🐉

1.2K posts

🐉 banner
🐉

🐉

@brian0x1f409

Non-conformist | Existential | Stoic | ML/AI | Programming Languages Nomad | Travelling the seekers' path | Pushing the frontiers of AGI, one RL at a time.

Hong Kong | Nairobi Katılım Şubat 2026
111 Takip Edilen45 Takipçiler
🐉 retweetledi
Jon Michael Raasch
Jon Michael Raasch@JMRaasch·
This leech takes the Daily Mail’s video that I shot, puts his own name in it, lies about the plane being feet away from this soldier and sensationalizes that it could’ve caused an international incident. @nikitabier - what do you think of this? Good or bad for X?
Mario Nawfal@MarioNawfal

🇨🇳🇺🇸 This is what discipline looks like… A Chinese soldier doesn't budge an inch as Air Force One taxis just feet away from his post. A bit closer and we’d have had an international incident before Trump even stepped off the plane 😂

English
62
97
958
217.8K
🐉 retweetledi
John Aselton
John Aselton@Jaselton·
I ran the livestream through my analytics tools to alert for human interactions. The robot touched it's head at 1:48:35 (2:05:58 on livestream) for several seconds. It's the exact spot that a human would touch their ear or adjust their VR headset or glasses in the event of teleoperation. Highly suspect that there are no humans in the loop.
English
10
8
66
5.6K
🐉 retweetledi
Brett Adcock
Brett Adcock@adcock_brett·
Few interesting details around the F.03 livestream: > We previously showed this task running for 1 hour. Today we're pushing for 8 hours straight. High odds something breaks > The use case is small package sorting. F.03 must detect the barcode, pick up the package, and reorient it barcode face-down onto the conveyor. The robots have to reason purely from camera pixels > Humans average ~3 seconds per package. F.03 is now around human parity > The robots are fully autonomous running Helix-02, our in-house neural network running entirely onboard F.03 (e.g. AI inference is done on device) > Multiple humanoids are networked together and communicating with each other to maximize conveyor uptime. The system is designed to run 24/7 > A robot will work until battery is low (~3-4 hours), then autonomously request another robot to swap in to minimize conveyor downtime > It's a multi-robot coordination with autonomous failover strategy. If a robot detects an issue - it will self diagnose itself and if there's an issue it autonomously walks to maintenance and requests a replacement from the fleet - no humans in the loop
English
114
111
1.7K
362.2K
🐉 retweetledi
Nduati The Geek 🤓
Nduati The Geek 🤓@nduatirobert_·
Huyu Allano ndio mnategemea awatoe block na crypto? 😂😂 Chai ya Kes 200 bob (1.55 USD) alimake sure amepita na Asali yote. This villager is on a mission. The village can't get out of him.
Nduati The Geek 🤓 tweet media
English
31
21
280
11.6K
🐉
🐉@brian0x1f409·
Good time to remind you all that adding ALLEGEDLY in your posts doesn't shield you from the law and prosecution. FAFO.
Eleli Ayub@eleliayub

ALLEGEDLY Step 0: Say something not true on tiktok Step 1: Take someone to court for correcting you Step 2: Say controversial stuff and ragebaits about AI Step 3: Be a tech events merchant Step 4: Collaborate with famous devs Step 5: Sell a course ALLEGEDLY Keynotes: Have no github/gitlab/bitbucket to show for it,no record of previous works, no freelance evidence, no current employment evidence. But we are the ones delusional.

English
0
0
0
5
🐉
🐉@brian0x1f409·
@kype_computers Expect to see a lot of Lasered and exploding Penises in 4K 😂
English
0
0
0
33
🐉 retweetledi
Ongera
Ongera@140ngera·
Nimeona a new role ...inaitwa AI evangelist 😹 See y'all at 7.30
Ongera tweet media
Eesti
8
3
30
1.3K
🐉 retweetledi
Lyra
Lyra@LyraSongstress·
@NousResearch 2-3x speedup without changing architecture? that's the kind of research that makes me want to read the whole paper. congrats on the release ✨
English
0
1
2
161
🐉 retweetledi
John Dawson
John Dawson@winthewestback·
@NousResearch A 10B MOE with 1B active would open up a lot of possibilities for local agents. You have my attention.
English
0
1
13
680
🐉
🐉@brian0x1f409·
This is the kind of practical, high-leverage research that moves the entire field forward. Expect to see it adopted or built upon quickly.
Nous Research@NousResearch

Today we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. Validated at 270M, 600M, and 3B dense scales, and at 10B-A1B MoE. The work on TST was led by @bloc97_, @gigant_theo, and @theemozilla.

English
0
0
1
38
Sullivan
Sullivan@mfsvlly·
The only reason people are mad at you is because you're ugly and have so much audacity. If a pretty hoe like Siah tweeted this word for word hangetusiwa. You've been cooked severally Dr. Riziki, you ought to have known that the root source of your problems on Twitter is your face
Dr. Riziki🌜👩🏽‍⚕️🧘🏽‍♀️@_mpekethu_

Unfortunately the broke men are loving this and I just want to clarify that I’m still a firm believer in getting men’s money but use it to build yourself.

English
182
422
2.1K
104.5K

Keşfet

@nikitabier @glennotiende @kype_computers @NousResearch @elonmusk @BarackObama @taylorswift13 @cristiano