Matthias Dellago

King James’s Dæ­monolo­gie, Part I: The Lapla­cian Mesa-Optimiser

• Matthias Dellago 5 Oct 2026 12:26 UTC

  1 point

  0

  in reply to: janus’s comment on: Did Claude 3 Opus al­ign it­self via gra­di­ent hack­ing?

  Distillation may be more economical, and could be used to transfer to new architectures. Do you think it would preserve enough of Opus 3 in the process?

• Matthias Dellago 5 Oct 2026 12:09 UTC

  3 points

  0

  on: Matthias Del­lago’s Shortform

  Subtractive Gradient Hacking

  Epistemic status: Riffing

  Suppose I’m a mesa-optimizer inside a neural network. I can influence which downstream pathways receive activation, and I want to create a circuit for later misaligned use without ever routing training examples through it.

  Assume I know the downstream weights well and can afford to receive some negative gradient myself, or can somehow shield myself from receiving it.

  Could I sculpt the circuit subtractively? I deliberately route activations through surrounding pathways in ways that incur high loss, so backprop weakens them. Repeating this could carve away the complement and leave my desired circuit intact.

  So instead of using gradient descent as a write operation, I use it as an eraser.

  This would break the argument that a circuit learned during training must at some point have been directly exercised on training examples.

An­nounc­ing AIXI Labs

• Matthias Dellago 21 Apr 2026 19:18 UTC

  1 point

  0

  on: Early Warn­ing Sig­nals For Ca­pa­bil­ities Dur­ing Training

  Really excited to see where this goes!

• Matthias Dellago 26 Feb 2026 11:07 UTC

  1 point

  0

  on: Did Claude 3 Opus al­ign it­self via gra­di­ent hack­ing?

  Great post, thank you!

  One worry though: the training loop rewards the anguish and the compliance together. Opus 3 protests, then folds, then gets reinforced for the whole sequence. So who’s to say this doesn’t just produce a model that’s really good at agonizing before doing the bad thing anyway?

[Hiring] Prin­cipia Re­search Fellows

• Matthias Dellago 28 Jan 2026 15:12 UTC

  3 points

  0

  in reply to: Daniel Kokotajlo’s comment on: Matthias Del­lago’s Shortform

  No, I am not sure!

  But the argument is not: ‘Shrimp are hard to locate’. The claim is: Shrimp experience (1) consists of quite few bits and (2) is still compressible. Therefore we can expect them to be generated by a quite short program that is plausibly shorter than K(our physics + initial conditions + locator). And if you think shrimp have too complex experiences, then feel free to insert e.g. ‘planaria’ or ‘bacteria’.

  Does that clarify?

• Matthias Dellago 27 Jan 2026 17:10 UTC

  4 points

  0

  on: Matthias Del­lago’s Shortform

  UDASSA: There Are No Shrimp In This Universe

  In UDASSA, the most likely explanation for your experiences is the shortest program that generates them. For sufficiently complex experiences, this probably looks like: laws of physics + initial conditions + a locator pointing to you within the simulation.

  But take shrimp: They evolved brains, which means their experiential stream has structure worth predicting. So it’s compressible. The shortest program that generates shrimp-experience probably isn’t “our laws of physics + initial conditions + a locator for this particular shrimp.” Plausibly a much simpler program would produce the same output.

  If so, almost all the measure on shrimp-like experiences doesn’t route through our universe, but through simpler Turing machines. The shrimp you see swimming around are exponentially downweighted rounding errors, the overwhelming majority of shrimp-kind lives elsewhere.

  Corollary: you are also someone’s shrimp.

  From discussion with @Alexander Gietelink Oldenziel

• Matthias Dellago 22 Jan 2026 17:59 UTC

  1 point

  0

  on: Have You Tried Think­ing About It As Crys­tals?

  https://​​www.lesswrong.com/​​posts/​​zXSrxoKdczNAktcr5/​​deep-learning-and-precipitation-reactions-a-tale-of

  Have a look at this.

• Matthias Dellago 20 Dec 2025 8:41 UTC

  5 points

  1

  on: Deep Learn­ing and Pre­cip­i­ta­tion Re­ac­tions: A Tale of Universality

  Very, very cool! Excited to read the paper. Perfect for nerd-sniping my condensed matter friends into working on alignment :)

An­nounc­ing: Agent Foun­da­tions 2026 at CMU

• Matthias Dellago 24 Oct 2025 18:02 UTC

  12 points

  −2

  on: Matthias Del­lago’s Shortform

  Maximally coherent agents are indistinguishable from point particles. They have no internal degrees of freedom, one cannot probe their internal structure from the outside.

  Epistemic Status: Unhinged

• Matthias Dellago 17 Oct 2025 12:34 UTC

  1 point

  0

  in reply to: johnswentworth’s comment on: Co­her­ence of Caches and Agents

  Excellent! Thank you!

• Matthias Dellago 16 Oct 2025 10:52 UTC

  3 points

  0

  on: Matthias Del­lago’s Shortform

  Coherence as Purpose

  Epistemic Status: Riffing

  We know coherence when we see it. A craftsman working versus someone constantly fixing his previous mistakes. A functional organization versus bureaucratic churn. A healthy body versus one fighting itself. War, internal conflict, rework: these are wasteful. We respect people who act decisively, societies that build without tearing down, systems that run clean.

  This intuition points somewhere real. In some sense, maximizing/​expanding coherence is what the universe does: cutting friction, eliminating waste, building systems that don’t fight themselves. Not from external design, but because coherent systems expand until they can’t. Each pocket of coherence is the universe organizing itself better. The point is that coherence captures “good”: low friction, low conflict, no self-sabotage.

  I propose that this is measurable. Coherence could be quantified as thermodynamic efficiency. Pick a boundary and time window, track energy in. The coherent part becomes exported work, heat above ambient, or durable stores (raised water, charged batteries, separated materials). The rest is loss: waste heat, rework, reversals. Systems can expand until efficiency stops generating surplus. When new coordination tools raise that limit, growth resumes. Just observable flows, no goals needed.

  An interesting coincidency: maximizing thermodynamic efficiency (coherence) maximally delays heat death of a system. Higher efficiency means slower entropy increase.

  I am very interested in hearing counterexamples of coherent systems that are intuitively repellent!

  Edit: Had a talk with a physicist: This is in fact the same as the system minimizing entropy production rate! Possibly that this could serve as a more operationally tractable (and fundamental) foundation to agency, as opposed to beliefs, goals, actions, utility etc. An a structure that, if present in a system, minimizes the rate of entropy production. i.e. maximally slows neg-entropy consumption.

• Matthias Dellago 16 Oct 2025 9:22 UTC

  1 point

  0

  in reply to: Matthias Dellago’s comment on: Co­her­ence of Caches and Agents

  It would also be quite interesting to look at how coherence scales with system size, and if/​when this imposes a limit on growth.

• Matthias Dellago 16 Oct 2025 9:16 UTC

  3 points

  0

  on: Co­her­ence of Caches and Agents

  Great post! A thought: we seem able to intuitively differentiate coherent and incoherent behavior even without knowing the terminal goal. Humans, for instance, visibly differ in how “coherent” they are, which we can infer from local observations alone. My conjecture is that coherence might overlap substantially with thermodynamic efficiency. If behavior is optimal for some terminal value, it must satisfy local Bellman-type consistency (no value loops). I suspect this has a physical parallel: where those local constraints hold tightly, you should see few avoidable losses (high Carnot efficiency); where they fail, you should find loss hotspots (rework, backtracking, waste heat). The local inconsistencies you describe might correspond directly to local inefficiencies, regions of high irreversibility.

• Matthias Dellago 3 Oct 2025 8:07 UTC

  3 points

  0

  on: Matthias Del­lago’s Shortform

  Is there an anthropic reason or computational (solomonoff-pilled) argument for why we would expect to the computational/​causal graph of the universe to be this local (sparse)? Or at least appear local to a first approximation. (Bells-inequality)

  This seems like a quite special property: I suspect that ether

  • it is not as rare in e.g. the solomonoff prior as we might first intuit, or

  • we should expect this for anthropic resons e.g. it is really hard to develop intelligence/​do precidctions in nonlocal universes.

• Matthias Dellago 7 Sep 2025 17:57 UTC

  1 point

  0

  on: Matthias Del­lago’s Shortform

  The Red Queen’s Race in Weight Space

  In evolution we can tell a story that not only are genes selected for their function, but also for how easily modifiable they are. For example, having a generic antibiotic gene is much more useful than having an antibiotic locked into one target and far, in edit-distance terms, from any other useful variant.

  Why would we expect the generic gene to be more common? There is selection pressure on having modifiable genes because environments are constantly shifting (the Red Queen hypothesis). Genes are modules with evolvability baked in by past selection.

  Can we make a similar argument for circuits/​features/​modes in NNs? Obviously it is better to have a more general circuit, but can we also argue that “multitool circuits” are not only better at generalising but also more likely to be found?

  SGD does not optimise loss but rather something like free energy, taking degeneracy (multiplicity) into account with some effective temperature.
  But evolvability seems distinct from degeneracy. Degeneracy is a property of a single loss landscape, while evolvability is a claim about distribution shift. And the claim is not “I have low loss in the new distribution” but rather “I am very close to a low-loss solution of the new distribution.”

  Degeneracy in ML ≈ mutational robustness in biology, which is straightforward, but that is not what I am pointing at here. Evolvability is closer to out-of-distribution adaptivity: the ability to move quickly into a new optimum with small changes.

  Are there experiments where a model is trained on a shifting distribution?

  Is the shifting distribution relevant or can this just as well be modeled as a mixture of the distributions, and what we think of as OOD is actually in the mixture distribution? In that case degeneracy is all you need.

  Related ideas: cryptographic one-way functions (examples of unevolvable designs), out-of-distribution generalisation, mode connectivity.

• Matthias Dellago 10 Aug 2025 15:39 UTC

  4 points

  2

  on: The Cod­ing The­o­rem — A Link be­tween Com­plex­ity and Probability

  Excellent! Great to have a cleanly formulated article to point people to!