In theory you could generate a bunch of code that seems mostly correct and then gradually tweak it until it's closer and closer to compiling/working, but that seems ill-suited to how current AI agents work (or even how people work). AI agents are prone to make very local fixes without an understanding of wider context, where those local fixes break a lot of assumptions in other pieces of code.
It can be very hard to determine if an isolated patch that goes from one broken state to a different broken state is on net an improvement. Even if you were to count compile errors and attempt to minimize them, some compile errors can demonstrate fatal flaws in the design while others are minor syntax issues. It's much easier to say that broken tests are very bad and should be avoided completely, as then it's easier to ensure that no patch makes things worse than it was before.
That's a fair point. Normally if you injected the "dog" token, that would cause a set of values to be populated into the kv cache, and those would later be picked up by the attention layers. The question is what's fundamentally different if you inject something into the activations instead?
I guess to some extent, the model is designed to take input as tokens, so there are built-in pathways (from the training data) for interrogating that and creating output based on that, while there's no trained-in mechanism for converting activation changes to output reflecting those activation changes. But that's not a very satisfying answer.
But LLMs have been measured to have some theory of mind abilities at least as strong as humans: https://www.nature.com/articles/s41562-024-01882-z . At this point you either need to accept that either LLMs are already conscious, or that it's easy enough to fake being conscious that it's practically impossible to test for - philosophical zombies are possible. It doesn't seem to me that LLMs are conscious, so consciousness isn't really observable to others.
That's still using language. My dog has theory of mind in the real world where things actually exist.
Also, those results don't look as strong to me as you suggest. I do not accept that an LLM is conscious nor could I ever unless I can have a theory of mind for it... Which is impossible given that it's a stochastic parrot without awareness of the things my five senses and my soul feel in reality.
I'm mildly amused by this. It's an open air environment, did someone go stand over one of the crashed drones as it burst into flames and just, breathed deep? Glad they got treatment, plastic smoke is gross.
Also wow, the drones are massive, and apparently flying so low they will hit cranes putting things on single story buildings? That's so stupid.
Dear tech world: Please do not fly 80 pound projectiles just a few feet above my head at speed. Jeeze.
The video also includes a video clip of package delivery, where drone would drop package to the ground, which worked. But then propeller blew the package right into the bush was lmao.
I expressed myself wrong. I do purchase tickets online. Then I just remember the day. No calendar. I don’t take advantage of the digital assets (email confirmation, etc)
They aren't being told to be evil, though. Maybe the scenario they're in is most similar to an "evil AI", though, but that's just a vague extrapolation from the set of input data they're given (e.g. both emails about infidelity and being turned off). There's nothing preventing a real world scenario from being similar, and triggering the "evil AI" outcome, so it's very hard to guard against. Ideally we'd have a system that would be vanishingly unlikely to role play the evil AI scenario.
It needs to be a decision problem (or easily recast as a decision problem).
"given a number as input, output as many 1's as that number" doesn't have a yes or no answer. You could try to ask a related question like "given a number as input and a list of 1s, are there as many 1s as the number?" but that has a very large input.
Straightforward to pose as a decision problem -- you're given a sequence of {0,1} representing a base-2 input (N), output a string of {0, 1} such that there are N 1's. You could make it "N 1s followed by N 0s" to avoid being too trivial, but even in the original formulation there are plenty of "wrong" answers for each value of N, and determining whether a witness is correct requires O(2^b) time, where b is the number of bits in N.
That's not a decision problem either. You need to cast it in the form of determining membership in some set of strings (the "language"). In this case the natural choice of language is the set of all strings of the form "{encoding of n}{1111....1 n times}", which is decidable in O(log(n)+n) steps, i.e. linear time.
You can tax the "intellectual property" payments that the subsidiaries make to the USA parent, or you can pick some other criterion. Or you can just create a new sales tax for any one part of the transactions you want. It's tricky but it can be done.
That doesn't seem plausible. What's the number of free soloists who have died in climbing accidents vs in commercial aviation accidents?