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
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!
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
@@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.
Sam wasn't talking about foundation models but about all startups building solutions that will get solved by future iterations of the foundation models.
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
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.
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.
The future of Ai in startups is indeed exciting. Instead of limiting innovation with regulations, countries should choose to support Open Source Ai initiatives.
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.
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 .
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'
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.
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.
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. 🌐💡
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.
@@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
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.
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.
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.
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
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
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
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.
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?
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.
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.
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?
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
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.
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!!
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.
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
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