minjune
@minjunesh
Followers
310
Following
899
Media
61
Statuses
733
Explore trending content on Musk Viewer
Halloween
• 2233044 Tweets
Mark Cuban
• 507795 Tweets
Thug
• 462152 Tweets
Sant Rampal Ji
• 219686 Tweets
Corinthians
• 152814 Tweets
Yuri
• 127890 Tweets
Raj Musix Kannada
• 125845 Tweets
Racing
• 122943 Tweets
Jets
• 105672 Tweets
Cazzu
• 85796 Tweets
Fate
• 84475 Tweets
Adams
• 73924 Tweets
Mariah
• 65320 Tweets
#GMMTVStarlympics2024
• 59483 Tweets
BABYMONSTER DRIP D-DAY
• 56169 Tweets
#WelcomeHomeJAYB
• 50117 Tweets
Nodal
• 47067 Tweets
Dame
• 41528 Tweets
Texans
• 41282 Tweets
あと2ヶ月
• 40845 Tweets
#BAEMON_DRIP_OUTNOW
• 40399 Tweets
Juanfer
• 37795 Tweets
#EliminaçãoAFazenda
• 29028 Tweets
Rodgers
• 24396 Tweets
Memphis
• 23731 Tweets
Garrett Wilson
• 23025 Tweets
Alito
• 22972 Tweets
ガンダムW
• 22358 Tweets
Garro
• 21792 Tweets
Giannis
• 18207 Tweets
Stroud
• 15545 Tweets
モンハン
• 15163 Tweets
Shah Mastana Ji Maharaj
• 13723 Tweets
Ramon Diaz
• 13722 Tweets
危険運転
• 12002 Tweets
Depay
• 10124 Tweets
results from running
@RekaAILabs
reka-vibe-eval on 8 smol (1.86-8B) vision language models
Top 3 on all
- Qwen-VL-Chat
- xtuner/LLAVA-llama3-8b
- qresearch/llama-3-vision-alpha
Top on hard
-
@qtnx_
research/llama-3-vision-alpha
2
6
43
I'm releasing Monolog, an app that transcribes and summarizes your voice recordings.
It's great for brainstorming and journaling.
Speak your thoughts, and let Monolog extract key points for you.
2
1
31
@brentsmith1024
@JohnnyNi13
@KeelVentures
credential gatinng is cringe, expected nothing less from hardvard bro
0
0
27
@PatrickJBlum
GptZero is a joke, openai itself stated it is impossible to distinguish ai from human
1
0
26
Jailbreak library for Llama3-8B & 70B coming soon
Jailbroken 8B by
@Nottlespike
is here:
2/ Activation Ablation Augmentation
Research team that identified how to nullify activation layers in Llama3 responsible for censorship
@song_minjune
@IanSears96
@Nottlespike
🥉 3rd place
5
18
249
2
0
17
are u kidding me claude? this is 1984 levels of information gating
3
0
17
> opens hw
> there's 1 problem
> it's "just implement MuZero"
> immediately begins hw
1
0
11
More on CMU’s Poker AI hackathon: unlike in perfect information games like chess or go where you reason over state spaces, in Poker you must reason over information states. An info state is your hand + game board + opponent actions + all possible opponent cards. The value of your
0
1
10
1
0
9
LLM steering via representation engineering (golden gate claude) is cool, but emotions don’t “pop” cuz they’re embedded in text. steering real time audio/video models will really hit, you can hear/see effect of ur steering instantly. plus, audio and video encapsulates text AND
Heads up comparison,
@ElevenLabsio
vs
@cartesia_ai
Cartesia almost completes before Elevenlabs starts lol
realtime SSMs are going to be -wild-
congrats
@krandiash
@_albertgu
19
22
288
0
0
10
@jheitzeb
@juberti
@AITinkerers
@FixieAI
correction: currently, it is speech to text not speech to speech
1
0
7
@Gradient_AI_
@huggingface
needle in the haystack tests mean little in practice, let’s see how it holds up against:
0
0
7
@BenRatkaj
space of all possible videos is pixels x audio x time. if right video, H(video) >> H(all else)
1
0
7
0
0
6
@powerbottomdad1
Wallace was a world architect. His descriptions are incredibly high res and eerily lifelike.
0
0
4
@siddrrsh
what are some long context evals that target real-world usefeleness that i can run this thru? dont even say haystack
2
0
5
@PradyuPrasad
from my convo w groq vp of eng: ~74 groq lpus to load 7B model into RAM. massivelly parallel “supercomputer” optimized for inference.
0
0
3
sleep pisses me off because it kills momentum. u need absolutely solid chunks of time to get shit done. i’d take a 72hr monster session of unbroken focus over a week of “balanced” work any week.
0
0
4
some days i wake up and ask myself if i’m full of shit. am i full of shit? i don’t know. i may be full of shit.
- my favorite prof, this guy can rly think
2
0
4
LLM libraries would be colossal
Llama3 8B would be ~1.6 million 1000 page books
for comparison the New York Public Library houses 7.5 million books
original GPT2 would be 300,000 books, not inconceivable that someone in the future will print this
The driving models are only ~50MB, so that's only ~10 1000page large books. A few layers per book!
4
0
87
0
0
5
@LukeFarritor
elite college students are trained to max the proxy of signal, not signal itself.
0
0
4
@teortaxesTex
he’s the final form of a gamer, he’s playing Civ IRL and taking it very seriously.
0
0
5
@itsandrewgao
@Tesla
India is the highest test of competence: if u optimzie FSD to work in India chances are it’ll work everywhere else.
0
0
3
Kearm is my CTO
Happy to report a Third Place finish at the Meta AI Hackathon! Thanks to all my teammates!
2
0
9
1
1
5
was excited to throw big neural nets at large state spaces for poker ai hackathon. one 1gb memory limit announcement later and everyone’s “AI” is a python dictionary.
0
0
4
@arjunkhemani
If there was no death, we’d still be bacteria. Forever unchanging, unadapting, frozen in time.
2
0
3
@ansonyuu
@socraticainfo
@AryanAfrouzi
@hud_zah
@b_ptrsn
@ElijahKurien
@BrookeaJoseph
@alexstratmoen
very looking forward to it
0
0
3
it's an incredible feeling making something from scratch. hit the raw html & css stack today
0
0
4
@RekaAILabs
@qtnx_
of course, evals are actually enjoyable to run when they are well designed!
0
0
3
@justinwangx
@freemanjiangg
@gazooksapp
@hud_zah
@edwardjxli
@usrbinishan
@taziksh
@_yashkarthik
yeah its good
0
0
4
@qtnx_
btw if you need linear probe vectors to lobotomize LLMs across all range of human psychospectrum to simulate average twitch chat, hmu
0
0
2
Phi 1.618B
"We are eternal, all this pain is an illusion."
This clip is from the music video for Parabola by Tool, often used as a meme or GIF.
The song is from the album 'Lateralus', in which many of its songs speak of the number Phi, which refers to the 'divine proportion,' 1.618.
140
82
514
0
0
4
@hubermanrules
Optimal schedule:
8 wake up
8:01 cold shower
8:05 coffee
8:10 begin work
the rest is easy
0
0
3
Just read "Surely You're Joking Mr. Feynman." Dude was a perfect mix of batshit crazy and genius. One thing that struck me was how extroverted his attention was - on the world, not on himself. He didn’t spend much time obsessing over his social reputation or his legacy or his
0
0
2
@jvboid
sounds exactly like my prof . a lot of cool new robotic demos are possible due to returns from moore's law - the underlying hardware remains much the same.
0
0
3
@DionysianAgent
bio/acc will be irrelevant when we switch substrates to silicon/other. the final form of mimetic agents will update neural/embodied architecture at electron speed, variation will be the mean. Hand Moravec wrote a lot on this
0
0
1
I was a voice memos power user.
But, I found myself rarely revisiting any of my recordings.
Simply, re-listening and re-comprehending was too much work.
I wanted something that made revisiting recordings effortless.
Monolog is the solution I built.
1
0
3
@prmshra
honorary mention to Sanger. he sequenced the base pairs of the X174 virus AND developed the tools to do it.
0
0
1
@3ameam
@yacineMTB
why would one do this? would the same reasons apply to writing about open source ai in public?
1
0
2
@R0b0tSp1der
@personofswag
its very well documented and follows pytorch standards. expected nothing less from an Apple team
0
0
2