DeepMind AI is as fast as humans at solving previously unseen tasks
newscientist.com
newscientist.com
"Unveiling the Crucial 5 GDPR Obstacles of ChatGPT That Can’t Be Ignored" - https://news.ycombinator.com/item?id=34709482
I think AI would be so important that Europe couldnt afford to not have AI. Wonder how this would resolve.
One could hope that this would cause a rebellion, but recent history suggests that the populace will go along with anything their lords decree.
Longer reply: I love the EU so much. One of the few institutions taking big tech to task. I can roam cell operators at no extra cost, ensure that companies cant data mine me without my express permission and other wonderful tech oriented regulation.
It is extremely narrow minded to believe that throwing all principle out the window is the only way to «not stifle innovation».
«The lords decree», where does one even begin. The biggest fight against bigtech involving amongst others the cloud act that lets the us govt spy on anyone in complete secrecy is literally spearheaded by a common man [0].
This entire post is either satire, and if so I ate it hook line and sinker… or it is some kind of privacy exploitation stockholm syndrome.
No one is banning books about gay people in my kids school for example.
Either way curtailing big tech with regulation has nothing to do with big brother. They are regularing companies, not individuals.
Check out some of the schools in Poland, or Hungary.
I’m Canadian, by the way. We now can go to prison for wanting to call our children the name we gave them at birth.
As a EU citizen: You really shouldn't.
They do occasionally pass decent legislature, but i fear GreedClarifies is likely correct that it's probably more about them wanting the power big tech is currently centralizing for themselves.
For examples check the recent news from Belgium as they've uncovered recent corruption issues in the European Parliament. There is even an organization funded by the EU which is unapologetically treating political refugees as prisoners (putting them behind bars with literal cameras in their "living" space.
There was a pretty good report on German state media about that topic, should have English subtitles though if you can't understand it https://youtu.be/tJMLNMlJkPw
I am glad they're currently pushing back against big tech though, as FAANG is already speedrunning our society into a total dumbsterfire, but love to that organization is just misplaced. (Or should we call it MAMAA now? Meta, Apple, Microsoft, Alphabet Amazon)
EU will do a lot in the name of its subjects, but in the end it comes down to power and money. Anyone thinking it won’t end up as a totalitarian forced unification of the member states are sorely mistaken, or viewing it within the context of a minuscule time frame.
Is the rebellion funded by Google and Microsoft? Because that is a conspiracy theory with some legs...
Will it be effective? Dunno, but LLMs aren't as easy to duplicate or execute as films are to pirate or watch, and yet copyright law still gets enforced somewhat.
For a sense of scale - the EU has 0 (zero) of the 14 largest companies by market cap. Out of the next 25, only a handful are in the EU.
If your prediction happens, then EU based startups/companies that'll fill the void will booom.
If FAANG announced leaving EU, then somebody's stock would go up hard, at least on the beginning.
Android is a suitable alternative for iOS.
The absence of European tech giants might reflect the consolidated nature of software rather than the competence of the EU.
And there isn't sufficient venture capital to build up new ones either. Maybe they can all adopt Yandex and VK.
Which one?
The moment of transition from “data-centric” (big centralized system/database) to “agent-centric” (locally stored/run systems and identity, sharing arbitrary data/storage) has arrived.
GDPR, LLM “guard rails”, … — only the plebs will be affected by those.
Remember when OpenAI claimed that was too much power to allow open access to?
I agree. Use cases such as LLMs-as-backend-of-my-web-stack will move a lot of compute back to the edge of the network, eventually.
what happens in cases like this is a slightly different dynamic what it seems on the surface. Apple, coming of huge success with the Apple ][ using the anemic but cheap 6502 8 bit to build a low-end (compared to Xerox) mass market (compared to Xerox) product, was well positioned to capitalize/productize the next generation of more capable 16/32 bit chips without changing their business model or distribution channel. It's right place right time, plus "easier to improve from the bottom, than downgrade from the top".
This idea was identified in a famous McKinsey study of the British motorcycle industry's loss of market share to low cost Japanese competition in the 1960's and 70's. The post WWII Japanese market developed to serve people who needed transportation but could not afford the leading British brands, not to mention autos. Once Japan had a successful motorcycle industry it was natural for them to export inexpensive bikes to Southeast Asia and South Asia. The British companies (Triumph, BSA, Norton) did not make much profit on the cheap bikes, they made their profit on the powerful luxury models, so they abandoned competing in the cheap sectors.
But then another force comes into play: if you manufacture a large number of something (there is always more of the cheap things) you get all sorts of manufacturing advantages. If you figure out a way to use achieve sturdy construction with fewer nuts and bolts, you get to save those nickels over many many bikes which makes it worth your effort to be good at that. If 1 out of 100 of your bikes leaks oil, and you sell a million, you get a lot of complaints, and you fix it. (these are called "learning curve advantages", and they tend to be logarithmic, so by being 10x bigger, you get +something better)
But who especially wants to buy sturdy, reliable transport (everybody) and is willing to pay a premium for it (rich people in the form of high margins)? So being the largest (and by definition the best) manufacturer of a product leaves you perfectly positioned to be the best high margin luxury supplier.
Xerox was not asleep at the switch, they were just not a high volume low cost manufacturer with a presence in the consumer market, at the time when these learning curve manufacturing/marketing ideas had just been developed so they weren't used to thinking that way. Nor was it a case of the suits not listening to the engineers; the engineering ethos at Xerox was not "can we squeeze this on the smallest chip possible", it was "omg let's leverage Moore's law onto even bigger high end chips, compile into microcode!"
I know it's popular to hate on MBAs here, but this is an example of what they learn, and why VC's might like to see an MBA on the team, this is the kind of talk they want to see in the business plan rather than an impossible dream.
From what I've been reading/tinkering, this expertise is essential bc these models are very useful only if the UX paradigms accommodate for them. Getting their hands dirty with consumer-facing applications may keep them a step ahead of Google despite the research 'disadvantage'.
Deepmind loves doing research and getting accolades from a fawning press while never allowing commoners to access their dear technology for fear that the rabble could somehow misuse it.
Google, you lost your way.
Here is for AlphaFold. I mean what you are going to do with this? It looks cool but I don't have the background to even contemplate doing anything with this information: https://alphafold.ebi.ac.uk/entry/L8XZM1
What is most interesting to me is the sociology of people completely writing off the monstrosity of resources and brain power that is Google. Speaking as if they are stubbornly committed to some outdated technology as opposed to the organization that published Transformer: A Novel Neural Network Architecture for Language Understanding.
They sit on their golden goose and don’t want anything to happen to it.
When people understand how much they know about each one of us (with AI) they will be regulated.
For example to improve the current a.i/bot in the game ?
https://github.com/deepmind/alphastar
Same with AlphaFold.
Now people here complain that they don't sell APIs to proprietary models like OpenAI. They are a research lab after all.
But what would anyone do with a bot that beats almost all human players in Starcraft? Its was mainly a demo to show how far they can take their reinforcement learning algorithms after AlphaGo.
AlphaFold has much greater utility and its trained model parameters can be downloaded.
Makes us wonder where Google would be if Larry and Sergie didn’t open it up to the public and we just kept hearing through the press that “there is this amazing search engine that works better than every other search engine”.
After the past decade of every damn startup “guru” advising to “just ship it”, Google just couldn’t follow that advice.
Ultimately it doesn’t matter how good chatgpt or Google’s AI is, all that matters is who shipped it first and how they incorporating feedback in to the model.
That can be said about Google as well, and I think that's what parent meant to express.
Especially with software, pretty much every programmer knows that sky is the limit, and that's why promises mean so little.
In a way it's a similar problem to pre-orders.
People need to try it to truly understand it and its value.
Eg.: That new MMORPG looks amazing, but what's the gameplay like, because at the end of the day gameplay > graphics, if there's nothing to do except just the visuals people will get bored.
And people usually forget the boring stuff.
This is the right response for Google's AI news, but not Deepmind.
Building products isn't Deepmind's purpose. They have essentially "demo'd" it by releasing the paper and simulation videos. They don't build any products. This research is absolutely irrelevant for most people right now, but hopefully will be used by other researchers to make progress.
It's like saying "Stanford should build products and do demos instead of just releasing papers".
I wonder what Larry and Sergei think of all this. Has Sergei said anything lately? Larry has been on a remote island in Fiji for years now. It seems with all this exciting geeky AI stuff they should be passionate about the business, but as far as I can tell they aren't very engaged.
Edit: Here's a recent article. Apparently, Sergei is back working on things for the first time since 2019. Larry and Sergei also checked in on the AI strategy in December: https://www.moneycontrol.com/news/technology/google-co-found...
The only thing they can "demo" is the simulation of the model, which they have published as videos on YouTube. Here's the paper if you're curious btw: https://arxiv.org/pdf/2301.07608.pdf
That’s the standard chatgpt set
Without some real examples, this is just PR talk. I could just as well say that "the Stockfish AI" is able to solve previously unseen tasks on a virtual board, moving pieces around, thinking ahead to navigate around and outsmart any human and non-human opponent. Doesn't say much.
The important questions are (1) can the computer do it in the first place and (2) at what financial cost?
Perhaps there’s only a few search algorithms and humans (and now this AI) are optimized to do it as well.
One of my shower thoughts are that we’re effectively recreating the wheel with AI. Yes it can be immensely powerful and outperform humans for specific tasks. However, humans are generalized and evolved to interact with the real world independently. I imagine an AI would effectively match a human in this regard someday, but it’s worse as it can’t generate energy from eating some plants or animals (ie naturally occurring solar generated food). The interconnected nature will let AI have an advantage for many tasks, no doubt, but humans are pretty well optimized for the real world. Imo it’ll be hard to beat.
And while the plants themselves may be 'solar powered', the tractor used to plant them isn't. Neither is the factory where they made the fertilizer etc.
In terms of subject matters, ChatGPT is way more generalized than humans.
It can tell you often convincing stories about how to do those things, but it has no means to enact any of it nor any internal motivation to try.
This is what like half the people in first world countries do all day.
> It can’t wipe a floor...
There is no reason to suspect this is a technical limitation of the models we can build now, rather than a consequence of the fact that no one has bothered trying due to limitations on mechatronics.
> but it has no means to enact any of it nor any internal motivation to try
Any claims to this end are both/either A) purely speculative and B) ontologically nonsensical. It's also not really relevant.
I think the point still stands, if mechatronics turn out to be a more difficult problem than intelligence: we will need new models to overcome these limitations.
Not by choice. That is an artifice of our institutions.
> Any claims to this end are both/either A) purely speculative and B) ontologically nonsensical. It's also not really relevant.
It's speculative, and it's nonsense, and it's not relevant. This claim itself is ontologically nonsensical.
I have seen publications where a LLM like ChatGPT was mated with a robotic arm, and taught to give instructions to the arm to do whatever it wanted to do. I think this may be it: https://sites.research.google/palm-saycan
Plop Excel on a computer and see what it does on its own.
Not seeing what ChatGPT can be tooled to do is simply a lack of imagination.
A machine can be worked 24/7 with no regard for their wellbeing whatsoever though.
AI will never be a free lunch.
> The AI, called Adaptive Agent or AdA, works in a 3D virtual world where it is asked to solve tasks that involve navigating, planning and manipulating objects.
Not in real world.
In 2016, their system could locate objects in a scene[0]. In 2017, it could answer questions about a scene[1]. In 2018, it could infer how a scene looks from any angle using a couple photos[2].
Now, it can perform inference tasks in a 3D scene.
Going to the real world will require more work! But they will be helped by converging existing advances[3].
[0]: https://arxiv.org/pdf/1603.08575.pdf
[1]: https://arxiv.org/pdf/1706.01427.pdf
[2]: https://www.deepmind.com/blog/neural-scene-representation-an...
[3]: https://www.deepmind.com/blog/stacking-our-way-to-more-gener...
EDIT: Found these with Google:
However, there is a big question of how much the results are actually transferable to the real world.
I hope the volumes of worthless content it generates destroy the internet as we know it, so we can finally move on to better things and start building communities again.
ChatGPT even today performs some tasks well e.g., extracting the intent from a natural language query, or generating text that on the surface looks human. It is often inaccurate on most tasks but it can I believe become better on some tasks to have a great economic effect.
Most most likely we won't get AGI though, and that's great because AGI that is not restricted by the human scalp is an existential threat. I'm not talking jobs, It is about surviving as a species.
It's the same as: all juniors can write code. Only senior can write code that's clean, simple to understand.
Mind you, talking to ChatGPT is easier to talking to real human. Because most of the time, it understands your intent very well.
Also it has an unlimited patience as well as unlimited time to dedicate exclusively to you. You can ask the same question from different point of views and it puts the same effort into each answer every time, even taking into consideration what it already told you and build upon that.
How parents start to get annoyed when kids ask "Why?" in a recursive loop to every answer you give them, or a teacher or coworker would have to tell you "I don't have any more time for this", this technology can really answer your question.
Ok, at this time the answers still contain too many mistakes, but once they start being correct more often than the ones from humans, something really powerful will happen to the society.
Imagine kids being able to ask any question and getting the answer immediately, them getting used to it, how fast they will be able to learn. Also be able to learn to ask the right questions.
Volunteering sounds like what you should look into. It happens in the real world with real people, small communities of people doing worthy things. It is a thing.