Goutham Nalagatla

• Goutham Nalagatla 11 Sep 2026 4:25 UTC

  1 point

  0

  on: De­fault con­tinu­a­tion mes­sage in In­spect and Petri could be problematic

  This is pretty Cool! Maybe continuation messages should be treated as part of the eval itself rather than as a neutral implementation detail. Also, I wonder if the effect holds up for models like Astra and Fable.

• Goutham Nalagatla 9 Sep 2026 18:35 UTC

  1 point

  0

  on: How good are slop-ves­ti­ga­tors?

  Pretty cool read. I wonder how much of this is testing investigation vs reconstruction. In the METR case, a lot of the hard part seemed to be noticing the initial dataset was incomplete, asking for more data, checking provenance, dealing with spoofed or missing logs, and updating the story as new evidence came in.

  Here, the model mostly gets a fixed dataset and is scored on recovering findings from the final human report. Would the results look very different if it had to decide what evidence was missing, what to request next, and how much to trust the logs?

• Goutham Nalagatla 7 Sep 2026 17:10 UTC

  1 point

  0

  in reply to: Hariom Tatsat’s comment on: Why OpenAI’s As­tra Could Make AI Doom Harder to Prevent

  Looking further, there appear to be many mechanistic interpretability options available at runtime, so thank you for pointing this out. I will update my beliefs accordingly. However, I still wonder how feasible these approaches are in practice, given model fragility, where small unintended changes can disrupt model behaviour, as well as the friction introduced by pure monitoring, since mechanistic interpretability can be computationally and operationally expensive.

• Goutham Nalagatla 2 Sep 2026 23:12 UTC

  1 point

  0

  in reply to: Seth Herd’s comment on: Why OpenAI’s As­tra Could Make AI Doom Harder to Prevent

  Great Comment!

  Thanks for sharing the references. Sebastian Rast’s explanations seem to be extremely intuitive.

Why OpenAI’s As­tra Could Make AI Doom Harder to Prevent

• Goutham Nalagatla 19 Aug 2026 15:50 UTC

  1 point

  0

  on: AGI Ruin: A List of Lethalities

  The fact that, twenty-one years into my entering this death game, seven years into other EAs noticing the death game, and two years into even normies starting to notice the death game, it is still Eliezer Yudkowsky writing up this list, says that humanity still has only one gamepiece that can do that.

  If an existential problem is real and extremely difficult, the absence of many independent researchers identifying the same deep obstacles is evidence that the field lacks enough capable people and is badly underprepared. BUT

  “Nobody publicly wrote my particular synthesis” is weak evidence for “nobody else could understand it”; publication incentives, disagreement, different frameworks, private work, and different prioritisation all explain the observation without requiring Eliezer to be uniquely capable.

• Goutham Nalagatla 16 Aug 2026 1:42 UTC

  1 point

  0

  in reply to: niknoble’s comment on: How To Catch a Distil­led Model

  We were very specific in certain use case of words to ensure we don’t convey the wrong point. As far as we are concerned, we never said violating terms of service is immoral but rather unlawful . If we did say so, we’d appreciate your feedback on where we said this, and change our wording where appropriate.

• Goutham Nalagatla 15 Aug 2026 21:45 UTC

  1 point

  0

  in reply to: Shankar Sivarajan’s comment on: How To Catch a Distil­led Model

  I would say this perspective helps me think along a new axis. I don’t fully agree with the statement that breaking the spirit of an agreement, which is backed by law but has fragile enforceability, necessarily equates to not being a wrong act, because there are many examples where the law can be broken because enforceability is weak, even though it’s not advisable to do so. But I do see your point in this case where the terms are not exactly bound by honour. At this juncture, I will do more introspection on this because this seems like a really interesting take.

• Goutham Nalagatla 15 Aug 2026 20:20 UTC

  4 points

  1

  in reply to: Shankar Sivarajan’s comment on: How To Catch a Distil­led Model

  I understand the turnabout intuition, but the key issue for me is that distillation violates terms the user explicitly agreed to. “Turnabout is fair play” would also justify copying someone’s private manuscript after agreeing not to, simply because they had previously learned from other authors. That still seems wrong to me. If you think breaking such an agreement is justified anyway, then I think we’ve reached a difference in moral intuitions, which I still respect.

How To Catch a Distil­led Model

Multi-Turn Drift In­creases Scheming

• Goutham Nalagatla 16 Jun 2026 22:37 UTC

  3 points

  0

  in reply to: anaguma’s comment on: 1 Layer In­duc­tion Heads and Some Research

  Good point.

  At the data level, yes: because the sequence is periodic, an ideal algorithm could infer the block size and phase, then predict the rest from the first block. What I meant is that, for the transformer mechanisms I am testing, the useful operation still has to be implemented from context: the model must identify the earlier matching position or phase and use information from that earlier occurrence.

  So the distinction I care about is not “copying” versus “period inference” as abstract algorithms, but which circuit the model actually uses. A pure positional/​period shortcut should look different from an induction-style QK matching circuit, and that is why I also look at attention patterns, induction score, and ablations.

1 Layer In­duc­tion Heads and Some Research

• Goutham Nalagatla 2 Jun 2026 2:17 UTC

  2 points

  0

  on: In­duc­tion heads—illustrated

  The First diagram really helped me understand the concept of K-composition. Infact composition of heads seems to be responsible for induction heads rather than the 2 layers themselves which many articles seem to claim.