I don't think we have the same understanding of "insider". There are dozens to several hundred insiders at any one time. There aren't enough for them to reliably trade against ones another. To day nothing of how this system would fail when a company decided to raise capital. Like, is the CTO going to trade with the CEO because they have different views on the company? When they need to raise capital, are they going to get it from an accountant at E&Y who's auditing their books?
I said similar, not identical. The primary role of a futures market is for consumers and producers of commodities to hedge out their risk so that they don't go bankrupt if there's volatility in the market. Speculators (by which I do not mean gamblers) make money in futures markets by acting like insurance companies and absorbing that risk. Hedgers take "losing" trades that are equivalent to paying an insurance premium.
Prediction markets could be interesting to hedge a bunch of weird little things. Alas, not how things have gone so far really.
Lots of Unix stuff uses hooks. cron, init, bash has multiple different hook-shaped files (eg .bashrc). The Unix philosophy isn't a sacred cow, purity over pragmatism was the Multics philosophy.
NPM has at least had the good sense to use namespaces so that it isn't entirely a free for all and is less of a high stakes game of Mavis Beacon. (You could typosquat a namespace too, of course.) Unlike AUR but also pip, cargo, etc.
"Better" depends on the application but more precisely it would be "real time" and may include things that are measured but rarely if ever published. A prediction market doing poorly might be a negative result, "this is a martingale and our ability to forecast is limited." That would be my suspicion for a sports event but it could also mean that the participants are gamblers and so not making useful predictions, entirely possible.
The other value proposition would be that you could hedge your risk against all sorts of things that would otherwise be very difficult, like someone who sells funnel cakes hedging the risk of poor at the state fair due to rain. I think that's potentially interesting but was always far fetched and is definitely not as common as gambling.
Insider trading by, say, having direct access to the particular wind vane that grounds the market and being able to manipulate it is a problem. Being better at trading because you have better models or better computers isn't per se a problem. Any more than it's a problem if someone is a better poker player.
In a broader context this may contribute to income inequality and regulatory capture, it can have negative externalities. But it isn't on it's face a problem. And the alleged positive externalities (realtime forecasts and a liquid market for hedging) are driven by such skilled traders.
Not all prediction markets are equally easy to manipulate. The bar was supposed to be (according to my hazy recollection) that a market had to be run in collaboration with an outside authority who, along with delivering a verdict that was already part of their normal duties, would help you exclude insiders from the market. That's a defensible position (not morally, tactically).
But "will so and so tweet about such and such" is a total joke. That's indefensible. That I do think that should never have been allowed or should be regulated as gambling. If there's no conceivable natural hedger, it isn't a real financial market and doesn't have a function (or the function is gambling, whichever).
Don't get me wrong, I think this turned into an end run around gambling regulations (that's certainly what happened with PredictIt), but I don't think that was always inevitable or will necessarily always be the case. That's a strong attractor for sure but it was pushed that way by the COVID era meme stock frenzy, the deregulation of online gambling, and ultimately competitive pressure from the sketchier markets. Once those forces were aligned it was inevitable, but it's possible that in 10 years there will have been a crackdown and tightening down of regulation and things will be very different. I'm not that optimistic, but it's a real possibility.
As far as I know it is "P2P", a continuous double sided auction like any proper exchange. If there's something fishy going on with the order book I do not personally know about it. (I don't think there is but I'm also not going to stick my neck out for them.)
Tl;Dr there's an order book with prices set by supply and demand. So you're at the whims of the market rather than the whims of someone who is your broker, exchange, clearinghouse, and taking the other side of your trade. I don't know much about Polymarket but iirc Kalshi is only your broker and exchange, so dealing with them is materially less risky in my mind.
If you're going to gamble, it's probably for the best that your counterparty doesn't also control the platform. I'm not saying that justifies being able to gamble frictionlessly, but it is marginally less exploitative. Eg, back in the day bucket shops (which sold binary options, like prediction markets do) would increase the spread in proportion to their assessment of your skill such that you would lose even if you were more skilled. In a proper market the platform makes the same amount of money whoever wins.
So, not all that different, but marginally less exploitative. I've never looked at Polymarket but Kalshi and PredictIt slid steadily from things of at least plausible real economic value (a market where it was conceivable [if unlikely] someone would be hedging their insurance contract or something) into total nonsense with no non-gambling function like whether someone would tweet a certain word.
I think prediction markets could serve a similar to a futures markets and have a functional purpose in the economy. It could be useful to generate real time estimates of the probability of some events that no one can control and have real economic consequences, like a hurricane. But evidently that's not where the money is.
If you could write that reward function you wouldn't need an LLM, you'd just query the reward function to answer any question. You can create a benchmark and check that automatically, but you can't solve this in the general case. The model can do well on the benchmark but still give overconfident answers in areas the benchmark doesn't cover.
You can definitely tune a model to say "I don't know" more often but it will cost you performance, the model will reject some questions that it could answer meaningfully. In the degenerate case the model could collapse predicting that sequence always or almost always.
What's burned me before is iterating over hash maps. B-tree maps (or hash maps that are guaranteed to iterate in insertion order, or any fixed order) are your friend.
It definitely wasn't causation, the causation flows the other way (criminalizing a proxy for a political belief as an end run around the first amendment).
Yeah, I agree. It works as an argument for why science can't rely on wealthy benefactors, not for understanding what is possible with a given budget at a given point in time. And biology and chemistry are good counterexamples where capabilities are getting smaller and cheaper.
I would point out that that's on the back of a huge amount of research funded by grants and performed in national labs, but it doesn't impact your argument.
That sounds like excellent grounds for suspicious but I don't know what you mean. We were talking about Peter Higgs for example. I don't think Peter Higgs could have self funded CERN. I don't think a thousand Peter Higgs could have. Nation state level resources are the table stakes for fundamental research into particle physics, because everything beneath that barrier has already been explored - I don't think that's really controversial.
It's definitely an exaggeration to say that all science on a shoestring budget has already been accomplished, there are new frontiers out there. But once they start gaining momentum, the low hanging fruit will be consumed in due course. Methodically searching a domain works and works from the most tractable end up until it is at the frontier of what is tractable given our current technology/constraints.
I don't really understand the alternative hypothesis. That there's an infinite amount of low hanging fruit? What's this 100% failure rate?
What you're describing is (pre-)training. Distillation requires richer labels, the probability distribution over tokens (it would be logits rather than probabilities but that's not important). From a chat transcript you can only understand the argmax/most likely token of that distribution (and only if the API allows you to set the temperature to 0). It's not impossible for an API to give you that but they won't if they don't want you distilling their models.
The intuition is that distillation exploits not only the "right" answer but the relationship between answers (what's the second most right answer? the third? etc).