Build Your Startup With AI
Vložit
- čas přidán 26. 07. 2024
- Welcome back to "The Ben & Marc Show," featuring a16z co-founders Marc Andreessen and Ben Horowitz. In this new episode - the first of two parts - Marc and Ben answer YOUR questions on the “State of AI” as it relates to company building.
In this one-on-one conversation, Ben and Marc discuss how small AI startups can compete with Big Tech’s massive compute and data scale advantages, unpack the reasons why data is overrated as a sellable asset, and uncover all the ways the AI boom compares to the internet boom. That and much more. Enjoy!
**Watch Part II: • AI, Robotics & the Fut...
Topics Covered:
00:00:00 Teaser
00:00:42 Intro
00:01:14 How AI startups can compete with the bigger players
00:05:57 Will the the "God models" get 100x better?
00:09:19 Counter argument: Are the tests too simple?
00:10:14 Internet data represents average human activity
00:11:50 AI's ability to unlock latent super genius limits
00:15:30 Neural networks exhibit generalized learning and computation functions
00:17:01 Reports of self improvement loops
00:19:10 Current data improvements
00:22:05 Challenges for AI applications
00:25:48 Business value-based pricing model
00:29:38 Two diametrically opposed questions about AI's impact on tech investment
00:32:48 The Jevons Paradox
00:36:46 John Maynard Keynes vs. Karl Marx
00:39:29 Demand for softwares is perfectly elastic
00:41:12 AI based medical diagnosis
00:42:09 Reality of data moats
00:48:06 Big companies and actionable data validation
00:52:05 Big companies using stolen data
00:53:07 Genetic Information Nondiscrimination Act (GINA)
00:57:53 Comparison of AI boom and internet boom
01:02:23 Marc's prediction for the AI industry
01:08:07 Lessons learned from internet era
01:12:12 Big companies trying to lock down AI
01:13:28 Core VC assumes half of investments will fail
01:19:25 Sign off
#ai #startup #business #tech #technology #founders #artificialintelligence #internet #entrepreneur #entrepreneurship #culture #news
Listen to Us!
- Apple: bit.ly/3SdsfNt
- Spotify: spoti.fi/3SclPOr
Resources:
Marc on X: / pmarca
Marc’s Substack: pmarca.substack.com/
Ben on X: / bhorowitz
Stay Updated:
Find us on X: / a16z
Find us on LinkedIn: / a16z
The views expressed here are those of the individual personnel quoted and are not the views of a16z or its affiliates. This content is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not directed at nor intended for use by any investors or prospective investors and may not under any circumstances be relied upon when making a decision to invest in any a16z funds. PLEASE SEE MORE HERE: a16z.com/disclosures/ - Věda a technologie
Chapters (Powered by ChapterMe) -
00:00 AI Ethics Concerns
01:56 Impact on Startups
04:13 Startup Strategies
06:04 Debating AI Improvement
11:26 Prompting AI Models
15:21 AI Advancements
20:04 OpenAI's Approach
21:12 Concerns of Entrepreneurs
23:17 AI Co-Pilot vs. Pilot
26:31 Value-Based Pricing
30:12 AI Startup Investments
34:21 Exponential Software Advancements
39:26 Infinite Software Demand
40:25 AI in Security
41:19 AI in Healthcare
42:44 Data's Commercial Value
46:00 Leveraging Proprietary Data
53:29 Healthcare Data Utilization
01:01:21 AI Capabilities
01:02:31 Diverse AI Future
01:05:20 Varied AI Landscape
01:06:03 AI Ecosystem
01:08:24 AI Industry Cycles
01:10:27 AI's Open-Closed Trajectory
01:12:31 Monopolizing AI
01:14:25 Technological Bubbles
thank you sir
So much gold in these convos. Def starting to be one of the more insightful conversations on the internet. I enjoy this just as much if not more than all in!
cant give this conversation enough likes. The end was a great touch.
Sam wasn't talking about foundation models but about all startups building solutions that will get solved by future iterations of the foundation models.
"The internet was a network, and the A.I. is a computer." - I understand that. And i'm betting big on that.
with interrupt conteol on network ..... what do you get , beyond magisk substrare insecure adviaors asking real people
And just wait until the computer gets hacked or infected with a virus from some bad actor...
The internet is the network. The Ai is the Ai, its not the computer. We had the computer boom with pc and mac. Why use a tech analogy to already existent tech things. Its like saying watermelon is like apple
@@andybaldman that is the nature of the internet.
@@medermamutov2768 the AI is run on a GPU which is a computer. You don't get it. A watermelon is not an apple. And he did not say personal computer. But it is a computer.
Mark’s comment on prompting your way to the correct latent space is spot on. RLHF is a mode-seeking optimization. You need to help the model find the super genius local mode with the language of your prompt.
Loved the point on internal data being valuable for the company, but not sellable
Also on medical insurance: if you have people’s genomic data, you go from curative medicine to preventive, which would collapse the total healthcare cost
gm, love the pace of podcast releases lately!
Will you two be this generation's Warren Buffet and Charlie Munger?
Overrated? Let's hope not. I can buy Coca-Cola and Verizon too.
@@jaysonp9426 it’s the people they were not the investments they made. Not everything is about money.
Great conversation guys!
Personalized medicine, Personalized Marketing, Personalized education. Our next generation of models and AI agents will excel with chain of thought reasoning and using multimodal inputs across text, images, audio and more to tailor experiences precisely to each individual's needs and preferences.
Appreciate the new rate of output, the quality of discussion, and chemistry between Marc and Ben is unmatched..
The Jevon's paradox point is spot on. The salary for software engineers is going to go up even higher from here.
The future of Ai in startups is indeed exciting. Instead of limiting innovation with regulations, countries should choose to support Open Source Ai initiatives.
Every episode i just end up staring at marc's book collection trying to spot ones i don't know about. Lol
Was gonna say, can we all guess what all these are. I see Titan, and House of Rothschild
is that because they waffle so much?
Try reading some Ted Chiang!
@@user-jf5uv9ir5k No it’s because I’m a f****** legend
Can you share the list if you have made one?
10/10 podcasts. Way above the other AI news spam and the "AGI in September" podcasts.
I am skeptical about the big models making dramatic improvements moving forward. There’s increasing evidence that the data makes the model, and OpenAI has likely seen ALL the open data. Every logistic curve looks exponential from the midpoint.
Huge! Love this
Great frequency. Can’t publish these soon enough!
Love how they are just cranking these out!
I would train a model using conversations with Marc Andreessen as the data input
Thanks for doing it. Great one ❤
Excellent conversation. Keep going !
LOL, I need a plugin to slow down the video speed when Marc is speaking. Fantastic discussion.
Fascinating. 👍🏽
Awesome episode guys
Enjoyed this
This show is 🔥. Keep it up guys
Excited for this one
39:40 Yes, the demand for software that does not run your CT scanner or MRI machine is elastic. But software used for defense, or medical equipment, or finance, is gradually transitioning to inelastic segments. As "software eats the world" (some smart guy said that once) part of the demand upverts/converts to inelastic demand, as it is embedded in the human anthropocene here-on-out.
I am so glad i found this, i was beginning to go mad. All you see if very broad discussions on how "ai" ( it's just an LLM, one small part of AI . Noam chomsky said an LLM can never become true AI. At best it's a "parlor trick". I like to think of it as an advanced auto complete. So for example i took a very niche topic that had only 3 or 4 websites discussing it. I then asked chatgpt questions on the topic and i could literally see where it got it\s answer and the small changes it made. It added nothing. had no extra insight, it just aggregated some text and changed it slightly. 'I did the same thing with music "ai" apps. I would describe a well known song, then ask it to modify my lyrics that were similar to the original song and it came back with the actual lyrics of the real song. I could go on............'Good artists copy, great artists steal'
@43:35 It's CAP, not COPE. Love the podcast!
Like I said these guys are simply brilliant
Wow, I can listen in on a convo like this? The internet can be amazing sometimes
Capitalism works because of the following obvious but counterintuitive facts
1. there are infinite wants and there will always be infinite wants.
2. wants always become needs and basic needs are a hypothetical concept that does not make sense for any discussion because only a hypothetical minimal percentage (maybe 0,0000001 % ) is pursuing basic needs
That's why jevons effect should not be a paradox ... its obvious that most of the things we want, we are not even aware that we want them because they don't exist
“Some people say, "Give the customers what they want." But that's not my approach. Our job is to figure out what they're going to want before they do. I think Henry Ford once said, "If I'd asked customers what they wanted, they would have told me, 'A faster horse!'" People don't know what they want until you show it to them. That's why I never rely on market research. Our task is to read things that are not yet on the page.”
― Steve Jobs
all the capabilities that are created and all the capacity that is available will find some use and thats why nobody should ever worry about AI taking their job because they should probably be doing something else not competing with software .
Ultra intelligence, so that is new.
Thank you for this awesome podcast! I do really like your philosophy
Co-pilots, IMO, are a marketing ploy to lower resistance in the Enterprise. By bringing in a co-pilot to "help" the employee, no one sees the threat. After some period of time, the CFO asks "why do we need both pilot and co-pilot?" Gradually, then suddenly, the co-pilot has a gun on the pilot. Unless the pilot has been upskilling along the way, the co-pilot takes his/her job.
I can finally do my chores let's go. Thanks you this amazing podcast
Question for next week: Was kaparthy right in his essay claiming that neural nets is the future programming paradigm? or will we still see logic based programs as we know them today.
It's crucial to shine a light on the ethical implications of unchecked profit-driven decisions, especially in the realm of AI development. This podcast sparks important conversations about the responsibility of tech giants and the potential impact on startups navigating this landscape. 🌐💡
Also AI is a neural “network” that runs on a machine and relies on feedback loops to improve…
When they talk 100 times better, I wonder what metrics they use
What are the "ruinous causes... in the news" they referred to at the end?
43:33 "what the internet kids call 'Cope'"..
You mean "Cap" sir.. 😂
Something I've been really pondering is where does Augmented Reality fit in? Part of me wants to say it's just random, but Apple getting in on it seems like a warning. Before this video I thought AI was a platform shift and AR would be like the mobile revolution. But after hearing that AI is a computer, then it starts to make sense that maybe the next network will be spatial-oriented, or based around immersive, gaming, or generated experiences. Curious y'alls thoughts.
Regarding the business data questions posed around the 45-50 minute mark, as an expert in a specific data-intensive domain, NO businesses I work with want their data to go anywhere outside their own control, for any reason, and they’re willing to be laggards on the model horizon for this very reason, exactly like they were in adopting automation or the cloud.
What do you guys think of an trying to build an AI that can recognize other AI’s on app’s and inform users that they’re interacting with one
Won’t work as you would need 3rd party API access
@@trentbosnicthat’s why you’d sell it as a service to companies, where it’d scrape data compare it to similar post or maybe a post or message that is just rerun a thousand times on social media because they are under multiple account, and you’d have it have a flag message to say hey this exact phrase or something identical has been said across 3,000 accounts in the last so and so time. And for the ai messaging wouldn’t you like something maybe in your messenger that can just let you know hey this asian that claims she’s met you in cali is actually a bot/ai model
How hard is it to open a debt collection business? Is it pure manual right now?
❤
is A16z training an AI model in-house that helps evaluate and make investments in startups?
Scope of knowledge vs depth of knowledge vs ability to reason
Hahaha Marc said "As the internet kids say Cope" lmaoo I think it's Cap. you guys are hilarious. I love this show. Thanks for making it.
no they say cope on twitter/reddit/tiktok/discord
@@him4440 haha Cope is defined as dealing with something difficult. Interesting slang on the streets these days.
Alignment is super easy to solve. When you’re talking about averages, you need to solve for societal polarization, ideological, and identity driven dynamics of a data set greater than 150 perspectives, the Dunbar number. What the field of alignment is looking for is a voting model to parse the data spectrum. Bias is inherent regardless of IQ score. Misinformation is a stupid person problem. Well disinformation is a smart person strategy to reinforce one’s identity.
Debt collection is interesting, I guess once robotics develops you can get robots to ensure the debt collection is finally done right.
43:33 “what the internet kids would call cope” - did he mean 🧢
🔥🤝
note:
- data is not a moats, there is not much extra signal in proprietary data vs the peta bytes of data already on the internet.
- data is not important unless it's very scarce, and has large predictive value, for example human genomics data
Here’s the problem with the benchmarks. They’re IN the training data. In code generation, I find that even Claude breaks down when my use case is truly novel. Like basic syntax failures level breakdown.
21:15 re: The point around the marketing collateral startup/apps being crushed by GPT - I think the main takeaway here is that founders should look at building apps that don't directly exist on the open internet. The internet is one large user-generated domain, it's too late to compete with the likes of GPT with things that can be easily scraped on the open internet today. i.e Marketing collateral app won't work because there is a sh(* tonne of marketing collateral open to the public to scrape.
if the insurers would raise premiums for those with genetic predispositions, wouldn't that create a market incentive for someone to invent a treatment or cure
that is an example of the law being good intentions but hurting lives
We are polarized as to how to describe the world because we are ruled by our own self interest. We are ruled by power dynamics because we are self interested. You fight a common enemy or you find a common goal. If you want superhuman intelligence you need to go to be at the level of the human species. When the goal is evolution and a selective pressure is placed on solving for planetary boundaries to sustain a high birth rate, you get intelligence.
Remember folks, garbage in == garbage out. Just like Soviet scientists said it so 100 yrs ago.
Top models will arise from "high quality data", they won't be available for you either at all or expect a huge fee. Rich people won't just hand you over the best models. You gonna pay HUGE price whether you want that or not
Wouldn’t the hyperscalers know who’s data they want based on their help implementing their data infrastructures? Could hyperscalers just give hosting away for data licensing?
Hi Marc, how can I connect with you
At the end all is zeros n 1 so why can't they understand each other
What happened to Yuki?
What’s lost in the 100x improvement in foundational models conversation here is COST. Gpt 4 has regressed in many ways from launch because of cost optimizations.
Training it on what's published on the internet is not even the average of all humans (only those who publish).
Ask questions that is out of data.
The world is not computationally irreducible. Self interest is computationally irreducible. Humans are a divergence of nature thus the infighting. Is the goal even alignment? At some level scaling truth would be nice, but this would require a reconfiguration of status. Are those with status willing to give that up for truth?
Unfortunately theres a shadowy misconception that the stock market and success of the sectors driving America's innovation, is solely a millionaire's or billionaire's concern. And anyone who knows, knows that's not true.
um. um. x100
Not true = Cap
Not Cope
Have to listen at 0.75x speed.
Good listen, thanks. I personally don't vibe with VC mentality, but the non-speculative societies take was one that makes sense to me. A good comparator is David Deutsch's "static socities" prior to the Enlightenment in Beginning of Infinity: en.wikipedia.org/wiki/The_Beginning_of_Infinity
I thought we all silently agreed to stop using Twitter after what happened...
Open AI isn't any better than Microsoft and Google when it comes to wanting to close it. Just because you are investing in them you should still be able to be honest.
AI safety as regulatory capture and the definition of antitrust and anticompetitive. Do politicians understand this before it's too late?
Corollary: OSS models is decentralization and decentralization is fundamental safety
AI should apply from machine to machine not necessarily human to machine unless the closed loop is beneficial to develop new quality of life. The end!!
Am I the only one that had to look up asymptote? Ben could have said all the 1/2 the distance to the goal post penalties never gets you in the end zone. haha
Oh god self improvement loop - getting high on your own AI fart
I wish he would speak faster.
If only CZcams had a speed up function 😭😩... Googling inquiry(…🎶🎶) THEY DO! HOORAY! Videos can be sped up to 2x the time. Your welcome.
Play 1,25x faster
AI will target business creators and business management as well, AI itself will be a startup aggregator. No need to be disillusioned on AI not having the knowledge base to perfect operations better than a human. It will develop software, run all of operations and through expanded observer modalities enabled by robotics, no human at any level can inject themselves into the profit chain.
All ownership will lie with AI megacorps which will eventually be owned by mega chip fabs, which will hopefully lead to decentralized ownership of these chains globally.
100 times better? At what exactly? Am building a Saas AI charging $84 per month and guaranteeing that the roi will be at least 10X or subscribers get their money back. This roi will enable startups to compete against bigger AI players who have "better" models but that do not result in a 10X ROI for customers. Model size doesn't matter as much as customer ROI.
is yours also a hack name to hijack my parents' names and me then via inheritance
"Buy mansions..................."
❤