DeepSeek: America's AI Sputnik Moment? Tech Veterans Weigh In
HTML-код
- Опубликовано: 9 фев 2025
- Two words have caught the Internet by storm. DeepSeek.
The Chinese reasoning model r1 is rivaling others at the frontier with an open-source MIT license, methods that some claim may be 45x more efficient, an alleged $5.6m cost, the release of reasoning traces, a follow-on image model, and the fact that all of this was released by a hedge fund China.
Many are already referring to this as a Sputnik moment. If that’s true, how should we - whether founder, researcher, policy maker - not just react, but act? Joining us to tease out the signal from the noise are a16z General Partner Martin Casado and a16z board partner, Steven Sinofsky. Both Martin and Steven have been on the frontlines of prior computing cycles, from the switching wars to the fiber buildout, and have witnessed the trajectories of companies like Cisco to AOL to ATT - even Worldcom.
So what really drove this DeepSeek frenzy and more importantly what should we take away? Today, we answer that question through the lens of Internet history.
Timecodes:
00:00 - DeepSeek's release
04:06 - Teasing signal from noise
09:19 - OS license and reasoning traces
11:22 - Monetizing layers of the stack
18:42 - What’s different this time around?
22:09 - Scaling up vs scaling out
29:24 - Changing benchmarks
31:48 - Building defensibility
35:27 - AI’s Sputnik moment?
41:17 - Impact on frontier companies
43:00 - Participating in the next wave
Resources:
DeepSeek Has Been Inevitable and Here's Why (History Tells Us): hardcoresoftwa...
Why DeepSeek Is a Gift to the American People: www.thefp.com/...
Follow Martin on X: x.com/martin_c...
Follow Steven on X: x.com/stevesi
Stay Updated:
Let us know what you think: ratethispodcas...
Find a16z on Twitter: / a16z
Find a16z on LinkedIn: / a16z
Subscribe on your favorite podcast app: a16z.simplecas...
Follow our host: / stephsmithio
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
📌 Timestamps:
00:00 - DeepSeek's release
04:06 - Teasing signal from noise
09:19 - OS license and reasoning traces
11:22 - Monetizing layers of the stack
18:42 - What’s different this time around?
22:09 - Scaling up vs scaling out
29:24 - Changing benchmarks
31:48 - Building defensibility
35:27 - AI’s Sputnik moment?
41:17 - Impact on frontier companies
43:00 - Participating in the next wave
If an open source reasoning language model helps us towards collective terrestrial intelligence, CTI, then I am all for it.
CTI would be realized with an open source, fact checking global platform that can merge the knowledge and sentiment expressed in public conversations with billions of people around the world.
Key Takeaways & Timestamp
- AI advancements are accelerating, with unexpected breakthroughs from global players like China
- Restrictive policies, such as export controls, are ineffective and hinder domestic innovation
- The U.S. should invest in research and development to maintain a competitive edge in AI
- AI's impact will be similar to the internet's, requiring adaptable business models and open collaboration
- The focus should shift from controlling technology to enabling innovation and application development
00:00:00 🚀 The AI Race: A New Era
00:04:06 🇨🇳 China's AI Breakthrough: A Closer Look
00:09:19 🔍 Unpacking the Deep Seek Model's Impact
00:11:22 🌐 Learning from the Internet Era
00:18:42 💡 Capitalizing on AI's Potential and Infrastructure
00:22:09 📈 Scaling Strategies: Up vs. Out
00:29:24 🏁 Redefining AI Benchmarks
00:31:48 🧠 The Shift to Workflow-centric AI and Applications
00:35:27 🌍 AI's Geopolitical Implications and Regulatory Insights
00:41:17 🔄 Innovation from Unexpected Places: The Role of Hedge Funds
Full Summary: www.digestly.me/digests/cm4gzntid0009ns8ssfevt1tx
I wholeheartedly agree with you both gentlemen: Export control does not work. We should embrace competition. There are always more solutions than problems. The next step for Ai is developing best app in all sectors.
I love that he’s wearing a Clippy hoodie
👀
This is a great history lesson by two tech veterans. So much important learning here. Terrible title and facilitation. The “Sputnik moment” is just wrong phrase. You want America to be more competitive and the government to have more urgency - not warmongering.
Attiva la traduzione audio simultanea
I used a CBM PET back in 1977 to run weapon simulations and also to design hybrid analogue-digital chips, .. all coded in 8k bytes of BASIC!
The programs ran for up to a week at a time!
This is AI's Cambrian Explosion moment.
In 10 years times we shall look back on OpenAI etc with fondness, although they will have faded away.
FWIW I have a very effective Deepseek (70B) running on a $1000 home PC.
I can leave it running overnight or longer to solve tricky problems.
A few years ago I could have sold this system for a fortune - but now AI is a commodity.
What kind of setup are you running for that one?
Incorrect. This is basically a non-event. Deepseek did exactly what you would expect a good lab to do: they took all the accumulated knowledge--some of which is simply that o1 and o3 exist and that they spend a lot of tokens talking to themselves--and they applied that knowledge. Then, because they didn't have to do very much experimenting (they have the advantage of KNOWING about o1 and o3), they could train V3 and R1 very cheaply. Now, since the release of R1, we have three reproductions, three validations (at least) of deepseek's "breakthrough." And two (Berkeley, Stanford/UofW) have been accomplished for less than $50. Not $50 million or $50 thousand, $50. So, now, who's the genius? Who has the breakthrough? Deepseek did a very good thing, a very impressive thing, AND they open-sourced it. We should stand up and cheer deepseek: thank you VERY much! But they didn't give us any sort of Sputnik moment. The people who overreacted because they are not able to understand gave us a pseudo Sputnik moment.
@@tanker242 I have it working on an i3 with SSD drives, 32GB RAM and a 4GB GPU under Ollama.
I get about 1 - 5 tokens per second.
Ollama has a cache system and the model is a Mixture Of Experts so only a portion of the 70GB is in use at any one time.
The 70B model feels very similar to the full size version.
It did however take 1 hour to create the code for the Snake game .. which worked first time!!!
get that woman a watch band that fits!
0:25 not Twitter but X. 😂
Hi everyone
Say hi back plz
🌎🌍🌏
My god, there are a lot of dumb regulators in the United States! With all the suggestions about control and export controls! Imagine trying to censor math.. definitely a wake up call..
I just hope it speeds all up... I wanna Upload my mind on a synthetic entity asap to become Immortal. No joke.
Silly wabbit u already are ~
Revelation 9:6
♥️🩸
@fr7nkyph7llyj7ne5 🤣🤣🤣