The Truth About Building AI Startups Today
Vložit
- čas přidán 14. 05. 2024
- In the first episode of the Lightcone Podcast, YC Group Partners dig into everything they have learned working with the top founders building AI startups today. They share the ideas that are working particularly well, mistakes to avoid, and take a look at the competitive landscape among the current AI giants.
Chapters (Powered by bit.ly/chapterme-yc) -
00:00 Intro
00:41 The Lightcone Podcast
01:53 YC Recent Batch
03:34 College Students and AI Startups
04:44 AI Startup Success Factors
05:53 Opportunities in Mundane AI Tasks
07:30 Beware of "Tarpit Ideas"
09:36 AI Integration into UIs
10:30 Avoiding the "Checkbox Mentality"
11:54 Focus on Genuine Use Cases
13:45 Fine-Tuning Open-Source Models
15:20 Data Privacy Concerns
17:09 Purpose-Trained AI Models
18:36 AI Models for Prototyping
19:45 Surge in Startup Ideas
20:54 The "GPT Wrapper" Term
22:02 Importance of UX
23:39 Focus on Specific Problems
25:03 AI Agents and Open-Source AI
26:58 Resurgence of AI Researcher-Founders
30:15 Periodic Dismissal of Emerging Tech
32:06 Outro - Věda a technologie
![The Possibilities of AI [Entire Talk] - Sam Altman (OpenAI)](/img/n.gif)
What are you building with AI right now?
Ecommerce shopping agent & AI ethics research.
Props AI - A cost monitoring tool for Open AI spend
Artificial English tutor that understands your situation and needs - writes email for your English speaking boss, prepares you for a job interview in your specific industry, etc.
How do I get funding? I want to bring AI powered Vtubers to the masses. I am also working on custom chatbots
Mitra - build AI teams that do anything
I love how this is pretty off the cuff. Gives me a good idea of the actual, unscripted personalities of the group partners, and what they’re like in a group. Awesome guys!
They feel more human haha!
30 minutes is about the perfect length for a podcast episode 👌
4 you. i perfer 3 hours.
A lot of info
Love the energy and maturity around the topic! Waiting for Ep2!
love the modeling around "if a company isn't buying your co-pilot, just build their company with co-pilot and beat them" because if your co-pilot can't level up a company enough to be better than your potential client, then you probably aren't providing enough value.
The hubris in that idea.
@@ryanslab302 Exactly! These people are so delusional. What they are saying is if your little chatbot tool is good enough, you can become better and bigger than your potential client. So if I am trying to sell a chatbot to Tesla, and they don't buy it. All I should do is start another Tesla because my chabot brings so much value that I can successfully out compete any company without any experience, passion, expertise, or domain knowledge.
It's a moronic take.
then the company just implements the co-pilot once you start getting traction...
@@janirobe ok, so if it took them that long to catch up on one thing, you use your smaller company speed to keep ahead.
It's not everday that YC launches a new podcast! Love the name and the content, keep em coming!
Excited to see where this podcast goes! lots of things to learn from you guys! thanks!
This is very insightful! Keep up the good work, team YC!
What a great ending, man!
as well as the whole podcast
Best video I watched in a while. Thank you! It's good to have validation that what we are doing is correct.
Chapters (Powered by ChapterMe) -
00:00 - Intro: Differentiating Startup Ideas
00:48 - The Light Cone" Podcast
01:53 - Y Combinator's Recent Batch
03:34 - College Students and AI Startups
04:44 - AI Startup Success Factors
05:53 - Opportunities in Mundane AI Tasks
07:43 - Beware of "Tarpit Ideas"
08:30 - AI Copilot
09:36 - AI Integration into UIs
10:30 - Avoiding the "Checkbox Mentality"
11:54 - Focus on Genuine Use Cases
13:45 - Fine-Tuning Open-Source Models
15:20 - Data Privacy Concerns
16:31 - Purpose-Trained AI Models
18:36 - AI Models for Prototyping
19:45 - Surge in Startup Ideas
20:54 - The "GPT Wrapper" Term: Importance of UX
22:20 - Building a billion-dollar AI company: Focus on Specific Problems
24:16 - AI-Powered Voice Agents
25:50 - Advocacy for Open-Source AI
26:58 - Resurgence of AI Researcher-Founders
29:36 - Returning YC to Its Roots
30:20 - Periodic Dismissal of Emerging Tech
30:55 - Classic Hacker News Essay: The New Cycle of Tech Geeks
32:06 - Outro
Very good summary
@@rembautimes8808 thank you 😊
Clever ;)
@@nathanbrannan5228 Thanks 🙃
now i can skip to whatever complete trash their taking about
Thanks guys. This video inspired me to implement LLM to a feature I’m currently building for my startup.
How come i didn’t think about it that way 😊
"Mundane tasks and boring work"! Love this!
Thanks for the inspiring conversation
This is an excellent conversation straight from the source of one of the best places tech is born.
Thanks for starting this Y Combinator Team.
I do so many tech startup interviews myself as a host for Grit Daily mag its refreshing to see what Y Combinator sees as important to discuss.
Thank you. Interesting times. Look forward to more episodes!
Absolutely phenomenal, cant wait for the next episodes
Great episode, looking forward to more!👌🏿
Absolutely amazing talk very bright panel
Really like the format and the wisdom here!
Another day, another YC Classic.
Thank you very much guys for that amazing episode , helpful insights ❤
Yo plz don't give up on this podcast. Looking forward to ep 100!
yes sir
The analogy of LLM as FPGA of idea prototyping is quite apt !
I love this conversation! It's about time
Thank you for this conversation!
This is the best thing happening on the internet right now!🔥
Working in an AI startup, I can vouch for these observations. They were on point, to which we are working on right now!
love this! keep it up ,please!!
Love the genesis of the Lightcone name, Jared :) Looking forward to more of these chats.
Thank you for this podcast.
Thank you for this podcast, you gave me the idea of taking 3 Google Cloud certifications: Cloud Architect, Data Engineer, Machine Learning Engineer to create one of the companies that you recommended to do. Your reviews are really helpful. Thank you 😃
I feel like this is the start of something special.
same, in no other circumstance you have this level of intelligence converge on a podcast for any reason.
Great to hear your insights. I look forward to more!
Good content - Go YC! Great to see Diana there too!
thanks for this Gary 👍
very insightful, keep it coming! thanks for sharing
yes
Dang! This was great pod! Keep it going guys. :-)
I love the episode, it's really insightful and would love to have more.
But I would like it if most of the acronyms used could be outlined to help those unfamiliar with them add them to their vocabulary. Thank you
This is a gem. every minute is packed with good ideas.
This is an incredible episode
The difference is: MySQL doesn’t have a consumer facing product.
And in SaaS, the user interactions are much more complex.
In Chat based software, it’s the same interaction. And there’s a consumer facing app already. 22:45
This means the barriers to entry are low, which means endless substitutes, which means you'll compete on price (race to the bottom)... low margins, low top line... this isn't rocket science.
Agreed
That works both ways though. It means open source will likely win in the end since all it takes is one massive LLM to be trained so OAI and the like will lose their moat.
I actually think access to chips will likely be the limiting factor, and I can see arguments for both OSS and closed source there.
Honestly, I feel like the gold rush around AI and LLMs is creating this negative pressure on the rest of the ecosystem. I think it's pretty likely that we're going to see a huge consolidation in the LLM space in the coming years and there will be a few huge winners and lots of small losers. My hot take is that now is paradoxically the best time to found startups that *aren't* banking on AI and LLMs. Lots of people are going to waste lots of time trying to shoehorn LLMs into spaces they don't belong and fail, and it's a great time to get ahead. You were talking about AI tarpits, I think it's not that there are a few tarpits, but that AI is a giant tarpit with a few paradise-like islands. Build something great while others are swimming in the tar.
Every startup can be a tar pit if you can’t raise funding 😅
I agree completely, it seems like the pressure is making people ignore huge red flags, such as the fact that so many startups are based on the behaviour of openAIs models that can change on a daily basis.
PromptArmour is actually changing the game! Can't wait to see what they do next!
Key takeaway: "SaaS is just a DataBase Wrapper", golden phrase right there.
Jeez, the subtitles kept showing Rapper, first I was like what ! Then I rationalized it saying as GPTs are language models spewing words they in a light heart calling it Rapper
TY, really interesting discussion
Best 32 minutes i attended in a while.
A half hour is very brief. I'd love to hear more about how you see the world.
Don’t worry we post a new episode of Lightcone every two weeks
Most of the problems are like ages old. When tech meet use cases, they seem like strange to each other. So many mundane tasks are around us, and just a little bit calibration, these job would intrigue people's potential rather than make them age faster. Talk to the people in different industries, understand them, are so important.
Great job guys/gals very inspiring stuff.
I love the FPGA vs SOC analogy.
Great episode!
I love your videos, continue pls
The Video Just started and ver talked about the general theory of relativity, oh boy, this be fun.
This was excellent.
The tarpit idea concept is pretty interesting takeaway.
As a german startup founder I directly recognized this intro jingle 😅 - shout out to all OMR education listener ✌❤
00:02 AI startup ideas are everywhere.
02:10 YC funds a significant number of AI companies due to the preference for funding smart founders, not specific industries.
05:53 Finding a niche problem can lead to successful startup ideas
07:57 AI startups often get stuck in untested ideas, leading to challenges in customer adoption
11:46 AI startups face challenges with product market fit and actual use case adoption.
13:53 Demand for fine-tuned open source models driven by cost and customization needs
17:57 Customized models are winning in specific domains.
19:48 Startups in YC often pivot quickly to find a new idea
23:27 AI can transform the scope of software and reimagine existing applications with AI capabilities
25:18 AI voice agents are being used as receptionists by companies.
29:06 Returning to the roots of funding hardcore technical founders
30:54 Geeks drive tech innovation despite fluctuations
Love this podcast. One quick suggestion, the light cone is not exactly the definition you mentioned in the beginning. It's related to time and light speed so that have the cone shape. Just a suggestion.
Great content that just answer my question!
Light cone is such a cool and appropriate name, and the explanation deserves a little expansion: Alpha centauri is 4 light years away, so any choice I make today will only affect alpha centauri 4 years from now, at the earliest, because nothing can travel faster than light. If you plot a spacetime graph of all the events throughout the universe that I could theoretically have an impact on, that graph will form a cone shape, which is why it's called a light cone. A very apt metaphor for starting a company that will shape the future!
Love thisssss!!!
So is AI a solution in search of a problem?
It is very interesting how AI is changing the way we find information and work around the world so exponentially. Investing in large companies dedicated to AI is a smart move. Many companies, are and will continue to migrate to this new technology, it saves money if a robot can do a job instead of a person.
Such an underwhelming explanation of the lightcone. It’s the space-time region that any physical entity can be causally connected to! Such a cool podcast title!
This is incredible. Everyone of you looks and sounds great too.
00:00 - Intro: Differentiating Startup Ideas
00:48 - The Light Cone" Podcast
01:53 - Y Combinator's Recent Batch
03:34 - College Students and AI Startups
04:44 - AI Startup Success Factors
05:53 - Opportunities in Mundane AI Tasks
07:43 - Beware of "Tarpit Ideas"
08:30 - AI Copilot
09:36 - AI Integration into UIs
10:30 - Avoiding the "Checkbox Mentality"
11:54 - Focus on Genuine Use Cases
13:45 - Fine-Tuning Open-Source Models
15:20 - Data Privacy Concerns
16:31 - Purpose-Trained AI Models
18:36 - AI Models for Prototyping
19:45 - Surge in Startup Ideas
20:54 - The "GPT Wrapper" Term: Importance of UX
22:20 - Building a billion-dollar AI company: Focus on Specific Problems
24:16 - AI-Powered Voice Agents
25:50 - Advocacy for Open-Source AI
26:58 - Resurgence of AI Researcher-Founders
29:36 - Returning YC to Its Roots
30:20 - Periodic Dismissal of Emerging Tech
30:55 - Classic Hacker News Essay: The New Cycle of Tech Geeks
32:06 - Outro
Thanks for doing this.
How would you do it for my vids?
Just curious, it's interesting how every one of you mentioned LLMs, but no one talked about CNNs (Convolutional Neural Networks). Is it because there are so few computer vision-related startups applying or generally building in this area?
Great first episode, by the way. Always love your content, can't wait for future episodes - it's going to be epic!
There are just fewer focused there but they do exist Eg Standard AI or Flock Safety
@@ycombinator Thanks for the clarification and quick response.
Exciting start! The Lightcone Podcast's debut episode, featuring insights from YC Group Partners working with top AI startup founders, promises a treasure trove of valuable lessons.
First, great show. I appreciate getting into the nuts-n-bolts, but also appreciate higher level talk shows like this. Thank you. Secondly, what is the name of the Government contracting AI company/tool that you mentioned? I almost fell off my seat when I heard someone was doing this. I started this very same thing, however, it's a 50/50 split of code and AI, as I've found at least 50% of a Sources Sought response or Market Survey response is a regurgitation of what can be found through code. In any case, if I can pay for an already operational SaaS for this, I would gladly do that in place on finishing what I've been working on.
Highly informative.
AI is changing the way we access information and the way we work. Even just investing in the major public players is a great way to get exposure to AI. It simply makes sense economically if you can have a robot do what you need to pay a human to do.
Who's gonna buy all the junk that robots are making, when the average person can no longer afford to pay for their housing and food?
Noting a namespace collision: Lightcone Infrastucture (often just called Lightcone) is the parent organization of LessWrong.
Many such cases
Its not a legal problem so I say go ahead.
Gary Tan is such a great listener
I’m a GPT-4 customer. I’m frustrated with the overconfidence in responses that end up being wrong. I’m also frustrated with its laziness and not knowing where the laziness starts and ends. I asked it to give me a list of all the countries in the world and it was adamant that I just go to a Wikipedia page.
Thanks, Y Combinator, for the knowledge shared. I'm exploring a niche within the LLM-based AI tool space, similar to what Agent GPT offers, but more focused. Given that broader AI tools like the upcoming GPT-5 may not cater specifically to every unique problem or audience, how viable do you think a specialized, niche AI solution is? Can targeting a specific audience with tailored solutions still be a sustainable business model despite the broader capabilities of general AI platforms?
No for long term. The general use AI is growing and developing rapiy and becoming close to the solution people need
2 Things I've advised our clients to consider with AI:
> Focus on the picks and shovels: Enable AI rather than create it.
> Double click on a niche: Don't just focus on an industry but a niche within a single industry. e.g. Not just antiques but antique books.
About the first, I think a lot of clients want all-in-one observability platforms that typically just get built by the biggest players (e.g. Azure), it seems like niche ideas may win in that field as well
Note for myself:
Popular ideas for AI startups that are not working(tarpit ideas)
AI copilot: Build a copilot for someone's product or service.
Finetuning open-source models.
Ideas working: LLM security.
I think you are right it's crowded place there is more ai startups than demand
Why does an AI copilot ain't work?
@@ephreamjudegeorge8063 there are many industry giants already like ibm unless you want to get f up.Better do something on other products
Thank you
3:55 - I don't believe in coincidences. I decided to finally go back to college & earn a CS degree I started years prior. And about 3 months before graduating, an old friend asked me to partner with him on an AI startup. So I'm grateful to say not only am I getting in on this once-in-a-lifetime opportunity, but doing it with a fresh current Computer Science education. Granted the timing worked out for me, but I do believe there is more benefit to earning that degree while still jumping into this world. Great video, thanks!!
Data privacy is the main driver for customers looking into hosting their own LLMs specially in regulated markets in the EU
we need less products with GPT wrappers, instead it is good to see open source AI model with training dataset, but it can contain some danger. anyway looking forward for appearing new AI driven companies observed - backed by YC ⭐
Well from what I understand that AI as it stands is very difficult to implement, but even more difficult to get people to use, which is closer to my experience.
The solution to this is to design an AI that the user doesn't necessarily interact with, but functions in the background (usually as an auto-complete).
I think that's probably the direction to go in first, and then once we figure how to use AI as a really sophicated auto-complete, then we can properly move on to auto-gen bots that can just use the tools without supervision.
Finally once it understands how to just use the tools as good as a human, then we can go into full automation, where we just give it a task, and it will complete it, as well as a human can.
for others closer to my age (mid 40s and up), I have a course on skool called prompt-engineering , I taught kindergarten 9 years then went to computer science/ data science school/ bootcamps with the kids, now 8 years in tech, working as an ai integrated specialist
On skool, I have a course called prompt-engineering , I do live zoom on Saturdays if people have questions. Not free but free stuff, you can find anywhere online
A light cone in special relativity does does not refer to the spatial cone shape produced by a flashlight. Maybe im confused abt what he meant. It actually is related to the cone shape that all possible trajectories of light produce on a local spacetime diagram. For a given starting point it defines a boundary that no object can ever cross.
I am a geek and love innovations and cutting edge in AI.
Can you code?
Still percolating on "GPT Wrappers" and subsequent discussion.
Great startup idea lying on the ground: make it easy for viewers to fix youtube transcript (so it can pass the touring test)
lol exactly what I thought 😂
Building a startup should start from problems.
Building an AI startup should start from problems, too, not AI.
Ha! I literally thought lightcone was a reference to the lamp shades hanging over their heads😂
I suspect that the term 'smart founder' might scare off the young, talented but shy potential entrepreneurs.
It could be a term invented by those who have already 'made it' and who have more than a touch of hubris.
So the truth is you all funded AI without having any idea what the use case is and now everyone is stuck in a bubble.
That's really accurate
These guys literally became millionaires in the era where explicitly unprofitable businesses were given tens of millions in funding… what do you expect lol
It sounds like a bunch of marketers knowing nothing about tech and especially about llms
@@abesari29so true.
If models become local (stored and processed on prem for mobile and desktop)
Then applications serve as a literal UI layer.
Apps become more lightweight going forward?
would be cool if someone could merge Real-Time Computer Vision and a Large Language Model to develop a real-time conversation-based AI co-worker equipped with vision.
Instead of solely focusing on replacing tasks, which might not happen as quickly in some areas, perhaps someone could concentrate on enhancing productivity and aiding individuals to become more proficient in utilizing AI effectively.
Amazing
I like the lady's point there at 11:54: Maybe we don't need shovel, maybe we need entirely new tool. Who knows.
I enjoy your channel and am becoming a fan. However, I often find it challenging to grasp everything because of the frequent use of abbreviations and technical terms. This approach might work well if your intended audience is intermediate to advanced users, but it might not be as inclusive for beginners. Considering beginners today could be the experts of tomorrow, I suggest making your content more accessible to a wider audience. By doing so, you're likely to attract more followers. This is just my opinion, but I do believe you're onto something great. I'm eager to reach a point where I can understand more than 80% of your content.
Hope they notice this 😅 . help me out comments community 😁
it's a causal horizon. a light cone, it's the maximal radius that light spreads out over the existence, of the expansion of that particular light horizon, over the course of its existence as it is plotted on a two-dimensional surface, usually a chalk board or dry erase board. the actual light cone would be three dimensional in our observational universe.
For businesses, leveraging private data with approval. But consumers face leveraging (mostly public) data without approvals (ie. using copywrited data)
I used AI to make a summary:
Highlights
🚀 AI is permeating every aspect of society, creating opportunities for startups.
📈 Large language models (LLMs) are being utilized by many startups.
🌟 Boring tasks can be the basis for successful businesses.
⚠ “Tarpet” ideas can trap founders if they are not careful.
💡 Focus on solving concrete problems and providing practical solutions.
🤖 AI co-pilot concept is promising but faces challenges with user adoption and finding product-market fit.
Curious how much should LLM API costs be for B2B SaaS built using LLMs - 10%?