A less charitable interpretation might be that many of these ideas are actually not workable -- and never will be -- in the real world where you can't just insert special cases whenever you need to. (I've seen this sooooo often in research code.) It's also not just a question of computing resources.
<rant incoming, be warned>
I'll use the Mario example from one of his first widely disseminated talks as an example: In the Mario example you see him run time backwards. That's actually reasonably achievable by just recording all previous states or (more plausibly) all state changes and then "just" replaying backwards. Then he runs time forwards again after having changed a setting (or some such). This is, again achievable within limits, see e.g. "rr".
The problem comes when he shows a projected future path of Mario. Because of a) Turing Completeness (TC) + the Halting problem, and b) the fact that much code actually interacts with other systems which are not able to just rewind (and/or themselves interact with systems that ...).
The only plausible scenario in which this is even remotely possible in practice is if the programs adhere to strict rules about side effects, have built-in "project into the future" (or have no side effects so that we can run them to see what the results are _without affecting anything external_.).
That is, it's not about fancy IDEs -- it's about programming languages and turning away from Turing Completeness and/or non-abstract/non-formalized effects. FRP/(React+Flux) sort of work for this if you adhere strictly to purity, but then again break down when it comes to things like "what is the focused element" if the user clicks around when you're replaying.
I don't want to be a party-pooper, but good luck changing the programming paradigm of everyone to achieve this vision. For it to work everything has to work in this FRPish way, and that's just never going to happen -- for example, networks (for real-world reasons) cannot behave this way, etc. etc. I WISH we could have everything behave this way, but we can't -- billions of e.g. browser installations which have to have backward-compatible DOM implementations and millions (billions?) of installations of C/C++-based operating systems say so.
Now, I'm not against Imagining Things How They Could Be, but if there's mathematical proof that "thing X" is not possible one either needs to explain how you are not doing "thing X" (perhaps not using a TC language, or disallowing non-pure state) or shut up. Brett Victor and many "visionaries" like him are just very bad at stating their fundamental assumptions up front (which is perhaps part of the reason we don't get to see the code, it's probably just simple pure FRP-like code) so that we can evaluate the vision against the assumptions. (Light Table just seemed to assume that the assumptions were plausible and see where that got them. Funded but with very little actual technological advancement to show for it.)
... and also don't get me started on the whole "Kill Math" thing. Ugh.
(Sorry about the rantish nature of this -- I should probably just write this up properly somewhere so I could properly flesh it out. I just get tired of this "What A Genius!" and "Well, it would take too much time away from inspiration to actually make the thing!" credulous nonsense.)
Do you think it might be possible to integrate non-FRP style programs (basically stateful, imperative style) by running them in a VM which can isolate the effects and snapshot/restore the entire system state? In theory at least, it seems something like a web browser could rewind the state of a web app and replay with different input.
Snapshots/diffs are not enough to be able to "project into the future".
In the web browser this is mostly achievable if the whole app is programmed in a very strict React/Redux style where everything the React portion is handled through props. (Focus being an exception that seems just untamed, so far. I may very well have missed developments here.)
I can confirm that dealing with focus in react hasn't really advanced. There's useful libraries for working around this like react-hotkeys but that's about it.
I worked with Chris on Eve, which was the project he started after he hit a wall with Light Table, that wall being the limitations of programming languages you outline here. Eve did have strict no side effects requirements, and did have the ability to log and rewind programs. The bulk of our work was in nailing down a good set of semantics, and also in figuring out a good way to present those semantics to users (syntax, UI, etc.).
After writing countless programs in this style, which is very different from conventional programming, I have to say the Mario example is a nice demo, but in reality we didn't find time-travel useful in scenarios like that. What was infinitely useful though, was being able to trace the provenance of different values in your program for debugging. Since we kept information around that the compiler usually throws away, we could tell you exactly how a value or UI element was calculated, and give you tools for interacting with that history that just aren't possible in other languages. For instance, you could encounter a bug, pack up the state of your session, send that over to another dev, and he/she would see exactly the environment in which your code was executing.
You bring up the good point that not everything is written in this way, and so interacting with external code can be tricky. For example, you might have some code that sends out an email every time the system memory reaches a threshold. Now, you might get an e-mail you didn't expect, but you could go back and see exactly why that email was sent, and deduce the bug through the provenance tree.
As I said earlier, after writing programs in this style for 3 years, I'm utterly convinced we won't be programming in the future the way we program today. There were versions of Eve where we wrote very complex applications, from webapps to robotics, that didn't even feel like programming -- it felt more like forming a shape out of clay, where the clay in this case is a digital "material".
In the end, Eve couldn't get more funding because it wasn't a venture that was really set up to fit into the VC mold (and I'm still not sure how Chris and Rob convinced VCs to fund them in the first place), but I'm absolutely sure the current way of doing things will fade with time. It's going to take a lot more research and a lot more convincing, because there are a lot of people out there that for some reason don't want to see projects like this work, but as someone who went down the rabbit hole and gotten a glimpse of this future, I really just can't fathom how humans could be working with languages held back with so much baggage from the early days of computing in the year 2100 for example.
I'm curious though, why you suggest for a language to work like this you have to abandon Turing Completeness. This wasn't my experience at all, and we implemented a Turing Machine in Eve just to prove we were TC (http://incidentalcomplexity.com/2014/12/01/nov/).
I think we agree that FRP-ish is the way of most of the future, but it must integrate well into the existing ecosystem. Which basically means no "projection into the future" and no "fast-forward". I think side effects are unavoidable for all practical purposes -- the trick is how to integrate them cleanly, if that's even possible. Even being able to trace backward would be a huge achievement for most of the programming world. To be clear, it's about 80% achievable right now by just programming your whole application in the FRP style until you reach the "transactional" or "IO" boundary.... Btw, it's not a coincidence that this sounds similar to CQRS/Event Sourcing and Bounded Contexts from DDD. I think CQRS/ES approach has an advantage in that it "realizes" that there's such a thing as "oops, I wrongly charged customer X for item Y, let's refund that". Plus, it's it's basically what we do in the real world, e.g. in banking.
It's obviously not enough to be able to "project into the future", but it's pretty awesome for looking at what happened. It's also a great way to maintain discipline for your code -- we're already off on a wild tangent here, so I won't elaborate.
> I'm curious though, why you suggest for a language to work like this you have to abandon Turing Completeness
You don't necessarily have to, as long as you a) assume complete purity (as in e.g. Haskell's ST monad), and b) are resigned to having to actually run all the computations when running "forward".
Keep in mind that some of those computations may be EXPTIME... but as long as you can interrupt them, then I suppose that's OK.
I think the conclusion is that anything TC that doesn't do side-effects would basically be OK (as in: not actively wrong), but not necessarily practical in our lifetime. Timeouts might work here, actually.
It isn't clear to me what the problem with projecting into the future is. You just simulate the system or just let it execute normally (but under the control of some kind of "harness"). This will give you the correct result up to the point where some user input (missing in this simulation or execution) would have changed the behaviour.
> You just simulate the system or just let it execute normally (but under the control of some kind of "harness")
"Simulate the system" means "execute normally" for all practical purposes. That's what TC + Halting Problem[1] essentially boils down to, absent other knowledge about the program. Hence my suggestion to let the program itself do "projection" because it could potentially be hugely more efficient, and possible (but perhaps with bad "fidelity" in the case of side-effecting computation).
Anyway... to the larger point and to expound a little:
If everything is pure, then there's basically no problem (except computational cost), but imagine our Mario is taking part in a multiplayer game.
What do you think should happen?
(Obviously, this is not realistic for the reason that you usally don't run "debuggers" or "IDEs" in a 'multiplayer' scenario, but I think the analogy holds in general. Imagine that the other players' actions are I/O operations like "what's the current time?" or "talk to a REST service", etc.)
Most interesting programs do this kind of thing all the time. Of course, there are interesting programs/functions which are purely algorithmic and rewind/fast-forward should work for them, but they are mostly pretty niche compared to most of the programming that currently happening.
A very interesting example that's somewhere "in between" is Haskell's STM monad which does have very controlled side effects and where you can just "retry" and particular computation in case synchronization fails, but it still doesn't allow you to "project".
[1] Alright, I suppose it's just undeciability for non-trivial properties of TCs languages, but Halting Problem is a convenient shorthand.
I was following closely LightTable and EVE. Very inspiring projects.
> that wall being the limitations of programming languages you outline here
Exactly. I was thinking about the same a while ago. EVE is an attempt to build a new reality where those limitations are gone. But it didn't work quite well -- in order to get support, money, users it should be somehow connected to production.
There is a platform where all this complexety with IO (the mail send in your example) is abstracted quite nicely and can be manipulated/mocked: the BEAM platform (Erlang, Elixir).
You can record IO for particular process (it just messages), replay it if you want.
Also, modules can be compiled in any order, functions in module can be defined in any order. It's much easier to treat function as smallest unit of code. To create a parser that could transform code into data, allow it to manipulate (structured editing and all) and transform it back to text form, to be used by other programmers.
I think if LightTable was aiming at BEAM platform it could go a bit further. Utilize growing popularity of Elixir, and create a new development platform.
<rant incoming, be warned>
I'll use the Mario example from one of his first widely disseminated talks as an example: In the Mario example you see him run time backwards. That's actually reasonably achievable by just recording all previous states or (more plausibly) all state changes and then "just" replaying backwards. Then he runs time forwards again after having changed a setting (or some such). This is, again achievable within limits, see e.g. "rr".
The problem comes when he shows a projected future path of Mario. Because of a) Turing Completeness (TC) + the Halting problem, and b) the fact that much code actually interacts with other systems which are not able to just rewind (and/or themselves interact with systems that ...).
The only plausible scenario in which this is even remotely possible in practice is if the programs adhere to strict rules about side effects, have built-in "project into the future" (or have no side effects so that we can run them to see what the results are _without affecting anything external_.). That is, it's not about fancy IDEs -- it's about programming languages and turning away from Turing Completeness and/or non-abstract/non-formalized effects. FRP/(React+Flux) sort of work for this if you adhere strictly to purity, but then again break down when it comes to things like "what is the focused element" if the user clicks around when you're replaying.
I don't want to be a party-pooper, but good luck changing the programming paradigm of everyone to achieve this vision. For it to work everything has to work in this FRPish way, and that's just never going to happen -- for example, networks (for real-world reasons) cannot behave this way, etc. etc. I WISH we could have everything behave this way, but we can't -- billions of e.g. browser installations which have to have backward-compatible DOM implementations and millions (billions?) of installations of C/C++-based operating systems say so.
Now, I'm not against Imagining Things How They Could Be, but if there's mathematical proof that "thing X" is not possible one either needs to explain how you are not doing "thing X" (perhaps not using a TC language, or disallowing non-pure state) or shut up. Brett Victor and many "visionaries" like him are just very bad at stating their fundamental assumptions up front (which is perhaps part of the reason we don't get to see the code, it's probably just simple pure FRP-like code) so that we can evaluate the vision against the assumptions. (Light Table just seemed to assume that the assumptions were plausible and see where that got them. Funded but with very little actual technological advancement to show for it.)
... and also don't get me started on the whole "Kill Math" thing. Ugh.
(Sorry about the rantish nature of this -- I should probably just write this up properly somewhere so I could properly flesh it out. I just get tired of this "What A Genius!" and "Well, it would take too much time away from inspiration to actually make the thing!" credulous nonsense.)