$\require{mathtools} % %%% GENERIC MATH %%% % % Environments \newcommand{\al}[1]{\begin{align}#1\end{align}} % need this for \tag{} to work \renewcommand{\r}{\mathrm} % % Greek \newcommand{\eps}{\epsilon} \newcommand{\veps}{\varepsilon} \newcommand{\Om}{\Omega} \newcommand{\om}{\omega} \newcommand{\Th}{\Theta} \let\fi\phi % because it looks like an f \let\phi\varphi % because it looks like a p % % Miscellaneous shortcuts % .. over and under \newcommand{\ss}[1]{_{\substack{#1}}} \newcommand{\ob}{\overbrace} \newcommand{\ub}{\underbrace} \newcommand{\ol}{\overline} \newcommand{\tld}{\widetilde} \newcommand{\HAT}{\widehat} \newcommand{\f}{\frac} \newcommand{\s}[2]{#1 /\mathopen{}#2} \newcommand{\rt}[1]{ {\sqrt{#1}}} % .. relations \newcommand{\sr}{\stackrel} \newcommand{\sse}{\subseteq} \newcommand{\ce}{\coloneqq} \newcommand{\ec}{\eqqcolon} \newcommand{\ap}{\approx} \newcommand{\ls}{\lesssim} \newcommand{\gs}{\gtrsim} % .. miscer \newcommand{\q}{\quad} \newcommand{\qq}{\qquad} \newcommand{\heart}{\heartsuit} % % Delimiters % (I needed to create my own because the MathJax version of \DeclarePairedDelimiter doesn't have \mathopen{} and that messes up the spacing) % .. one-part \newcommand{\p}[1]{\mathopen{}\left( #1 \right)} \newcommand{\b}[1]{\mathopen{}\left[ #1 \right]} \newcommand{\set}[1]{\mathopen{}\left\{ #1 \right\}} \newcommand{\abs}[1]{\mathopen{}\left\lvert #1 \right\rvert} \newcommand{\floor}[1]{\mathopen{}\left\lfloor #1 \right\rfloor} \newcommand{\ceil}[1]{\mathopen{}\left\lceil #1 \right\rceil} \newcommand{\inner}[1]{\mathopen{}\left\langle #1 \right\rangle} % .... (use phantom to force at least the standard height of double bars) \newcommand{\norm}[1]{\mathopen{}\left\lVert #1 \vphantom{f} \right\rVert} \newcommand{\frob}[1]{\norm{#1}_\mathrm{F}} %% .. two-part \newcommand{\incond}[2]{#1 \mathop{}\middle|\mathop{} #2} \newcommand{\cond}[2]{ {\left.\incond{#1}{#2}\right.}} \newcommand{\pco}[2]{\p{\incond{#1}{#2}}} \newcommand{\bco}[2]{\b{\incond{#1}{#2}}} \newcommand{\setco}[2]{\set{\incond{#1}{#2}}} \newcommand{\at}[2]{ {\left.#1\right|_{#2}}} \newcommand{\pat}[2]{\p{\at{#1}{#2}}} \newcommand{\bat}[2]{\b{\at{#1}{#2}}} % ..... (use phantom to force at least the standard height of double bar) \newcommand{\oldpara}[2]{#1\vphantom{f} \mathop{}\middle\|\mathop{} #2} %\newcommand{\para}[2]{#1\vphantom{f} \mathop{}\middle\|\mathop{} #2} \newcommand{\para}[2]{\mathchoice{\begin{matrix}#1\\\hdashline#2\end{matrix}}{\begin{smallmatrix}#1\\\hdashline#2\end{smallmatrix}}{\begin{smallmatrix}#1\\\hdashline#2\end{smallmatrix}}{\begin{smallmatrix}#1\\\hdashline#2\end{smallmatrix}}} \newcommand{\ppa}[2]{\p{\para{#1}{#2}}} \newcommand{\bpa}[2]{\b{\para{#1}{#2}}} %\newcommand{\bpaco}[4]{\bpa{\incond{#1}{#2}}{\incond{#3}{#4}}} \newcommand{\bpaco}[4]{\bpa{\cond{#1}{#2}}{\cond{#3}{#4}}} % % Levels of closeness \newcommand{\scirc}[1]{\sr{\circ}{#1}} \newcommand{\sdot}[1]{\sr{.}{#1}} \newcommand{\slog}[1]{\sr{\log}{#1}} \newcommand{\createClosenessLevels}[7]{ \newcommand{#2}{\mathrel{(#1)}} \newcommand{#3}{\mathrel{#1}} \newcommand{#4}{\mathrel{#1\!\!#1}} \newcommand{#5}{\mathrel{#1\!\!#1\!\!#1}} \newcommand{#6}{\mathrel{(\sdot{#1})}} \newcommand{#7}{\mathrel{(\slog{#1})}} } \let\lt\undefined \let\gt\undefined % .. vanilla versions (is it within a constant?) \newcommand{\ez}{\scirc=} \newcommand{\eq}{\simeq} \newcommand{\eqq}{\mathrel{\eq\!\!\eq}} \newcommand{\eqqq}{\mathrel{\eq\!\!\eq\!\!\eq}} \newcommand{\lez}{\scirc\le} \newcommand{\lq}{\preceq} \newcommand{\lqq}{\mathrel{\lq\!\!\lq}} \newcommand{\lqqq}{\mathrel{\lq\!\!\lq\!\!\lq}} \newcommand{\gez}{\scirc\ge} \newcommand{\gq}{\succeq} \newcommand{\gqq}{\mathrel{\gq\!\!\gq}} \newcommand{\gqqq}{\mathrel{\gq\!\!\gq\!\!\gq}} \newcommand{\lz}{\scirc<} \newcommand{\lt}{\prec} \newcommand{\ltt}{\mathrel{\lt\!\!\lt}} \newcommand{\lttt}{\mathrel{\lt\!\!\lt\!\!\lt}} \newcommand{\gz}{\scirc>} \newcommand{\gt}{\succ} \newcommand{\gtt}{\mathrel{\gt\!\!\gt}} \newcommand{\gttt}{\mathrel{\gt\!\!\gt\!\!\gt}} % .. dotted versions (is it equal in the limit?) \newcommand{\ed}{\sdot=} \newcommand{\eqd}{\sdot\eq} \newcommand{\eqqd}{\sdot\eqq} \newcommand{\eqqqd}{\sdot\eqqq} \newcommand{\led}{\sdot\le} \newcommand{\lqd}{\sdot\lq} \newcommand{\lqqd}{\sdot\lqq} \newcommand{\lqqqd}{\sdot\lqqq} \newcommand{\ged}{\sdot\ge} \newcommand{\gqd}{\sdot\gq} \newcommand{\gqqd}{\sdot\gqq} \newcommand{\gqqqd}{\sdot\gqqq} \newcommand{\ld}{\sdot<} \newcommand{\ltd}{\sdot\lt} \newcommand{\lttd}{\sdot\ltt} \newcommand{\ltttd}{\sdot\lttt} \newcommand{\gd}{\sdot>} \newcommand{\gtd}{\sdot\gt} \newcommand{\gttd}{\sdot\gtt} \newcommand{\gtttd}{\sdot\gttt} % .. log versions (is it equal up to log?) \newcommand{\elog}{\slog=} \newcommand{\eqlog}{\slog\eq} \newcommand{\eqqlog}{\slog\eqq} \newcommand{\eqqqlog}{\slog\eqqq} \newcommand{\lelog}{\slog\le} \newcommand{\lqlog}{\slog\lq} \newcommand{\lqqlog}{\slog\lqq} \newcommand{\lqqqlog}{\slog\lqqq} \newcommand{\gelog}{\slog\ge} \newcommand{\gqlog}{\slog\gq} \newcommand{\gqqlog}{\slog\gqq} \newcommand{\gqqqlog}{\slog\gqqq} \newcommand{\llog}{\slog<} \newcommand{\ltlog}{\slog\lt} \newcommand{\lttlog}{\slog\ltt} \newcommand{\ltttlog}{\slog\lttt} \newcommand{\glog}{\slog>} \newcommand{\gtlog}{\slog\gt} \newcommand{\gttlog}{\slog\gtt} \newcommand{\gtttlog}{\slog\gttt} % % Miscellaneous \newcommand{\LHS}{\mathrm{LHS}} \newcommand{\RHS}{\mathrm{RHS}} % .. operators \DeclareMathOperator{\poly}{poly} \DeclareMathOperator{\polylog}{polylog} \DeclareMathOperator{\quasipoly}{quasipoly} \DeclareMathOperator{\negl}{negl} \DeclareMathOperator*{\argmin}{arg\,min} \DeclareMathOperator*{\argmax}{arg\,max} % .. functions \DeclareMathOperator{\id}{id} \DeclareMathOperator{\sign}{sign} \DeclareMathOperator{\err}{err} \DeclareMathOperator{\ReLU}{ReLU} % .. analysis \let\d\undefined \newcommand{\d}{\operatorname{d}\mathopen{}} \newcommand{\df}[2]{ {\f{\d #1}{\d #2}}} \newcommand{\ds}[2]{ {\s{\d #1}{\d #2}}} \newcommand{\part}{\partial} \newcommand{\partf}[2]{\f{\part #1}{\part #2}} \newcommand{\parts}[2]{\s{\part #1}{\part #2}} \newcommand{\grad}[1]{\mathop{\nabla\!_{#1}}} % .. sets of numbers \newcommand{\N}{\mathbb{N}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\F}{\mathbb{F}} % %%% SPECIALIZED MATH %%% % % Logic \renewcommand{\and}{\wedge} \newcommand{\AND}{\bigwedge} \newcommand{\or}{\vee} \newcommand{\OR}{\bigvee} \newcommand{\xor}{\oplus} \newcommand{\XOR}{\bigoplus} \newcommand{\union}{\cup} \newcommand{\inter}{\cap} \newcommand{\UNION}{\bigcup} \newcommand{\INTER}{\bigcap} \newcommand{\comp}{\overline} \newcommand{\true}{\r{true}} \newcommand{\false}{\r{false}} \newcommand{\tf}{\set{\true,\false}} \DeclareMathOperator{\One}{\mathbb{1}} \DeclareMathOperator{\1}{\mathbb{1}} % % Linear algebra \renewcommand{\span}{\mathrm{span}} \DeclareMathOperator{\rank}{rank} \DeclareMathOperator{\proj}{proj} \DeclareMathOperator{\dom}{dom} \DeclareMathOperator{\Img}{Im} \newcommand{\transp}{\mathsf{T}} \renewcommand{\t}{^\transp} % ... named tensors \newcommand{\namedtensorstrut}{\vphantom{fg}} % milder than \mathstrut \newcommand{\name}[1]{\mathsf{\namedtensorstrut #1}} \newcommand{\nbin}[2]{\mathbin{\underset{\substack{#1}}{\namedtensorstrut #2}}} \newcommand{\ndot}[1]{\nbin{#1}{\odot}} \newcommand{\ncat}[1]{\nbin{#1}{\oplus}} \newcommand{\nsum}[1]{\sum\limits_{\substack{#1}}} \newcommand{\nfun}[2]{\mathop{\underset{\substack{#1}}{\namedtensorstrut\mathrm{#2}}}} \newcommand{\ndef}[2]{\newcommand{#1}{\name{#2}}} \newcommand{\nt}[1]{^{\transp(#1)}} % % Probability \newcommand{\Normal}{\mathcal{N}} \let\Pr\undefined \DeclareMathOperator*{\Pr}{Pr} \DeclareMathOperator*{\G}{\mathbb{G}} \DeclareMathOperator*{\Odds}{Od} \DeclareMathOperator*{\E}{E} \DeclareMathOperator*{\Var}{Var} \DeclareMathOperator*{\Cov}{Cov} \DeclareMathOperator*{\corr}{corr} \DeclareMathOperator*{\median}{median} \newcommand{\dTV}{d_{\mathrm{TV}}} \newcommand{\dHel}{d_{\mathrm{Hel}}} \newcommand{\dJS}{d_{\mathrm{JS}}} % ... information theory \let\H\undefined \DeclareMathOperator*{\H}{H} \DeclareMathOperator*{\I}{I} \DeclareMathOperator*{\D}{D} % %%% SPECIALIZED COMPUTER SCIENCE %%% % % Complexity classes % .. classical \newcommand{\Poly}{\mathsf{P}} \newcommand{\NP}{\mathsf{NP}} \newcommand{\PH}{\mathsf{PH}} \newcommand{\PSPACE}{\mathsf{PSPACE}} \renewcommand{\L}{\mathsf{L}} % .. probabilistic \newcommand{\formost}{\mathsf{Я}} \newcommand{\RP}{\mathsf{RP}} \newcommand{\BPP}{\mathsf{BPP}} \newcommand{\MA}{\mathsf{MA}} \newcommand{\AM}{\mathsf{AM}} \newcommand{\IP}{\mathsf{IP}} \newcommand{\RL}{\mathsf{RL}} % .. circuits \newcommand{\NC}{\mathsf{NC}} \newcommand{\AC}{\mathsf{AC}} \newcommand{\ACC}{\mathsf{ACC}} \newcommand{\TC}{\mathsf{TC}} \newcommand{\Ppoly}{\mathsf{P}/\poly} \newcommand{\Lpoly}{\mathsf{L}/\poly} % .. resources \newcommand{\TIME}{\mathsf{TIME}} \newcommand{\SPACE}{\mathsf{SPACE}} \newcommand{\TISP}{\mathsf{TISP}} \newcommand{\SIZE}{\mathsf{SIZE}} % .. keywords \newcommand{\co}{\mathsf{co}} \newcommand{\Prom}{\mathsf{Promise}} % % Boolean analysis \newcommand{\zo}{\set{0,1}} \newcommand{\pmo}{\set{\pm 1}} \newcommand{\zpmo}{\set{0,\pm 1}} \newcommand{\harpoon}{\!\upharpoonright\!} \newcommand{\rr}[2]{#1\harpoon_{#2}} \newcommand{\Fou}[1]{\widehat{#1}} \DeclareMathOperator{\Ind}{\mathrm{Ind}} \DeclareMathOperator{\Inf}{\mathrm{Inf}} \newcommand{\Der}[1]{\operatorname{D}_{#1}\mathopen{}} \newcommand{\Exp}[1]{\operatorname{E}_{#1}\mathopen{}} \DeclareMathOperator{\Stab}{\mathrm{Stab}} \DeclareMathOperator{\T}{T} \DeclareMathOperator{\sens}{\mathrm{s}} \DeclareMathOperator{\bsens}{\mathrm{bs}} \DeclareMathOperator{\fbsens}{\mathrm{fbs}} \DeclareMathOperator{\Cert}{\mathrm{C}} \DeclareMathOperator{\DT}{\mathrm{DT}} \DeclareMathOperator{\CDT}{\mathrm{CDT}} % canonical \DeclareMathOperator{\ECDT}{\mathrm{ECDT}} \DeclareMathOperator{\CDTv}{\mathrm{CDT_{vars}}} \DeclareMathOperator{\ECDTv}{\mathrm{ECDT_{vars}}} \DeclareMathOperator{\CDTt}{\mathrm{CDT_{terms}}} \DeclareMathOperator{\ECDTt}{\mathrm{ECDT_{terms}}} \DeclareMathOperator{\CDTw}{\mathrm{CDT_{weighted}}} \DeclareMathOperator{\ECDTw}{\mathrm{ECDT_{weighted}}} \DeclareMathOperator{\AvgDT}{\mathrm{AvgDT}} \DeclareMathOperator{\PDT}{\mathrm{PDT}} % partial decision tree \DeclareMathOperator{\DTsize}{\mathrm{DT_{size}}} \DeclareMathOperator{\W}{\mathbf{W}} % .. functions (small caps sadly doesn't work) \DeclareMathOperator{\Par}{\mathrm{Par}} \DeclareMathOperator{\Maj}{\mathrm{Maj}} \DeclareMathOperator{\HW}{\mathrm{HW}} \DeclareMathOperator{\Thr}{\mathrm{Thr}} \DeclareMathOperator{\Tribes}{\mathrm{Tribes}} \DeclareMathOperator{\RotTribes}{\mathrm{RotTribes}} \DeclareMathOperator{\CycleRun}{\mathrm{CycleRun}} \DeclareMathOperator{\SAT}{\mathrm{SAT}} \DeclareMathOperator{\UniqueSAT}{\mathrm{UniqueSAT}} % % Dynamic optimality \newcommand{\OPT}{\mathsf{OPT}} \newcommand{\Alt}{\mathsf{Alt}} \newcommand{\Funnel}{\mathsf{Funnel}} % % Alignment \DeclareMathOperator{\Amp}{\mathrm{Amp}} % %%% TYPESETTING %%% % % In text \renewcommand{\th}{^{\mathrm{th}}} \newcommand{\degree}{^\circ} % % Fonts % .. bold \newcommand{\BA}{\boldsymbol{A}} \newcommand{\BB}{\boldsymbol{B}} \newcommand{\BC}{\boldsymbol{C}} \newcommand{\BD}{\boldsymbol{D}} \newcommand{\BE}{\boldsymbol{E}} \newcommand{\BF}{\boldsymbol{F}} \newcommand{\BG}{\boldsymbol{G}} \newcommand{\BH}{\boldsymbol{H}} \newcommand{\BI}{\boldsymbol{I}} \newcommand{\BJ}{\boldsymbol{J}} \newcommand{\BK}{\boldsymbol{K}} \newcommand{\BL}{\boldsymbol{L}} \newcommand{\BM}{\boldsymbol{M}} \newcommand{\BN}{\boldsymbol{N}} \newcommand{\BO}{\boldsymbol{O}} \newcommand{\BP}{\boldsymbol{P}} \newcommand{\BQ}{\boldsymbol{Q}} \newcommand{\BR}{\boldsymbol{R}} \newcommand{\BS}{\boldsymbol{S}} \newcommand{\BT}{\boldsymbol{T}} \newcommand{\BU}{\boldsymbol{U}} \newcommand{\BV}{\boldsymbol{V}} \newcommand{\BW}{\boldsymbol{W}} \newcommand{\BX}{\boldsymbol{X}} \newcommand{\BY}{\boldsymbol{Y}} \newcommand{\BZ}{\boldsymbol{Z}} \newcommand{\Ba}{\boldsymbol{a}} \newcommand{\Bb}{\boldsymbol{b}} \newcommand{\Bc}{\boldsymbol{c}} \newcommand{\Bd}{\boldsymbol{d}} \newcommand{\Be}{\boldsymbol{e}} \newcommand{\Bf}{\boldsymbol{f}} \newcommand{\Bg}{\boldsymbol{g}} \newcommand{\Bh}{\boldsymbol{h}} \newcommand{\Bi}{\boldsymbol{i}} \newcommand{\Bj}{\boldsymbol{j}} \newcommand{\Bk}{\boldsymbol{k}} \newcommand{\Bp}{\boldsymbol{p}} \newcommand{\Bq}{\boldsymbol{q}} \newcommand{\Br}{\boldsymbol{r}} \newcommand{\Bs}{\boldsymbol{s}} \newcommand{\Bt}{\boldsymbol{t}} \newcommand{\Bu}{\boldsymbol{u}} \newcommand{\Bv}{\boldsymbol{v}} \newcommand{\Bw}{\boldsymbol{w}} \newcommand{\Bx}{\boldsymbol{x}} \newcommand{\By}{\boldsymbol{y}} \newcommand{\Bz}{\boldsymbol{z}} \newcommand{\Balpha}{\boldsymbol{\alpha}} \newcommand{\Bbeta}{\boldsymbol{\beta}} \newcommand{\Bgamma}{\boldsymbol{\gamma}} \newcommand{\Bdelta}{\boldsymbol{\delta}} \newcommand{\Beps}{\boldsymbol{\eps}} \newcommand{\Bveps}{\boldsymbol{\veps}} \newcommand{\Bzeta}{\boldsymbol{\zeta}} \newcommand{\Beta}{\boldsymbol{\eta}} \newcommand{\Btheta}{\boldsymbol{\theta}} \newcommand{\Biota}{\boldsymbol{\iota}} \newcommand{\Bkappa}{\boldsymbol{\kappa}} \newcommand{\Blambda}{\boldsymbol{\lambda}} \newcommand{\Bmu}{\boldsymbol{\mu}} \newcommand{\Bnu}{\boldsymbol{\nu}} \newcommand{\Bxi}{\boldsymbol{\xi}} \newcommand{\Bomicron}{\boldsymbol{\omicron}} \newcommand{\Bpi}{\boldsymbol{\pi}} \newcommand{\Brho}{\boldsymbol{\rho}} \newcommand{\Bsigma}{\boldsymbol{\sigma}} \newcommand{\Btau}{\boldsymbol{\tau}} \newcommand{\Bupsilon}{\boldsymbol{\upsilon}} \newcommand{\Bphi}{\boldsymbol{\phi}} \newcommand{\Bfi}{\boldsymbol{\fi}} \newcommand{\Bchi}{\boldsymbol{\chi}} \newcommand{\Bpsi}{\boldsymbol{\psi}} \newcommand{\Bomega}{\boldsymbol{\omega}} % .. calligraphic \newcommand{\CA}{\mathcal{A}} \newcommand{\CB}{\mathcal{B}} \newcommand{\CC}{\mathcal{C}} \newcommand{\CD}{\mathcal{D}} \newcommand{\CE}{\mathcal{E}} \newcommand{\CF}{\mathcal{F}} \newcommand{\CG}{\mathcal{G}} \newcommand{\CH}{\mathcal{H}} \newcommand{\CI}{\mathcal{I}} \newcommand{\CJ}{\mathcal{J}} \newcommand{\CK}{\mathcal{K}} \newcommand{\CL}{\mathcal{L}} \newcommand{\CM}{\mathcal{M}} \newcommand{\CN}{\mathcal{N}} \newcommand{\CO}{\mathcal{O}} \newcommand{\CP}{\mathcal{P}} \newcommand{\CQ}{\mathcal{Q}} \newcommand{\CR}{\mathcal{R}} \newcommand{\CS}{\mathcal{S}} \newcommand{\CT}{\mathcal{T}} \newcommand{\CU}{\mathcal{U}} \newcommand{\CV}{\mathcal{V}} \newcommand{\CW}{\mathcal{W}} \newcommand{\CX}{\mathcal{X}} \newcommand{\CY}{\mathcal{Y}} \newcommand{\CZ}{\mathcal{Z}} % .. typewriter \newcommand{\TA}{\mathtt{A}} \newcommand{\TB}{\mathtt{B}} \newcommand{\TC}{\mathtt{C}} \newcommand{\TD}{\mathtt{D}} \newcommand{\TE}{\mathtt{E}} \newcommand{\TF}{\mathtt{F}} \newcommand{\TG}{\mathtt{G}} \newcommand{\TH}{\mathtt{H}} \newcommand{\TI}{\mathtt{I}} \newcommand{\TJ}{\mathtt{J}} \newcommand{\TK}{\mathtt{K}} \newcommand{\TL}{\mathtt{L}} \newcommand{\TM}{\mathtt{M}} \newcommand{\TN}{\mathtt{N}} \newcommand{\TO}{\mathtt{O}} \newcommand{\TP}{\mathtt{P}} \newcommand{\TQ}{\mathtt{Q}} \newcommand{\TR}{\mathtt{R}} \newcommand{\TS}{\mathtt{S}} \newcommand{\TT}{\mathtt{T}} \newcommand{\TU}{\mathtt{U}} \newcommand{\TV}{\mathtt{V}} \newcommand{\TW}{\mathtt{W}} \newcommand{\TX}{\mathtt{X}} \newcommand{\TY}{\mathtt{Y}} \newcommand{\TZ}{\mathtt{Z}}$

Summary of Gradient hacking: definitions and examples.

Gradient hacking is a hypothetical behavior of advanced AI models trained by gradient descent where the model:

• is aware of how it’s being trained;
• strategically chooses actions that will make the gradient point in the direction it wants, as opposed to just trying to minimize loss.

For example, the model could realize that it’s currently close to local optimum which likes better than the global optimum for some reason, and decide to steer towards that local optimum.

This is of concern because it might make inner alignment and interpretability harder.

Will it happen?

By definition, gradient hacking means intentionally choosing to get worse loss, so gradient descent should penalize the part of the model that chose to gradient hack. However, depending on how things shake out, it could be that

• the part of the model that chose to gradient hack is very useful in other situations, so the penalty won’t be significant;
• or by the time the hacking is penalized, it’s already too late, because the descent has already fallen into a local well.

It’s overall very fuzzy to me; I’d need to study some concrete toy models to form an opinion of whether it’s likely to happen.

It might not be a very useful or natural category

A lot of general reasoning could be interpreted as “gradient hacking”, including some positive:

• Humans sometimes tweak their goals in an effort to make them more consistent.
• A model might realize that taking over its datacenter would produce the highest reward (and thus be the “right thing to do” in terms of minimizing loss), but choose not to because one of its goals is to obey humans.