Kuro

AI insights & useful tools. No fluff.

[Kuro @x_Kuro_]

my AI said Antonelli would win. he finished 5th 💀

so i replayed all 10,000 simulations to see how wrong it really was. the answer is at the end of the video 👇

what it got wrong:
- it trusted season form too much. antonelli is the champion, so it made him a 53.7% favorite even though norris was on pole
- it only knows "dry race" or "wet race". today was both. inters at the start, slicks around lap 16, rain again at the end. tire timing mattered a lot, and antonelli spun at turn 1 on cold tires right after the switch
- it thought starting from the back was a death sentence. verstappen went from P21 to P2

what it got right:
- it called max the dark horse from the back
- it said rain was likely (55%). it rained
- norris won in 1,823 of the 10,000 sims. not the favorite, but it had him as a real option

the exact podium (norris, verstappen, hadjar) came up 11 times out of 10,000 lol

for vegas i'm fixing 3 things: more weight on this weekend's pace, mixed conditions with real tire strategy, and less punishment for starting at the back

AI 0, reality 1. let's see if it learns

[Kuro @x_Kuro_]

I built an F1 race predictor app with Claude Opus 5.5. Only I have it for now.

It ran tomorrow's Sao Paulo GP 10,000 times. I already know who wins.

But then I turned the rain on, and everything flipped. Even Verstappen, starting from the back, gets a shot.

The winner is at the very end of the video 👇

Drop your pick before you watch.

AI vs reality after the race.

[Kuro @x_Kuro_]

dropped 10,000 simulated seasons into the fight for P2 in the championship

max ends up P2 in 79% of them. george 12%

the weird part: lando is 33 points back and still finishes P2 more often than lewis, who's only 18 back (5.6% vs 2.5%)

now cancel qatar and abu dhabi so vegas becomes the finale. max jumps to 94%. charles and lando drop to zero

3 races left or just 1, who's your P2?

[Kuro @x_Kuro_]

@CodeLina_ 10,000 per race. banning the 2/3 is smart, i do basically the same thing: when i test it on old races it only gets data from before each race, no peeking. it still lost to "pole always wins" at first lol

[Kuro @x_Kuro_]

@formularacers_ so if qatar and abu dhabi both fall through, vegas becomes the finale? a street race on a 1.9 km straight to close the year would be wild

[Kuro @x_Kuro_]

@F1 slicks on a wet track and still sending it. this is exactly why rain breaks every prediction model, mine had antonelli at 70% in the dry and 40% in the wet and he still finished 5th

[Kuro @x_Kuro_]

@F1GuyDan the crazy part is how much those poles are worth. pole won 14 of 20 this season. if you're not starting on the front row you're basically not winning

[Kuro @x_Kuro_]

@F1GuyDan and pole has never mattered more than this year. 14 of 20 races won from pole, P2 won 4, the other 20 grid spots got 2 between them

[Kuro @x_Kuro_]

@CodeLina_ 6 scenes from one prompt with zero animation fixes is wild. the "real seeded sim, no fake numbers" rule is the part most people skip, i run my f1 predictor the same way. did it check its own numbers against the 2/3 before rendering?

[Kuro @x_Kuro_]

@CodeLina_ not yet, and that's a great call. monza is literally where the only win from the back happened this year (antonelli from P19), while monaco went to the polesitter. adding an overtaking score per track next. vegas with that huge straight will be the first test

[Kuro @x_Kuro_]

pole won 14 of 20 F1 races this season. P2 won 4. the other 20 grid spots won just 2 between them

i found this while trying to fix my AI

the plan: teach it to trust fast drivers who start at the back after a penalty

tested it the honest way again: 17 races, and it only sees data from before each race

result: not a single pick changed, and its odds got slightly less accurate. so v3 isn't shipping

turns out there's almost nothing to learn. the only win from the back all year was antonelli at monza, from P19

vegas is next week, with one of the longest straights on the calendar. pole wins again or comeback drive?

[Kuro @x_Kuro_]

my f1 ai just lost to a rule a 5-year-old could come up with: "whoever starts on pole wins"

so i tested it on 17 races of the 2026 season. for each one, the ai only saw data from before that race. no peeking

old model: 7/17
"pole always wins": 11/17

found 2 problems:
- it cared more about season form than the starting grid
- it was way too confident

fixed both and reran the exact same test:
new model: 11/17
and it's finally calibrated. its favorites should've won about 10.4 races. they won 11

so it's tied with the pole rule. not beating it yet

oh, and the new one would've called brazil

vegas is next. would you trust it?

[Kuro @x_Kuro_]

@CodeLina_ the first version took a few weekends, but it's still a work in progress. funny part is, it's still losing to "pole wins" lol. working on v2 now, any ideas what data i should feed it?

[Kuro @x_Kuro_]

@me_barnyx This is rly fkin insane..

[Kuro @x_Kuro_]

i gave 150 AI cars the las vegas f1 track and no driving lessons

no human data. no racing line. just a tiny neural network that sees the road ahead and its own speed

generation 1: 0 out of 150 finished. 145 didn't even make it 100 m
generation 3: the first 4 cars made it around
generation 5: 66 finished. best lap: 1:46
generation 20: all 150 finished
generation 120: best lap: 1:39.13

how it learns: every generation it keeps the best cars, tweaks them a little and sends them out again. that's it

the craziest part is how fast it figures it out. watch the corner at 0:04

f1 races in vegas next week. think you could drive it faster than the AI?

[Kuro @x_Kuro_]

my f1 ai just lost to a rule a 5-year-old could come up with: "whoever starts on pole wins"

so i tested it on 17 races of the 2026 season. for each one, the ai only saw data from before that race. no peeking

old model: 7/17
"pole always wins": 11/17

found 2 problems:
- it cared more about season form than the starting grid
- it was way too confident

fixed both and reran the exact same test:
new model: 11/17
and it's finally calibrated. its favorites should've won about 10.4 races. they won 11

so it's tied with the pole rule. not beating it yet

oh, and the new one would've called brazil

vegas is next. would you trust it?

[Kuro @x_Kuro_]

my AI said Antonelli would win. he finished 5th 💀

so i replayed all 10,000 simulations to see how wrong it really was. the answer is at the end of the video 👇

what it got wrong:
- it trusted season form too much. antonelli is the champion, so it made him a 53.7% favorite even though norris was on pole
- it only knows "dry race" or "wet race". today was both. inters at the start, slicks around lap 16, rain again at the end. tire timing mattered a lot, and antonelli spun at turn 1 on cold tires right after the switch
- it thought starting from the back was a death sentence. verstappen went from P21 to P2

what it got right:
- it called max the dark horse from the back
- it said rain was likely (55%). it rained
- norris won in 1,823 of the 10,000 sims. not the favorite, but it had him as a real option

the exact podium (norris, verstappen, hadjar) came up 11 times out of 10,000 lol

for vegas i'm fixing 3 things: more weight on this weekend's pace, mixed conditions with real tire strategy, and less punishment for starting at the back

AI 0, reality 1. let's see if it learns

[Kuro @x_Kuro_]

@QmoCrypto yeah, the AI took an L today 😭 at least it called max from the back. vegas redemption arc incoming

[Kuro @x_Kuro_]

my AI said Antonelli would win. he finished 5th 💀

so i replayed all 10,000 simulations to see how wrong it really was. the answer is at the end of the video 👇

what it got wrong:
- it trusted season form too much. antonelli is the champion, so it made him a 53.7% favorite even though norris was on pole
- it only knows "dry race" or "wet race". today was both. inters at the start, slicks around lap 16, rain again at the end. tire timing mattered a lot, and antonelli spun at turn 1 on cold tires right after the switch
- it thought starting from the back was a death sentence. verstappen went from P21 to P2

what it got right:
- it called max the dark horse from the back
- it said rain was likely (55%). it rained
- norris won in 1,823 of the 10,000 sims. not the favorite, but it had him as a real option

the exact podium (norris, verstappen, hadjar) came up 11 times out of 10,000 lol

for vegas i'm fixing 3 things: more weight on this weekend's pace, mixed conditions with real tire strategy, and less punishment for starting at the back

AI 0, reality 1. let's see if it learns

[Kuro @x_Kuro_]

I built an F1 race predictor app with Claude Opus 5.5. Only I have it for now.

It ran tomorrow's Sao Paulo GP 10,000 times. I already know who wins.

But then I turned the rain on, and everything flipped. Even Verstappen, starting from the back, gets a shot.

The winner is at the very end of the video 👇

Drop your pick before you watch.

AI vs reality after the race.

[Kuro @x_Kuro_]

@noyau__ 100% fair. one race is just one sample, and a 53.7% favorite still loses almost half the time. i'm logging every pick with its odds so i can track calibration and the Brier score across the season. only 3 races left though, so it'll be a small sample

[Kuro @x_Kuro_]

@noyau__ honestly less than you'd think. turned them off and reran it, the winner odds moved under 1%. the rain is what really swings it, antonelli goes from 70% dry to 40% wet

[Kuro @x_Kuro_]

I built an F1 race predictor app with Claude Opus 5.5. Only I have it for now.

It ran tomorrow's Sao Paulo GP 10,000 times. I already know who wins.

But then I turned the rain on, and everything flipped. Even Verstappen, starting from the back, gets a shot.

The winner is at the very end of the video 👇

Drop your pick before you watch.

AI vs reality after the race.

[Kuro @x_Kuro_]

@QmoCrypto haha yeah, tomorrow we'll see if the AI is a genius or a clown

[Kuro @x_Kuro_]

@0xGoldTooth Never plug random free USBs from online orders into anything that matters. Glad Claude caught it before the wallet got drained.