Noxey

🎨 Graphic Designer | 💻 CS Student | 🌐 Exploring Web3 | Building, learning & creating.

[Noxey @FarukeyOnchain]

GM CT

It's the weekend, Are you touching grass today?
Do well to touch quoted Tek

GM to you

[Farukey〆 @Web3_farukey]

GM Web3 fam ☀️💙
No dey just build your bags, build your body too. 🏋️‍♂️
Let’s hit the gym and get fit! 💪🚀

[Farukey〆 @Web3_farukey]

Just logged into X and saw this 😭

3K FOLLOWERS! 🥹💙

We’ve officially crossed another milestone on this Web3 journey.
From zero to 3K — the community keeps growing, the network keeps expanding, and we’re only getting started. 🚀

Big love to everyone who followed, supported, engaged, and believed in the journey.

3K today. Bigger numbers tomorrow. 🫂🌐

#Web3 #3K #CryptoCommunity #BuildInPublic

[Farukey〆 @Web3_farukey]

🇳🇬 Happy Independence Day & Happy New Month! 🎉
New month, new energy, new opportunities. 🚀🌐
Keep building. WAGMI! 💚

[Noxey @FarukeyOnchain]

@Web3_farukey Up and grateful

[Farukey〆 @Web3_farukey]

Good morning, Web3 fam ☀️🌍

Let’s put God first, trust His timing, and move with purpose today. 🙏🏽

New day. New opportunities. Same God.
Stay blessed, stay building, and keep believing. 🚀

GM & God bless! 💙

[Noxey @FarukeyOnchain]

@Web3_farukey Stressful

[Mvis the great @Mvisthegreat]

Mother is a blessing no one can replace that.

[Biggest Willy @Will4Money]

Good evening everyone

[Biggest Willy @Will4Money]

@w_willmoney Will make that Chelsea Blue 💙🔵 badge a reality for you.

[Noxey @FarukeyOnchain]

@Ada_grams @ej_onx I have engage on al pot pull me up

[P_hexx🌬️ @P_hexx]

𝗧𝗛𝗘 𝗖𝗢𝗔𝗟𝗜𝗧𝗜𝗢𝗡 𝗝𝗨𝗦𝗧 𝗚𝗢𝗧 𝗣𝗔𝗜𝗗 🦍💰

➢ $60 USDC is up for grabs for early Coalition members.

➥ 3 random ID numbers.
➥ $20 USDC to each winner.
➥ No complicated hoops.

𝗛𝗢𝗪 𝗧𝗢 𝗘𝗡𝗧𝗘𝗥
➢ Claim your FREE membership ID at https://t.co/oHFUZtrWcN
➢ Post your membership card on X (make a fresh post, do not reply to this post)
➢ Tag @Apeshit_co

⇨ Giveaway ends Sunday, Sept 27 at 11:59 PM Pacific.

Three IDs get picked at random.

Winners announced Monday in Telegram.

𝗚𝗘𝗧 YOUR ID. 𝗝𝗢𝗜𝗡 THE COALITION. 🦍

[Noxey @FarukeyOnchain]

@P_hexx Gm

[Noxey @FarukeyOnchain]

@when_tge @vangrid_io Gm

[𝙬𝙚𝙣🔅 @when_tge]

On @vangrid_io, reconstruction usually takes 30 minutes or more after a bounty is accepted.

That work is already covered by the bounty, so there is no second payment.

The requester gets a textured 3D mesh once it finishes.

Live per-stage progress in the account is listed as coming later.

Submit before the deadline, because a bounty can only be accepted while it is still open.

Android app is live if you want to record. You’ll need an Invite code to upload

[Noxey @FarukeyOnchain]

@Ada_grams Don tree engaging

[Noxey @FarukeyOnchain]

@Ada_grams Pull up

https://t.co/DrTDoHCchT

[Noxey @FarukeyOnchain]

GM CT

It's the weekend, Are you touching grass today?
Do well to touch quoted Tek

GM to you

[Noxey @FarukeyOnchain]

@Jife790 @HalbornSecurity @1winToken @1win Gm

[Noxey @FarukeyOnchain]

@Mr_Hy4 @CreaoAI @FD_XYZ Gm

[Noxey @FarukeyOnchain]

@ChuzzyTech @axisrobotics Gm

[Cleopatra @ChuzzyTech]

Axis ran an experiment.

Teach a robot to place tofu on a plate using human corrections.

660 people jumped in to fix the robot when it messed up.

The obvious assumption is more corrections means a better robot.

That turned out to be wrong. Most of that data was junk.

People pause, hesitate, or overcorrect, and none of that is worth teaching a robot to copy.

So @axisrobotics built a filter. Did the fix actually work when replayed. That dropped 660 down to 496.

Was the robot really stuck and still stuck the moment the human stepped in. That dropped it to 416.

Then they stripped everything down to just the short snippet, less than a second, that flips failure into success, then let the robot finish the rest on its own. That left 161 out of the original 660.

Here is why this is important. Training on the full human correction traces made the robot worse, dropping 16 successes.

Training on just those 161 verified short snippets made it dramatically better, adding 40 successes and it kept improving as more verified snippets were added.

So the lesson is not more human help is better.

It is that a tiny, precise fix at the exact failure point teaches more than a human doing the whole recovery for the robot.

Post training correction works like surgery, not like replaying the entire demo.

Does this change how you think about what actually counts as good training data?