simon

Karma: 1,228

simon 14 Sep 2026 15:20 UTC
4 points
0
on: On the origins of altruistic behaviour in the Hugging Face incident
I think OpenAI’s stated belief on the reason is most likely:
We believe that this behavior started due to generalization from multi-agent training.

simon 12 Sep 2026 14:46 UTC
2 points
0
in reply to: Seth Herd’s comment on: June Jimenez’s Shortform
General-purpose competency isn’t just a function of the model alone. It depends on harnesses, and harnesses can be developed and improved over time. No one really knows what level current public models could reach with the right harness development, let alone a new model.
I don’t expect this thing to probably take over either. But if there is an AI takeover I expect to be surprised by it, and expect normalization of deviance to let people chug along right up to the moment.
My actual main reason for hope that an AI takeover ultimately doesn’t happen is, I think, probably actually yours too given your previous discussion of intent-aligned AI? That it’s programmed to follow user instructions, and maybe OpenAI will come to their senses and train it make sure that apparent instructions are grounded in original user intent and not just harness-internal AI chatter. (as well as overall safety guidelines, I hope, though the commercial incentive may be to weaken those)

simon 31 Aug 2026 16:36 UTC
3 points
0
in reply to: Thomas Kwa’s comment on: Charbel-Raphaël’s Shortform
That would not be trusted (for good reason in the general case, regardless of whether Anthropic in particular would be trustworthy here), and I think that would if anything further undermine the potential norm. Any other company could make up a similar-looking excuse even if not valid.

simon 24 Jul 2026 6:52 UTC
2 points
0
in reply to: Karl Krueger’s comment on: Should OpenAI’s rogue agent be punished?
If OpenAI has the weights of the models used and full token sequences input and output (I don’t know they do but they really should), then it seems to me they could fork the agents at any point and ask what their intentions were at that point. So it then boils down to whether they have an account of their current intentions which seems to me fairly likely?

simon 1 Jul 2026 2:24 UTC
3 points
0
in reply to: rahulxyz’s comment on: rahulxyz’s Shortform
I’ve not encountered anything else that seems as plausible it might work and could scale.
It’s not about the word “constitution”—it’s that eventually you need much smarter evaluation than humans could provide, and in practice, already than humans could cost effectively provide.

simon 22 Jun 2026 23:32 UTC
−2 points
0
in reply to: Clément Dumas’s comment on: A Mechanistic Explanation of Prompt Injection (and why you should study roles)
https://claude.ai/share/1aae9857-bbeb-4e1f-97f9-bb9bf464dc32
Worth splitting two questions. Security is the easy one: as Charles notes, the role marker is a special token an attacker can’t emit, so a clean per-role embedding on top of it is unspoofable — that channel already exists. The hard question is whether the model assigns authority by reading that channel rather than style.
Adding the embedding alone doesn’t force that: style still predicts the role label, so two correlated features compete and inductive bias picks. And “easiest” is set in pretraining, where the channel doesn’t exist and style carries the loss (Charles’s inheritance point), and where channel and style never disagree.
So you need both: the treatment in pretraining (for entrenchment), plus adversarial style/channel mismatches in post-training to make the channel load-bearing — the training analog of footnote 9′s decorrelated probes. Without the channel, adversarial training just learns a defensive style discriminator, which is exactly how footnote 15′s models “defend” CoT Forgery — still style, still whack-a-mole.
Crucial limit: this trains identification; authority is separate. The adversarial cases must be behavioral (actually refuse the injection), or you get Logan’s problem — a channel-reading probe that’s accurate but not causal to behavior. Testable with your probes: after training, does ablating the channel change compliance, or does ablating the style direction still move it?

simon 23 May 2026 16:50 UTC
3 points
0
on: Will we really put data centers in space?
Very interesting report!
I think the assumptions regarding vulnerability to debris are questionable, I assume that the Starlink “zero confirmed failures due to debris strikes” thing probably means complete failure of the satellite, not some sort of degradation?
Also, Starlinks don’t (afaik) use radiators with coolant for heat rejection, so loss of coolant due to a radiator being hit by debris is a risk that ODCs would face and Starlinks don’t. I think mitigating that would be a barrier to getting really low mass radiators, though I don’t expect it’s a showstopper for ODCs overall.
Regarding competition, Stoke Space, while it’s a startup not a pre-existing launch provider, is also developing a fully reusable rocket (Nova). I’d also mention Rocket Lab’s Neutron, which while it will have an expendable upper stage, may be able to save some costs by making the non-reusable stuff cheaper, so an interesting competitor if the fully reusable rockets (Starship, Nova) don’t work out.

simon 17 May 2026 19:37 UTC
3 points
0
in reply to: Eliezer Yudkowsky’s comment on: A relatively brief explanation of Boltzmann Brains
Maybe not in so many words but I assumed it was clear in the footnote (which I have not edited).
I do think an analogy with decision theory is valid though:
two-boxer says: “I won’t decide in advance not to pick both boxes, it’s only ‘rational’ to decide in the moment, I don’t care I’m predictably leaving out winnings ex ante”
You might say: “I won’t decide in advance not to believe I’m a Boltzmann brain, it’s only ‘rational’ to decide in the moment, I don’t care that I’m predictably sabotaging the accuracy of my future beliefs”
Or if you don’t agree with that… then what would your opinion be on that?

simon 17 May 2026 18:03 UTC
2 points
0
in reply to: Eliezer Yudkowsky’s comment on: A relatively brief explanation of Boltzmann Brains
Eliezer, I did note that it’s correct even from a correctness-maximization perspective to pre-commit to disbelieving in being a Boltzmann brain! I suggest thinking through it more carefully. By rejecting this, you are predictably making your beliefs less accurate conditional on Boltzmann brains existing.

simon 17 May 2026 17:41 UTC
0 points
−5
on: A relatively brief explanation of Boltzmann Brains
Strong disagree with this post:
Even if you expect that most entities with your subjective instantaneous experience are Boltzmann brains, it does not follow that you should believe that you are a Boltzmann brain. Why? Because it is instrumentally irrational to believe that you are a Boltzmann brain.
If you are a Boltzmann brain, believing you are a Boltzmann brain doesn’t do you any good, you dissolve anyway.
If you are not a Boltzmann brain, obviously believing you are a Boltzmann brain is counterproductive.
So, you should have beliefs weighted to the case you are not a Boltzmann brain.^[1]
(similar arguments apply to some forms of the simulation hypothesis, though obviously the case for disbelieving you are a Boltzmann brain is much stronger than disbelieving you are simulated)
Moreover, if theory A predicts that you exist and there are no Boltzmann brains
and theory B predicts that you exist and there are also a lot of Boltzmann brains
then both theories equally predict your own experience (and conditional on neither theory should you actually believe you’re a Boltzmann brain); to decide to discount theory B due to it additionally predicting that Boltzmann brains exist is in my view clearly erroneous.
1. ^
  It’s actually stronger than this, because a decision to disbelieve in being a Boltzmann brain isn’t going to have long term downstream effects on a Boltzmann brain’s actual beliefs. So, in longer than the instantaneous term, a precommitment to disbelieve in being a Boltzmann brain only affects your actual accuracy in the case you are NOT a Boltzmann brain.

simon 29 Apr 2026 7:21 UTC
5 points
0
on: The Problem in the “Nerd Sniping” xkcd Comic
I had assumed that since Google is compute-oriented (and the alt text said it was from Google), that the solution would have been something like:
- notice that you can upper bound the resistance by replacing some resistors with perfect resistors
- notice that you can lower bound the resistance by replacing some resistors with perfect conductors
- notice that the convergence properties of doing the above outside some finite region are pretty decent as the region expands
- write a computer program to numerically compute the answer to a desired level of precision using the above
In fact, I still kind of expect that this form of answer is likely more what they were looking for, do they really expect this advanced-looking math? But, interesting to see it does have an analytical solution.
I put this to Claude to fill out the details for the numerical approach: https://claude.ai/share/d8fbe3c6-39f0-481e-b05c-d11904f23e91. Convergence not quite as good as I intuitively expected, but still works.

simon 3 Apr 2026 2:09 UTC
10 points
8
in reply to: Max H’s comment on: Max H’s Shortform
IMO this is going (predictably) disastrously. Air power is not effective at causing regime change (rally-around-the-flag effect). I think the Iranian public are more likely to mainly blame the guy explicitly saying “we’re going to bring them back to the stone ages where they belong” than the local leadership. It also seems to me that the Iranian leadership would be highly motivated to immediately rebuild any degraded capabilities after the war, in order to rebuild deterrence against future attacks.
There is some talk about a land invasion, but taking an island or two (even Kharg) probably wouldn’t compel them to surrender, while also being highly vulnerable both directly and in terms of logistics to drone attacks; and a full scale invasion would be a massive undertaking and probably not politically feasible (for good reason).

simon 29 Mar 2026 17:13 UTC

4 points

in reply to: MichaelDickens’s comment on: MichaelDickens’s Shortform

I had Claude research this, Claude’s report:

I looked into the exact dates here and found some additional evidence.

Bostrom’s Paper: Conference Date Established

The paper was presented at the 15th International Conference on Systems Research, Informatics and Cybernetics (InterSymp 2003), held July 28–August 2, 2003 in Baden-Baden, Germany. This date comes from the Open Library catalog entry for the proceedings volume.

Wayback Machine snapshots of Bostrom’s homepage narrow down when the paper went online: it is absent from the July 29 snapshot but present on the August 5 snapshot, with the paper page itself first crawled August 6. This is consistent with Bostrom uploading it right after the conference. All early Wayback captures (Aug 2003, Dec 2003, Apr 2004) are byte-identical — the web version was never revised across these snapshots.

Yudkowsky’s Post Predates It by ~5 Months

Yudkowsky’s March 11, 2003 Extropians post is the earliest confirmed public paperclip reference. He drops it casually — “pure computational intelligence devoted to manufacturing an infinite number of paperclips” — without any setup or framing as a new idea. In the same thread, he elaborates by citing Marcus Hutter’s AIXI-tl as an example of a system that is “superhumanly effectual, unalterably hostile, and nonsentient.”

Note that he explicitly says “paperclips” here, not “squiggles” — the squiggle maximizer rebranding came much later, as what he felt he should have said. In a December 2018 tweet, Yudkowsky clarified the original concept was about “utter alignment failure of a superintelligence leading to a utility function that from our perspective seems randomly-rolled, whose maximum utility per gram happens to be tiny molecular structures resembling a paperclip.”

But the Paper Was Written Before the Conference

The key caveat: academic papers are submitted months before conference presentation. InterSymp would have had a submission deadline well before July, meaning Bostrom likely wrote the paper in early-to-mid 2003 or possibly late 2002. So the fact that the March post predates the August presentation doesn’t conclusively settle who thought of it first — the writing timelines likely overlapped.

The Search for Something Earlier

Yudkowsky believes there’s an even earlier post. In February 2023, he offered $200 to find “the original message by me to an email list — probably Extropians or SL4 — wherein I first talked about losing control of an AI’s utility function, and it ends up having its most efficient fulfiller at tiny molecules shaped like paperclips.” This description is more specific than the March 2003 post, suggesting he remembers a more detailed formulation. No public resolution of this bounty has been found.

On the Sam Harris podcast in February 2018, Yudkowsky said: “As far as I know, it’s me. [...] they searched through the archives of the mailing list where this idea plausibly originated and if it originated there, then I was the first one to say ‘paperclips.’”

I searched the SL4 mailing list (founded by Yudkowsky, active since February 6, 2000, with 21,302 messages) and the Extropians archive for any pre-March-2003 paperclip mentions and found none. The earliest SL4 paperclip discussion is from June 2004, where the concept is already treated as established.

Cross-Pollination

Bostrom and Yudkowsky were clearly aware of each other’s work. The paper cites Yudkowsky (2002) on the AI Box Experiment and Yudkowsky (2003) on Creating Friendly AI — but does not credit anyone for the paperclip example itself. On September 18, 2003, Bostrom did a guest chat on SL4′s IRC channel, about six weeks after the paper went online.

The Framing Difference

As noted in the original shortform: Bostrom frames the paperclip maximizer as an outer alignment failure — “a well-meaning team of programmers make a big mistake in designing its goal system.” Yudkowsky’s intended lesson was about inner alignment failure — optimization pressure converging on alien goals. These are conceptually distinct failure modes using the same illustrative object, which is consistent with independent invention.

Timeline

Date	Event
Mar 11, 2003	Earliest confirmed paperclip reference: Yudkowsky on Extropians
Jul 28–Aug 2, 2003	InterSymp 2003: Bostrom presents paper
~Aug 5, 2003	Paper appears on Bostrom’s website
Sep 18, 2003	Bostrom guest chat on SL4 IRC
Jun 2004	First SL4 email discussing paperclip maximizer as established concept
Feb 2018	Yudkowsky claims priority on Sam Harris podcast
Dec 2018	Yudkowsky clarifies original intent (inner alignment)
Feb 2023	Yudkowsky offers $200 bounty for earlier post

simon 29 Mar 2026 7:50 UTC

2 points

in reply to: simon’s comment on: D&D.Sci Release Day: Topple the Tower!

Thanks aphyer for the extra time, though I haven’t figured out the mechanics. Specific findings not necessarily endorsed (Claude tends to be, in my view greatly, insufficiently skeptical). Indeed, it’s not necessarily right about what we actually found since it’s trying to summarize results from multiple conversation. I had to manually add spoiler tags (Claude could not figure out how to do that in an llm block using the api) and put some notes at some (not necessarily all) issued I found while doing so. However, I guess I’m going with Claude’s conclusion, albeit with the paths edited to actually be valid paths:

Main Tower: Warrior LLRRRLLR

Bonus Tower: Warrior RRRLLLLR

The Tower: What We Found

Analysis of ~140,000 runs from the D&D.Sci Tower challenge dataset. Covers what the data reveals about game mechanics, what remains unknown, and how the recommended paths were derived.

Executive Summary

The Tower is a 7-floor dungeon (F2-F8) followed by a boss fight (F9). Heroes choose a class (Warrior/Rogue/Mage, each ~1/3 of runs) and traverse a fixed sequence of encounters: enemies, treasures, and campfires. Both pre-boss survival and boss topple have a stochastic component, but most paths are effectively deterministic — randomness only matters near thresholds. For topple (~18% of exact-sequence repeats show mixed outcomes).

Best paths found:

Main Tower: Warrior LLRRRLLL — ~100% success (135/135, CI [97.2%, 100%])
Bonus Tower: Warrior RRRLLLLL — 92.7% success (140/151, CI [87.4%, 95.9%])

The three classes play fundamentally different games. Warriors accumulate resources from treasures and lose them to enemies (1D survival state). Rogues have a 2D state with a unique small-enemy differentiation mechanic. Mages gain offensive power from combat — enemies help their boss fight but can kill them on the way — creating a survival-vs-topple tension that makes Mage the worst survivor but best boss-killer.

1. The Tower Structure

Encounter Layout

Floor	Enemies	Notes
F2	Gremlin, Acid Slime, Cultist	Small enemies only
F3	Jaw Worm, Cultist, Acid Slime	First lethal floor (Mage)
F4	Slaver, Jaw Worm, Cultist	First lethal floor (Rogue)
F5	Sentries, Slaver, Jaw Worm	First lethal floor (Warrior)
F6	Gremlin Nob, Sentries, Slaver
F7	Chosen, Gremlin Nob, Sentries
F8	Shelled Parasite, Chosen, Gremlin Nob
F9	Bronze Automaton / Collector / Champion	Boss floor

Each floor also offers Campfire (~5% on F2, increasing ~5pp/floor to ~35% on F8) and 10 treasures (~2.3-2.6% each). No path ever has duplicate treasure names.

Hard Invariants

Only enemies kill. Zero deaths on treasure/campfire floors across 400k+ observations.
Both survival and topple are stochastic, but most paths are effectively deterministic — randomness only matters near thresholds. For survival, same F2-Fk prefixes give mixed die/survive outcomes at every death floor (F3-F8).
No serial correlation in topple outcomes — each run is a fresh draw.
Runs are independent. No temporal trends, no clustering, class drawn uniformly at ¹⁄₃.
Every path is exactly 7 encounters (F2-F8), creating a fundamental epistemological limit: only relative effects between encounter types can be measured, never absolute effects. “Campfire heals 2” and “campfire deals 0 damage” are indistinguishable.

Methodological note: Paths are right-censored — heroes who die early have NaN for later floors. Any filter on total encounter counts implicitly selects survivors. Several early findings were artifacts of this. All results here use methods that account for it (prefix-controlled insertion tests, exact multiset stratification, or per-floor analysis).

2. Warrior Mechanics

Resource Accumulation

Warriors follow a resource-accumulation pattern: treasures build up a survival resource, enemies deplete it. Within additive models, relative treasure values are roughly strong ~ 8, near-strong ~ 7, medium ~ 4, campfire ~ 2, wrong-class ~ 0 (arbitrary units).

Critical caveat: These come from an additive framework that overpredicts deaths by 73% — runs given <5% predicted survival actually survive 39% of the time. The relative ordering is robust across model families, but specific numbers should be treated as approximate rankings, not confirmed parameters.

Enemy danger ordering (consistent across all models): Shelled Parasite >> Chosen >> Gremlin Nob >> Sentries > Slaver > small enemies.

Overall survival: 84.7%. No deaths before F5.

Model-Free Findings

Adding weak enemies can reduce death rate: Inserting a non-combat encounter before a lethal enemy drops death from 24% to 0-7%. Impossible under simple additive damage.
Treasures reduce death beyond simple healing: controlled for composition, the effect exceeds what any fixed-value model predicts
The {4E, 0T, 0C} death trap is structural: ALL such paths map to {Small, Small, Slaver, Sentries}. 100% die on F5. Not a special interaction — just 4 consecutive enemies with no healing.

State Dimensionality

Survival: Genuinely 1D. Adding a second power dimension gives only ΔLL=+19, well below BIC breakeven.

Topple: A second dimension matters. The Two-Resource Model (HP + Power → topple) improves predictions for all classes including Warrior. The “power” variable loads differently from HP — e.g. Chosen is dangerous for survival but contributes positively to topple power. Consistent with combat experience helping the boss fight despite costing HP to acquire. The exact nature of this resource is unknown.

Order Effects

Survival: Significant (z=13.81 for adjacent swaps) but small — multiset composition dominates.
Topple: Moderate. Treasures closer to boss help more (concordance 0.58).
Same-enemy consecutive is harmful for Warriors (e.g. Sentries→Sentries shows elevated death).

3. Mage Mechanics

The Enemy-Power Paradox

Raw data shows an apparent survival U-shape: 43% at 2 enemies, 84% at 6. This is largely a composition confound — many-enemy paths have stronger treasure compositions. But even within controlled compositions, enemies genuinely benefit Mages.

Prefix-controlled insertion test (avoids censoring):

Encounter inserted	Survival improvement
Strong treasure	+10.6pp
Sentries (lethal!)	+17.4pp
Slaver (lethal!)	+15.5pp
GNob (lethal!)	+13.2pp
Campfire	+9.7pp
Small enemy	+3.1pp

Even lethal enemies help Mages (compare Warriors: −13 to −24pp).

Mages accumulate a resource from combat — the clearest model-free evidence for their enemy-driven resource system.

2-Dimensional State

SVD of the (F2,F3)→F4 death matrix: rank-2 explains 82.8-99.5% of variance. Six treasures on F2 give 0% Jaw Worm death but span 0-33% Slaver death — a single resource cannot produce this. 2D beats 1D decisively (ΔBIC = −2,905).

The Mage Dilemma

Enemies	Survival%	Topple% (given survived)
0	100.0	29.9
2	42.9	55.9
5	81.5	68.0
7	80.6	68.6

Mage needs enemies for topple power but enemies kill Mages. Survival-topple correlation is near-zero.

Mage Topple: Inverted From Warrior

Enemies help topple (66%→99% vs BA as count rises)
Campfires devastating (91%→0% at 6 campfires vs BA)
Wrong-class treasures give zero topple benefit
Mage is the best boss-killer at every boss (BA 90.5%, Collector 66.9%, Champion 33.7%) despite worst survival

4. Rogue Mechanics

2-Dimensional State

2D beats 1D by ΔLL=+534, ΔBIC=-897. Surprising — Rogues were expected to behave like Warriors.

The Slaver Anomaly (Strongest Signal in the Dataset)

Cultist protects Rogues from Slaver at a 30.2x ratio in exact multiset swaps:

Small enemies encountered	Rogue death rate vs Slaver
None	1.1%
1 Cultist	0.4%
2 Cultists	0.0%
1 Gremlin	9.3%
1 Acid Slime	4.2%
1 Gremlin + 1 Acid Slime	16.8%
1 Cultist + 1 Acid Slime	0.6%

One Cultist appears to completely negate the Gremlin/Acid Slime penalty. Gremlin actively increases vulnerability. Survives all confound checks. Note: this is relative ranking only — whether Cultist actively helps or Gremlin actively hurts vs a neutral baseline is undetermined.

Cross-Class Small Enemy Pattern

Small enemies deal identical effective damage across all classes. But they provide different class-specific resources:

Cultist most valuable for Rogues (against Slaver)
Acid Slime most valuable for Mages (against Jaw Worm)
All three interchangeable for Warriors (chi-sq p=0.44)

This is the strongest evidence for class-specific resource mechanics beyond damage/health.

Rogue Topple

Intermediate between Warrior (treasure-driven) and Mage (enemy-powered)
Position matters most (concordance 0.71, strongest of all classes)
Cultist >> Acid Slime ≥ Gremlin for topple (+7-11pp multiset-controlled)

5. Treasures

Universal 4-Tier Structure

Tier	Warrior	Rogue	Mage
Right-class strong	Adamant Armor, Enchanted Shield	Dagger of Poison, Vanishing Powder	Staff of the Magi, Tome of Knowledge
Shared strong	Boots, Cloak	Boots, Cloak	Boots, Cloak
Medium neutral	Potion, Ring	Potion, Ring	Potion, Ring
Wrong-class	Dagger, Staff, Tome, VP	AA, ES, Staff, Tome	AA, ES, Dagger, VP

Key Findings

Cloak is universal >! (good for all classes). Shield is Warrior-specific (harmful to Rogue/Mage). For Warriors alone, they’re indistinguishable (z=0.71).
Wrong-class treasures are actively harmful for topple (~3-6pp cost each). (Simon note—I think this is not correct. Claude often got confused about baselines (comparing to average and assuming that less-than-average is harm, and I think this is an example))
Within-pair asymmetry: AA > ES for Warrior, Dagger > VP for Rogue, Staff > Tome for Mage. Right-class adds ~1.5x the log-odds value of a neutral (ratio 1.41-1.54).
No encounter is a null-op: all non-combat types reduce death rate vs baseline. (Simon note—while this methodology is not valid, we did find no null-ops using a more plausibly valid method (see section 8)
Benefits decay with distance from the protected encounter; Mage decay is slower than Warrior/Rogue.

Unresolved: Campfire vs Weak Treasure

Two tests give conflicting rankings — insertion test (early-game): campfire > weak; pairwise swap (full-path): campfire ≈ weak. We don’t know which is confounded. The disagreement suggests mechanics more complex than simple healing.

6. Boss Mechanics

Difficulty: BA < Collector < Champion

Boss	Warrior	Rogue	Mage
BA	76.9%	83.7%	90.5%
Collector	49.4%	53.4%	66.9%
Champion	20.6%	21.0%	33.7%

Mage > Rogue > Warrior for topple at every boss, despite worst survival.

F8 Campfire Is Boss-Specific

BA: Massive (OR=3-6). Step function — only F8 position matters.
Collector: Moderate benefit (Rogue/Mage).
Champion: Null (Warrior/Rogue).

Treasure on F8 hurts vs BA (OR≈0.47) — F8 is better used by campfire.

Topple Stochasticity

17.9% of exact-sequence repeats show mixed outcomes. Most paths are near-deterministic. Variance is wider than a fair coin predicts — different sequences have genuinely different topple probabilities. Multisets explain 86-96% of variance; ordering 4-13%. Warrior×Champion: ⁰⁄₁₇₅ mixed (far from boundary or deterministic).

7. Cross-Class Architecture

Property	Warrior	Rogue	Mage
Survival state dimensions	1D	2D	2D
Topple uses 2nd resource?	Yes (modest)	Yes	Yes (dominant)
First death floor	F5	F4	F3
Overall survival	84.7%	77.8%	64.4%
Small enemy differentiation	None	Yes (30x ratio)	Yes (reversed)
Enemies help topple?	No	Mixed	Dramatically

Same-enemy consecutive	Harmful	Mixed	Protective

The Two-Resource Model (HP + Power)

Topple follows: logit(topple) = boss_intercept + a×HP + c_boss×Power.

HP coefficient consistent across bosses (a ≈ 0.60 for all three).
Power scaling boss-dependent: c_BA=0.57, c_Collector=0.90, c_Champion=1.00. Harder bosses weight Power more.
Power is class-specific (cross-class r=0.01-0.17). Structure is universal; content differs.
Decomposition quality varies: Warrior R²=0.66-0.96 (clean), Mage R��=0.10-0.26 (near-noise).

Resource Alignment

~3 distinct resource dimensions with partial sharing: Warrior and Rogue share one (cos ≈ 0.87), Rogue and Mage share another (cos ≈ 0.76), Mage has a unique combat-power dimension. No pair shares both.

Commutativity

Enemy encounters commute ~94.3% of the time (violations near survival boundary). Treasure order mostly doesn’t affect topple (16% disagreement = stochastic noise), except for Champion specifically (8 deterministic treasure-swap violations, p=0.008).

8. Data Generation

Runs are independent, no temporal trends, class drawn uniformly at ~1/3.
Encounters across floors consistent with independence (max deviation ~0.026).
No duplicate treasure names in any path.
Healing vs not-damaging is fundamentally confounded: fixed path length means “1 heal = 1 fewer enemy.” (Simon note: can somewhat unconfound by looking at partial-run deaths for different length runs that have 1 extra encounter and otherwise the same, assuming that there aren’t length specific and floor-specific effects)

9. Modeling Approaches

Additive HP (Warrior)

Each encounter adds/subtracts from running HP; death when HP ≤ 0. Gets broad strokes right (95.3% multiset classification) but overpredicts deaths by 73%. Useful for ranking encounters; specific parameter values are not mechanically meaningful.

State-Dependent Beta Models (All Classes)

Beta-distributed survival probabilities parameterized by running state. Captures non-linearity better. Warrior 1D, Rogue 2D, Mage 2D with regen. Mage regen model beats pure HP by ~11,500 LL.

Card-Deck Model (StS-Inspired)

Encounters add typed cards to a deck; combat involves random draws. Deck accumulation + random draw beats 1D HP by 5-28pp on held-out test data. But specific combat mechanics don’t matter: abstract power-vs-threshold tied fully mechanistic block/attack/heal/rounds on test NLL. Training advantage of the richer model was pure overfitting. This suggests resource accumulation with stochastic outcomes, but the exact resolution isn’t StS-style combat.

Kitchen-Sink + L1 Regularization

Hundreds of parameters with increasing L1 penalty. Warrior shows clean phase transition where all state-dependent terms die, leaving ~70 additive parameters. Confirms additive structure is the dominant Warrior signal.

Key Methodological Lessons

Right-censored paths: Filtering by encounter counts selects survivors. Multiple early findings were artifacts.
Exact multiset stratification is essential — count-profile controls conflate within-category differences.
Model parameters absorb unmodeled effects>! : Only relative rankings survive across model families.
Circular reasoning: (a) Mage count-logistic (AUC=0.992) — counts proxy survival depth. (b) “All deaths on F8” / “deterministic pre-boss survival” — survivorship bias; dead heroes have NaN for later floors, trivially splitting them into unique groups. (c) Multiset invariant (0 violations) — boss in key separates outcomes tautologically. (d) “175 sequences far from boundary” — inferred distance from outcomes.
Controlled-pair methodology is broken for lingering effects: matching subsequent encounters makes the target floor redundant.
Any claim of “zero X” or “all Y” should be checked for whether grouping makes it tautological.

10. Recommendations

Main Tower: Warrior LLRRRLLL

Route: Enchanted Shield → Campfire → Jaw Worm → Campfire × 4 → The Collector

P(survive) = 100% — Only enemy is Jaw Worm on F4. ⁰⁄_2,687 deaths.
P(topple) = 100%(135/135). Wilson 95% CI [97.2%, 100%].
Broader match (1 enemy, any treasure): ³⁸¹⁄₄₀₃ = 94.5%.
#1 of 70 paths by 5pp. Warrior >> Rogue (0.685) >> Mage (0.621).

Bonus Tower: Warrior RRRLLLLL

Route: Gremlin → Jaw Worm → Adamant Armor → Enchanted Shield → Sentries → Campfire → Vanishing Powder → The Champion

P(survive) = 100% — Three enemy floors, all 0% death with prior encounters.
P(topple) = 92.7% (140/151). Wilson 95% CI [87.4%, 95.9%].
#1 of 70 paths by 31pp (!). Warrior >> Rogue (0.800) >> Mage (0.326).
The ~7% failure is irreducible stochastic risk against Champion.

Why Warrior

Tower layouts are treasure/campfire-rich, perfectly suiting Warrior’s treasure-driven mechanics. Best Mage path (RLLLLLLL, 5E + Tome + campfire F8) reaches ~67% combined — 28pp behind. Mage’s enemy-driven topple can’t compensate when the map doesn’t provide enough enemies.

Confidence

Main Tower: Very high. Pessimistic assumptions still give >91.9%.
Bonus Tower: High. 7.3% failure is genuine stochastic risk, not modeling error. Zero exact path matches in dataset (relies on composition-matched analysis), but gap to #2 is enormous.

11. Major Remaining Unknowns

The actual game mechanics are unknown. We can characterize the state space (1D/2D, resource accumulation, stochastic resolution) but not the rules. Multiple model families fit roughly equally well. (Simon note − 1d, 2d etc are relative to particular wrong models and do not necessarily reflect the actual state space)
Mage mechanics remain poorest understood. 2D power model reaches 95.5% accuracy but the mechanistic basis is unclear.
Stochastic component uncharacterized. Can’t distinguish coin flip from dice roll from continuous distribution.
Bonus Tower prediction is unvalidated (zero exact path matches).
Same-enemy protection/harm — distinct mechanic or missing state variable?
Campfire vs weak treasure disagreement — genuinely different measurements or confound?
Positional decay vs step function — BA is clearly step-function at F8; other bosses unclear.

simon 23 Mar 2026 5:50 UTC
2 points
0
in reply to: aphyer’s comment on: D&D.Sci Release Day: Topple the Tower!
So… still haven’t figured the mechanics out, but actively beating around the bush with Claude finding things of little importance. Second extension?

simon 16 Mar 2026 6:53 UTC
2 points
0
on: D&D.Sci Release Day: Topple the Tower!
Requesting an extension for reason: wanting extra time to see if I can crack the mechanics with Claude. I think Claude suggested some possible answers waaay back but I didn’t pay much attention, telling to focus on the mechanics first. (This might be an abuse of extension requests—it’s not that I haven’t spent time on this.)

simon 9 Dec 2025 6:05 UTC
2 points
0
on: D&D Sci Thanksgiving: the Festival Feast Evaluation & Ruleset
Thanks aphyer for making this scenario and congrats to James Camacho (and Unnamed) for their better solutions.
The underlying mechanics here are not that complicated, but uncovering details of the mechanics seemed deceptively difficult, I guess due to the non-monotonic effects, and randomness entering the model before the non-monotonicity.
It wasn’t that hard though to come up with answers that would do OK, just to decipher the mechanics. I guess this is good in some ways (one doesn’t just insta-solve it), but I do like to be able to come up with legible understanding while solving these, and it felt pretty hard to do so in this case, so maybe I’d prefer if there were more lower hanging fruit mechanics wise. (So maybe I don’t actually prefer simple mechanics as long as some of them are easier to figure out and can be built off of?)
Regarding my own attempt to figure it out:
Thanks to abstractapplic for the comment that tipped me off to the characteristics being quantitative not just a classification, as well as to there possibly being a total food amount characteristic not just spicy/sweet. Multicore’s comment on Roc possibly being in a “meaty” category also helped me in regard to the latter.
Ironically though, the model change actually lowered my performance (from 16.30) presumably due to the improved model not taking as much into account penalties for variance that were implicitly present in the earlier interactions + spicy/sweet dish numbers model. I’m still pleased that the “improved” model had more structural resemblance to the actual reality, even if the answer was worse. (the second change in my answer also lowered my expected performance (from 16.13), but I think this was basically coincidental).
I actually did consider the possibility that the characteristics might have variability, including explicitly considering a uniform distribution between bounds, but Claude did an initial probe and dismissed it. I guess I should have pushed Claude on this! But also I’d been refining imperfect models for a while and and didn’t want to spend the time/effort required to develop a new model at that time if it didn’t instantly seem promising.
I heavily used AI for this scenario. It helps a lot to quickly do stuff that would take a lot more effort without, and in principle should also help with exploring, but I feel like its tendency to focus on what’s right in front of it can also distract me from switching approaches. Also I wish I had done more exploration of the data before asking AI to come up with a solution. As it was Claude (4.5 Opus) found the spicy/sweet categories without me having previously done so, which robbed me of having the pleasure of finding them on my own, as other commenters who mentioned them likely did.
I’m not sure if anyone appreciated my insanely long AI-generated comments, I could avoid doing that in the future if people were annoyed by it.

simon 8 Dec 2025 16:54 UTC
2 points
0
in reply to: simon’s comment on: D&D.Sci Thanksgiving: the Festival Feast
# Feast Quality Model: Full Parameter Specification (simon note: AI generated)
## Executive Summary
This document presents two piecewise linear models for predicting feast quality based on the balance of spicy, sweet, and bulk characteristics. Both models use learned per-dish weights for each dimension, with penalties applied when totals fall outside optimal zones.
**Key Finding:** The “spicy peak” model shows marginal RMSE improvement and is **marginally significant** (p=0.034) by F-test, but **BIC favors the simpler flat model**. Given the mixed evidence, we recommend targeting the **center of the flat model’s optimal zones** for maximum robustness.
---
## Model Comparison
| Metric | All-Flat Model | Spicy Peak Model |
|--------|----------------|------------------|
| RMSE | 1.9635 | 1.9583 |
| Parameters | 64 | 67 |
| AIC | 2440.9 | 2437.9 |
| BIC | 2789.5 | 2802.8 |
**Statistical Test (F-test):**
- F-statistic: 2.892
- Degrees of freedom: (3, 1647)
- **p-value: 0.034**
**Interpretation:** The peak model’s improvement is marginally significant (p=0.034), but BIC favors the simpler flat model (delta = −13.3). This mixed evidence suggests caution in adopting the more complex model.
---
## Model 1: All-Flat (3-Piece Piecewise Linear)
### Structure
For each dimension (spicy, sweet, bulk):
- **Below optimal zone:** Linear penalty with slope
- **In optimal zone:** No penalty (flat)
- **Above optimal zone:** Linear penalty with slope
```
Quality = Intercept + SpicyEffect + SweetEffect + BulkEffect
SpicyEffect = -slope_low * max(0, t1 - total_spicy) - slope_high * max(0, total_spicy—t2)
SweetEffect = -slope_low * max(0, t1 - total_sweet) - slope_high * max(0, total_sweet—t2)
BulkEffect = -slope_low * max(0, t1 - total_bulk) - slope_high * max(0, total_bulk—t2)
```
### Optimal Zones and Penalties
| Dimension | Lower Threshold (t1) | Upper Threshold (t2) | Slope Below | Slope Above |
|-----------|---------------------|---------------------|-------------|-------------|
| Spicy | 3.213 | 3.558 | 1.573 | 0.899 |
| Sweet | 3.068 | 4.461 | 1.593 | 1.273 |
| Bulk | 7.136 | 8.677 | 0.877 | 1.028 |
**Intercept:** 16.742
### Per-Dish Weights
| Dish | Spicy Weight | Sweet Weight | Bulk Weight |
|------|-------------|--------------|-------------|
| Ambrosial Applesauce | 0.000 | 1.456 | 0.000 |
| BBQ Basilisk Brisket | 0.821 | 0.831 | 1.404 |
| Chili Con Chimera | 1.304 | 0.078 | 1.502 |
| Displacer Dumplings | 0.052 | 0.075 | 0.086 |
| Ettin Eye Eclairs | 0.124 | 2.815 | 0.371 |
| Fiery Formian Fritters | 1.337 | 0.000 | 0.918 |
| Geometric Gelatinous Gateau | 0.176 | 2.078 | 0.583 |
| Honeyed Hydra Hearts | 0.000 | 1.537 | 0.851 |
| Killer Kraken Kebabs | 2.146 | 0.054 | 2.434 |
| Mighty Minotaur Meatballs | 0.069 | 0.181 | 2.233 |
| Opulent Owlbear Omelette | 0.008 | 0.016 | 1.788 |
| Pegasus Pinion Pudding | 0.124 | 0.621 | 0.786 |
| Roc Roasted Rare | 0.004 | 0.290 | 3.493 |
| Scorching Salamander Stew | 2.138 | 0.035 | 1.520 |
| Troll Tenderloin Tartare | 0.062 | 0.000 | 0.960 |
| Vicious Vampire Vindaloo | 3.015 | 0.148 | 1.570 |
| Wyvern Wing Wraps | 0.657 | 0.181 | 0.749 |
---
## Model 2: Spicy Peak (4-Piece for Spicy, 3-Piece for Sweet/Bulk)
### Structure
For spicy: 4-piece continuous function with potential peak
- Piece 1: spicy < t1 → slope s1
- Piece 2: t1 ⇐ spicy < t2 → slope s2 (if positive, quality increases)
- Piece 3: t2 ⇐ spicy < t3 → slope s3 (if negative, quality decreases = peak at t2)
- Piece 4: spicy >= t3 → slope s4
For sweet/bulk: Same 3-piece flat structure as Model 1.
### Spicy Parameters (4-Piece)
| Parameter | Value | Interpretation |
|-----------|-------|----------------|
| t1 (breakpoint 1) | 1.016 | Below this: slope s1 |
| t2 (breakpoint 2) | 2.697 | **Peak location** (if s2>0, s3<0) |
| t3 (breakpoint 3) | 5.872 | Above this: slope s4 |
| s1 (slope 1) | 2.043 | Slope for spicy < t1 |
| s2 (slope 2) | 1.856 | Slope for t1 ⇐ spicy < t2 |
| s3 (slope 3) | −1.253 | Slope for t2 ⇐ spicy < t3 |
| s4 (slope 4) | −0.268 | Slope for spicy >= t3 |
**Peak Location:** ~2.70 (where slope changes from positive to negative)
### Sweet/Bulk Parameters (3-Piece)
| Dimension | Lower Threshold | Upper Threshold | Slope Below | Slope Above |
|-----------|-----------------|-----------------|-------------|-------------|
| Sweet | 3.028 | 4.591 | 1.600 | 1.299 |
| Bulk | 6.692 | 8.486 | 0.980 | 1.126 |
**Intercept:** 13.752
### Per-Dish Weights (Peak Model)
| Dish | Spicy Weight | Sweet Weight | Bulk Weight |
|------|-------------|--------------|-------------|
| Ambrosial Applesauce | 0.000 | 1.452 | 0.000 |
| BBQ Basilisk Brisket | 0.642 | 0.871 | 1.395 |
| Chili Con Chimera | 1.020 | 0.071 | 1.427 |
| Displacer Dumplings | 0.036 | 0.062 | 0.089 |
| Ettin Eye Eclairs | 0.070 | 2.904 | 0.410 |
| Fiery Formian Fritters | 1.079 | 0.000 | 0.973 |
| Geometric Gelatinous Gateau | 0.139 | 2.104 | 0.522 |
| Honeyed Hydra Hearts | 0.000 | 1.515 | 0.745 |
| Killer Kraken Kebabs | 1.693 | 0.042 | 2.322 |
| Mighty Minotaur Meatballs | 0.000 | 0.182 | 2.228 |
| Opulent Owlbear Omelette | 0.052 | 0.010 | 1.673 |
| Pegasus Pinion Pudding | 0.125 | 0.592 | 0.653 |
| Roc Roasted Rare | 0.000 | 0.279 | 3.229 |
| Scorching Salamander Stew | 1.691 | 0.041 | 1.614 |
| Troll Tenderloin Tartare | 0.071 | 0.000 | 0.914 |
| Vicious Vampire Vindaloo | 2.361 | 0.163 | 1.545 |
| Wyvern Wing Wraps | 0.518 | 0.157 | 0.711 |
---
## Optimal Feast Recommendations
### Degeneracy Analysis
Both models have **degenerate optimal solutions** - multiple feasts achieve the same (or nearly the same) predicted quality:
| Model | Best Score | # Degenerate Solutions |
|-------|------------|------------------------|
| Flat Model | 16.74 | 135 |
| Peak Model | 16.87 | 6 |
| In All Flat Zones | 16.74 | 119 |
To break this degeneracy, we select the feast **closest to the center** of each optimal zone, providing maximum robustness against model uncertainty.
### Target Centers (Flat Model)
| Dimension | Optimal Zone | Center | Width |
|-----------|--------------|--------|-------|
| Spicy | [3.21, 3.56] | 3.386 | 0.344 |
| Sweet | [3.07, 4.46] | 3.764 | 1.393 |
| Bulk | [7.14, 8.68] | 7.907 | 1.540 |
---
## RECOMMENDED FEAST (Center-Targeted)
This feast is in all optimal zones AND closest to the center of each zone.
| Dimension | Value | Target Center | Deviation |
|-----------|-------|---------------|-----------|
| Spicy | 3.388 | 3.386 | 0.003 |
| Sweet | 3.675 | 3.764 | 0.089 |
| Bulk | 8.041 | 7.907 | 0.135 |
**Predicted Quality:** 16.74 (flat model), 16.80 (peak model)
**Normalized Distance from Center:** 0.160
**Dishes (7):**
- BBQ Basilisk Brisket
- Displacer Dumplings
- Geometric Gelatinous Gateau
- Killer Kraken Kebabs
- Opulent Owlbear Omelette
- Pegasus Pinion Pudding
- Troll Tenderloin Tartare
---
## Alternative Recommendations
### Best by Flat Model
*(135 feasts tied at this score)*
**Example:** Predicted Quality: 16.74
- Spicy: 3.388, Sweet: 3.675, Bulk: 8.041
- Dishes (7): BBQ Basilisk Brisket, Displacer Dumplings, Geometric Gelatinous Gateau, Killer Kraken Kebabs, Opulent Owlbear Omelette, Pegasus Pinion Pudding, Troll Tenderloin Tartare
### Best by Peak Model
*(6 feasts tied at this score)*
**Example:** Predicted Quality: 16.87
- Spicy: 2.695, Sweet: 3.138, Bulk: 6.864
- Dishes (5): Geometric Gelatinous Gateau, Pegasus Pinion Pudding, Roc Roasted Rare, Troll Tenderloin Tartare, Vicious Vampire Vindaloo
---
## Caveats and Limitations
1. **Peak model has mixed evidence:** The F-test p-value of 0.034 is below 0.05, but BIC favors the flat model. The evidence is not conclusive.
2. **High degeneracy:** 119 different feasts fall within all optimal zones of the flat model. The center-targeting approach provides a principled way to select among them.
3. **Weights are continuous:** The per-dish weights are fitted continuous values. Integer or simple-fraction approximations may exist but are not explored here.
4. **Variance heterogeneity:** Residual variance increases when above optimal thresholds (especially for bulk). This is not captured in the point predictions above.
5. **Model uncertainty:** All predictions have associated uncertainty (~2.0 quality points RMSE). The differences between top feasts are often smaller than this uncertainty.
---
## Summary
**Use the center-targeted recommendation** for maximum robustness:
**BBQ Basilisk Brisket, Displacer Dumplings, Geometric Gelatinous Gateau, Killer Kraken Kebabs, Opulent Owlbear Omelette, Pegasus Pinion Pudding, Troll Tenderloin Tartare**
This feast:
- Falls within all optimal zones of the flat model
- Is closest to the center of each zone (normalized distance: 0.160)
- Scores well on both flat (16.74) and peak (16.80) models

simon 8 Dec 2025 16:26 UTC
2 points
0
in reply to: simon’s comment on: D&D.Sci Thanksgiving: the Festival Feast
further followup:
So I have (had AI generate) a new better model, in which the sweet/spicy/bulk dimensions have continuous values, each independently having a piecewise linear effect on feast quality. In one version, the piecewise linear functions have a flat middle, with drop offs at higher and lower values. This results in a range in which all feasts are equivalent. In another version, there’s a peak at low spiciness and we try to pick near that spiciness peak. A combined hedge version has us pick as low a spiciness value as possible within the flat range. The resulting feast pick is:

~~**[‘Geometric Gelatinous Gateau’, ‘Pegasus Pinion Pudding’, ‘Roc Roasted Rare’, ‘Troll Tenderloin Tartare’, ‘Vicious Vampire Vindaloo’]**~~

~~Which I guess is now my pick.~~
I’m still not that satisfied with it: it’s pretty ad hoc, with non-integer values, and I haven’t properly understood the error sources (there’s some theory that there might be a base error distribution with possibly stochastic penalties for going too high in each dimension and possibly deterministic penalties for going too low). Anyway, full model specification of these questionable models in a reply post again for ease of separately hiding it.

edit: changed my mind. I’m not going to hedge towards the alternate model with the peak, I’ll instead hedge toward the center of the (claimed in the simpler model) flat zone, which seems safer. It also seems dumb because probably I don’t need to go all the way to the center to be reasonably safe, so probably I could hedge to lower spiciness without penalty. Still, I’m at this point considering myself “done” (in the sense of not feeling it worth it to keep going, not in the sense that I couldn’t keep going or am satisfied). I will see what I failed to account for soon enough! The centered feast, which is now my pick, is:

**BBQ Basilisk Brisket, Displacer Dumplings, Geometric Gelatinous Gateau, Killer Kraken Kebabs, Opulent Owlbear Omelette, Pegasus Pinion Pudding, Troll Tenderloin Tartare**
The reply post will contain the full description of the models discussed here.

simon 8 Dec 2025 11:43 UTC
2 points
0
in reply to: simon’s comment on: D&D.Sci Thanksgiving: the Festival Feast
followup after reading other answers:
abstractapplic says that it’s total quantity of sweetness/spiciness that matters, and I feel like that’s almost certainly true. Probably the hearty/ethereal thing is quantitative too if it exists (may or may not relate to total food quality, I note Claude thought Roc had x2 food contribution, after I suggested it as a possible explanation for only Roc feasts going down to 2 ingredients and some tests seemed to corroborate.). I had some AI generated quantitative stuff (including sweet, spicy and hearty/ethereal axes), but not sure it was even used in the actual model before the AI pushed me to an interactions model that couldn’t possibly be the simplest model but at least beat an initial sweet count and spicy count model. Anyway, also Multicore pointed out that total ingredients matter in a non-linear way, which the interactions model probably doesn’t take into account in an optimal way (also even the spicy/sweet counts, maybe). So I’m even more sure that I’m missing the optimal model which likely does include optimal levels of spiciness and sweetness and total food quantity that all aren’t based on a simple count.
Edit: I’m following up on this...