From my shallow understanding, it seems that human training is involved heavily in the post-training/fine-tuning stage, after the base model has been solidified already.
In that case, how is the notion of truthiness (what the model accepts as right or wrong) affected during this stage , that is affected by human beings vs. it being sealed into the basic model itself, that is truthiness being deduced by the method / part of its world model.
Hypothetically, AI is going to move us from development to validation. Think about writing more unit tests, integration tests, e2e tests. Spend more time verifying, really carefully reading these pull requests.
Development is moving towards quality assurance. Because that's what matter eventually. You have a product that works reliably and fast, and you can quickly get it to the market. You don't really care how the code is written.
Of course some people will continue to write "better software" than AI, more readable, or more elegant, bringing some diminishing marginal value to the table, the market doesn't really care about.
I don't think AI is there yet, but realistically speaking, it's gonna get there in 5 to 10 years. Some of us will adjust, some not. Reaction is real.
When LLMs write 100% of the code and we humans are only tasked with validating and verifying its function, programming languages won't be needed (prog langs are for people).
I wonder if at some point we will have an LLM that basically understands English and say, Java bytecode or V8 bytecode. So in goes English descriptions and comments and out goes program bytecode implementing the required functionality.
Also for LRMs.. why use English for the reasoning part? Could there be a more succinct representation? Like Prolog?
The next evolution is you don't need applications at all. Applications are for automation speed, nothing else.
Prior to computers, processes were completed by human to human communication. Hard to scale, impossible to automate. So then we had applications, which force fairly strict processes into a funnel.
But they're extremely restrictive and hard to make.
If you already have God AI, you just don't need an application. I don't go to an airlines website and book a flight. No, I ask my assistant to book me a flight, and then I have a flight.
The assistant might talk to the airline, or maybe hundreds of other AI. But it gets it done instantly, and I don't have to interface with a webpage. The AI have a standard language amongst themselves. It might be English, it might not be.
Is creating a giant wobbly world of unreliable AIs all talking to each other in effort to get their own tasks accomplished, is that leaving us much better off than humans doing everything themselves?
Better yet, if you have an application that does exactly what you want, why would you (or an AI representing you) want to do anything other than use that application? Sure you could execute this binary and get what you want, OR you could reach to the AI-net and make some other AI do it from scratch every time. With inherently less reliable results.
Sorry I should of specified, this is assuming a world with perfect AIs.
The world right now is pretty strict just because of how software has to be, which has a lot of upsides and downsides. But there's some wobblyness because of bugs, which break contracts.
But I think then in the future you have AI which doesn't make mistakes and you also have contracts.
Like the airline agent booking your flight (human or AI) has a contract - they can only do certain things. They can't sell you a ticket for one dollar. Before applications we just wrote these contracts as processes, human processes. Human often break processes. Perfect AI won't.
And to us, humans, this might even be completely transparent.
Like in the future I go to a website because I want to see fancy flight plans or whatever and choose something.
Okay, my AI goes to the airline and gets the data, then it arranges it into a UI on the fly. Maybe I can give it rules for how I typically like those UI presented.
So there's no application. It works like an executive assistant at a job. Like if I want market research, I don't use an application for that. I ask my executive assistant. And then, one week later, I have a presentation and report with that research.
That takes a week though, perfect AI can do it instantly.
And for companies, they don't make software or applications anymore. They make business processes, and they might have a formal way for specifying them. Which is similar to programming in a way. But it's much higher level. I identify the business flow and what my people (or AI) are allowed to do, and when, and why.
Oh I have no expectation of this ever coming true, I'm just day dreaming or thinking out loud. Like, how could our systems evolve in such a world and what would that look like? I think it's a fun thought experiment.
disagree. Programming languages are useful at minimizing context for humans as well as AI. Much easier to call preg_replace rather than implement a regex engine.
Me neither. Considered we are talking about collaborative network applications, you are loosing the single-source-of-thruth (the server database) with the local first approach. And it just adds so much more complexity. Also, as your app grows, you probably end up to implement the business logic twice. On the server and locally. I really do not get it.
It’s difficult to ensure that it’s always the same logic when client and server software versions can get out of sync. You can try to force the client to reload whenever the server redeploys, but realistically you’ll probably end up dealing with cases where the client and server logic diverge.
The latency is off the critical path with local first. You sync changes over the network sure, but your local mutations are stored directly and immediately in a local DB.
This kind of announcement without an appropriate demo to verify their claims is pretty common with DeepMind at this point. They barely even discuss their limitations, so as always, this should be taken with a grain of salt.
Most of the big labs never go into their models' limitations. OpenAI does it best, despite their inveterate hype-building. Their releases always have a reasonable limitations section, usually with text/image/video examples of failures.
Google does a good job with that too usually. Which makes their last two announcements (IMO success and Genie 3) being a bit light on details is somewhat surprising.
> World models are also a key stepping stone on the path to AGI, since they make it possible to train AI agents in an unlimited curriculum of rich simulation environments.
I don't think Humans are the target market for this model, at least right now.
Sounds like the use case is creating worlds for AI agents to play in.
In that case, how is the notion of truthiness (what the model accepts as right or wrong) affected during this stage , that is affected by human beings vs. it being sealed into the basic model itself, that is truthiness being deduced by the method / part of its world model.