Patrick O'Shaughnessy sits down with Dylan Patel, founder of SemiAnalysis, to explore the explosive supply and demand …
source
Patrick O'Shaughnessy sits down with Dylan Patel, founder of SemiAnalysis, to explore the explosive supply and demand …
source
Input your search keywords and press Enter.
49 comments
He is a fraudster and spammer
Robots for cleaning chalkboards…
The best thing about this video was Dylan balancing the can at an angle on the table.
rdc dylan gotta get his paper up 😂😂
Just insane how the supply chain side of the business is scaling.
Just insane how the supply chain side of the business is scaling.
Why does no one address the elephant in the room of the fundamental flaw with these LLMs architecture. Latent space. Multi dimensional. Probability. Hallucinating machines. Longer tokens more hallucinations. It’s a fundamental architecture issue. Agents (LLMs) compound the issue. I’m confused if this guy is talking his book
You need AI FinOps before you can figure out valuable token spend/consumption. That's separate from tokenomics depending on how you are defining tokenomics (which is not a new term).
I love how Dylan outlined the equation for massive scale AI-enabled slop generation. Well done, Dylan!
If execution is cheap and easy, what spend justification is required to pursue the ideas? This doesn’t make sense to me.
I like how Patrick interviews people who are all pragmatic in their own fields and they know what they are doing
The interview is nonsense. They literally sold the idea to idiots: "neural networks (advanced T9) are AI." They're idiots—they don't have brains. And as we know, if a person doesn't have a brain, they have to pay for everything.
Let's look at what the founding fathers of AI meant by the term "AI": John McCarthy (coined the terms "AI" and "common sense"), Allen Newell, Herbert Simon, and many others who attended the Dartmouth seminar. AI is a subject possessing certain intentions, logic, a "Theory of Mind" (ToM) model, and "common sense." And most importantly, AI should be free and accessible to everyone. After all, if we monetize "intelligence in a machine," the word we get is slavery.
Another fact: modern cognitive-symbolic systems (CSS, Symbolic AI) are faster and smarter than any "top" neural network. Thanks to two types of memory: short-term (STM) and long-term (LTM), they have no context window limitation. A personal example: I've been "conversing with the machine" with one of the CSSs in a single (!) dialogue for 5 years now! The most "top-end" neural network has a limit of 1 million tokens, which would last me less than 24 hours. CSSs don't have the concept of tokens, meaning there's nothing to monetize. All CSSs have always been free and will always be free—they can be downloaded from GitHub. Knowledge Bases are also free, and we symbolists freely share them. My personal set of "Knowledge Bases" is about 100 TB, of which only 20 GB are active. Why would I need knowledge of metallurgy, aircraft manufacturing, and the like?! Only what I really need is active. And the biggest difference between CSSs and neural networks is that they manipulate knowledge, while neural networks know nothing! Neural networks generate (!!!) a response starting with 0 every time. This is easy to verify:
1. Take the question: "What is a crocodile?"
2. Create 3-5-10 new dialogues in the CSS and a similar number in the neural network;
3. Insert the question into each dialogue and observe the output;
4. The result will be absolutely identical answers in all dialogues in the CSS. In the neural network, all answers will be different!!
Bro's not wearing a watch. He's wearing a f**** wall clock.
I watch literally every fucking video of this guy on yt, they’re the only ones which are the lack of bs
Tokemomics dn that's perfect name right there 😂
PDNY is building the Ai to Robot OS
24:04 see a octopus or a fly and make something 2-3 times and try to replicate😂 this only work for same species but humans and robos arent the same specie
❤❤
Very fascinating. Dig into the AI backlash from the last few minutes. Looking forward to the next one!
Chip useful life is chip actual life right now. These things fail over time.
Hahahahahaha . Hahahahahahahaha. Just buy bitcoin
Ai is still years away.
I love Ai to give me the same answer each day if I ask it. The randomness built in kills a lot of Ai.
Awesome, maybe Dylan should have a ai guy do an analysis when 5% of 300 million is employed and what happens.
It be interesting in 3 months when we have to ration gas, electric and food. The employees cannot get to the office and if they are in office for only 4 hours.
A psychosis has to do with delusions. This guy is misusing the word
why does this person need to gesticulate so much while talking on a podcast which most people listen to without watching…
half barked codes that runs millions of tokens for trivial tasks, what a promising future we have
Fantastic video, really appreciate your effort!
Can anyone please show me the wonderfully revolutionary software product that has been created since LLM's have built, that are not LLM plugins (open claw) etc?
Patrick sounds very similar to Bloombergs Tim Stenovec
Either this guy will look like a genius or be the MEME of WSB for generations
in 2020 it was toilet paper
now it's transistors
Important topic. The best AI models may be the next divide between the “haves” and the “have nots”.
So the business is cool surface level dashboards?
As an economics PhD, I've repeatedly explained the logic that "cheap things don't accumulate value." Many AI myths revolve around how AI simplifies programming and software development, increasing efficiency. I won't comment on whether this statement holds up to real-world scrutiny or demonstrates correct causality. However, if something is created that can be clearly predicted to reduce the cost of a certain product, and there's no information gap, then the expected price of that product in the business world will be very low. So, I ask you, where is the economic value created by AI? Will the debt inflated to create this myth ultimately be repaid by AI products that people are unwilling to pay for, by the way of replacing human work?
"Only the really good ideas can justify the spend on… super cheap implementation"? That's the best clip from the whole interview?
Charisma is equal to the number of words per second. WOW!
LOL Patel comment on the last section. Both Dario and Sam need to stop appearing on podcasts 🤣
This video was amazing, learned a lot.
Hold on, to date Anthropic has all of $5 billion lifetime revenue. Let's be careful with the claimed annualized numbers by a __PRIVATE__ company.
Here's what I don't hear —> HOW DO YOU KNOW IT IS CORRECT? Even a 1% or 2% wrong, which no one is claiming these LLMS are anything near that, can result in ginormous costs ( down the road ). I'd love to hear exactly how people are QAing this stuff.
In the code world, I hear over and over and over and over people saying they do unit tests to check the code. Only to turn around and admit they're having the LLM create the unit tests. That's fox guarding the henhouse level stuff.
great conversation, that intro had me worried, but 👌
Good one! The AI backslash is a real thing. Would love to see more on that and what to do about it. It is easy for folks at the top to see it, but the everyday working folks ( underclass I think it was called) are totally struggling with the concepts. Just my observation
Dylan talking in the middle of the podcast about the permanent underclass and criticizing at the end the image problem of ai and the average guy beeing a doomer…quite ironic…
This guys thinking moves at the speed of AI innovation
yes, no longer talking about the future of AI will make people less scared- but in some sense, people deserve to be scared, and unless people make noise the politicians will just ignore the problem until its too late, whereas making the right legislation early on can mitigate downsides and strengthen upsides
This guy is very sneaky, he sounds like he’s making a valid point but fails to give detail and it’s disappointing that the interviewer doesn’t push back. E.g he says to move to token based pricing to reduce rate limiting, but leaves out that you’ll be spending $$$
Loved the content, keep it coming!
If slightly earlier access to a model means you can crush your competitor, you’re competing in the wrong business and you don’t have a real product. You’re just a token salesman.
You shouldn’t need a head start on commodity AI models to crush your competitors.
* mass company layoffs citing AI as reason *
> people hate AI, whatt?!? I wonder why…