// A Fucking Side Project

Just Fucking Use DeepSeek

A hedge fund manager's weekend hobby that destroyed $589B in market cap

2023
幻方创始人 (High-Flyer Founders) 3人
梁
Liang Wenfeng May 8, 2023 2:15 PM
ok so i started an AI company
陈
Chen Wei 2:16 PM
bro you literally run a $8B hedge fund
梁
Liang Wenfeng 2:16 PM
it's just a fucking side hustle
王
Wang Jun 2:17 PM
what's it called
梁
Liang Wenfeng 2:17 PM
DeepSeek
👍 1 🤷 1
2024
幻方创始人 (High-Flyer Founders) 3人
梁
Liang Wenfeng Dec 26, 2024 11:30 AM
we trained a model
陈
Chen Wei 11:31 AM
how much did it fucking cost
梁
Liang Wenfeng 11:31 AM
$5.6 million
王
Wang Jun 11:32 AM
OpenAI spent like $100M on GPT-4
梁
Liang Wenfeng 11:32 AM
sounds like a them problem
💀 2 🔥 2
陈
Chen Wei 11:33 AM
wait how many GPUs
梁
Liang Wenfeng 11:33 AM
2,048 H800s
王
Wang Jun 11:34 AM
bro Meta used 16,000 H100s
梁
Liang Wenfeng 11:34 AM
skill issue tbh
😭 3 🐐 2
January 2025
幻方创始人 (High-Flyer Founders) 3人
陈
Chen Wei Jan 27, 2025 9:15 AM
LIANG WHAT THE FUCK
梁
Liang Wenfeng 9:16 AM
what
王
Wang Jun 9:16 AM
NVIDIA JUST LOST $589 BILLION
梁
Liang Wenfeng 9:17 AM
oh that
陈
Chen Wei 9:17 AM
YOUR APP IS #1 IN THE US APP STORE
梁
Liang Wenfeng 9:18 AM
cool
💀 3
王
Wang Jun 9:19 AM
you beat ChatGPT. the app that's been #1 for like 2 years
梁
Liang Wenfeng 9:19 AM
neat
梁
Liang Wenfeng 9:19 AM
anyway gtg, Xi Jinping wants to meet
陈
Chen Wei 9:20 AM
WHAT
梁
Liang Wenfeng 9:20 AM
kinda busy though, training new model
😭 3 👑 2 🐐 2
📱 Liang Wenfeng has set status to Do Not Disturb
What's happening
Trending in Tech
Meanwhile, the typical anti-China vibe on X is going exactly how you'd expect...
142K posts
A
AIBro2024 @AIBro2024 · Jan 27
Yeah but DeepSeek is CENSORED. Can't even ask about Tiananmen Square. Checkmate China.
24 156
F
FreedomTechGuy @FreedomTechGuy · Jan 27
I would simply NOT use an AI that won't write Tank Man fan fiction. This is why America wins.
12 45
V
VCPartner @VCPartner · Jan 27
Sure DeepSeek is "efficient" but can it refuse to help me with my taxes in 47 different creative ways? Didn't think so.
89 312
I
InfraEngineer @InfraEngineer · Jan 27
DeepSeek just cost-optimized my entire career. I was the "AI infrastructure guy." Now I'm just "guy."
891 2.1K
R
ActualResearcher @ActualResearcher · Jan 27
Guys they just won gold at both the International Math Olympiad AND the Informatics Olympiad. Their model beats GPT-5. Maybe we should focus on—
67 234
A
AIBro2024 @AIBro2024 · Jan 27
CENSORED. MODELS. DON'T. COUNT.
56 23
R
ActualResearcher @ActualResearcher · Jan 27
bro stop with the AI slop replies you dumb fuck
2.1K 8.7K
C
CorporateEmployee @CorpEmployee · Jan 27
I tried to use DeepSeek but my company blocked it
12 34
R
ActualResearcher @ActualResearcher · Jan 27
What company?
8 12
C
CorporateEmployee @CorpEmployee · Jan 27
Microsoft
89 156
R
ActualResearcher @ActualResearcher · Jan 27
The company that spent $13B on OpenAI?
456 1.2K
C
CorporateEmployee @CorpEmployee · Jan 27
...yes
567 2.1K
S
SFFounder @SFFounder · Jan 27
Just raised $200M at $2B valuation for my AI wrapper startup. This DeepSeek news is... concerning.
234 567
R
NVDAHolder @NVDAHolder · Jan 27
My portfolio is down $50k but at least my AI has DEMOCRATIC VALUES
234 567
R
ActualResearcher @ActualResearcher · Jan 27
Their paper is open source. They literally showed their work.
123 456
S
SFFounder @SFFounder · Jan 27
But did they show it in a $50M/year office in SF???
789 2.3K
A
AIBro2024 @AIBro2024 · Jan 27
They just did distillation bro. Anyone could do that. It's not real research.
34 12
R
ActualResearcher @ActualResearcher · Jan 27
Then why didn't you?
1.2K 4.5K
S
SFFounder @SFFounder · Jan 27
Can someone explain what this MLA formula means?
DeepSeek @deepseek_ai
cKV = WDKV · h (latent compression)
d_c = 512, d_h = 4096 → 93% KV cache reduction
234 567
R
ActualResearcher @ActualResearcher · Jan 27
It means your Series B is cooked.
2.1K 8.9K
New Year's Eve 2025

Some people have hobbies.
Liang Wenfeng rethinks transformer architecture on New Year's fucking Eve.

$5.6M
Training Cost
-$589B
NVIDIA Loss
🥇🥇
IMO + IOI Gold
>GPT-5
Performance
📄
mHC: Manifold-Constrained Hyper-Connections arXiv:2512.24880 · Submitted December 31, 2025
x' = Ax + B · layer(Cx) A ∈ DSn×n (doubly stochastic manifold) Sinkhorn-Knopp projection → stable training at scale

Manifold-Constrained Hyper-Connections with Doubly Stochastic Matrices

They published this on New Year's Eve. While you were watching the ball drop, Liang was dropping doubly stochastic manifold constraints.

Just Fucking Use DeepSeek

It's free. It's open-source. It won gold medals. It dropped a paper on New Year's Eve. US labs are currently googling 'how to un-spend $500 billion.'

Use It While OpenAI Copes →

// Not affiliated with DeepSeek. This chat log is satirical. The $589B loss was real. The New Year's Eve paper drop was real.

PART OF THE JFU NETWORK
justfuckinguseclaudecode.com justfuckingusecopilot.com justfuckingusegemini.com justfuckingusenanobanana.com justfuckingusefigma.com justfuckinguseperplexity.com