Tuesday

Wednesday

  • post AGI equilibria
  • Why can’t models understand themselves ; what is the fixed point ;
  • ” In many cases, it will be easier for AI to convince humans it has a proof than to come up with a correct mathematical argument, and I believe that we as mathematicians are not sufficiently prepared for this. Melanie marchett wood ;

Tuesday

54.5 ; 62 lbs

Monday

  • Ollie Olsen ;
  • Carpet and cork and underfloor heating ; wood burner ; move in in July ;

-ToDo: a very specific model and asks for labs (1) the three numbers ; (2) the adjustments to model ; - elasticity funding ;

Avoid authoritarian government by

  • Olivia cybersecurity – what capabilities do defenders need
  • Kevin –
  • Praise reminder

Returns to tokens: - can use best of n ; lower bound on efficient ; but might not be lower bound on a bad harness ; - Should look at redwood note ; Toby ord ;

Fra: - can we share the harness - Do the agents have internet access ; skeptical about contamination behind the frontier ( - What compute counts as training run - if you abort training run ; - Salaries – base salary - 3

  • time for some galaxy brain shit ; what are the forces that set us where we’re at ; evolution ;
  • Phil: it could be rotating membership ; could be economists@constellation ; he was pleased that Arjun was explaining rsi logic to his table mate ;

Saturday

  • airbnb for Joe visit

  • observations

  • most discussion alignment not recursive self improvement

  • the 3pra was a bit let down

  • whether models shouldn’t do tobacco

  • they disagreed on abstract but didn’t disagree on specifics ;

  • Constitution vs spec ;

  • Joe – whether

  • Boaz : i want corrigibility ; so I give up on a model petrov ; who would refuse to bomb the world ;

  • Giving examples to be

  • joe : the models will get crazy good and have power ;

  • Amnesia of models is so important ; that you can wipe memory ; Groundhog Day ;

  • rnd ;

Wednesday

  • Memo recursive
  • Beth : Tom Davidson
  • Beth : talk to neev ;
  • Sidebar toc for cursor
  • Economists @ constellation – less excited than I thought ; makes me want to do my own thing ;
  • Overleaf diagrams
  • credit default swaps on coreweave and oracle # Theura
  • crashing waves : (1) abc ; (2) def.
  • Feedback on Joel proposal ;
  • Floor 3 ;
  • 1 gallon per goat per day ; 1 pound of cheese ; # Wednesday
  • macrame / 3-4
  • Divya
  • Naomi
  • Nick Otis
  • Anna makanju
  • Wojciech # Monday
  • think about earning 10 dollars and rogue deployment ; # Sunday
  • observe that some share of academic gossip is not just cattiness , but also sharing stories and quotes of dunking on each other ;
  • Price of compute falling not rising – shifts left ;

Saturday

  • https://rodneybrooks.com/why-todays-humanoids-wont-learn-dexterity/

  • The false prince ;

  • Micro - shrunken to tiny people ; # Friday

  • theee outcomes – all land , all compute , all labor ;

  • Stella shimonoka email

  • Low hanging fruit – agents can price the long tail of markets ; can price in new information very quickly ; bad at weird cases unusual stuff ;

  • best@2 among multiple runs

  • Marginal value

  • Email Erik b

Recursive confidence – https://www.recursive.to/ # Wedneday - Andy hall - Abhishek – scale humans vs parallel ; - Erdos problems – graph them over time ; - X minus one stories - Tuesday adventure ; - Concave returns – in markdown - L having second thoughts about job - “we have deprioritized this threat model… We are now studying these capabilities under the heading of general autonomy.” - https://arxiv.org/abs/2305.15324

  • fill out the wednesday adventures

Monday

  • time with Joel
  • despite scaling returns are still very concave compared to human time
  • Evidence that it’s got less concave are weak
  • AGI could have concave returns

Sunday

  • countervailing : will likely spend way above expected marginal returns because of (1) option value and (2) fixed costs / learning curve
  • Put in Chris painter thread
  • perfect substation
  • Task model should plot costs together and identify the outcome as cost minimization ;

Ok, does this work? Okay, we’re going to make a fairly general set of related claims. The key question is about the value that companies are going to get from AI versus the expenditure on AI. In general, you’d expect the value to be above expenditure in order to be a net benefit, and the question is what their ratio is. If they’re spending $1,000 on AI, are they getting $2,000, $3,000, or $10,000 in value?

Another way we can put that is the elasticity of the returns to AI, because that’s the ratio of the marginal value and the average value. Immediately we can say that if we hold everything else fixed and we’re just spending on AI, then the optimal expenditure will be where the marginal value is equal to the average value. Then we can turn to what happens if you also adjust other variables. We’ll leave that for a little bit later, and we’ll have a section on that below. The critical claim is that it’s pretty useful to know what the elasticity of returns to expenditure on AI is. Secondarily, we made the claim for many applications. In many cases, it looks like it’s relatively inelastic, which means that the returns diminish pretty rapidly. This means the ratio of expenditure to value is going to be pretty low, or the ratio of value to expenditure is high.

Nice! Why is this interesting? Two concrete things are: 1. Looking for RSI signs when the labs have got recursive self-improvement. Should we expect expenditure on AI on agentic agents to explode? 2. Or, more concretely, suppose the lab is spending a billion dollars a year on AR research salaries, which is probably a reasonable ballpark. Should we expect the value of our genetic AI to be equal to the value of our AI? Will our genetic AI be higher, lower?

Let’s leave that aside. Two big questions: 1. Monitoring lab expenditure on genetic AI to see where RSI is happening. 2. Monitoring firm expenditure on genetic AI to figure out the value beginning.

Two baseline cases for this electricity trick is that, in the Cobb Douglas, each factor is a little bit less. OK. What’s the evidence on disaster returns? It’s pretty simple. We want to know; ideally, we want to measure the value produced and map out that curve of value to expenditure between expenditure and value. We know what the ideal thought experiment is, which is to take a bunch of different people, take a bunch of different organizations, give them different budgets for agentic AI, and see how much value they produce and how quickly the value diminishes. We don’t have good data on that. We have almost no data; we don’t have any data to speak directly to that. What we mostly rely on instead are these scaling curves showing the returns to agent expenditure, either on a cardinal scale or on an autonomous agentic AI. That doesn’t tell us elasticity, but if we additionally see human expenditures on that ordinal scale, we can rank the elasticity.

Okay, there’s one thing still confusing me, and I’ll just try speaking out loud, which is this cross-task substitution. Okay, here’s the confusion: if the price, if you want to get a nice, crisp merit advantage, you don’t need difficulty to be perfectly correlated between human nations. Difficulty can be uncorrelated, but the ratio is monotonic. What’s it going to look like in a scatter plot, mean, for the ratio to be, if there’s just a way to, since we can, without loss of generality, always rank any tasks by their ratio of costs okay. Yeah, see a good costs. Costs. Great, great, great, great. Distributional task costs. Inventor human, distributional task costs. Oop! Okay, you can think of the two axes. Suppose we plot about a hundred tasks and we plot human costs versus agent costs on either axis. Then the two marginal distributions, the distribution of costs for each of those, has an elasticity. Sweet! Good!

Then we can think about sweeping a line starting at zero across the whole plot. That gives you the relative cost, and there’s going to be some line which indicates agents on the right-hand side versus humans on the left-hand side, which is determined by the relative cost of agents and the cost of humans. What is that question again? I forget.

To recap: the distribution x-axis, distribution y-axis, sweep across here, and this is something like that. Suppose that the cost of should I do it this way? Okay, suppose that the perfect order. Let’s go.

Distributional task costs are going to be 50 points; they will be perfectly correlated, and it makes the cost of human tasks uniform for now for simplicity. The cost of agents looks like a concave curve, and the human cost is the distribution, which I guess means a very skewed distribution of agent costs. # Saturday - Chris and Charles ; - Anjana data centers – water and electricity ; - Invitation to dinner ;

Friday

  • ideal is drop in the real world and what can it achieve ; humans could achieve much more if they had no downside ; if you’re in a video game can u learn and evaluate strategy from playthoughs?
  • Jason and Phil
  • Ben recht : union theorem on performance on holdout , depending on holdout set ; but actually we hill climb on holdout set ;

Wednesday

  • correlation of capabilities

  • for David: https://x.com/aakashgupta/status/2046860622160371869

  • Josh clymer ;

  • alpha school

  • Bates from seminar to lecture

Tuesday

  • token scaling curves , after cheating detection ; could do parallel lines ;
  • no nervousness for a week
  • Detox week went well – she hardly complained – and nice hanging out without audiobook ;
  • https://docs.google.com/document/d/1Dz-s5nIa1M_TkJAbRScYJ4X74ZKeUzNWaVLduMccsTk/edit?usp=drivesdk
  • nervous about time to swimming pool
  • https://www2.law.ucla.edu/Volokh/cheap.pdf

Sunday

  • too soon about capabilities but correct about effects ; every science fiction story ; skeptics plump and smug ;
  • Blurry line of capabilities and effects ;
  • Idaho high end retirement community
  • lookup table : Claim and corollaries
  • We have a good time
  • Richard Dawkins would say this is beautiful glorious ;
  • Kuia things

And Saturday - implies no negative spillovers across - Shelter in place - I shut down things – uplift ; time horizon ; inputs and outputs - Ben jones suggestions for evals - Two dimensions of rareness (1) obscurity of facts ; (2) distance from the manifold ; - Existential discussion: will we be able to do more? Need opinionated take on rsi ; # friday - Apple-picking model of time horizon

Wednesday

  • David cesarini reading
  • Econtai at uva ;
  • https://digitaleconomy.stanford.edu/publication/what-is-generative-ai-worth/

Monday

  • mailto:peligrietzer@gmail.com
  • doctor Lee
  • red dots are Cherry
  • Forehead dot took biopsy could be precancerous
  • Froze two spots on my nose

  • artisans : time spent preparing food has halved ; (market work isn’t important , time spent is important ) ;

  • Descartes : a clock that is out of order is not an exception to the laws of horology but a different mechanism which obeys its own laws

  • Kant: I used to love truth for its own sake ; Rousseau woke me , to love things for their ability to shed truth ;

  • livijg in anticipation of war ; Living in anticipation of fascism ; living in anticipation of the bomb ;

  • The moon will be setting directly behind you and you’ll see the faintest blush from the sun under the horizon in front of you

  • Simone Weil: Sophocles , fellowship in love not fellowship in hate ; Christianity is the religion of the slaves that slaves can do nothing but adhere to it ; Simone Weil saintly but also enormously selfish – like pascal ; other women Christian martyrs – stories of martyrs eagerly consumed ; well blamed France for Wwi and for provoking Germany into wwii – anger makes you see the person you’re angry with as responsible for everything ; Montaigne : war always arises from triviality and then grows ;

  • put it in a google doc , “instructions for mums crib”

  • to do: weights ; write in notevooo for an hour each day ; bookmark with priorities on it ;

  • Post civilization equilibria :;

  • rummikub ; seal bones ;

  • Luther Johnson to Bruno

  • Freaky Friday ;

Sunday

  • room 226

  • Phil’s substack: labor share will fall

  • Alex imas: the human touch , wages will be high ;

  • gollancz , Paul, long man Paul

  • https://www.longangle.com/research/high-net-worth-spending

  • https://blog.aifutures.org/p/ai-as-profoundly-abnormal-technology?utm_campaign=post&utm_medium=web

  • https://technologik.substack.com/p/musings-on-recursive-self-improvement

  • https://x.com/i/status/2029649040649146487

  • If you hear tires on the drive then it’s probably my mum arriving back u expectedly and you’ll need to hide as quick as you can.

  • 708 harington

  • Feminist publisher with an iron as the logo ;

  • ird number

  • Mr McKinney : Peter Barnett

  • Mum phone code 290575

  • Auckland massage # Friday

  • multiplicative in apples

  • constant elasticity for human effor , diminishing elasticity for machine effort ;

  • papercraft

  • marinescu physical things are very hard ;

  • secret password ;

  • Respirator

  • Evaluation of ai safety forecasts ;

  • July 13 start date

  • Should be inference cost not training efficiency ;

  • I’ll try to keep this doom posting to once/week but I think worth saying my beliefs: I think you’re all fiddling while Rome is about to burn ;

Tuesday

Cumulative :

,

  • devils half acre
  • body corporate
  • Kids talking over each other
  • swans in bond street

Herodotus

  1. I hold it then in every way proved that Cambyses was very mad; else he would never have set himself to deride religion and custom. For if it were proposed to all nations to choose which seemed best of all customs, each, after examination made, would place its own first; so well is each persuaded that its own are by far the best. It is not therefore to be supposed that any, save a madman, would turn such things to ridicule. I will give this one proof among many from which it may be inferred that all men hold this belief about their customs:—When Darius was king, he summoned the Greeks who were with him and asked them what price would persuade them to eat their fathers’ dead bodies. They answered that there was no price for which they would do it. Then he summoned those Indians who are called Callatiae,1 who eat their parents, and asked them (the Greeks being present and understanding by interpretation what was said) what would make them willing to burn their fathers at death. The Indians cried aloud, that he should not speak of so horrid an act. So firmly rooted are these beliefs; and it is, I think, rightly said in Pindar’s poem that use and wont is lord of all.2

Monday

  • letter to citlali
  • What to measure? Not speedip ; not cost of human ; quantity instead to track
  • Thanks to Eli ;
  • Eli : tradeoff bw human and ai researchers ; effects are additively separable so don’t need state variable ;
  • monkey economist – reasonable calibrated models of monkey enclosures in zoos around the world, typical shocks like adding a new fruit tree , then asked to predict what will happen ; curious whether people say bad analogy or bite the bullet and prepared to say this wouldn’t even be the biggest shock ; an Aztec economist predicting Cortez arrival ; a zookeeper predicting what happens with humans in monkey cage ; 1930s physicist anticipating effect of nuclear energy on human society ;
  • candle
  • Electric water pump
  • Scenario a

Sunday

  • offer in the low 400s, do painting yourself

  • Tenant moving out on the 8th ; next door neighbors have the key

  • Dan and Donny

  • Christine – husband was Roger belton – carer coming and going ;

  • 1880 Gladstone terrace ;

  • bosshards – Patricia and kobi;

  • Sarah – 48 – has the wall taken out ; was Patricia’s earlier ;

  • Man and wife we met – writer and teacher ;

  • ?

  • ?

  • Christine

  • Dan and Donny

  • two 5000L tanks ; folly tank not attached but better quality ; Felix ran out of water and had to refill them – Elle said to get the tanks cleaned , cleaned last winter ; yellowness is the walnuts ; can get low in the summer ;

  • Hot water little lpg gas bottles ; has 3 bottles , weigh them ; always have two joined ; a bottle will last mum a month ;

  • Backup is cook on the range ,

  • range: Chimney swept last year , but sweep says he says he couldn’t sweep the range chimney ;

  • Wood – howie broad , previous commissioner of police, sells wood for charity , got 3 meters of wood , should last a winter ;

  • Sewerage in the garden ; after storm and slips ; Elle dug a trench ; got 6 inches high around the house but not inside ; since then it’s been smelly , water has bubbled up near the cherry tree ; Ruben the plumber said contact Gary ; Elle didn’t like Gary ; Gary wanted to flush the drains with water ; Gary has been out of touch ; Ruben going to talk to Gary ; cherry tree is planted over the drain , roots could be corrupting the drains ;

  • Septic tank : sits and solids drop and water drains out ; was cleaned 5 years ago ; different people do the field drains ;

  • Brian did 5000 work to slow down water ;

  • Denby stoneware mug

  • Nico Nz graduate earnings

  • Alz – https://x.com/alz_zyd_/status/2040436073461878982?s=20

  • apple tree ; chess and fizzing ;

  • Preoccupied ;

Friday

  • https://post-agi.org/
  • gauti eggertson
  • https://x.com/GautiEggertsson/status/2039580991929688333?s=20
  • Roses
  • what would you need for 10% economic growth?
  • reminder of discourse falling into gutter
  • Economists think they’ve been vindicated ; radiologists ; gpts paper ;

Wednesday

  • https://x.com/sebkrier/status/2038959774553833845
  • two Bailey tank 5000 liters ; manufactured 2009 ;
  • emphatic ;
  • https://x.com/saffronhuang/status/2039433283797451254?s=20
  • Linda Cullen ;
  • Turkish place by George street normal ;
  • Haiti

Tuesday

  • thoughts been turning gloomy ;
  • setting up mum: tell her it’s for the next pandemic
  • Three things (1) pandemic ; (2) all infrastructure down ; (3) killer robots – won’t be able to protect against them
  • To do (1) water tank ; (2) food ; (3) fuel ; (4) solar panel ; (5) respirator ; (6) build a shed for storing food and fuel
  • Ombra

Monday

  • bill:
  • Vanessa’s talking over isabel
  • kate stevens west ;
  • Walking around thinking what I would’ve been if I hadn’t left 25 years ago ;
  • heebie jeebies (1) at 10 outside portrait gallery said she was jealous of attention that L was giving to Arthur ; (2) in the library looking at warhammer suddenly
  • Lyle Keith :
  • 11 intercontinental ;
  • Rodrigo at godfathers barber ; Brazilian skinny dude excellent haircut ; #4 sides ; $55 ; mestre morals and muestra jinja ;
  • Hester says we are not a prank household
  • I want to tell Hester to be gentle with will
  • Two types of Nz men –
  • Sals pizza kid like an anemone seize up ;
  • Teachers w bright white hair # Sunday
  • intellectually right but sociallt wrong ; will said firing Wellington people lowered tax take and that cycles through ;

Saturday

  • radio shows from point of view of bad guy ;
  • Susan hamish
  • Susan and Hester weirdly ambivalent about the Paul’s book arcade event ; they say it’s weird that someone knows more about your family than you know yourself ;
  • Nz things so natural ; makes me want to bring citlali up here ; Lynn had a rough time last time ;
  • master chef : both made Bruschetta
  • Lockbox code 1975
  • Intellectual right but socially wrong ;

Starter - originality - Lynn - Presentation - will and girls - Taste - will - spice

Main - originality - Lynn meatballs and chips - Presentation - will for rostis - Taste – Lynn and Ava for chips

Dessert - originality - draw - Presentation - Lynn and Ava - Taste - will and girls

Friday

  • Mac cheese frijol cookie time yoghurt potato chips
  • Schmidt $800k ;
  • Nancy pelosi tech policy person ;
  • Think tank live or die by comms ;
  • Institute for progres is a good comparison – could be advaiaors ; Heidi ;
  • basil: Annie Snyder – Schmidt sciences ai
  • Rose matafeo
  • rationalists see me as clumsy irrationalist ;
  • From point

Wedneday

  • Joel - follow-up w dude

  • summarize with two numbers ;

  • what are apples? Research taste ; time horizon ; jagged frontier ; could align with human difficulty but may not – jagged ;

  • Where are we? We can probably get equivalent to 1 month work , but it doesn’t add up, can’t stack in the same way as humans ; small relative to aggregate optimization machine of OOMs ; speed up is more relevant now ; but labs likely have internal machines ;

  • expenditure isn’t the only thing ; want to discount ;

  • Shape of trees

  • Send to Eli ;

  • Suggestive evidence

  • Add equation at top with exponential ;

  • Post about holiday ;

  • We are not sure how far the curve goes ;

  • which apples are high ;

  • Implications: can summarize with human time for most problems but not additive ; only works against human optimized problems , not additive against machine optimized problems ;

Tuesday

  • tease L not me

  • Car brushes against my face

  • Everyone talking about low hanging fruit - karpathy , Tao , Nathan ,

  • keeper of the lost cities

  • Land of stories

  • Moritz hardt – https://podcasts.apple.com/us/podcast/latent-space-the-ai-engineer-podcast/id1674008350?i=1000658401616

  • why are some apples high? Discussion on Twitter ; make it concrete ; wide but not deep, knows a lot but can’t go far from frontier ;

  • Sunday

  • Tao erdos low hanging fruit ; breadth but not depth ; think about landscape and small vs big jumps :

  • Formalizing this helps highlight not o virus implications (1) when a new model comes out you now the lawn ;

  • Emphasize that can interpret this diagram as many years of progress ;

  • Directed search – couldn’t be isoelastic and multiplicative ; need to keep track of state if allocation shifts endogenously (within period); between periods

  • Two sources of uncertainty (1) uplift ; (2) asymptotic contribution ;

  • Focus on 3x compression , not on physical limits

  • desiderata for diagram – show rsi loop ;

  • Unquestionably if we have 10,000 geniuses then ai r&d we’d see a 3x compression ;

  • Mining for coal

  • Best guess (1)

  • Apple picking model qualitatively different from (1) augmentation (2) automation.

  • Tao – problems that haven’t received much attention ;

firday

  • order a box of artline markers
  • Vals ai –
  • Dustin
  • making a mistake vs minimizing the mistake ;

Wedneday

  • I’ll pay half the rent ; 460 ;
  • Where does the time go
  • Draw the scaling curves

Tuesday march 206

  • 3 equations
  • Complication: bottlenecked on data
  • read ajeya report
  • BART on the quarter hour
  • Andy haupt to Anna
  • Molly kinder : senior people , partners seem to be getting more valuable ; how to resolve that contradiction
  • Vision: everyone now has all expertise ;
  • The case for prohibiting vice ;

Monday

  • plan for Saturday

  • Rent 1080 a week

  • Hirschman

  • 449 replacement iPad Air ; # Sunday

  • abc

  • https://outofcontextreasoning.com/

  • Girl Scouts

  • Do the matrix of saturation

  • Scaling debate – horizontal cs vertical ;

  • Incentives to do exploration is low ;

  • Hervir jeebies ;

  • told my mum I would spend a week with her ;

Saturday

  • graph from rebench ;
  • applied to economic research
  • Ide Talamas ; mention difference is cumulative progress unstead of binary success ;
  • What is the quantity we want to measure
  • Gave up $10m to do this ;
  • https://x.com/RyanPGreenblatt/status/2032624966123008066?s=20
  • Pranks
    • walkie talkie pecas
    • pajama day Deb franzen ;
    • sees candy carrots ;
  • I will never eat boba ; why do they keep asking me like wake up?

Friday

  • three models of tasks (1) facts ; (2) compound facts ; (3) logic navigation of a tree but you only have stylized pieces ;
  • Hammerspoon ; window manager ;
  • Vanessa
  • Barbara and shame ; respect and dignity;
  • Sayash kapoor ;
  • email to dinner people

Thursday

  • no saying anything about food
  • someone saying alex imas and john Horton are thoughtful about ai ;
  • Invite Tarun to constellation happy hour ;
  • Tarun doesn’t belive in superhuman stimuli
  • salary 250 to 160 ;
  • Hammerspoon – https://news.ycombinator.com/item?id=47367932
  • https://news.ycombinator.com/item?id=47367932

Wednesay

  • weenie ; fair play at games don’t cheat ; don’t be annoying for the sake of being annoying ; follow the rules ; not precious about silly little things ; don’t complain behind peoples backs ; don’t complain ; follow through on your promises ;
  • elasticity wrt effort not useful parameter
  • Can’t summarize state with a scalar
  • strawberry creek park in Berkeley

Tuesday

  • time horizon not right ;
  • Deep shallow distinction difficult with LLMs because they’re so much imitation learners ; not objective shape of optimization landscape but qualitatively whether they’re new ;
  • produce big clouds of material – long theses on topics – that never emerge
  • Pool – napoleon – Marco Polo – art materials –

Sunday

  • c is bored
  • Brake pads on Mac n cheese ;
  • good holidays : forestville ; otakou ; Arizona w kishan ; London Rome Zurich ; soames island ;

Saturday

  • baseline optimization model of ai r&d: 3x/year ; each component as deep or shallow ;
  • Claire – came to dads funeral ; came to Dunedin to have Kevin ; met her previous sweetheart in Australia ; he died before Clare was born ; lived w her parents and brought up Clare ; Clare has a couple of kids

Friday

  • set a deadline – 1 week
  • simile
  • bill Manhire
  • Predict scaling on optimization tasks
  • Sean Taylor says he’s sick of talking about ai
  • Sean – everyone at OpenAI has their own idiosyncratic setup and no sharing ;
  • Jeeves and Wooster
  • Summer intern unambiguously better than me in every way
  • Comparative advantage not right it’s absolute
  • henrik karlsson – https://www.henrikkarlsson.xyz/p/unfolding
  • Peter Wildeford
  • Ai models escaping

Thursday

  • andrey fradkin talk very bad

  • tkwa note

  • human saturation definition

  • Email bill Manhire ;

  • more justification of the spline

  • Megan doesn’t look at me when she talks just looks at Nate and Beth ;

Wednesday

  • Sean Kai – human data markets ; making long horizon tasks that reflect human work ;

Tuesday

  • economists freaking out
  • function of intellectuals is to reassure people that everything is for the best / that king was chosen by god ; that Stalin ; that invisible hand ; you don’t discover this until well into career as intellectual ;
  • Economists don’t know what their motives are – you could drop them in
  • jasmine

March 2 2026

  • email Tarun
  • Sigmoid ,
  • Clear thinking about floor and ceiling effects , effect of each
  • Problem is if you get a score of exactly 0 or 1 then you get assigned a time horizon of 0 or infinity.

Feb 28 2026

[gone]