eternal singularity

*unpops your popped bubble*

DeepSeek also wrote some PTX (NVIDIA’s intermediate assembly language). Low-level GPU programming is the way to go, folks. The more you optimize, the more you reduce costs or increase your performance budget for further advancements elsewhere at no additional cost

game recognizes game

release the nsfw sora model

they deployed Chinese trade waifus to fight against our twinks

semiconductor fund managers after seeing 4 memes about deepseek

Jevons Paradox Strikes Chip Stocks Overnight After DeepSeek Takes First Place in Stores

The release of DeepSeek sent ripples through the chip market, as engineers proved that optimizing code could maximize GPU efficiency without increasing hardware demand. The result? A major hit to chip stocks:

📉 Arm ($ARM): -5.5%
📉 Nvidia ($NVDA): -5.3%
📉 Broadcom ($AVGO): -4.9%
📉 Super Micro ($SMCI): -4.6%
📉 Taiwan Semi ($TSM): -4.5%
📉 Micron ($MU): -4.3%
📉 Qualcomm ($QCOM): -2.8%
📉 AMD ($AMD): -2.5%
📉 Intel ($INTC): -2.0%

Why the market panic?

Instead of relying on raw compute power, engineers behind DeepSeek focused on highly efficient code optimization, reducing dependency on high-end hardware.

FAQs About DeepSeek's Success:
Q: How did DeepSeek get around export restrictions?
A: They didn’t. Instead, they optimized their chips for maximum memory efficiency. With perfectly tuned low-level code, they avoided bottlenecks entirely.

Q: How did DeepSeek train so efficiently?
A: They used predictive formulas to determine which tokens the model would activate and trained only those tokens. This approach required 95% fewer GPUs than Meta by focusing training on just 5% of parameters for each token.

Q: Why is DeepSeek’s inference so much cheaper?
A: They innovated by compressing the KV cache, a breakthrough from earlier research that dramatically cut costs.

Q: How did they replicate o1?
A: Through reinforcement learning. They tested the model with verifiable, complex tasks (like math and code), updating it only when it got the answers correct.

The Bottom Line: DeepSeek’s success is a testament to software-driven innovation. Engineers are proving that efficiency can outpace brute force—and the market is feeling the impact.

OpenAI vacuumed the whole internet, while DeepSeek vacuumed the o1 models and karpathy warned us about this a month ago

it’s so over

you can just do things

this app is free

yes mr. president, once we build the 500B AGI/ASI supercluster eggs will be so cheap, you won't believe how cheap they're going to get

OpenAI announcing they’re teaming up with tech giants and dropping half a trillion dollars over four years to build a massive AGI/ASI

nvidia publicly supporting the first trump administration while criticizing biden's policies for their impact on ai is a major vibe shift

https://blog.samaltman.com/reflections

how do you deal with these windows bloatware moments

Revised 12 days of OpenAI predictions

1. Tsunami ✅
2. Locusts
3. Rivers of blood
4. Bird flu
5. Golden Rings, Sauron
6. Frogs
7. GPU famine
8. Nuclear fallout
9. Zombies
10. Death of first born sons, rip GPT-4
11. Rats, eleven pipers piping
12. Eternal darkness in a pear tree

we got agi at home

we are so back

cyberbullying

None of my intelligent (130+ IQ) friends use GPT-4o. They only use it selectively and rarely e.g. for voice or a DallE, but almost never use it spontaneously in their own time. This has been a long term consistent observation, but today confirmation came. A new meta-analysis showed time using ChatGPT increases exponentially with lower IQ, with drop off to near 0 minutes at 145+. ChatGPT is mostly appealing as a way to fill an otherwise vacant mind that has no interesting thoughts and minimal goal-directed behavior. Intelligent people find stimulation from Claude 3.5 Sonnet, not from marinading in vaguely pleasant GPT-4o slop like a soggy potato sitting in oil in a dirty oven tray

she gets it

found this gem today

tldr: neuralink needs a compression solution for electrode data from a non-human primate's motor cortex. Data rate is ~200Mbps; must compress >200x to ~1Mbps. Solution must run in real time (<1ms) at low power (<10mW)

thank you copilot

already deleted

It was only a matter of time until micros would bloat-in-tegrate AI into their OC.
So here we are:
"Recal" - a system that runs 24/7 in the background, taking screenshots of all your activities every 5 seconds and sending your data to their backend servers.
"Copilot Plus" - runs in the background 24/7, with real-time monitoring, meaning it's going to read everything on your screen.

An absolute nightmare for people who care about data privacy. Everything is closed-source, so it's basically "you can trust us"

never gonna use "i will tip you 200$" ever again

the programming language you should learn first is talking to women

Ilya left OpenAI