Day 22: Monkey Market
Megathread guidelines
- Keep top level comments as only solutions, if you want to say something other than a solution put it in a new post. (replies to comments can be whatever)
- You can send code in code blocks by using three backticks, the code, and then three backticks or use something such as https://topaz.github.io/paste/ if you prefer sending it through a URL
FAQ
- What is this?: Here is a post with a large amount of details: https://programming.dev/post/6637268
- Where do I participate?: https://adventofcode.com/
- Is there a leaderboard for the community?: We have a programming.dev leaderboard with the info on how to join in this post: https://programming.dev/post/6631465
Uiua
It’s been a while since I posted one of these, but I thought this would be straightforward in Uiua. Turns out that bitwise operations are a bit (haha) of a pain, so the
Rng
operation is very slow at 4sec for live data.I took this as an opportunity to play with the
⧈(stencil)
operator which probably slowed things down too.Data ← 1_2_3_2024 Xor ← °⋯◿2⬚0+∩⋯ # Bitwise xor of two numbers. Rng ← ⊙◌◿,Xor×2048.◿,Xor⌊÷32.◿,Xor×64.⊙16777216 Runs ← ⍉(⇌[⍥(Rng.)])2000 Data # Should be constant? Firsts ← ( ⊟⊂0⧈₂/-.◿10 ↘¯1 # Build run, gen pair diffs ⊢⧈(⊟⊙⊣/(+×40+20)°⊟) 2_4 # Convert 4-diff into key, collect. ⊕⊢⊛⊙⍉⊙◌°⊟.⍉ # Only keep first of each key. # ⍜(map°⊟⍉⇌|∘) failed. ) &p /+≡⊣.Runs &p /↥⊕(/+)+1⊛°⊟⍉/◇⊂wait≡spawn(□Firsts) # Group by key, sum prices, return highest.
Rust
Nice breather today (still traumatized from the robots). At some point I thought you had to do some magic for predicting special properties of the pseudorandom function, but no, just collect all values, have a big table for all sequences and in the end take the maximum value in that table. Part 1 takes 6.7ms, part 2 19.2ms.
Solution
fn step(n: u32) -> u32 { let a = (n ^ (n << 6)) % (1 << 24); let b = a ^ (a >> 5); (b ^ (b << 11)) % (1 << 24) } fn part1(input: String) { let sum = input .lines() .map(|l| { let n = l.parse().unwrap(); (0..2000).fold(n, |acc, _| step(acc)) as u64 }) // More than 2¹⁰ 24-bit numbers requires 35 bits .sum::<u64>(); println!("{sum}"); } const N_SEQUENCES: usize = 19usize.pow(4); fn sequence_key(sequence: &[i8]) -> usize { sequence .iter() .enumerate() .map(|(i, x)| (x + 9) as usize * 19usize.pow(i as u32)) .sum() } fn part2(input: String) { // Table for collecting the amount of bananas for every possible sequence let mut table = vec![0; N_SEQUENCES]; // Mark the sequences we encountered in a round to ensure that only the first occurence is used let mut seen = vec![false; N_SEQUENCES]; for l in input.lines() { let n = l.parse().unwrap(); let (diffs, prices): (Vec<i8>, Vec<u8>) = (0..2000) .scan(n, |acc, _| { let next = step(*acc); let diff = (next % 10) as i8 - (*acc % 10) as i8; *acc = next; Some((diff, (next % 10) as u8)) }) .unzip(); for (window, price) in diffs.windows(4).zip(prices.iter().skip(3)) { let key = sequence_key(window); if !seen[key] { seen[key] = true; table[key] += *price as u32; } } // Reset seen sequences for next round seen.fill(false); } let bananas = table.iter().max().unwrap(); println!("{bananas}"); } util::aoc_main!();
Also on github
How have I never noticed that
scan()
exists? Very handy.I liked the zipping of the offset prices, neater than my helper method.
Rust
Not too hard today, apart from yesterday’s visit to a cocktail bar leaving me a little hazy in the mind.
code
use std::{fs, str::FromStr}; use color_eyre::eyre::{Report, Result}; use gxhash::{HashMap, HashMapExt}; const SECRETS_PER_DAY: usize = 2000; const SEQ_LEN: usize = 4; type Sequence = [i8; SEQ_LEN]; fn produce(n: usize) -> usize { let n = (n ^ (n * 64)) % 16777216; let n = (n ^ (n / 32)) % 16777216; (n ^ (n * 2048)) % 16777216 } #[derive(Debug)] struct Buyer { prices: [u8; SECRETS_PER_DAY + 1], changes: [i8; SECRETS_PER_DAY], } impl Buyer { fn price_at_seq(&self, seq: &Sequence) -> Option<u8> { self.changes .windows(SEQ_LEN) .position(|win| win == *seq) .and_then(|i| self.price_for_window(i)) } fn price_for_window(&self, i: usize) -> Option<u8> { self.prices.get(i + SEQ_LEN).copied() } } struct BananaMarket { buyers: Vec<Buyer>, } impl FromStr for BananaMarket { type Err = Report; fn from_str(s: &str) -> Result<Self, Self::Err> { let buyer_seeds = s .lines() .map(|s| s.parse::<usize>()) .collect::<Result<Vec<_>, _>>()?; let buyers = buyer_seeds .into_iter() .map(|seed| { let mut prices = [0; SECRETS_PER_DAY + 1]; let mut changes = [0; SECRETS_PER_DAY]; let mut secret = seed; let mut price = (seed % 10) as u8; prices[0] = price; for i in 0..SECRETS_PER_DAY { let last_price = price; secret = produce(secret); price = (secret % 10) as u8; prices[i + 1] = price; changes[i] = price as i8 - last_price as i8; } Buyer { prices, changes } }) .collect(); Ok(Self { buyers }) } } impl BananaMarket { fn sell_with_seq(&self, seq: &Sequence) -> usize { self.buyers .iter() .map(|b| b.price_at_seq(seq).unwrap_or(0) as usize) .sum() } fn maximise_bananas(&self) -> usize { let mut cache: HashMap<Sequence, usize> = HashMap::new(); for seq in self .buyers .iter() .flat_map(|buyer| buyer.changes.windows(SEQ_LEN)) { let seq = seq.try_into().unwrap(); cache.entry(seq).or_insert_with(|| self.sell_with_seq(&seq)); } cache.into_values().max().unwrap_or(0) } } fn part1(filepath: &str) -> Result<usize> { let input = fs::read_to_string(filepath)? .lines() .map(|s| s.parse::<usize>()) .collect::<Result<Vec<_>, _>>()?; let res = input .into_iter() .map(|n| (0..SECRETS_PER_DAY).fold(n, |acc, _| produce(acc))) .sum(); Ok(res) } fn part2(filepath: &str) -> Result<usize> { let input = fs::read_to_string(filepath)?; let market = BananaMarket::from_str(&input)?; Ok(market.maximise_bananas()) } fn main() -> Result<()> { color_eyre::install()?; println!("Part 1: {}", part1("d22/input.txt")?); println!("Part 2: {}", part2("d22/input.txt")?); Ok(()) }
Haskell
I have no Idea how to optimize this and am looking forward to the other solutions that probably run in sub-single-second times. I like my solution because it was simple to write which I hadn’t managed in the previous days, runs in 17 seconds with no less than 100MB of RAM.
import Control.Arrow import Data.Bits (xor) import Data.Ord (comparing) import qualified Data.List as List import qualified Data.Map as Map parse :: String -> [Int] parse = map read . filter (/= "") . lines mix = xor prune = flip mod 16777216 priceof = flip mod 10 nextSecret step0 = do let step1 = prune . mix step0 $ step0 * 64 let step2 = prune . mix step1 $ step1 `div` 32 let step3 = prune . mix step2 $ step2 * 2048 step3 part1 = sum . map (head . drop 2000 . iterate nextSecret) part2 = map (iterate nextSecret >>> take 2001 >>> map priceof >>> (id &&& tail) >>> uncurry (zipWith (curry (uncurry (flip (-)) &&& snd))) >>> map (take 4) . List.tails >>> filter ((==4) . length) >>> map (List.map fst &&& snd . List.last) >>> List.foldl (\ m (s, p) -> Map.insertWith (flip const) s p m) Map.empty ) >>> Map.unionsWith (+) >>> Map.assocs >>> List.maximumBy (comparing snd) main = getContents >>= print . (part1 &&& part2) . parse
Haha, same! Mine runs in a bit under 4s compiled, but uses a similar 100M-ish peak. Looks like we used the same method.
Maybe iterate all the secrets in parallel, and keep a running note of the best sequences so far? I’m not sure how you’d decide when to throw away old candidates, though. Sequences might match one buyer early and another really late.
Haskell
solution
import Control.Arrow import Data.Bits import Data.List import qualified Data.Map as M parse = fmap (secretNums . read) . lines secretNums :: Int -> [Int] secretNums = take 2001 . iterate (step1 >>> step2 >>> step3) where step1 n = ((n `shiftL` 06) `xor` n) .&. 0xFFFFFF step2 n = ((n `shiftR` 05) `xor` n) .&. 0xFFFFFF step3 n = ((n `shiftL` 11) `xor` n) .&. 0xFFFFFF part1 = sum . fmap last part2 = maximum . M.elems . M.unionsWith (+) . fmap (deltas . fmap (`mod` 10)) deltas l = M.fromListWith (\n p -> p) $ flip zip (drop 4 l) $ zip4 diffs (tail diffs) (drop 2 diffs) (drop 3 diffs) where diffs = zipWith (-) (tail l) l main = getContents >>= print . (part1 &&& part2) . parse
Rust
Part 2 is crazy slow, but it works, so thats cool :D
Edit: Gonna fix this, because pt2 is stupid.Much better, 2.4s. Still slow, but not 6 minutes slow.
#[cfg(test)] mod tests { use std::collections::HashMap; use std::iter::zip; fn step(start: usize) -> usize { let mut next = start; next = ((next * 64) ^ next) % 16777216; next = ((next / 32) ^ next) % 16777216; next = ((next * 2048) ^ next) % 16777216; next } fn simulate(initial: usize) -> usize { let mut next = initial; for _ in 0..2000 { next = step(next); } next } #[test] fn test_step() { assert_eq!(15887950, step(123)); } #[test] fn test_simulate() { assert_eq!(8685429, simulate(1)); } #[test] fn day22_part1_test() { let input = std::fs::read_to_string("src/input/day_22.txt").unwrap(); let initial_values = input .split("\n") .map(|s| s.parse::<usize>().unwrap()) .collect::<Vec<usize>>(); let mut total = 0; for value in initial_values { total += simulate(value); } println!("{}", total); } #[test] fn day22_part2_test() { let input = std::fs::read_to_string("src/input/day_22.txt").unwrap(); let initial_values = input .split("\n") .map(|s| s.parse::<usize>().unwrap()) .collect::<Vec<usize>>(); let mut all_deltas = vec![]; let mut all_values = vec![]; for value in initial_values { let mut deltas = String::with_capacity(2000); let mut values = vec![]; let mut prev = value; for _ in 0..2000 { let next = step(prev); values.push(next % 10); deltas.push((10u8 + b'A' + ((prev % 10) as u8) - ((next % 10) as u8)) as char); prev = next; } all_deltas.push(deltas); all_values.push(values); } let mut totals = HashMap::with_capacity(100000); for (delta, value) in zip(&all_deltas, &all_values) { let mut cache = HashMap::with_capacity(2000); for j in 0..delta.len() - 4 { let seq = &delta[j..j + 4]; let bananas = value[j + 3]; cache.entry(seq).or_insert(bananas); } for (key, value) in cache { *totals.entry(key).or_insert(0) += value; } } let max_bananas = totals.values().max().unwrap(); println!("{}", max_bananas); } }
Six minutes? 😅 I was feeling crappy about my 30 seconds (my naive big O cubed(?) logic means my code spends most of its time testing array equalities - 72 billion samples in the flamegraph!)
Most of my time is wasted on hashmap stuff. And the processing into the string, which really isnt needed anymore. :/
Have you tried gxhash or one of the other non-cryptographic hashers?
I probably should give that a try. Looks like it can just drop in, so might try it later. I see FxHash is pretty popular here as well.
Go
Re-familiarizing myself with Go. The solution to Part 2 is fairly simply, the whole packing of the sequence into a single integer to save on memory was an optimization I did afterwards based on looking at other solutions. I thought it was cool.
package main import ( "bufio" "fmt" "os" "strconv" ) type SequenceMap struct { Data map[int32]int } func PackSeq(numbers [4]int8) int32 { var packed int32 for i, num := range numbers { packed |= int32(num+9) << (i * 5) } return packed } func UnpackSeq(packed int32) [4]int8 { var numbers [4]int8 for i := range numbers { numbers[i] = int8((packed>>(i*5))&0x1F) - 9 } return numbers } func NewSequenceMap() SequenceMap { return SequenceMap{make(map[int32]int)} } func (m *SequenceMap) Increment(seq [4]int8, val int) { pSeq := PackSeq(seq) acc, ok := m.Data[pSeq] if ok { m.Data[pSeq] = acc + val } else { m.Data[pSeq] = val } } func (m *SequenceMap) Has(seq [4]int8) bool { pSeq := PackSeq(seq) _, ok := m.Data[pSeq] return ok } type Generator struct { Secret int64 LastPrice int8 ChangeSequence []int8 } func NewGenerator(Secret int64) Generator { var ChangeSequence []int8 return Generator{Secret, int8(Secret % 10), ChangeSequence} } func (g *Generator) Mix(value int64) *Generator { g.Secret = g.Secret ^ value return g } func (g *Generator) Prune() *Generator { g.Secret = g.Secret % 16777216 return g } func (g *Generator) Next() { g.Mix(g.Secret * 64).Prune().Mix(g.Secret / 32).Prune().Mix(g.Secret * 2048).Prune() Price := int8(g.Secret % 10) g.ChangeSequence = append(g.ChangeSequence, Price-g.LastPrice) g.LastPrice = Price if len(g.ChangeSequence) > 4 { g.ChangeSequence = g.ChangeSequence[1:] } } func ParseInput() []int64 { if fileInfo, _ := os.Stdin.Stat(); (fileInfo.Mode() & os.ModeCharDevice) != 0 { fmt.Println("This program expects input from stdin.") os.Exit(1) } scanner := bufio.NewScanner(os.Stdin) var numbers []int64 for scanner.Scan() { line := scanner.Text() num, err := strconv.ParseInt(line, 10, 64) if err != nil { fmt.Printf("ERROR PARSING VALUE: %s\n", line) os.Exit(1) } numbers = append(numbers, num) } return numbers } func main() { numbers := ParseInput() m := NewSequenceMap() sum := int64(0) for i := 0; i < len(numbers); i += 1 { g := NewGenerator(numbers[i]) tM := NewSequenceMap() for j := 0; j < 2000; j += 1 { g.Next() if len(g.ChangeSequence) == 4 { if !tM.Has([4]int8(g.ChangeSequence)) { tM.Increment([4]int8(g.ChangeSequence), 1) if g.LastPrice > 0 { m.Increment([4]int8(g.ChangeSequence), int(g.LastPrice)) } } } } sum += g.Secret } fmt.Printf("Part One: %d\n", sum) var bestSeq [4]int8 bestPrice := 0 for pSeq, price := range m.Data { if price > bestPrice { bestPrice = price bestSeq = UnpackSeq(pSeq) } } fmt.Printf("Part Two: %d\n", bestPrice) fmt.Printf("Best Sequence: %d\n", bestSeq) }
Dart
Well, that was certainly a bit easier than yesterday…
I know running a window over each full list of 2000 prices rather than looking for cycles etc means I’m doing a lot of unnecessary work, but it only takes a couple of seconds, so that’ll do.
import 'package:collection/collection.dart'; import 'package:more/more.dart'; rng(int i) { i = ((i << 6) ^ i) % 16777216; i = ((i >> 5) ^ i) % 16777216; i = ((i << 11) ^ i) % 16777216; return i; } Iterable<int> getPrices(int val, int rounds) { var ret = [val]; for (var _ in 1.to(rounds)) { ret.add(val = rng(val)); } return ret.map((e) => e % 10); } int run(int val, int rounds) => 0.to(rounds).fold(val, (s, t) => s = rng(s)); part1(lines) => [for (var i in lines.map(int.parse)) run(i, 2000)].sum; part2(List<String> lines) { var market = <int, int>{}.withDefault(0); for (var seed in lines.map(int.parse)) { var seen = <int>{}; for (var w in getPrices(seed, 2000).window(5)) { var key = // Can't use lists as keys, so make cheap hash. w.window(2).map((e) => e[1] - e[0]).reduce((s, t) => (s << 4) + t); if (seen.contains(key)) continue; seen.add(key); market[key] += w.last; } } return market.values.max; }
C
Really proud of this one! Started with with an O(n^atoms in the universe) scan which took 44s even after adding a dedup check.
But iterating on a trick to encode the deltas for the dedup check, using it to build a mapping table here, a lookup there etc brought it down to a very fast, fairly low memory, linear complexity solution!
Code
#include "common.h" #define STEPS 2000 #define NCODES (19*19*19*19) int main(int argc, char **argv) { static int8_t prices[STEPS]; static int8_t by_deltas[NCODES]; static int sums[NCODES]; uint64_t p1=0, secret; int p2=0, i; if (argc > 1) DISCARD(freopen(argv[1], "r", stdin)); while (scanf(" %"SCNu64, &secret) == 1) { memset(by_deltas, 0, sizeof(by_deltas)); for (i=0; i<STEPS; i++) { secret = (secret ^ secret << 6) & 0xFFFFFF; secret = (secret ^ secret >> 5); secret = (secret ^ secret << 11) & 0xFFFFFF; prices[i] = secret % 10; } /* * Build a deltas->price map for the buyer. Deltas are * encoded as an integer for easy indexing. Iterating * backwards ensures the stored price is the _earliest_ * occurence of that sequence. */ for (i=STEPS-1; i>=4; i--) by_deltas[ (prices[i-3] - prices[i-4] +9) *19*19*19 + (prices[i-2] - prices[i-3] +9) *19*19 + (prices[i-1] - prices[i-2] +9) *19 + (prices[i] - prices[i-1] +9) ] = prices[i]; for (i=0; i<NCODES; i++) sums[i] += by_deltas[i]; p1 += secret; } for (i=0; i<NCODES; i++) p2 = MAX(p2, sums[i]); printf("22: %"PRIu64" %d\n", p1, p2); return 0; }
day22 0m00.04s real
https://codeberg.org/sjmulder/aoc/src/branch/master/2024/c/day22.c
Haskell
A nice easy one today; shame I couldn’t start on time. I had a go at refactoring to reduce the peak memory usage, but it just ended up a mess. Here’s a tidy version.
import Data.Bits import Data.List import Data.Map (Map) import Data.Map qualified as Map next :: Int -> Int next = flip (foldl' (\x n -> (x `xor` shift x n) .&. 0xFFFFFF)) [6, -5, 11] bananaCounts :: Int -> Map [Int] Int bananaCounts seed = let secrets = iterate next seed prices = map (`mod` 10) secrets changes = zipWith (-) (drop 1 prices) prices sequences = map (take 4) $ tails changes in Map.fromListWith (const id) $ take 2000 (zip sequences (drop 4 prices)) main = do input <- map read . lines <$> readFile "input22" print . sum $ map ((!! 2000) . iterate next) input print . maximum $ Map.unionsWith (+) $ map bananaCounts input
Python3
Hey there lemmy, I recently transitioned from using notepad to Visual Studio Code along with running a local LLM for autocomplete(faster than copilot, big plus but hot af room)
I was able to make this python script with a bunch of fancy comments and typing silliness. I ended up spamming so many comments. yay documentation! lol
Solve time: ~3 seconds (can swing up to 5 seconds)
Code
from tqdm import tqdm from os.path import dirname,isfile,realpath,join from collections.abc import Callable def profiler(method) -> Callable[..., any]: from time import perf_counter_ns def wrapper_method(*args: any, **kwargs: any) -> any: start_time = perf_counter_ns() ret = method(*args, **kwargs) stop_time = perf_counter_ns() - start_time time_len = min(9, ((len(str(stop_time))-1)//3)*3) time_conversion = {9: 'seconds', 6: 'milliseconds', 3: 'microseconds', 0: 'nanoseconds'} print(f"Method {method.__name__} took : {stop_time / (10**time_len)} {time_conversion[time_len]}") return ret return wrapper_method # Process a secret to a new secret number # @param n: The secret number to be processed # @return: The new secret number after processing def process_secret(n: int) -> int: """ Process a secret number by XORing it with the result of shifting left and right operations on itself. The process involves several bitwise operations to ensure that the resulting number is different from the original one. First, multiply the original secret number by 64 with a left bit shift, then XOR the original secret number with the new number and prune to get a new secret number Second, divide the previous secret number by 32 with a right bit shift, then XOR the previous secret number with the new number and prune to get another new secret number Third, multiply the previous secret number by 2048 with a left bit shift, then XOR the previous secret number with the new number and prune to get the final secret number Finally, return the new secret number after these operations. """ n ^= (n << 6) n &= 0xFFFFFF n ^= (n >> 5) n &= 0xFFFFFF n ^= (n << 11) return n & 0xFFFFFF # Solve Part 1 and Part 2 of the challenge at the same time @profiler def solve(secrets: list[int]) -> tuple[int, int]: # Build a dictionary for each buyer with the tuple of changes being the key and the sum of prices of the earliest occurrence for each buyer as the value # At the same time we solve Part 1 of the challenge by adding the last price for each secret last_price_sum = 0 changes_map = {} for start_secret in (secrets): # Keep track of seen patterns for current secret changes_seen = set() # tuple of last 4 changes # first change is 0 because it is ignored last_four_changes = tuple([ (cur_secret:=process_secret(start_secret)), -(cur_secret%10) + ((cur_secret:=process_secret(cur_secret)) % 10) , -(cur_secret%10) + ((cur_secret:=process_secret(cur_secret)) % 10) , -(cur_secret%10) + ((cur_secret:=process_secret(cur_secret)) % 10) ] ) current_price = sum(last_four_changes) # Map 4-tuple of changes -> sum of prices index of earliest occurrence for all secrets for i in range(3, 1999): # sliding window of last four changes last_four_changes = (*last_four_changes[1:], -(cur_secret%10) + (current_price := (cur_secret:=process_secret(cur_secret)) % 10) ) # if we have seen this pattern before, then we continue to next four changes # otherwise, we add the price to the mapped value # this ensures we only add the first occurence of a patten for each list of prices each secret produces if last_four_changes not in changes_seen: # add to the set of seen patterns for this buyer changes_seen.add(last_four_changes) # If not recorded yet, store the price # i+4 is the index of the price where we sell changes_map[last_four_changes] = changes_map.get(last_four_changes, 0) + current_price # Sum the 2000th price to the total sum for all secrets last_price_sum += cur_secret # Return the sum of all prices at the 2000th iteration and the maximum amount of bananas that one pattern can obtain return last_price_sum,max(changes_map.values()) if __name__ == "__main__": # Read buyers' initial secrets from file or define them here BASE_DIR = dirname(realpath(__file__)) with open(join(BASE_DIR, r'input'), "r") as f: secrets = [int(x) for x in f.read().split()] last_price_sum,best_score = solve(secrets) print("Part 1:") print(f"sum of each of the 2000th secret number:", last_price_sum) print("Part 2:") print("Max bananas for one of patterns:", best_score)
Bit odd having
main()
returning an actual value, probably would have named it something else, otherwise, nicely documented solution.I bet VSC is a lot nicer to work in than notepad :D
you have a point to call name it something else, but lazy to do that. should I simply call it
solve()
maybe, that would work fine.
I do want to note that having it return a value is not unheard of, it is just part of being lazy with the naming of the functions.
I definitely would not have the code outside ofmain()
be included in the main function as it is just something to grab the input pass it to the solver function( main in this case, but as you noted should be called something else ) and print out the results. if you imported it as a module, then you can callmain()
with any input and get back the results to do whatever you want with. Just something I think makes the code better to look at and use.While doing this is highly unnecessary for these challenge, I wish to keep a little bit of proper syntax than just writing the script with everything at the top level. It feels dirty.
Coding in notepad was a bit brutal, but I stuck with notepad for years and never really cared because I copy pasta quite a bit from documentation or what not.(now a days, gpt helps me fix my shit code, now that hallucinations are reduced decently) even with VSCode, I don’t pay attention to many of its features. I still kinda treat it as a text editor, but extra nagging on top.(nagging is a positive I guess, but I am an ape who gives little fucks) I do like VSCode having a workspace explorer on the side. I dislike needing to alt-tab to various open file explorer windows. Having tabs for open files is really nice, too.
VSCode is nice, and running my Qwen-coder-1.5B locally is neat for helping somethings out. Not like I rely on it for helping with coding, but rather use it for comments or sometimes I stop to think or sometimes the autocomplete is updated in realtime while I am typing. really neat stuff, I think running locally is better than copilot because of it just being more real-time than the latency with interacting with MS servers. though I do think about all the random power it is using and extra waste heat from me constantly typing and it having to constantly keep up with the new context.
The quality took a little hit with the smaller model than just copilot, but so far it is not bad at all for things I expect it to help with. It definitely is capable of helping out. I do get annoyed when the autocomplete tries too hard to help by generating a lot more stuff that I don’t want.(even if the first part of what it generated is what I wanted but the rest is not) thankfully, that is not too often.
I give the local llm is helping with 70% of the comments and 15% of the code on average but it is not too consistent for the code.
For python, there is not enough syntax overhead to worry about and the autocomplete isn’t needed as much.Its normal for main to return a value, its just usually a status thing, rather than actual data. But given python doesn’t really treat main as anything special, it hardly matters