I think he's saying, random mutation wouldn't produce all required components at once. One mutation gives you a bit of a flagella, another gives you bit of a nose, but how does the flagella mutation survive to coexist with the nose mutation that makes it useful.
I suspect the answer is that having flagella without a nose is still better than having no flagella. If so it suggests evolution isn't good at accessing groups of mutations that aren't individually beneficial.
Evolution doesn't produce 1st part of the flagellum, second part of the flagellum, third part of the flagellum.
It produces shitty flagellum, better flagellum, good flagellum.
But the problem is we don't see the intermediate forms. So right now you might see a complicated flagellum that has a lot of highly specialized parts that all need eachother, but that is merely a refinement that took place after all the pieces were already there. Like once an arch is complete, all the scaffolding that was holding it up is now vestigial and if it is removed the arch will remain standing.
It seems you may have misunderstood the original argument. The iterative approach suggests increments so minute at each step that they wouldn't significantly impact an organism's survival at any given time. Also given the extremely slow process of evolution and the relatively short number of iterations it is infeasible to suggest such a solution. If a person would like to create an iPhone it's easy to tell them to start with a shitty scrap of metal and work from there. You can make that sort of argument as a solution for creating anything but it is clearly not feasible.
No, I understand the argument, it is just built on a false assumption about how the iterations work. That a change is small does not make its effects insignificant. A single codon change could profoundly alter the protein it encodes, and even a small change to a protein or its expression can have a massive effect on the organism. It's not the structures of an orgnaism that mutate, it's the instructions that generate those structures which mutate. Imagine for example a typo on a blueprint - where there was supposed to be a " instead there is a ' and suddenly instead of an 8 inch air vent, now you have an 8 foot door. There is no intermediate step where you have a useless 2 foot hole.
Evolution is not a slow process, it is an irregular process. The odds of a useful mutation popping up at any given time is low, but once it pops up it's there immediately. Yes, an evolutionary process could never make an iphone, but no is claiming that evolution produced the iphone. The complex systems evolution produces are things where all the changes are individually useful.
>so minute at each step that they wouldn't significantly impact an organism's survival at any given time.
That's the thing. Evolution isn't "survival of the fittest" or even "driven by more efficient anything", evolution is simply; if you die before you pass on your genes, you don't pass on your genes. Over long enough time scales, with large enough populations, with tight enough tolerances and strict enough niches, the system roughly approximates a directed iteration of more efficient parts.
Nothing about evolution prevents carrying forward explicitly negative mutations! Nothing about evolution prevents carrying completely unused functionality and features! Nothing about evolution guarantees monotonically increasing fitness!
The giraffe has a certain nerve that goes from it's brain, all the way down around it's aorta, back up it's neck, to it's tongue. It does this, because in the fish we all evolved from, such a detour was less than a centimeter longer than an "optimal" path, and as each next generation went in different directions, it's just not that big a deal. A few hundred extra calories in development, and rare instances of a negative injury outcome are just not going to get fixed, because evolution is almost never vigilant. Most higher level animals have mating behaviors that explicitly favor "wasted" energy, including the long neck of giraffes! Sexual selection has a stronger influence on most animals than evolutionary pressure.
> Also given the extremely slow process of evolution and the relatively short number of iterations it is infeasible to suggest such a solution
This is silly. The vast majority of the ground work for complex life was developed by single celled organisms that produced a new generation every half hour, there were billions of these little creatures experiencing basically any possible mutation all the time, and a water droplet with a billion short lived single cells is exactly the kind of tight tolerance, competitive atmosphere where evolution is most prominent!
Evolution is not iteration. Evolution is pruning bad branches in your breadth first tree based algorithm.
Why not? People think in such a short time and amount scale such that we cannot comprehend trillions of cells spending billions of years, iterating. Even a small change can be significant at those scales.
I understand that, but it seems like even the MVP "shitty" flagellum would require many mutations that individually have no benefit. But I suppose with enough generations/parallelism you get enough stacking of useless mutations to reach the useful ones.
That's the thing most people have difficulty wrapping their head around. What you need to remember is it's not the structures evolving, it's the instructions evolving. If for example you have a small molecular pump that the cell uses to suck up sodium ions, and a mutation causes the part of the rotor sticking out of the cell to just be longer, which might be due to a single change to the gene controlling the length of the rotor, then congrats, you now have a shitty flagellum. The mutations don't even need to be useful for the eventual purpose. For example the highly dexterous fingers which enable complex tool use that humans used to conquer the world and with which I type this comment now started out as structural reinforcement for fish fins, absolutely useless for object manipulation. And those reinforcements in turn are just extremely bastardized version of a calcite growth which offered some protection to a soft body organism hundreds of millions of years before.
But again you're starting with a fairly complex system already, the molecular pump.
And my (limited) understanding is that changes that are not useful or helpful would get lost.
And additional to that, if an organism has a pump (which it needs to function properly) and that pump suddenly is no pump, it's a very bad flagellum, that organism has a very big problem. It's like if we swapped our arms for wings. Wings are cool, but we wouldn't be able to fly anyway, and we'd have serious problems as humans with no arms and hands.
> and that pump suddenly is no pump, it's a very bad flagellum, that organism has a very big problem
Cells has many duplicates of pumps, not just one. Switching one of those to a motor to move around the liquids so the pumps can get to new molecules to absorb will be extremely beneficial to the cell, now all pumps are more efficient at just the cost of a single pump.
> But again you're starting with a fairly complex system already, the molecular pump.
Yes, but it's less complex, and it in turn evolved from even simpler forms. The point is a single mutation doesn't need to create a working flagellum from scratch, it just needs to make it from what's already available. Flagella are complex structures that did not arise until after a lot of other things had already developed.
> And my (limited) understanding is that changes that are not useful or helpful would get lost.
This misunderstanding again comes from the distinction between the features and the instructions. If a mutation isn't harmful, it doesn't get reverted and in fact will spread throughout a sizeable fraction of the population. The thing is that without evolutionary pressure as more mutations occur, they will eventually break whatever the original mutation did, so the feature it coded will eventually disappear, though it can take a long time and it may change significantly before it does. There are some caveats though - a gene might code for multiple features, or may exist on a part of the DNA where further mutations are suppressed anyway, and thus even though the feature provides no advantage on its own there will still be evolutionary pressure to preserve the gene, and thus a neutral or even a slightly bad mutation might be retained indefinitely.
> And additional to that, if an organism has a pump (which it needs to function properly) and that pump suddenly is no pump, it's a very bad flagellum, that organism has a very big problem. It's like if we swapped our arms for wings. Wings are cool, but we wouldn't be able to fly anyway, and we'd have serious problems as humans with no arms and hands.
In this particular case, cells have many molecular pumps, so converting some to flagella is not a very big problem. The benefits of a shitty flagellum did outweigh the cost of losing some molecular pumps, but this is a very real limitation to what evolution can produce. Humans certainly won't evolve wings naturally without a lot of other changes happening first. But at the same time wings have evolved - in the case of birds their ancestors evolved an extremely efficient respiratory system more than 100 million years before they took flight, which helped them survive the great dying and subsequently take advantage of the oxygen rich Mesozoic era. Among these a mutation for hollow bones aided agility on the ground, among these adaptations for feathers helped with retaining body heat, among these adaptations for lunging their arms forward helped them grab prey, among these adaptations for tree climbing allowed them to become better ambushers, and it was among these that sacrificing some of their arm capabilities for shitty wings was a net gain.
> nobody at this point expects a 13B parameter model to succeed with the same accuracy at the broad range of tasks supported by what may be a 1T parameter model
I think a lot of people believe exactly that. To take one example from the "We Have No Moat" essay:
"It doesn’t take long before the cumulative effect of all of these fine-tunings overcomes starting off at a size disadvantage. Indeed, in terms of engineer-hours, the pace of improvement from these models vastly outstrips what we can do with our largest variants, and the best are already largely indistinguishable from ChatGPT." - https://www.semianalysis.com/p/google-we-have-no-moat-and-ne...
That essay works in a context of specific datasets and tasks, which are referenced in the surrounding sentences and paragraphs. They are saying that for a particular "emergent" capability you might reach with a giant LLM, you might get there more efficiently with distillation / LoRa.
My comment is about generality, which is the remaining advantage of giant models.
Unity and Unreal both support Android and PBR so I don't see an obvious benefit to being able to use this alternative renderer. Because of the license and the fact that it's a newish codebase written in simple C++ I can see it being popular as a starting point for in-house engines.
I think the interactive diagram is the most important artifact here, rather than the visual prototype. The synchronized combination of the two is even better.
This kind of thing isn't usually implemented as a formal state machine in "real" code, so even if designers were capable of writing the "real" code, they wouldn't be able to automatically generate an interactive diagram the way this thing can.
They use a gamma distribution which has more probability density near the origin, which causes samples around the origin and interpolations to be more like real input.
Your app looks amazing. I love how you can leave handwritten notes next to the audio control nodes. I know these kind of hybrid sketching/programming environments have been a HCI dream for a long time now but I haven't seen it look as good as in your demo.
What mattered was seeing the value of the capability, not predicting any particular applications
How can you separate the value of a capability from its potential applications? Surely he doesn't mean scientific/intellectual value so.. how else can a capability have value?
In product as a platform, you may understand that there are problems in a particular space, but you may not understand the problems. However, you may understand how a platform might enable the solving of problems, while not really understanding all the problems that will be solved themselves. Generally to be sure of this you should see at least 2 or 3 solutions that you could have conviction in, but understand that entablement would solve very many tangential problems to those you have conviction in is more important. You can somewhat use this to size a future ecosystem market, but you have to be wary about your predictions of the number of solutions you enable. You're also very much fighting too early or just right in terms of timing. Looking for solutions to problems will help you judge the value of the capability, but trying to predict all the applications of those solutions is a fool's errand.
That quote stood out to me too. My question is, if there is a new capability (Bitcoin is a good example) then how do you invest in that capability? Sure you can pick a company working on some application, but they will almost certainly miss the real value of the capability.
Missing real value can be the same as showing potential value. If you have conviction there is real value, then you're banking on many misses, some capturing most of it and a fewer number using it to it's potential. From there things like commoditization come into play. However, you should be mindful of tying capability to value, just because something is capable does not also mean it must be adopted as valuable. Plenty of things that are not particularly capable are overvalued, and many things that are incredibly capable are under valued. I think the important part to look at is does this capability accessibility solve very many problems. If so, then you can start to think about a few solutions, but it isn't meaningful to try and understand all the applications of the solutions (because you typically don't understand that abstract customer). Cloud is a good example of this.
It's not clear to me that you can. How would you have invested in air travel in the time of the Wright brothers? None of the companies that eventually came to dominate existed then.
I think in this light you either invest in an inventor or radical innovator, or a component of their supply chain. One is: I'm going to make this happen and all these things will happen, the other is: this thing is happening so we should should to these things. This is why the path from venture (paradigm shift) to public (shifted) is important, the best public companies indicate a fundamental in other businesses, they are enablers. However, over time you become a commodity item on a supply chain and run risk the "enabled" companies will out innovating you as they build their capabilities and margins .
The article talks about superpowers, which I take to mean the technology can do some general thing that could not be done before. In that case people can invent a great many specific things to do with it, and so in is likely inevitable some of them will be great successes. So the fact that when it is first invented a given individual can't think of a specific use that would be very valuable doesn't mean much.
I suspect the answer is that having flagella without a nose is still better than having no flagella. If so it suggests evolution isn't good at accessing groups of mutations that aren't individually beneficial.