Jeff Amico

@iampaulgrewal @coinbase @brian_armstrong @Emilemc Amazing run

After 6 years I’m leaving @Coinbase. I’ll be transitioning to an advisory role at the end of the month and continue my service on the Board of Coinbase National Trust Company. I will be a Coinbase ally for life and am grateful to @brian_armstrong, @emilemc and the Coinbase board for the opportunity of a lifetime. ⬇️

@JacobRobinsonJD @LawofCodeFM Eugene 🐐

Great report from @FalconXGlobal on Gensyn's AI oracles, and how they settle on-chain markets like @Delphi_fyi.

Enjoyed this conversation with @JacobRobinsonJD on prediction market oracles and how to fix them.

@hosseeb I was surprised Nemotron 3 didn't make more of a dent tbh

@_jamico 100% but that's obvious to everyone already

Once again: share of tokens is a stupid measurement. You want dollar-weighted. Problems with drawing inferences based on raw token usage on OpenRouter: * Chinese labs routinely launch new models with high subsidies or even free usage for the first couple of weeks. This attracts lots of "Groupon" style users jumping from free model to free model, inflating token usage, but spending nothing. * Small models like Qwen 3.5-27B cost like 1/100th of an Opus token. So if Qwen tokens double, but in a way that doesn't actually *displace* Opus market share, that would show up on this chart as open models gaining a lot of market share. Yet it's economically a rounding error. You want to analyze within model size weight classes in order to get useful aggregates. * It's economical to spend the same amount on a complex multi-agent system built on DeepSeek or GLM 5.2 as on a single frontier model like Opus or GPT-5.5 Pro, at equivalent performance. But the multi-agent setup burns far more tokens for the same spend. So if 5% of Opus usage shifts to one of these systems at 4x the tokens, a token-weighted chart shows ~18% share loss for Opus, while actual spend moved only 5%. These charts magnify low-value tokens. Tokens are not created equal! * Lastly, if you're using a single frontier lab, you don't want to use OpenRouter. Two years ago, the best frontier models were shifting around a lot (Google => Grok => Anthropic => OpenAI), so OpenRouter made sense to be able to adjust quickly. Today, Google and Grok have largely fallen out of the race at the frontier. If as a company you have just decided that Anthropic / OpenAI are your horse, then you'd shift away from using OpenRouter which charges a small surcharge on top of going direct. This would show up in this chart as a nominal decrease in US %, but the tokens are just going off-chart. TL;DR: OpenRouter is much more useful to look at market share WITHIN open models. But to measure open models vs closed models, it's the wrong tool.

@jaibhavnani Cool stuff 👍

@_jamico Yes! And getting paid a higher rate of return to do so (so in the FalconX market, they are getting paid 19.81% to cover the first losses)

Pretty neat that folks are paying a 12% risk premium to sit in a senior version of Pareto's AA FalconXUSDC. Tranching is cool.

@jaibhavnani Ah so junior is posting collateral here to cover that payment?

Let's use an example of apyUSD: Let's say it's at $1.20. Then there's a senior/junior tranche, getting the yield split accordingly. Then apyUSD drops to $1.10. The junior pays the senior to cover the losses. apyUSD holders down $0.10, junior holders down >$0.10, but senior holders stay protected.

@jaibhavnani Makes sense if it's implemented at the underlying vault level, but if it's a layer above like this then it's not senior to the rest of the vault investors - it's on par. Unless I'm misunderstanding the structure.

@_jamico If there are losses, the junior tranche loses first. For assets like apyUSD (drawn down significantly), the Royco sr is still safe Next month there will be a new feature where the sr tranche can pay for liquidity (same way they do to jr), to exit sooner for assets w duration

@jaibhavnani Why would someone invest in the senior tranche of this vs. investing directly in the underlying vault? The two would be on par no?

@_jamico Meet Royco Dawn: dawn.royco.org

@MaxResnick That makes sense as FDV is not a real thing.

These multiples are based on market cap not FDV. They also do not take into account tokenholder rights to the revenue or whether or not the revenue is likely to be passed on to the token holders.

@tombRaider_kw It doesn't actually - need AI-based oracle that can made judgments based on unstructured data. Not just a price feed.

@_jamico Chain link fixes this

RWAs need better oracles. Can't just relay the borrower's claims at face value. Need to independently verify collateral / covenants on a regular basis.

@kevinsxu x.com/_jamico/status…

@ThibaultJaigu @OpenRouter @RequestyAI interesting - which US models? and how many tokens?

@_jamico @OpenRouter We see pretty different numbers on @RequestyAI Gateway! Mainly enterprises and high volume startups still choosing US private models

Nemotron hasn't cracked the Chinese OS stranglehold, despite being free for 2 weeks. Would love to see user geo data from @OpenRouter. Possibly an East vs. West thing.

The way to keep open source legal is to make it even more open. Weights, data, recipe should all be open source and provable. Show you didn't train on harmful data or steer the model. Doesn't eliminate risks but strikes the right balance.

@drippy_eth .@Delphi_fyi permissionless market creation, settled by AI judges.

Who is building unique projects in the prediction market space right now? I'm not talking terminals, copy trading, or wallet analysis tools. Projects that are truly special and adding real innovation to the space.

@emollick ...that's exactly where all the money has gone

In large part that’s because it’s hard to invest in a world where the exponential continues and open models never close thr gap. The value js just chips, energy, data centers, and the labs themselves.

There is a ton of money riding on the hope that the exponential curve the Big Three Labs are on will end soon If that happens, small and open models become viable, businesses get time to react, costs drop & the world gets weirder more slowly. But that isn’t happening so far.