Every time we call something GPT we invariably would reference OpenAI's work. This tells me they don't want us to do that, and would instead like to control what can be called GPT and what cannot.
Great, they've showed their intentions and we should respond accordingly. Don't call it GPT, call it parroting (or literally whatever you want). I like this for its connection with the Stochastic Parrots work and because it is already a common usage of the word outside of ML/AI. AutoParrot, BabyParrot, ParrotAPI, etc. We train by first parroting a large natural language corpus and then fine tune the parrot model with RLHF.
Companies will do what companies do, but communities work better when they are free (as in thought).
This article fails the standard of rigor I would expect from a good scientific blog post. They don't define their terms, unsupported claims, relying on an appeal to authority, informal language, and absurd straw-man versions of others' work.
Presumably Google and/or Fitbit operating within the European Union depends on compliance with European law. So choosing to disregard the EU's decision on the matter could blow apart a large part of an international business.
My guess is Google/Fitbit could only proceed without the EU's approval if Fitbit stopped operating in Europe and deleted all Europeans' data.
Arguably: All European customers who bought a Fitbit are likely screwed anyways, Google has a poor history with supporting acquired product lines. While I am irritated about my data being acquired, personally, I'd also recommend against anyone buying a Fitbit while this acquisition is in flux, as Google might choose to terminate service to devices shortly post-acquisition anyways. And privacy loss is a harm as well.
In short, this acquisition is already harming everyone, and the EU is trying to prevent the most consumer harm.
We have no idea what Google's endgame for Fitbit is. Keep Fitbit OS or shove Android Wear on the brand? Focus on the Fitbit app or push everyone to Google Health.
Without knowing this, there's no safe bet in buying a Fitbit right now.
Anything the size of Google is actually a collection of multiple (probably hundreds) of separate legal entities in each country it has significant operations (e.g. a sales office) - all of these will be ultimately owned by whatever the top level listed company (presumably Alphabet). So Google almost certainly has companies in most of the countries of the EU.
It's pretty common for international companies to need approvals by multiple governments for mergers. For example, when German Bayer bought Monsanto, a US company, the Chinese government had to approve it.
"American Companies" is a pretty arbitrary measure these days - both domestic and abroad corporations are gone after by threatening (and carrying out depending on the judgement) to restrict domestic business and seize domestic assets. The US doesn't really have legal standing to seize assets of Google in their London offices (though the UK may be happy to run an assist) and the EC can't seize property in Mountain View - but Google is international and has assets in most countries that can be seized - for them the bigger threat is probably a restriction of business if google search were to be banned in the eurozone[1] then there would likely be a reverse network effect that led to a loss of traffic globally - it wouldn't be unlikely for a lot of browsers to pull google from being a default search engine.
So, the EC has teeth they can use to enforce here - but more importantly google also wants to minimize how much they piss people off. Hence a PR stance of "Hey, we're sorry you feel that way - is there anything we can do to make this go away faster?"
1. I'm not saying this is likely or makes any sense for this particular probe.
Both Google and Fitbit have major EU subsidiaries in order to serve the European market. They could spin off the EU subsidiaries and merge only the US headquarters, although that would probably be impractical.
Google Ads policy: "“Google Entity” means Google LLC (formerly known as Google Inc.), Google Ireland Limited or any other Affiliate of Google LLC"
Fitbit Privacy Policy: "Fitbit International Limited, an Irish company, is your data controller and provides the Services if you live in the EEA, UK or Switzerland"
Intriguingly, it appears TikTok is doing just this to deal with the US government. Selling their US business but having a presence there, managed by a different company, has apparently been seen as a worthwhile option. But yeah, there's a lot of potential issues, especially if said divestments/spinoffs are really only such on paper. Too much coordination and the government still may be upset about it.
Hi Will, thanks for the reply. One thing I'm curious about: this paper discusses how a neuron that uses a different slope for positive and negative updates will converge to an expectile of the reward distribution. And the behavior is very interpretable, in that a slope of 3 for negative updates and a slope of 2 for positive updates will lead the neuron to converge to the 60th expectile (3 / 3+2) if my understanding is correct.
But it seems that the more common approach in reinforcement learning is to estimate quantiles via regression rather than to get expectiles via asymmetric updates as in mice neurons. Do you have intuition for why this performs better? And is there an analog to the asymmetrically updating neuron in the quantile case?
We can think about asymmetric regression more generally. If you have an error and apply some 'response' function f to that error you change the estimator you learn. In the case of quantile regression f is a sign function, expectile regression it is identity.
In my opinion, and this is entirely speculation, I think with further experiments more completely studying the effect we found in our paper, that we will find the response function (f) in the brain is not linear, but a type of saturating function like if we smoothed the sign function out. We repeated our experiments in the paper using such a function, which has been proposed for dopamine neuron responses before, and the analysis continues to hold because the rewards are all quite small and likely simply in the linear region of a non-linear response function (we know firing rate saturates eventually so this isn't much of a surprise).
Regarding quantiles being more commonly used, it's actually the other way around. The Huber-quantiles we saw perform best in the QR-DQN paper, and which most often get used in the follow-on RL work, are actually more like the type of saturating non-linearity you might expect in the brain (although the Huber loss is not as smooth as you probably would expect the neuron response to be).
Hey Will-- congratulations on the publication. It's very cool to see an algorithm reflected in the brain, and the turnaround (from idea to implementation to experiment) in... what, three years? is astonishingly fast.
On a more technical note, I was curious about whether the distribution is being approximated "properly" (in the sense of probabilities summing to one, no negative probabilities) via expectile regression.
Does that hold here? I'm less than an expert on neurology so I am unsure if that would be necessary in vivo (since it's clearly not needed for good performance in silico).
Thanks! It felt like a very long time, but yes for neuroscience it is extremely fast and was only possible because Naoshige Uchida and his lab had already done the rodent experiments around probabilistic reward delivery.
There are a lot of open questions here, so anything I could say about the brain itself would be more of a guess. That said, for our proposed model no negative probabilities are needed, as the distribution is represented by a population of estimators for different predictors of value (in the general sense).
