Srreyansh Sethi @SSrrey97428]

@iScienceLuvr Did you ever hit diminishing returns on that Stable stage?

[Srreyansh Sethi @SSrrey97428]

@sueqian111 Did UniVG ever drop in T2I fidelity when you added referring-seg data?

[Srreyansh Sethi @SSrrey97428]

@andrew_ilyas @logan_engstrom CIFAR-10 is cool, but have you tried MAGIC on tasks with noisier loss landscapes (like RL or GANs), and if so, does the metagradient approach stay stable when gradients get jumpy?

[Srreyansh Sethi @SSrrey97428]

@Synced_Global Because Radial Attention uses a fixed, shrinking‐window mask to drop all token pairs beyond a certain space‐time “radius,” did that ever prune away an interaction that actually had a non-negligible attention weight and hurt the model’s output?