468 karma · joined July 14, 2014
My favorite trick when cutting down the search space of possible programs is to generate DAGs directly instead of an instruction sequence. This way, it's easier to write rules that avoid nonsensical programs and to some extent avoid generating multiple equivalent programs.
For example to avoid programs that overwrite calculations you make sure that each node in the generated DAG has something that depends on it.
In a suitable environment the DAG can then be executed directly without lowering to a sequence of instructions.
This is just fixing up the hash function, which should have a good enough distribution from the start. And if it did this remapping wouldn't be necessary.
The proper place to add this is in your hash function, to "fix up" the distribution. But then, if your hash function doesn't generate a good enough distribution without this trick you should probably redesign it.
https://www.statista.com/statistics/755052/population-of-the... tells us there are roughly 1 million childred aged 5 or less. Let's say 500 of these are this man's children.
If all of those 1 million kids form couples within the same age group we have 500000 couples.
Assume even split between the sexes, and assume heterosexuality.
Then each of the 500 kids will have a 250/500000=1/2000 chance of coupling with a half-sibling, or 1-1/2000 probability of not. Repeat this for all 500 kids we get a probability of 1-(1-1/2000)^500 = 22% of at least one half-sibling couple.
But where is the innovation? Where are the alternatives?
Is there simply not enough money in cheap 3D printing to justify investment?
> * According to a fun and non-scientific evaluation with GPT-4. Further rigorous evaluation is needed.
Come on... I love any effort related to LLM, but it's really disingenuous to claim high quality in the title and then immediately discredit it after click-through.
For example, let's say I have a short snippet of code that needs to run every H blanking period. I should just declare something like
@every_hblank Blabla_code()
Or scheduling some code for X pixels into a line.
Actual scheduling can be done however the compiler wants to, using interrupts or inserting dummy instructions for delays.
I guess in the end it's a constraint problem to solve where all these snippets fit in.
I imagine this makes it easier to write games/special effects as you remove the tricky cycle counting. Easier, but maybe less fun...
I assume one flash at the start and one at the end of movement
The slow progression of ChatGPT output is just a property of the output layer. The language engine doesn't work slowly like that, and once a token has been generated it can't backtrack.
There are special resistors for this purpose that are very low in resistance.
checksum ^= crc_of_removed_bytes
checksum ^= crc_of_added_bytes
Clickbait for sure, but it really delivered
This kind of bonus feature fits in a later iteration.