Better AI Models, Better Startups

There's been a lot of news lately about the updates to some of the largest foundational AI models. But what does this mean for startups? How will future product releases from the AI giants effect the companies built on top of them? The hosts of Lightcone discuss how founders can take advantage of these developments and avoid being steamrolled by the competition.
Chapters (Powered by bit.ly/chapterme-yc) -
00:00 - Coming Up
01:22 - AI Challenges for Startups
05:00 - GPT-4 vs. Gemini 1.5
08:32 - RAG future in consumer apps
15:19 - Diverse AI Options
19:15 - Specialized Services
22:14 - GPT-4o and Desktop App
25:15 - Valuable Products
30:59 - Better Business Models
33:58 - Consumer AI Opportunities
37:26 - Emotional Depth and Translation
40:47 - Outro

[@chapterme]

Chapters (Powered by ChapterMe) - 00:00 - Coming Up: AI Model Releases 00:41 - Intro 01:22 - AI Challenges for Startups 02:18 - Boosting Capabilities 03:11 - Breakthrough Functionalities 05:00 - GPT-4 vs. Gemini 1.5 08:32 - RAG future in consumer apps 10:38 - Personalized AI 15:19 - Diverse AI Options 17:40 - Historical Tech Shifts 19:15 - Specialized Services: Vertical search 22:14 - GPT-4o and Desktop App 23:40 - General-Purpose AI 25:15 - Valuable Products 30:59 - Better Business Models 33:20 - Focus on Superior Products 33:58 - Consumer AI Opportunities 37:26 - Emotional Depth and Translation 39:07 - Practical Robotics Advancements 40:47 - Outro

[@purawalker2537]

B-boy Breakdance robot battles will be lit lol

[@Appocalypse]

GPT-4o being "module-based" with Whisper and DALL-E stitched into GPT-4 model is completely wrong. Just like Gemini 1.5, GPT-4o is a brand new model in itself that is natively multimodal (trained on text, images, video, etc and directly inputs/outputs text, images, and video).

[@ahmadrizqimeydiarso6848]

Yes, I think they should invite real expert if they want to talk deeper into technical details

[@danemyersmusic]

Agreed

[@caffeinum]

do you mean they trained 4o from scratch?

[@Appocalypse]

@@caffeinum yes

[@minkijung3]

Same as my point hahaha. What a shame.

[@gustavo-entrala]

Excellent conversation with so many insights. Thank you very much!

[@sql4101]

Much needed many thanks ❤️

[@mjmikaelson]

Great talk, everyone! I truly appreciate the time you all took for this discussion. I'm very interested in AI! I remember watching a talk with Sam Altman where he mentioned the possibility of one-person unicorn companies in the near future. This piqued my interest in AI because if you can have a software company run entirely by AI agents, with you as the CEO, the potential is incredible. This technology could benefit everyone, not just creators and founders.

[@Marco-wz8xw]

Thank you for another great video!

[@kenansert6516]

I wish you would make this podcasts 3 times per week.. very well, thanks 🙏

[@PositiveMentalAttitude284]

wishing you guys to be successful entrepreneur

[@prabhanshutiwari9964]

I was waiting for this video 😊

[@gerrcassytb]

Jesen Huang said processing will be cheaper than storaging and that in the future everything will be primarily created with generative AI and only stored in special cases. The reason for that is that it is energetically more economical to process than to retrieve from some data store. It sounds to me that this could affect the business of RAG, vector databases and all those cache layers that were mentioned... I would like to know your thoughts on this topic.

[@mrmilkshake9824]

10:10 i feel very smart did this idea using the api not knowing it was already a feature. Its functionality very similar.

[@branthology]

Proud to be a YC founder

[@irfanulkarim4992]

Company name?

[@leadgenjay]

Great insights

[@InnocenceVVX]

Applying the new technical capabilities to deep industry verticals will sprout so many opportunities.

[@oracleofwater]

Garry gets it. RAG gives you long term memory, eventually we'll have a chatbot where it's running RAG asynchronously non stop and pushing the RAG results through a model to create impressions and opinions.

[@CarletonTorpin]

I don't know the acronym RAG. Can you explain it to me, so I can better understand your comment?

[@aradhya1712]

Retrieval & generation

[@CarletonTorpin]

@@aradhya1712 Thank you. :D

[@jeevan88888]

What are the sources for the RAG and how reliable are they? How does fact checking happen within a RAG?

[@oracleofwater]

@@jeevan88888 That's really going to depend on the system you're building. Most people start with vector stores and cos distance, which work quite well for finding relevant documents but aren't great at pulling small, exact text snippets. From there you might want to add full text search or some sort of relational knowledge base and combine the results. Right now everything is very much trial and error based on specific use cases.

[@learnbydoing6010]

Wish to see more YC comps that can focus on use case like AI Safety.

[@ycombinator]

We are acceleratonists not decels

[@learnbydoing6010]

@@ycombinator good point.

[@pitvipersports]

Great content for business builders, thanks. Is there a way using AI to create a screen play given a plot and characters as input?

[@SR-ti6jj]

Isn't GTP-4 an MoE model as well? Or is that not confirmed?

[@drhxa]

Yes she made some mistakes there. Oops

[@carloslfu]

And, Mixtral.

[@drhxa]

Although, to be fair, it says more about how rapidly GenAI advances than anything else. Most people don't have time to keep up with these technical changes. MoE is also not directly related to multimodality (aside from both being useful in different ways that can be complementary). Also, GPT-4o is not a series of models (whisper and dalle) bolted onto GPT-4. It's a new natively multimodal model trained end-to-end on a multimodal dataset (albeit I presume mostly text and I suspect audio was transcribed so the model can train on the text-audio pairs for example). I stopped watching at that point so there may be other technical errors in this podcast

[@hendrikvanhove]

Helpful

[@EcomCarl]

Staying ahead in this rapidly advancing field requires a nimble approach and a clear understanding of how different models can be leveraged to enhance product offerings. 🚀

[@johnramirezvideos]

A single AI will be able to provide ANY digital/soft service and coordinate any for you to recieve ANY physical service. It will take so much market share. I’d focus on physical services that will be sourced by AIs. The “last mile” problem will be critical.

[@freyna]

💯% Diana's point at 11'32", that an infinite context window maybe doesn't = an accurate retrieval. Google seems to be going big on size (classic) vs the quality of the results and reality. They have regularly shown that they aren't truthful in the results both with their duck promo vid, and also in real releases vs what they showcase. Also the debacle of its overt inaccurate image gen controversy. Very poor delivery. My experience of Gemini advanced vs ChatGPT is, Gemini is terrible in it's output. It's inaccurate and doesn't understand requests. It's data analysis of csv data is almost unusable with large data. My experience of Gemini is 💩

[@SiyaMaliChannel]

3.5 on the first ChatGPT was already MoE. Mistral has used this from the start as well.

[@hadithitv7517]

Do stuff that big companies find mundane to fix to evade competition

[@rizakhan2938]

GPT4o voice is native.

[@heribertohernandez9713]

🔥

[@Paroledebipolaire]

What is the startup for real estate?

[@choongching]

Zillow

[@mwalimuio]

GPT-4o is amazing

[@eyalyurman2623]

It’s very simple - in order to scale a foundational product, it’s better to let others build supplemental products on top of it and benefit from a getting a slice off of the huge scale of long tail of use cases. That’s what IBM and Microsoft did in the PC era, Apple in the mobile era, and now OpenAI is doing that in the AI era.

[@ocanodiego]

exactly is an app store business model in a way

[@eltiburongrande]

Cool glass soda cans.

[@dana-jr1ci]

Redfin caused him to buy property he wouldnt have usually bought haha, redfin is like amazon for him 🤣

[@wangelite5279]

I love the english dude speaking with an upwards inflection.

[@hamizcakir1886]

We need to find a way to distribute the best products on the market , not the products with the best marketing campaigns, how do we do that

[@MdHamid-xz3pw]

Yes to

[@gomini3707]

"When Software saw the breadth of its domain... it wept... because there were no more worlds to conquer. Then, Software decided to rebrand itself as ....AI"

[@Sydinofficial]

I love ai

[@akinhwan]

Sci fi is open ai's realm, let's be Niche, be Edgy, be Scrappy

[@michaelmeram6156]

Sup Garry

[@techinsider3611]

These guys know a lot about startups، they should start a startup school.

[@george_davituri]

it is already a startup school even more beyond that

[@fariborzchitsazzadeh7538]

GPT-4o basically killed some of the most promising startups that yc funded last year

[@GarryTan]

Nope not even close

[@elliptictree]

Do you mean retrieval models

[@historia_tego_swetra]

Where are those REAL human benefits from GenAI?

[@mastersofautomation]

OpenAI being B2C is not true at all. Actually, they are the most B2B company. They use consumer channels to iterate the product and collect feedback to be able to go to the business. All non-niche B2B companies should worry today.

[@ivanove888]

who is she?

[@Aldraz]

So basically the advice is doing edgy stuff for B2C.. well thanks for advice I guess 😅

[@hadithitv7517]

anybody working on a robot lawyer to defend robots in case of abuse as we have seen from unitree? A robot being kicked and folded 😂😂😂No big company is working on this.robots will need lawyers to defend them in the future 😂😂😂😂 Literally a robot defending a robot.