It feels like there should be some kind of tool analogous to valgrind/ASAN/TSAN/UBSAN for code that is just run on all the images of any kind of paper before it is published to check for these kinds of things. Some of them just seem to be honest mistakes like linking the same image in two different places and this would fix that so nobody needs to bother checking for it in the future. It would also have the byproduct of catching all the malicious image doctoring that some people do which is just horrific for science in general.
We can compare against another mode of human-controlled transportation. There are 1.37 deaths per 100 million passenger-miles driving in the US [1]. In comparison, there are ~0.2 deaths per 10 billion passenger-miles flying. Converting into the same units, there are 137 deaths per 10 billion passenger-miles driving. So you are 685X more likely to die while driving/riding in a car than flying. That's almost three orders of magnitude worse! Humans are pretty terrible drivers in comparison to how good we are at flying.
To be fair, it's never been easier to get access to thousands of GPUs in the cloud. It might be expensive, but that is an entirely different kind of barrier. Just a decade ago, it used to be that the only way to get access to thousands of GPUs was to get access to a supercomputer at a national lab. Now anybody with enough money can rent thousands of GPUs (with good interconnects too!) in the cloud. There's certainly a limitation on it from a money perspective, but access to the computational resources themselves is not a problem.
Alright, here's an idea: consciousness is spectrum that occurs when a system develops an automatic self-correcting mechanism for interacting with the external physical universe. In some sense all animals (including humans) wandering on this planet are conscious, because we all learn to build our actions around interactions with the external physical universe, e.g., we learn how to walk/swim/fly under the force of gravity without falling/crashing, we learn that the square peg fits in the square hole and not in the round hole, etc. The feedback from these interactions allows us to automatically adjust our future actions without external help. In this sense we learn what works, aka. what is "true" (at least under the laws of this universe). Some animals happen to have "higher" consciousness in that they interact with the universe in more sophisticated ways, learning "deeper" truths, but all animals possess some degree of consciousness under this definition (my cat is certainly consciousness, she has learned how to manipulate the external world, especially me, perfectly at this point). Consciousness is a matter of degrees, not a binary property that one can satisfy.
This definition also has the nice property of showing why current LLMs don't fit on the spectrum. They don't have any concept of learning what is true and automatically self-correcting. They will happily tell us things that are obviously not true, e.g., the square peg fits in the round hole, and then insist that they are right, when a basic physics experiment will disprove their assertions. Interestingly though, things like linear feedback control systems like we might find in an elevator do possess some degree of consciousness: they interact with the physical world, identify the true position of the elevator, move it where they want, and self-correct when necessary. They might be primitive, but I for one believe that they are certainly "conscious" at some level, and definitely more than LLMs. :)
Almost certainly this definition is incomplete and flawed in many aspects, but I think it's at least self-consistent.
I'm still using a Pixel 2, now 4.5 years after I got it (lack of security is easy: don't put or do anything I care about on the phone, turn off the power when having sensitive conversations nearby). It can go two days between charging.
The clever thing about interferometers is that they're actually measuring space and time in two different dimensions concurrently and then looking for changes in the interference pattern between light moving along each of the dimensions. Simplistically, imagine a gravitational wave propagating along one dimension of the interferometer (realistically gravitational waves will never be perfectly aligned with any direction of the interferometer). Space will be distorted in that dimension but not the other and we can notice the change in the resulting interference pattern. In practice, gravitational waves will come from all sorts of weird angles, but they will distort each of the two dimensions differently and allow us to figure what direction they were propagating by the interference pattern that's observed.
I never said it stopped all transmission, but it will stop a large fraction of it. We're playing games of probability. Anybody that deals in absolutes is living in a fantasy world.
> What again is the rush to vaccinate the lowest risk age group?
Because they still contribute to chain of transmission to the most vulnerable members of society, likely in a disproportional way since younger members of society tend to socialize more. Fighting a virus is a collective action. In order to stop transmission to the vulnerable members, you need to cut edges along all paths through the graph. Furthermore, additional spread, even among healthy people with no side effects, increases the probability of mutations that lead to more fit variants capable of causing even more sickness and death.
In my experience, the whole 10 year warranty seems a bit excessive for vehicles that are nigh close to indestructible [1,2,3]. My parents have had multiple Toyotas that made it to 300K-400K miles. My brother's Corolla was rear-ended by a semi doing 30mph and wasn't even close to being totaled. Eight years later and he's still driving it through swamps in south Florida. They are impeccable feats of engineering.
Please correct me if I'm wrong, but I think all of those system only work inside of a single process. Legion/Regent support distributed multi-node multi-process execution both on supercomputers and in the cloud.
Prior to working on Legion, I worked on a programming language called Sequoia that matches much of what you are describing [1]. In many ways Sequoia was the spiritual ancestor of Legion/Regent.
The main difference between Legion and TensorFlow is how and when the dataflow graph is constructed. In TensorFlow the graph is constructed lazily (no execution is performed until you've asked for it), it's optimized, and then distributed to processors (GPUs/TPUs) for execution. In Legion, the graph is built, distributed, and executed on the fly. What this means is that Legion can react to things like dynamic control flow (e.g. branches inside of loops) and analyze dependences at runtime to find task parallelism, in a similar way to how your out-of-order CPU extracts instruction level parallelism from a program. Doing things in the TensorFlow model works better when you can see your whole program up front and can "statically" optimize and schedule it because it has lower overheads, but it also has limits to the kinds of programs it can handle; the Legion approach works better when you have dynamic data-dependent behavior in your program and you need to react to it on the fly, but it does have some overhead to the dynamic analysis.
It's worth noting that Regent is the language that implements the Legion programming model. The Legion runtime system is just a C++ library with bindings for C, Fortran, Python, Terra, and Lua. Writing Regent code is much higher productivity than writing to the C++ Legion library directly, but if you want to you can drop down and write your tasks in any of the other languages that Legion supports. You can even mix and match tasks written in different languages.
Except they have no story for building composable libraries in a distributed setting. The story for distributed execution in Julia today is just use MPI, which is a terrible answer. Anyone who has ever use libraries backed by MPI in any language knows that they are inherently not composable. You can't just take an object partitioned across multiple nodes one way by one library and pass it into a second library that expects it to be partitioned a different way. As far as I can tell the Julia language has nothing to say about that, and that makes them a non-starter today for anyone trying to build composable libraries for distributed memory machines.
I had the same thought. The meta-customers here are the insurance companies. If you can prove that the tool improves the performance of doctors so that they get sued less when they use it then the insurance company will give discounts for doctors on their insurance for using it. Presto! We just created a market for doctors to pay for the service.