You’re not wrong, but AI being passably good at bullshitting (while mostly keeping a rhyme/meter) is still mindblowing to me.

Subjectively incredible; objectively flawed.

Edit:

I guess in this description dijkstras is more specific, in that dijkstras is the specific instance of A* with a zero heuristic.

But I think what HAL over here was saying is that you can use dijkstras in a superset of scenarios in which you can use (non-trivial) A*, so in that sense dijkstras is more general than A*, so it's not wrong.

For A*, everything depends on the heuristic, as you said. If h(v) = 0, A* is equivalent to Dijkstra's algorithm, so potentially, it can be as general. But the wrong heuristic (inadmissible, inconsistent) will lead to wrong results, and so calling it a "specialized" tool is correct. The heuristic gives you a specialized version of Dijkstra's algorithm which is faster on specific graphs.

> Dinkstras is a-star with a "min weight on shortest path first" heuristic.

I am not quite sure what you mean here. Dijkstra's algorithm is A* with no heuristic. I am not sure how a "shortest path first" heuristic would look like - do you mean that Dijkstra's algorithm chooses the node to expand next based on its shortest-path cost to the target t? Even if you would construct such a heuristic h(v) = c(v, t) by explicitly calculating c(v, t) each time, this is not what Dijkstra does, as it would basically be a perfect heuristic - you would then only visit nodes on the shortest path.

If you meant that Dijkstra's algorithm is A* with a "neighboring node with shortest path from start node first" heuristic, that's also not strictly true, because expanding the nearest node first is already built-in to A*. A* does not chose the next node v with the smallest h(v), but with the smallest g(v) + h(v), where g(v) is the shortest path cost from the source node to v.

I take issue with this statement. "any case" implies it's more general than A star, which is not true, it's a specialization of A star with (ok you win) h(v) set to nothing. It necessarily cannot solve as many problems as a star.

And Dijkstra's won't work in "any case" at all, esp graphs with negative edge weights.

The explanation struck me as glossing over too much, to the point of being misleading. Perhaps I'm being too pedantic, fine, but a more enlightened comparison would read more like our discussion than what was given.

Could you give an example of a positive-weight shortest path problem with A* can solve, but Dijkstra's algorithm cannot? I do not believe there is such a thing. Or did you mean something else?

> And Dijkstra's won't work in "any case" at all, esp graphs with negative edge weights.

A* won't solve those either.

they both solve the single-pair shortest-path problem in digraphs with nonnegative arc weights

a* which explores nodes in a different, better order thank dijkstra's algorithm, but can only be applied in cases where you can compute an admissible heuristic

that makes it less general than dijkstra's algorithm, which works to find shortest paths in any digraph with nonnegative arc weights, not just the ones where an admissible heuristic can be computed

any problem you can solve with a* can be solved with dijkstra's algorithm (usually more slowly) but the converse is not true

now, in a sense, a* with an inadmissible heuristic such as h(v)≡0 is 'more general' in the sense that it can emulate dijkstra's algorithm and also do other things; you could say that a* is a class of algorithms of which dijkstra's algorithm is one

It's not in the style of a psalm.

So I wonder if there was substantial fine-tuning for chatgpt specifically to reward it to generate poems, in a particular style.

And here, it's "over indexing" on that and still generating poems in that familiar style

That's chatgpt right now. To get it to confidently right will take years and years more

In less than a decade, I believe front line support chat support jobs and even graphic artist jobs will be made obselete

We can finally get the eighty hour long book-accurate Lord of the Rings adaptation we deserve.

But seriously, I can't wait to feed a novel into one of these things and get a comic book version in return.

It's already at the point where you start expecting any knowledge worker to be significantly more productive by leveraging these tools.

It's hard to imagine that it will be more than five years before AI tools are available that can handle almost all tasks in these types of jobs.

For example, on my website aidev.codes I just added preliminary knowledgebase support. It can reference the knowledgebase this to write code. I would say that with the code-davinci-002 model at least it seems about at the level of a junior software engineer already since it's pretty effective with close supervision by a senior peogrammer, except for the fact that it cannot interpret visual information.

Knowledgebases/embedding search can also be used right now with these models for answering support questions. The only thing holding it back from very very wide scale adoption is the problem of making up information. There are already solutions in progress for this. It's unlikely that will take more than a few years to roll out and replace the current generation of models. Google and Microsoft will probably roll out their internet-scale chat search interfaces this year even if they can't fully mitigate the hallucination problem immediately.

I would guess more like 2-3 years for many knowledge-based jobs. If you want employment/contracts you will need to be very good at leveraging AI, or people will just use the AIs instead.

Very human I must say.

https://www.theregister.com/2022/01/31/machine_learning_the_...

Even if you are right, you still acknowledge that ChatGPT is on the level of a non-specialist bullshitting human you can have a meaningful conversation with. This alone would be enough to freak me out. If you would've told me that this is possible 10 years ago I would've called you crazy.

While I wont deny that it has impressive summarization abilities in terms of making excellent Q&A if you're willing to vet the information, I wouldn't exactly say that chatGPT is capable of meaningful conversation. It has great powers of recollection but its capacitive powers to produce new interesting information feels highly formulaic.

  The sun will not harm you by day,
      nor the moon by night.

  My life is consumed by anguish
    and my years by groaning;
  my strength fails because of my affliction,[b]
    and my bones grow weak.

https://paperswithcode.com/paper/most-language-models-can-be...

Gwern cited it too!

It's not a crystal ball, it's a mirror.

There is obviously more to human intelligence than text based conversation but it is pretty humbling that such an aspect of ourselves can be replicated and be so convincing at such an early stage and perform better than some humans even when it makes stuff up: toddlers can't talk, kids are smarter but don't have the technical knowledge, most adults only have a few specific areas of expertise, etc.

but not always; And that's what makes it scary. You know people will use it and trusting it.

as far as i can tell it seems perfectly correct but admittedly my understanding of a* search is limited to messing with amit's visualizations of it

This is just generic poetry with a sprinkling of "ye's."

Notice that it sometimes struggles to rhyme?

Does it really matter when it comes to real life? Kids take medical advice from TikTok influencers. Half the country believes absurd news they see on Facebook. Half the country believes anything their president says.