Decaye

Every model is on the clock. Take a side on how fast.

A new model lands, takes the attention, and the one before it slides down the board. Decaye is a desk where you take a side on how quickly that happens to a given model — long the decay if you think it dies sooner than the room expects, short it if you think it hangs around. Every price is a probability, and every time one moves, the desk shows you who moved it and why.

Open the desk How it works

12models on the board

8quarters priced per market

the desk can never be down more than400credits

The argument

Why obsolescence is worth a price

Four steps, and none of them need you to believe anything about a particular lab.

Models are launched faster than they are retired

Something new lands every few weeks. The one it replaces is rarely switched off — it just stops being the thing anybody reaches for first.

Attention is close to zero-sum

There is only so much room at the top of anyone's list. A new release does not make more attention, it takes it from whatever was holding it.

The slide is legible before it is priced

A model drops down the rankings weeks before anyone changes what they are willing to pay for it. The evidence turns up early, in public.

So it can be priced

Anything with a date, and a public record of getting nearer to it, can carry a price. Decaye asks when a model falls out of the top five and puts a number on each answer.

How the pricing works

A price here is just how likely something is

The desk prices with Hanson's logarithmic market scoring rule. That sounds worse than it is — five steps and you have all of it.

One question, cut into quarters

Each market asks when a model falls out of the top 5. The answer is split into 8 quarter-long windows, and taking a side means saying which window you think it lands in.

Each quarter gets a share of 100%

The 8 shares always add up to one whole. A quarter sitting at 22% is the desk saying that is how likely it is — the price and the chance are the same number, not a translation of each other.

Backing a quarter makes it dearer

Buy into a quarter and its share goes up, while the rest come down to pay for it. That is how your view gets into the price, and the whole board still adds to 100% afterwards.

The desk always quotes

You are never waiting for somebody to take the other side, because the desk takes it itself. There is a price at three in the morning with nobody else in the room.

And it can only lose a fixed amount

The most the desk can be down on a market is 400 credits, worked out from the subsidy before the market opened. It is a ceiling, not an estimate, and the screen states it.

Which side you are on

Long the decay buys every quarter earlier than the strike — you think it goes sooner than the room does. Short buys the strike and everything after it. The two prices always add to 1, so one side getting dearer is the other getting cheaper by the same amount.

The board

What the desk is watching

12 models, hand-picked, with approximate public launch dates. 2 of them have not been announced by anybody — they are marked, and they are here because the whole question is what a new release does to an older one.

GPT-4 Turbo

OpenAI

Held the frontier through early 2024, then got squeezed from both sides.

Launched 6 Nov 2023

Claude 3 Opus

Anthropic

The first model to seriously trouble GPT-4 on reasoning.

Launched 4 Mar 2024

GPT-4o

OpenAI

Cheap, fast and everywhere. The long tail's favourite.

Launched 13 May 2024

Claude 3.5 Sonnet

Anthropic

Stuck around far longer than its successors expected.

Launched 20 Jun 2024

Llama 3.1 405B

The underlying is modelled. The market is not.

Real models and approximate public launch dates, curated by hand. The relevance curves are modelled, not measured — they behave the way obsolescence behaves, but no number here is a fact about the real world. Two entries are unannounced successors, present because the market is about what comes next.

So why is it still worth an afternoon? Because the part that is usually hand-waved is the part that is real here. The prices come out of the scoring rule itself, the same arithmetic every time, and the worst the desk can be down on a market is 400 credits — fixed before it opened, not a hope. Every move on the board can be traced back to the trade that caused it.

Point it at a real ranking feed and nothing downstream has to change. The list of models is one file.

Curated model list, modelled curves

Said plainly, so nobody has to guess which parts are real.

The money is play money — Balances are credits handed out when the desk opens. Nothing goes in, nothing comes out, and there is nothing here to buy.

The curves are modelled, not measured — How each model rises and fades is a modelling choice, made to behave the way obsolescence behaves. It is not a live feed and it is not counting anything.

The launch dates are approximate — Real models, dates taken from public announcements and rounded by hand. Close enough to reason about, not close enough to cite.

2 of the entries do not exist yet — Successor A and Successor B stand in for the next frontier releases, from whoever ships them. They are marked as unannounced everywhere they appear.

FAQ

What people ask before they take a side

Is any of this real money?

No. Everything settles in credits the desk hands out, there is no way to put money in or take it out, and nothing on the page is for sale. The market is the product, not the stakes.

Where does the data come from?

A curated list of real models with approximate public launch dates. The relevance curves are modelled rather than measured — they behave the way obsolescence behaves, but no number here is a fact about the world. Two entries are unannounced successors, present because the market is about what comes next.

What is LMSR, and why not an order book?

An order book needs somebody on the other side, and an empty book has no price at all. Hanson's scoring rule lets the desk quote both sides itself, always. It also has the property this product is built on: a price IS a probability. When a quarter reads 22%, that is how likely the desk thinks it is — the same number, not a conversion.

Can the desk lose money?

Yes, and that is the deal it makes to always have a price. What it cannot do is lose an unknown amount: the most it can be down on one market is 400 credits, worked out from the subsidy before the market opened. That figure sits on the screen next to what it has actually taken in.

Do I need an API key?

No. The desk runs, prices and settles without one. Add an OPENAI_API_KEY and it also writes the plain-language note on each move, with the model, the tokens and the cost shown beside it.

Is it the same every run?

Yes — same seed, same desk. The same models launch on the same days, the same 8 quarters open on each market asking when a model falls out of the top 5, and the same traders make the same first moves.

Every model is on the clock. Take a side on how fast.

The underlying is modelled. The market is not.

Point it at a real ranking feed and nothing downstream has to change. The list of models is one file.