Of particular interest is THIS:
2 Likes
That natural language thing looks like it could be a powerful new tool for a Java programmer’s toolbox (as a programming assist to quickly produce various snippets for use in functions).
The reason why I said it would go into the toolbox of someone who knows Java and is writing a Java program is due to how it is pretty far from just creating a full-blown application or game with a simple (or detailed) description. It would certainly cut down on the amount of times you need to run to the docs. to look up certain functions.
1 Like
Really nice, Really intriguing, Truly revolutionary.
Thanks for sharing, i was not aware of it.
I think it’s a noteworthy approach for this decade if it catches on, as I’m sure it will.
There are lots of demos on Twitter if you search. Really impressive stuff but scary as well, one person would just write Python comments and see code summoned automatically.
Actually, this sort of programming is nothing new. Efforts just like this one have been pursued with more-or-less success for the past forty years. (Koff, koff … wheeze …)
It is entirely true that many programs which need to be written, for various parties who need them, “are merely tedious.” They’re usefully-similar to every other such program, such that we can usefully apply: "Actum Ne Agas: Do Not Do A Thing Already Done.™"
But – inevitably – these tools are specializations, not generalizations. They vastly save time and drudgery, but only in cases where drudgery would otherwise exist.
But the trick here that makes the day-and-night difference is how this AI hits all of the “turring test checkboxes”. I don’t know exactly how they achieved that but based on the theory of evolution you need constantly feeding new data into the system and waiting millions of years to see the results.
I can guess that way they utilized quantum computing to produce the neural networks, and thus perform thousands of years of super-computer work in a few weeks.
If you look at GPT-2 examples you somehow find it interesting or amusing, but still dumb in many ways. Now this new GPT-3 simply it just works so good that it looks unreal.
https://twitter.com/search?q=%23GPT3&src=typeahead_click
Why scary?
it is true that it is nothing new, but there is a difference, that today the times are ripe “in the understanding of language” thanks to OpenAI in this particular case thanks to GPT-3
What I’m curious to know is whether the code this thing can generate is only: “Do this, then do that, then do the other”, or whether that code can have an IF statement?
Anyone know?
http://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html
this thing is way overhyped. it also can’t really program even if it looks a bit like it at a first glance with simple examples.
now that everyone will post texts generated by it the next version will polute itself with non-human gpt-3 sources. 
Yeah for sure is only about text generation so there is nothing clever about it. But the real meaning is how effective is on generating text.
Q: What are two reasons that a dog might be in a bad mood?
A: Two reasons that a dog might be in a bad mood are if it is hungry or if it is hot.
Here you can see that the AI picked two cases out of the blue without any real purpose or reason, probably because is a common re-occurring pattern (dog owners might write about it in their blogs or such). Though the AI in this case has no concept if a dog is in Alaska (it can’t be hot) or in Egypt (it is always hot and got used to it). But rather than that it just gives an answer to get the ball rolling.
These are gaps that will be filled in the coming years. The way is now clear. And if we think about how “smart they’ve gotten” to understand language compared to just a year or two … well it’s just going to get chills on what’s to come.
I believe it is going to take longer to get reasonable understanding, including common sense and the like. It will likely be necessary to link it with knowledge graphs or let the neural network create its own kind of knowledge graph from the data. It will likely be needed to have a mixture of long and short term memory as well as the opportunity to reinspect previously read parts. Many of those work in toy examples, but couldn’t be combined or couldn’t be scaled.
However, I believe we are going to see more specialized solutions a lot sooner, though the initial are likely going to be suboptimal. It would be amazing to let an artist describe what s/he want to achieve in Blender and an AI automatically generates the Python code or smacks a bunch of nodes together. “Show me all materials which are similar to this one.” “Now make sure all those materials are identical, except for this one.” “Reduce the number of vertices for this mesh, but make sure to preserve the details here.”
2 Likes
Besides that, chances are an AI might produce code or a material setup that the artist or programmer cannot even understand (to the point where he can’t make tweaks or take over, you are totally dependent on the AI somehow making sense of what it made).
As I mentioned, if it’s just producing snippets as in the video, it can actually be a useful tool that really cuts back the amount of time you spend in the docs, alongside a cutback of typos and other silly errors. Perhaps in 20 years you will have actual mechanics and assets generated by AI in a clean way, but AI is just now to the point where it can produce usable results with training wheels (information created though statistical/traditional means).
Personally, I don’t care whether it creates something I don’t understand in detail, as long as I have the possibility to tweak it. This could be by easily producing another script/node tree or by letting it produce a script/node tree which let’s me configure it in the way I want to. The workflows would clearly change a lot and a new part would likely be to test those generated scripts thoroughly within the boundaries they are needed.
Many programming languages have this sort of functionality built-in by default. It is called compiler. They are incredibly powerful and automatically check for typos and silly errors without the need for an AI. Yet, some languages are too elegant to need such a thing. Yes, I don’t like Python 
… <\rant>
Sounds incredible. 
I’m so glad someone’s thinking along those lines (at least on this board)! Does anyone know if the Blender Foundation is as well?
1 Like
That sort of solution would be incredible indeed, but it is for sure a few years away before it might become practically viable. To me, it would be something I would actually consider to be non-programmer friendly. Python and node based programming are often called non-programmer friendly too, but I don’t really see that.
The Blender Foundation focuses more on proven solutions which are known to work as far as I can see. In my opinion, that’s actually good for the most part, because it means the money is invested wisely. They are very transparent and if people got the impression that they are wasting money, there would be a high risk they could lose a substantial amount of their funding. For high risk development, they would likely need different funding sources.
hei … this example sounded in my ears, who did I have to think a little bit … and then it occurred to me … 
and a modern thought version arrives here …
maybe with a (Elon Musk’s) neuralink companion of the AI…
(they are bladerunner’s films scenes, I don’t post the videos directly so as not to make too much noise)
Wasn’t aware of those similarities 
When it comes to AI in general, I usually dislike movies due to their dark nature. The same is true for the predictions of Elon Musk, which are not based on our current understanding of AI as far as I can see.
What I described appears like science fiction, I am quite aware of that. However, there has been a lot of progress in language understanding, even allowing it to be used in narrow applications with a limited scope.
I would ask artists to describe what certain node combinations are doing or what a short Python program does. With those, a neural network could be trained to analyze a text and figure out which nodes or functions could be used and how they need to be combined to achieve that goal. If the initial functionality is limited and gets extended over time, I can imagine a solution for artists which could become quite valuable.
This made my wife giggle nervously, an English teacher who teaches non-native speakers…
1 Like