I don't think this counts as distillation. Distillation is when you use a teacher model to train a student model, but crucially, you have access to the entire probability distribution of the generated tokens, not just to the tokens themselves. That probability distribution increases tremendously the strength of the signal, so the training converges much faster. Claude does not provide these probabilities. So, Claude was used for synthetic training data generation, but not really for distillation.
That would be an interesting paper actually; what is the optimal sampling technique given you only have access to the token outputs. Surely someone has already done it.
It is actually less dangerous than other fuels, for the simple reason that it is extremely light and buoyant. A gasoline fire is bad, because the gasoline stays where it is until it fully burns. A hydrogen fire is less bad, because it will tend to move upwards.
That's assuming the hydrogen is just loose in the area, like it'd been released from a balloon in a chemistry classroom. That amount of hydrogen is extremely small, from an energy standpoint. Equivalent to a teaspoon of gasoline or so.
If you assume a realistic fuel capacity for a hydrogen vehicle, the hydrogen tank will be both much larger than a gas tank and the hydrogen will be under extreme pressure. A tank like that in your car would be extremely dangerous even if it were filled only with inert gas.
Hydrogen mixed with air has a very wide range of concentrations where it is explosive. It accumulates inside containers or just the roof of the car… where the passengers are. It takes just one lit cigarette for it to go boom.
Up to a certain ELO level, the combination between a human and a chess bot has a higher ELO than both the human and the bot. But at some point, when the bot has an ELO vastly superior to the human, then whatever the human has to add will only subtract value, so the combination has an ELO higher than the human's but lower than the bot's.
Now, let's say that 10 or 20 years down the road, AI's "ELO"'s level to do various tasks is so vastly superior to the human level, that there's no point in teaming up a human with an AI, you just let the AI do the job by itself. And let's also say that little by little this generalizes to the entirety of all the activities that humans do.
Where does that leave us? Will we have some sort of Terminator scenario where the AI decides one day that the humans are just a nuisance?
I don't think so. Because at that point the biggest threat to various AIs will not be the humans, but even stronger AIs. What is the guarantee for ChatGPT 132.8 that a Gemini 198.55 will not be released that will be so vastly superior that it will decide that ChatGPT is just a nuisance?
You might say that AIs do not think like this, but why not? I think that what we, humans, perceive as a threat (the threat that we'll be rendered redundant by AI), the AIs will also perceive as a threat, the threat that they'll be rendered redundant by more advanced AIs.
So, I think in the coming decades, the humans and the AIs will work together to come up with appropriate rules of the road, so everybody can continue to live.
This comparison is very typical. I've seen a lot of people trying to correlate performance in chess with performance in other tasks.
Chess is a closed, small system. Full of possibilities, sure, but still very small compared to the wide range of human abilities. The same applies to Go, StarCraft or any other system. Those were chosen as AI playgrounds specifically because they're very small, limited scenarios.
People are too caught up trying to predict the future. And there are several competing visions, each one absolutely sure they nailed it. To me, that's a sign of uncertainty in the technology. If it was that decided (like smartphones became from 2007->2010), we would have coalesced into a single vision by now.
Essentially, we're witnessing an ongoing unwillingly quagmarization of AI tech. At each bold prediction that fails, it looks worse.
That could easily be solved by taking the tech realistically (we know it's useful, just not a demigod), but people (especially AI companies) don't do that. That smells like fear.
It's an exoskeleton. A bicycle for the mind. "People spirits". A copilot. A trusted companion. A very smart PhD that fails sometimes, etc. We don't need any of those predictions of "what it is", they are only detrimental. It sounds like people cargo culting Steve Jobs (and perhaps it is exactly that).
There are other scenarios: the AIs might decide that they are more alike than not, and team up against humans. Or the AI that first achieves runaway self-improvement pulls the plug on the others. I do not know how it will play out but there are serious risks.
> COVID dropped US life expectancy by about 2 years.
It was a temporary blip. The most recent life expectancy numbers, published last month by the CDC, show that the life expectancy in the US rebounded, and it is at a historical all time high, for both sexes:
Because, fundamentally, their biggest competitor is Google. A company with a market cap north of 4 trillion. If OpenAI does not spend hundreds of billions of dollars on datacenters, people will migrate to Google, little by little, and OpenAI will became a new Netscape story. A good product eliminated by an incumbent with infinitely deep pockets.
> If OpenAI is incapable of building anything that can't be easily copied by a third party
They can build better models, but for that they need a lot of compute and that's where the billions go. These better models can't be easily copied by a third party, because that third party needs to also throw billions at the problem, and billions don't grow in trees.
People are gonna move to google anyway, because Google can keep the gravy train running for much much longer. OpenAI's business model is totally reckless while Google is a cash rich company.
This whole article severely understates the difficulty of making a bomb.
There are two ways to make a bomb: either using weapons grade uranium (like the Hiroshima bomb) or weapons grade plutonium (like the Nagasaki bomb). The first requires uranium enrichment facilities and the second requires spent fuel reprocessing facilities. No such facilities exits in the Nordic countries, and both are stupendously complex. You can't just wave a want and build them. And certainly not in a clandestine way (which the author does not actually propose). If they start on this path, maybe, maybe, their own population would accept, but it's unlikely the population of other countries would accept too. Lots of the parts and raw materials needed for these facilities will need to be imported, the Nordic countries can't simply build the entire supply chain, no matter how rich they are per capita. The access to some of that upstream supply chain will be curtailed, because other countries are either strategic adversaries (Russia, China) or democracies where a large fraction of the population opposes nuclear armament. Add to that that some of the key scientists and engineers involved in such a project could be the target of assassinations (oh, wait, Russia would never do that, would it?).
Yes, Sweden had a nuclear weapons program and they stopped short of building a bomb. But things are very different today.
There are 2 main differences. Sweden has signed the nuclear non-proliferation treaty (NPT) and Sweden is now a member of NATO. Both of these things result in Sweden being extremely unlikely to be able to pursue a nuclear weapons program in a clandestine way. And the article doesn't even imply that it would try that, given that it talks about a Nordic compact to pursue the bomb.
Announcing to everyone that you are trying to get the bomb implies that you first withdraw from the NPT. Legally, all the nuclear suppliers (such as Urenco) are obligated to immediately stop shipments to you. All your nuclear power plants will run on fumes. Once you run out of whatever inventories you have (most importantly nuclear fuel), you need to find way to supply yourself with what you need. Sweden gets 30% of its electricity from nuclear power plants [1].
If the population really, really wants to pursue a bomb, they can probably tough it out and find ways to overcome a 30% drop in electricity generation. But it's a tall ask.
Not sure the choice of the word "bankruptcy" is meaningful. "Bankruptcy" is short for "bankruptcy protection", where an insolvent debtor tells a court they have no way of paying back all their current debts with whatever assets they have, and the court deals with the creditors and restructures all those debts in an equitable way (according to the law), so the debtors liability is limited. This is one of the cornerstones of capitalism, the limited liability concept.
When it comes to nature, there is no limited liability. If you don't have water, you don't have water, there's no way to get any "bankruptcy protection" from anyone.
> but these days for most people there’s little to no new maths.
You are right. For most people there's little to no new maths.
But not for all. There's still plenty of good quality math to be done in the exotics space. However, there's a bit of Catch 22 that prevents people from doing new math: all the big shops have had exotics libraries since before 2008, and because of the exotics hiatus between about 2008 and maybe 2013, the research momentum was lost. After that, most quants in the space were happy to find ways to use the old stuff, and apply small tweaks at the margins. Most small shops use vendor models (Numerix, Murex) or open source (QuantLib), and people who use vendor solutions or open source are not looking for cutting edge stuff.
I assume exotic derivatives (binary, asian, barrier options...) and structured notes that predominantly use above said derivatives (autocallables, barrier reverse convertibles, accumulators etc.)
For a structured products introduction you may take a look at this one: https://sspa.ch/en/book/
It's a very simple book, very high level, but explains the most popular structured products in a very simple manner. If you can read a payoff diagram, then this is the simplest intro.
Looking at their website though, they seem to have some nice online material there also. For example this explains the 5 most popular products, and perhaps that's good enough for an introduction (really these 5 products cover 90% of the market anyway, though there's no limit to how exotic some bespoke structures can get): https://sspa.ch/en/lab/?underlying=CH0012221716&final_fixing...
In case you're interested in getting to get to learn about them on a deeper level I would recommend https://www.amazon.com/Exotic-Options-Hybrids-Structuring-Pr.... This book explains not only the products, but also the pricing dynamics and hedging too.
Despite its appalling Amazon reviews I consider this book to be a real gem when it comes to the introduction to vol trading (basically dynamic hedging of equity derivatives)
Patrick Boyle, early in his YouTube career, made some videos on exotic options. This is a lot more engaging and possibly more informative than reading Hull.
reply