dangerously addicted to kaggle, agents & post-training | more compute pls
Joined February 2025
- Tweets340
- Following232
- Followers381
- Likes2.1K
Pinned Tweet
Hardware build is finally done:
Threadripper PRO 9965WX
2x RTX PRO 6000 Blackwell Max-Q
192GB VRAM
192GB DDR5-6400 ECC
4TB WD_Black SN8100
ASUS WRX90E-SAGE
1650W PSU
This thing is a beast. Very excited.
I’ve been teaming with people on kaggle more and I’ve never felt better. I have more energy. My skin is clearer. My eye sight has improved.
chiefkeith retweeted
The ARC Daily Digest · Special edition
Team Turducken takes flight into third on ARC-AGI-3, up from fifth after an almost four-point gain.
The whole crew is airborne.
arc.markbarney.net/news#2026…
ALT AI-generated editorial fantasy: all seven approved Team Turducken crew characters and exactly two Yorkies fly a mechanical turkey containing a chicken containing a sunglasses-wearing duck above a moonlit snowy canyon. Mark, Son, black-capped yellow Keith, blonde DZ with pink bunny, crying banana-kitten NT, MRI-display Choryu and silver/cobalt Herminator are visible. Panel: Yi-Chia Chen second, 67.49; Team Turducken third, 52.82; October 18, 2026, 10:56 PM EDT. ARC Daily Digest branding.
Agents make it possible to compete in many kaggle competitions at once. I'm currently in 4, but can only make real progress in 2 because I'm running out of tokens and GPU compute. The bottleneck is no longer how much work I can do. It's how much compute I can afford.
poteto just gave me ultra + $220 in credits 😭
alright, straight into the kaggle mines. let’s see if grok bot can cook
dev day got crushed by its own hype.
rule #1 is underpromise and overdeliver.
doesn’t help that opus 5.5 is owning them right now either.
cloud agents were the one thing that actually got me excited. everything else felt pretty underwhelming.
ARC-AGI-3 isn’t saturated. Its public set is.
Most 90-100% scores come from tuned harnesses on inspectable games. Opus 5 scores ~30%. The best open model that fits on 1x RTX Pro 6000 is still under 4% on hidden games.
Show me 100% on the private set.
My AI agent used all my kaggle GPU quota trying to debug a notebook and now I can’t submit anything new for the rest of the week ☹️
Seems I overreacted to some leaderboard variance. Hopefully this holds over the next 2 weeks. I really thought I had a chance at gold but I think I chose too big of a model without enough compute
I thought I had silver locked in for kaggle pokemon and was pushing for gold. Not even sure I’m gonna get bronze now 😂
Pokemon kaggle getting crazy. Your rank will drop fast if you don’t keep improving. I expect it to only get crazier
Few days remaining of kaggle pokemon competition. Should be fun. Hopefully I can make a big jump at the end like I did in ROGII. I think the last day will be crazy with everyone’s best agents being submitted. I’ve seen some GMs hiding their best with dummy submissions recently
I am pretty surprised how many simple coding mistakes the top models still make. especially with all the unnecessary smoke tests they run