I wasnt expecting this much gifts even this time it was in only a cafe but still they got almost 10-15 gifts for each geust and winners (btw i didnt win 😭🙈)
@hacktoberfest @AwsKarachiUG #hacktoberfest2026
@hacktoberfest @AwsKarachiUG #hacktoberfest2026
0
0
0
3
I just earned my Hacktoberfest 2026 sticker pack! #Hacktoberfest hacktoberfest.com
0
0
0
4
i'm sabih, 15, from karachi. still in o-levels, already working as an ai engineer.
i'm trying to build a proper work setup for my career and i'm asking tech companies for help. even one item would mean a lot
i'm trying to build a proper work setup for my career and i'm asking tech companies for help. even one item would mean a lot
0
0
0
25
Palestine Chronicle LIVE BLOG - Palestine Chronicle palestinechronicle.com/blog/
0
0
0
18
⚽ My GitHub season as a player card: BRONZE tier, 60 OVR - “The Engine”
Check it out: githubfc.com/Sabih-sourcee
Check it out: githubfc.com/Sabih-sourcee
0
0
0
21
python 3.15 tip for ml scripts:
lazy import torch
torch only loads the first time your code touches it. so `python train.py --help` stops waiting on it.
catch: a broken install now fails at first use, not at the top.
worth it for cli tools?
lazy import torch
torch only loads the first time your code touches it. so `python train.py --help` stops waiting on it.
catch: a broken install now fails at first use, not at the top.
worth it for cli tools?
0
0
0
16
python 3.15 just shipped a built-in frozendict
one bug it helps with:
CONFIG = {"lr": 3e-4}
def train(cfg=CONFIG):
cfg["lr"] *= 0.1
call train() twice and your "default" lr is 3e-6
frozendict raises instead of letting it slide
what would you freeze first?
one bug it helps with:
CONFIG = {"lr": 3e-4}
def train(cfg=CONFIG):
cfg["lr"] *= 0.1
call train() twice and your "default" lr is 3e-6
frozendict raises instead of letting it slide
what would you freeze first?
0
0
0
13
new stanford study: companies using ai hired more seniors, fewer juniors.
simple way to see it:
ai = a fast intern
senior = knows when the intern is wrong
junior = used to be the intern
so if you're learning ml right now, how do you skip the intern stage?
simple way to see it:
ai = a fast intern
senior = knows when the intern is wrong
junior = used to be the intern
so if you're learning ml right now, how do you skip the intern stage?
0
0
0
12
routing apis all return confidence scores now. cheat sheet (save this):
confidence: what the model says
accuracy: what really happens
calibrated: 0.8 is right ~80% of the time
threshold: below it, human checks
ece: avg gap between them
a wrong 0.99 hurts more than a 0.6
confidence: what the model says
accuracy: what really happens
calibrated: 0.8 is right ~80% of the time
threshold: below it, human checks
ece: avg gap between them
a wrong 0.99 hurts more than a 0.6
0
0
0
8
semianalysis says pretraining dropped to about 7% of lab compute. post-training and rl eat most of it now
so if youre learning ml today, is building a tiny gpt from scratch still worth it, or skip straight to finetuning and rl on small models?
so if youre learning ml today, is building a tiny gpt from scratch still worth it, or skip straight to finetuning and rl on small models?
0
0
0
11
ml tip for the ai-math-papers week:
a metric is not a proof.
if you cannot re-run the eval from a pinned split + pinned seed and get the same number, you do not have a result. you have a screenshot.
what do you pin before you trust a score?
a metric is not a proof.
if you cannot re-run the eval from a pinned split + pinned seed and get the same number, you do not have a result. you have a screenshot.
what do you pin before you trust a score?
0
0
0
12
the silent ML bug that looks like a great model:
training data collected one way.
prod data arrives another way.
same feature names. different distributions.
offline metrics lie.
what do you check first when prod drops but offline looks fine?
training data collected one way.
prod data arrives another way.
same feature names. different distributions.
offline metrics lie.
what do you check first when prod drops but offline looks fine?
0
0
0
10
mental model for long context decoding: each new token is one query against the whole kv cache. so the gpu mostly waits on memory reads. flash-decoding splits the cache into chunks so more SMs read at once. where does that stop paying off?
0
0
0
12
hot take: if the same agent writes the code and the tests, green tests mean almost nothing. it just checked its own assumptions twice. write the test cases yourself, even rough ones, and let the agent make them pass
0
0
0
12
one python bug that shows up a lot in vibe coded apps:
def add(x, xs=[]):
xs.append(x)
return xs
the [] is made once, so every call shares it. fix:
def add(x, xs=None):
xs = [] if xs is None else xs
ruff rule B006 flags it
def add(x, xs=[]):
xs.append(x)
return xs
the [] is made once, so every call shares it. fix:
def add(x, xs=None):
xs = [] if xs is None else xs
ruff rule B006 flags it
0
0
0
13
been thinking about why models crush coding and math benchmarks but still feel shaky on real work. benchmarks hand you a clean spec and a grader. real tasks make you figure out what done even means first. is anyone actually measuring that part, or are we all just vibing it?
0
0
0
7
unsloth says a 0.8b qwen went from 20.7% to 74.3% on decision benchmarks on 4gb vram. cool if it holds.
first thing i'd check on a jump like that: does the train data look like the test set? near-dup prompts inflate scores fast.
how do you check?
first thing i'd check on a jump like that: does the train data look like the test set? near-dup prompts inflate scores fast.
how do you check?
0
0
0
4
apple ml research put one coding agent with plain shell access up against 4 fancy multi-agent harnesses on mle-bench. the simple one won, 62.5% vs 47.1% medal rate. before building a 6 agent swarm, try one strong model and a terminal. what's your setup?
0
0
0
5