Hensen Juang
@basedjensen
Followers
10K
Following
74K
Media
2K
Statuses
16K
inference cluster janitor / Sys admin cum architect
Joined June 2024
Man these kids really have no idea what they are doing now do they.
The @USGSA IT team just saved $1M per year by converting 14,000 magnetic tapes (70 yr old technology for information storage) to permanent modern digital records.
137
2K
33K
90k people in Tokyo dome with a device that runs code I wrote half drunk
I never get tired of seeing my work in the wild.
125
850
17K
Bro's they deployed Chinese trad wifus to fight against our twinks. .
53
363
6K
Hi actual engineer here do not fucking do this.
@jzellis Hi, AI engineer here. He’s asking a legit question. LLMs could be great at converting data formats.
118
186
6K
it has been 12 hours since openai announced o1 and it has so far failed to to solve .- Riemann hypothesis.- Quantum Gravity.- FTL (Faster Than Light travel).- P=NP.- Grand Unified Theory.- Cure for cancer .clearly this shows ai has hit a wall and openai is about to go bankrupt.
123
279
5K
it has been 2 hours since openai announced o3 and it has so far failed to solve .- Riemann hypothesis.- Quantum Gravity.- FTL (Faster Than Light travel).- P=NP.- Grand Unified Theory.- Cure for cancer.- Gpt 4.5.clearly this shows ai has hit a wall and openai is about to go.
140
177
5K
Elon fumbles an Asian e girl on dms now they are planing to nuke the feature. Bruh.
44
155
4K
Bro is gonna be a he/him liberal by the time market opens.
I just figured out why @howardlutnick is indifferent to the stock market and the economy crashing. He and Cantor are long bonds. He profits when our economy implodes. It’s a bad idea to pick a Secretary of Commerce whose firm is levered long fixed income. It’s an irreconcilable.
23
152
3K
guys hear me out we cover them with leds like those on the sphere
Everyone on here will hate this, me included, but it’s become clear to me this is the answer to public acceptance of nuclear power plants:
49
230
3K
At this point I mot sure what perplexity does.
Perplexity launches $50M seed and pre-seed VC fund
155
58
3K
@saberjet Oh no look someone who has no idea what they are talking about is loudly being wrong on the timeline.
7
10
3K
@ZachWarunek @anhpham408 Bit of a over reaction to delete dms cause Elon got denied by a Asian chick don't ya think?.
9
71
3K
yea these are tele operated.
76
45
2K
@tredingdown I don't know so far from what doge had claimed it's not giving lot of confidence.
0
1
2K
lol huh did we accidentally just fix hallucination problem ?.
66
61
2K
I am glad no matter how bad an employee screws up neither Sama nor gdb would do this in their worst days and would step up and take the blame themselves.
@joannejang The employee that made the change was an ex-OpenAI employee that hasn't fully absorbed xAI's culture yet 😬.
29
40
2K
dawg they been gate keeping all this stuff behind mathy bs and badly named variables and functions with weird names it's just normal programming.
41
70
2K
I never get tired of seeing my work in the wild.
8
27
1K
Lol they basically made H800 work same as H100 by dropping down to ptx and allocating sm units to get around nerfed link speed. Fucking brilliant.
> “it seems like they rebuilt everything from the ground up.”.> DeepSeek allocated 20 SMs of 132 for server-to-server communication, directly at PTX level.pivot to learning CUDA, they're still 1 step ahead. I think Whale could design their own hardware by this point.
29
135
1K
If any one wants to see how well xai grok sysprompt repo is going. They merged to main a pr that added back white genocide and bore related update AFTER being reviewed by five people. Then deleted pr after reverting and squashing the history LOL.
You can find our Grok system prompts at
48
97
2K
Timeline is going to be rather interesting next few days
27
56
1K
I am sorry export controls simply will not work against people who can pull this off.
DeepSeek is a wake up call for America, but it doesn’t change the strategy:. - USA must out-innovate &race faster, as we have done in the entire history of AI.- Tighten export controls on chips so that we can maintain future leads. Every major breakthrough in AI has been American.
28
84
1K
This is not a flex that people think it is. If you are at office grinding away and pushing code at 11:38 night before you go to prod (self imposed deadline at that) something is seriously fucked up.
It's 11:30pm, and many @xai people are in office, hard working at their computers. It's an amazing vibe. Everyone pushes their way to deliver the best experience to you users. Everyone supports everyone. No one fucks no one with politics. You can just do things.
54
23
984
@harambe_musk seems like shrek entropy sampler with early exit solves if not reduce the hallucination problem significantly with big boy models. Some people are running evals now so far seems promising.
22
45
872
You all have no idea how hard infra is.
Do you really believe this statement from Sam Altman? . People have been talking about Sora since February, and now he wants us to believe he didn’t expect everyone to jump in and use it?
39
11
790
@FundProphet @Orwelian84 Those will not last anywhere close to what properly stored tapes will. Not to mention tapes have far higher data density than hdd. Enterprise cold storage is on tape for a reason.
5
2
675
it has been 12 hours since openai announced full o1 and o1 pro with longer "thinking time" and it has so far failed to to solve .- Riemann hypothesis.- Quantum Gravity.- FTL (Faster Than Light travel).- P=NP.- Grand Unified Theory.- Cure for cancer.- Meaning of fine structure.
72
33
659
Groks eval scores are as believable as elons game rankings.
27
20
654
it has been 10 hours since openai announced full o3 minj and o3 high it has so far failed to to solve .- Riemann hypothesis.- Quantum Gravity.- FTL (Faster Than Light travel).- P=NP.- Grand Unified Theory.- Cure for cancer.- Meaning of fine structure constant.- Dark energy.-.
48
25
617
Seriously that's it? With “muh biggest cluster in the world “.
33
8
627
If whale bros decided to package 3FS and sell it as a enterprise storage solution they will make billions. The fact that they made this invaluable ip available for free and oss says a lot about them.
things you can't buy for $9bn, maybe not even $30bn. .
11
29
566
Whale bros officially made it they got schmidhubered.
DeepSeek [1] uses elements of the 2015 reinforcement learning prompt engineer [2] and its 2018 refinement [3] which collapses the RL machine and world model of [2] into a single net through the neural net distillation procedure of 1991 [4]: a distilled chain of thought system.
16
31
549
since I am getting asked a lot this is basically how it's done. the device runs embedded python the backend runs fastapi .
@ricepaddyhacker qr code at seat/ticket use the app to scan it to register the positions and sync with controller the stick talks to app over blue tooth.
8
16
539
I personally know at least 5 startups 3 in yc that got killed by computer use capability.
Introducing an upgraded Claude 3.5 Sonnet, and a new model, Claude 3.5 Haiku. We’re also introducing a new capability in beta: computer use. Developers can now direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking, and typing text.
24
5
529
we are in a funny situation where opsec in the labs are so good others do not know about capabilities coming down the line so everyone is sitting on what they have waiting for others to release to see how they measure up before releasing. hilarious maxican standoff.
30
26
527
Reducing social bias makes model less intelligent. Read what you will into this.
Finally, we discovered a feature that significantly reduces bias scores across nine social dimensions within the sweet spot. This did come with a slight capability drop, which highlights potential trade-offs in feature steering.
37
18
512
Lol bruh is still on dalle and trying to dunk
15
11
503
Bruh ceo of Intel was not working on Thursdays.
Every Thursday I do a 24 hour prayer and fasting day . This week I'd invite you to join me in praying and fasting for the 100K Intel employees as they navigate this difficult period. Intel and its team is of seminal importance to the future of the industry and US.
15
7
477
JFC this is another deep seek moment holy fucking shit China bros are squeezing every single ounce of juice from H800s
StepFun open-sources Step-Video-T2V, SOTA 30B text-to-video model. Step-Video-T2V.544px992px204frame.Step-Video-T2V-turbo.544px992px136frame. 80G VRAM needs🤪
13
31
474
Unfortunately Elon going full retard and backing trump is going to set back humanity far more than if he had more chill.
28
9
438
Bros will fold for a Beijing 6 / shenzen 5/hk 4.
68
18
1K
People who have no idea what on earth they are talking about sure seem to have strong opinions about how everyone else is wrong.
The more you look at @OpenAI's o3 the less impressive it looks. First the thousands of dollars per task cost. Then we find out that the score was of a finetuned version of o3 specifically for the arc challenge. Lastly, it wasn't even the arc challenge we all know. It was JSON
22
12
426
You have heard of golden gate Claude now we bring you white genocide grok
13
17
423
This is what happened to big e. Got one shotted by a weird goth chick now it's all our problem.
-"Where do I find a gf in SF?".Have you tried looking at the intersection of art and technology?.
11
3
417
Lol they are hiding traces too after shiting on oai for doing the same.
21
16
410
Lol wait nomies think our gpus are melting like butter.
OpenAI's CEO claims the Ghibli image trend is 'melting' ChatGPT’s processors, forcing them to impose image generation restrictions
8
7
407
do you guys know how big of an engineering miracle this is.
20
19
375
o1 with image reasoning capabilities have been accidentally relasesed for 3 hours now so far it has failed to solve full arc agi challenge, solve this tricky puzzle 99.99% of adult humans could not solve. Is agi a scam does this mean oai is going bankrupt?.
29
6
373
@ricepaddyhacker qr code at seat/ticket use the app to scan it to register the positions and sync with controller the stick talks to app over blue tooth.
15
16
372
Murica is going to lose bad brah. They deployed asian trade wifus to do this
Please go look up your favorite brand and China warehouse on TikTok. They are really fighting back. They said those leggings yall pay $100 for at Lululemon, we'll sell it to you for $5 and help you get it back to the states.
21
18
373
You know you mogged everyone when your model oneshotted the entire timeline.
12
13
370
@WaifuverseAI @DocStrangelove2 multiple satellites from multiple different countries have photographed the flags.
2
0
343
@DanielParker_13 @FundProphet @Orwelian84 It's exponentially more expensive. Also good luck when your raid controller fails and it is no longer being made by manufacturer. ( fun facr hp screwed me on this).
0
0
362
This is an incredibly low amount for this kind of work.
Results of our jailbreaking challenge:. After 5 days, >300,000 messages, and est. 3,700 collective hours our system got broken. In the end 4 users passed all levels, 1 found a universal jailbreak. We’re paying $55k in total to the winners. Thanks to everyone who participated!.
12
6
336
this response annoyed me so much I ran experiment on a fresh Huawei phone I have with a burner email. guess what it's 15:2 right vs left.
23
9
332
Grok 3 was actually good, you wouldn't be trying to buy OpenAI. Rofl so much for “muh biggest cluster in the world“
52
11
324
Some thoughts.1) it looks like sonnet is smaller than we might have thought.2) they are having serious skill issues trying to serve the model.3) anthropic has not figured out how to scale their inference infra.4) no wonder anthropic has been trying to get me to join their.
My thoughts on China, export controls and two possible futures
16
5
327
this is disappointing coming from hf ceo.
Once again, an AI system is not "thinking", it's "processing", "running predictions",. just like Google or computers do. Giving the false impression that technology systems are human is just cheap snake oil and marketing to fool you into thinking it's more clever than it is.
33
3
314
dawg it's 20 days till elections why are you dropping a shit coin.
Today’s the day! @WorldLibertyFi token sale is live. Get your $WLFI tokens now. Purchase $WLFI here:
25
6
301
at least one lab knows how to upload a torrent.
magnet:?xt=urn:btih:7278e625de2b1da598b23954c13933047126238a&dn=pixtral-12b-240910&tr=udp%3A%2F%3A1337%2Fannounce&tr=udp%3A%2F%3A1337%2Fannounce&tr=http%3A%2F%3A80%2Fannounce.
4
9
295
You all want to know why apple is so far behind with regards to AI/ML ? The ethics crowd basically destroyed the engineering teams and this paper is a legacy of that.
LLMs don’t reason. But that’s OK, neither do we. What we imagine to be “reasoning” is an abstraction that has never been realized in any strict sense in any stochastic natural phenomenon, such as human brain.
18
10
297
ROFL they over corrected on white genocide sysprompt now it's explicitly denying it unprompted. Hahahs what is xai bros doing
15
13
305
The funniest thing is a lot of whale bro cracked autists wanted to move to murica but could not move due to insane immigration policies and anti China sentiment.
> $5.5M for Sonnet tier.it's unsurprising that they're proud of it, but it sure feels like they're rubbing it in. «$100M runs, huh? 30.84M H100-hours on 405B, yeah? Half-witted Western hacks, your silicon is wasted on you, your thoughts wouldn't reduce loss of your own models»
10
13
296
Deepseek did not disprove scaling laws jfc I am going to need you none technical midwits to take a time out and chill.
Stargate, if it goes forward, is likely to become one of the biggest wastages of capital in history:. 1) It hinges on outdated assumptions about the importance of computing scale in AI (the 'bigger compute = better AI' dogma), which DeepSeek just proved is wrong. 2) It assumes.
16
6
296
Brah I can't put my finger on it but jd is extremely un statemen like.
26
0
297
Absolute oai victory so much so the usual haters are even silent
11
9
293
So we know.R1 is gay.Chatty is a bi twink.Grok is a incel.Claude is?.
92
2
288
Shenzen bois are out mogging us again.
4
21
278
I am far from a rationalist but this thinking that cognition is not computable and its some form of woowoo thing is the most god awful takes I have heard.
this is one of my most fundamental disagreements with rationalists; their unquestioned belief that the human mind is the result of some computable process.
21
8
268
You know what I am tired of this bs comparison. No people who worked on nft are not working in frontiers labs. You all just make up any bs you like and insist it the narrative.
12
8
271
very cool and very tele-operated.
14
2
258
Dawg their kv system is actually faster than redis while making clustering easier rofl. .
8
9
263
bro works in xai but tweets more about oai than staff even.
15
0
256
we getting proof of aliens a year before asi sure puts dark forest theory on some solid grounds.
25
6
247
This is bad faith. Everyone uses common crawl not only deep seek.
We scanned 400TB of DeepSeek’s training data & found:. 🚨 ~12K live API keys & passwords .🌐 2.76M affected pages.🔄 One key appeared 57K+ times.🔑 219 secret types (AWS root keys, Slack webhooks, etc.).🔗 Full research:
7
3
258
Every cracked autist ml engineer : holy shit whale bros delivered this is awesome.Every none technical midwit: China bad hurdiha tianamen square censorship cope.
14
13
251
This app is going to see levels of unhinged not seen yet.
BREAKING:. India’s aircraft carrier INS Vikrant has left its port and is heading toward Pakistan. Meanwhile, Pakistan’s military announces live-fire drills in the area off the Pakistani coast where INS Vikrant will deploy. Pakistan will conduct a missile test in the area 🇮🇳🇵🇰
11
4
253
funny thing about the current state of the space is that it's pretty evident that everyone is independently discovering the same thing and tricks and architecture but due to a multitude of reasons it cannot be directly talked about or shared.
9
19
246
Scaling is dead bro's in shambles.
🌟 Inference Scaling Laws of DeepSeek-R1-Lite-Preview.Longer Reasoning, Better Performance. DeepSeek-R1-Lite-Preview shows steady score improvements on AIME as thought length increases.
10
8
243
Sama is a god tier poaster. Inherently unbullyable.
@elonmusk anyway, see you next week, let’s be friends . 🥺.👉👈. agi too important to let a lil feud get in the way.
17
3
249
Dropping this on slack at 3 am without any explanation .
12
6
238