Chapters (Powered by ChapterMe) - 00:00 - Coming Up 00:29 - Intro 01:40 - Building a successful vertical AI company 06:05 - The unique challenges of law and AI 09:24 - The turning point for lawyers with ChatGPT 11:25 - Finding product market fit in legal 15:04 - Entering deep founder mode 20:40 - Approaching prompt engineering step by step 25:05 - Going beyond GPT wrappers 28:10 - Aiming for 100% accuracy 30:48 - Thoughts on o1’s capabilities 36:42 - Outro
VERY helpful and inspiring interview. Im also building an SaaS app. Initially, my entire focus was trying to find the right models and approaches to making the models work for me, but it became exhausting trying to keep up with the pace of innovation, so I opted to focus on building the application layer out and use the big LLMs/models to build out a prototype. Just recently finished the prototype, so I'm back focusing on the model selection and prompts, but seeing my app running, even with unreasonable costs of using the big LLMs/models, has just put more fuel in my tank. A mental game changer. Seriously recommend other entrepreneurs to do the same. That is, focus on building the application layer and build a level of abstraction to enable a plug and play setup.
This was an important video for me also. So is your advice. I have these gut feelings about what we should be doing, and videos like these have the proper language to allow me to intelligently articulate what my gut feelings are telling me.
@@jordanwestmoreland930haha you too! I'll shoot over some details in the near future after I get some solid testing done. Appreciate your interest! Need any feedback I can get!
As a founder in Nairobi, Kenya and I am working on Machine learning for payment security and fraud. In Africa, people are not in the loop about the effect of AI to life in general.
woooooow another Kenyan in these AI related spaces. where are you based. am in Karen, Nairobi. running an indigenous legume nutrition project for children. the org is called Center for nutrition Security-Kenya. we are now integrating REPLIT coding auto-agents and chatgpt-canvas in our administrative tasks. this helps us not hire any ICT or accounting staff-saving labor costs. AI agents will run most companies and waaaay too many Kenyans WILL LOSE THEIR JOBS AND THEY DONT EVEN KNOW IT. so sad.
Incredible overview with a clear eye on the market. Other comments agree but this is an absolute goldmine of insights for anyone building a business in this space
As someone who is building a project that integrates LLMs deeply, I was really very worried about someone stealing my idea until I started working on it, and then I realized how hard it actually was hahaha Great video!
As a founder building a HR tech SaaS infused with a LLM I really enjoyed this. However, I don't think the opportunities to build $B businesses like this are endless.
100% accuracy is not possible. It's a probability test and over fitting will give you results that are too good to be true. You need to test on new data to get more accurate results.
Theoretically, using algo's you can get so close near 100% but costs are the bottleneck. And also it's never truly 100% like you state. But you can get close. 99.9%~
14:00 They had access well before GPT-4 became public. This is why the average Joe cannot reproduce these kind of success stories. You must be at the right place at the right time with the right friends...
As a former attorney who is now in YC, this resonates. Hard to justify the $1k+ hourly rate when you can get 90%+ of the way there in pennies. Cant wait to do this for real estate agents.
Great video! One note: evals are great but I would be careful assuming that if it passes 100 evals you'll get to 100% accuracy, particularly in use cases where there's a potentially adversarial user element.
6:00 bruh if you're working on software on someone else's company time on their computer making software, they own it. Bro was playing a CRAZY game with a law firm 😭
Vertical is about speed. Expand your SaaS as fast as possible, board many users on, and dominate your niche. Quite interesting. Yc bringing another banger. Thanks!
00:04 First-hand experience with AI revolutionizing tasks 02:24 Successful transition to AI technology. 06:18 Starting a company requires time to find the right solution. 08:13 Recommendation algorithms power music platforms like Pandora and Spotify. 12:02 Achieving real product-market fit 13:53 Extended NDA allowed for pre-launch testing and positive feedback 17:27 Customer feedback was crucial in convincing people and driving quick change. 19:05 Accuracy and improvement in AI models 22:31 Legal research process involves diligent search and analysis 24:09 Importance of test-driven development in prompting 27:30 Bridging the gap between technical and non-technical users 29:13 Importance of thorough testing and avoiding 'raw dogging' in engineering 32:37 AI's ability to understand nuanced details 34:24 Investigating prompting AI to think strategically
Great video and insights. It is interesting that they are able to solve Vertical LLM integration into the domain with just Prompt Engineering. Integrating domain specific data into your LLM is still the biggest challenge that needs to be solved for
REPLIT is deploying an even more powerful auto agent with q* based reasoning capabilities. this will handle your concern. stuff will get very very crazy once Q* based reasoning AI agents come online by q1 2025.
Excellent Episode - lot of value. Prompting from a surface level may look simple but that is just tip of the iceberg - specially when it comes to Vertical / domain specific use cases. It is not about just asking questions but it is more of a Art & Science - right context, right question etc..This is very much like skills of Leadership - listening more, asking right question etc....New World, New Skills !!
Yes, of course, there’s LawLLM. Additionally, the most notable fine tuned is Harveyai, which I’ve heard is extremely impressive but I couldn’t get my hands on it to test. It’s based on ChatGPT 4 and could probably be the answer to the country’s lack of resources allocated to civil/family division.l have tried Casetext, but only had a week., and understandably, their tools are not feasible unless you have an established practice. Their tools are indeed impressive. I’m not an attorney, which makes me an even better test case, coming from the perspective of a victims of growing epidemic of pro se litigants in the US. In my case I had 3 attorneys then went pro se and now I need a bankruptcy attorney too
My big takeaway here is SPEED and CONVICTION. If you have a hunch that your edge is time bound, your job is to convince and inspire your team to act now and execute quickly. The other takeaway is leading from the front. When the founder builds something great, it inspires the team.
Yeah, agreed. The fact that they had early access to GPT4 under NDA also provided such an advantage. Seeing the technology and what is was capable of would have put a lot of pressure to create value from it before others.
I think this is how they trained o1. Take a question let model think and reach an answer. Now compare it's answer with correct answer(ground truth). The reasoning which reach correct answer Finetune model with those answers. It's based on paper called star: bootstrapping reasoning with reasoning.
What is really important about AI, like it was important with the internet 30 years ago is to dress in a cool simple fashion, ditch the tie, the suit, the chanel dress and start wearing a black or gray $200 t-shirt.
As a German and EU citizen, one key information is missing: What about data privacy? Is identifying info scrubbed before using it to train or fine-tune LLMs? How is this feasable at such a large scale?
If lawyers are being upgraded, what would stop them from launching legal offensives on more companies and individuals? Could it become economically viable to sue for minor things?
His experience with O1 is the exactly what I saw. The underlying question is with o1, you might need chain of thought prompt engineering anymore. If chain of thought prompt engineering is a big piece of Casetext IP, where does the new technology from the frontier models leave the AI products like Cast text. Case text, along with some others has early access to open AI models, which give them some early mover advantage. yet it still requires someone like Jake to recognize the opportunity, jump on on and even smarter, sell it within 12 months. Smart move!
There is no such thing as 100% accuracy with LLM based systems, that's not mathematically possible. Also in legal, 100% accuracy is not part of the field, it's very much opinion based.
What do you say about the displaced human souls accumulating on the streets? At what point do we return to the idea… … corporations must have some public benefit beyond profit for a few?
Inspiring interview! I'm developing an AI system that creates custom CRM prototypes. Both the agent data and the final CRM are embeded in a well known environment which has the most flexible organized data and on-the-fly code capabilities. The process involves interviewing business owners. Like Casetext, I'm translating domain expertise into AI capabilities. Anyone working on similar AI-driven business tools? I'd love to connect and exchange ideas. Any recommended communities for discussing such projects?
So all you need to build a business that's $650M is 10 years grinding, being a part of alpha programs with cool tech companies, be right in the middle of innovation and pivot right before a shakedown is about to go down. Oh and probably have had a career specialising in something niche would be necessary.
I‘m a lawyer myself building a legal tech product. It was especicially valueable to hear that it was a hard 10 year ride to get to the breakthroughs. You feel lonely on both sides. The legal colelleagues don‘t understand the Tech the Techies don‘t understand how important precision is in legal work. Good things usually don‘t come easy. Thanks for sharing!
@@ronilevarez901 idk maybe. The skeptical part of me knows that when there’s a profit motive, as exists with AI, there will be a lot of grift and hype; See Web 3, blockchain, Apple Vision Pro…
Considering language models are the most significant information technology in the history of the world, yes they are sort of interesting and far more interesting than practically anything else in the world of technology and commerce.
The specifics of actual use cases makes building a general case difficult or potentially impossible right now. The models conceivably can get better but it’s no guarantee
@@ycombinator Right now, sure. But mark my words -- this kind of business will become more and more commoditized. The ones who enable the framework will be the real sustainable business creators.
Me & my boy Sam were raw doggin' some sick prompts with our bros the other day, and he confided in me how o1 works. But he made me swear not to tell you guys. Sorry.
Obviously the term AI has existed for a long time. But even 5 years ago it wasn’t common in the industry to refer to an ML model as an AI. Generally models were specialized enough that they were referred to by the task they performed: sentence encoder model, ham/spam filter model, recommender model, etc. it’s just that LLMs and multimodal models have become so general purpose that describing them as AI has really caught on.
Read about the history of AI. All the AI winters, where AI was a forbidden word and people made fun of you when you mentioned it. Even the word neural networks was forbidden to mention, so they came up with the word deep learning. Machine learning was the term generally used.
Ideal customer profile. This is an avatar of the customer who makes you the most money, who is the best buyer compared to the other customer profiles you have.
I’ve lived through several technology revolutions in my lifetime. The AI revolution is the just the next one. Time 26 to 27 had one of the best insights that a founder needs.
You're going to need very powerful machines to run llama 400, probably costing in the order of several dozens if not hundreds of dollars PER HOUR, PER MACHINE. You'll want at least three in different availability zones. That alone already makes llama 400 out of reach for at least 90% of use cases, certainly +98% of startups. Even if spining up such an expensive compute ring is viable to a given project, you'll need specialized engineers which happen to be among the most in-demand and scarse right now - meaning upwards of USD300,000 / year. On the other hand, OpenAI or Anthropic offers you more capable LLMs, faster time to market, very high scalibility, less financial risks (pay as you go), for very little money in comparison.
@@ycombinator While this is indeed true, the cost of compute is reducing @ an accelerating rate, and there are a growing number of studies showing promising ways of reducing the load on compute, thus reducing the need for compute. The margins in the current game are attracting interest from capital allocators like sharks rushing to a chum ball. So the status quo is at risk, but these are still very early days.
Hey Y combinator , awesome video! Your unique style really stands out. I'm a video editing specialist focused on TH-cam growth, and I can help enhance your content with high-quality edits and engaging thumbnails. If you're ever open to exploring new ideas or strategies that can boost your channel further, I'd love to connect and see how we can collaborate. Let me know your thoughts!
Chapters (Powered by ChapterMe) -
00:00 - Coming Up
00:29 - Intro
01:40 - Building a successful vertical AI company
06:05 - The unique challenges of law and AI
09:24 - The turning point for lawyers with ChatGPT
11:25 - Finding product market fit in legal
15:04 - Entering deep founder mode
20:40 - Approaching prompt engineering step by step
25:05 - Going beyond GPT wrappers
28:10 - Aiming for 100% accuracy
30:48 - Thoughts on o1’s capabilities
36:42 - Outro
1.) ⭐⭐⭐⭐⭐
What is chapterme?
@@rightright6582 We are an AI tool for creating chapters/ timestamps / timecodes / table of contents for TH-cam videos
My respect instantly goes up when successful founders quickly acknowledge the role of luck and timing . And great video too ❤
and knowing how to scale quickly when you run into the lucky timing.
VERY helpful and inspiring interview. Im also building an SaaS app. Initially, my entire focus was trying to find the right models and approaches to making the models work for me, but it became exhausting trying to keep up with the pace of innovation, so I opted to focus on building the application layer out and use the big LLMs/models to build out a prototype. Just recently finished the prototype, so I'm back focusing on the model selection and prompts, but seeing my app running, even with unreasonable costs of using the big LLMs/models, has just put more fuel in my tank. A mental game changer. Seriously recommend other entrepreneurs to do the same. That is, focus on building the application layer and build a level of abstraction to enable a plug and play setup.
Cool name, friend.
I'd love to see this
This was an important video for me also. So is your advice. I have these gut feelings about what we should be doing, and videos like these have the proper language to allow me to intelligently articulate what my gut feelings are telling me.
@@jordanwestmoreland930haha you too! I'll shoot over some details in the near future after I get some solid testing done. Appreciate your interest! Need any feedback I can get!
Very sound thinking. Following the same approach myself
As a founder working on an LLM SaaS right now, this video just answered soo many questions.
SAME
What is ur LLM SaaS bussiness
Like which ones ?
As a founder in Nairobi, Kenya and I am working on Machine learning for payment security and fraud. In Africa, people are not in the loop about the effect of AI to life in general.
What s your startup name?
woooooow another Kenyan in these AI related spaces. where are you based. am in Karen, Nairobi. running an indigenous legume nutrition project for children. the org is called Center for nutrition Security-Kenya. we are now integrating REPLIT coding auto-agents and chatgpt-canvas in our administrative tasks. this helps us not hire any ICT or accounting staff-saving labor costs. AI agents will run most companies and waaaay too many Kenyans WILL LOSE THEIR JOBS AND THEY DONT EVEN KNOW IT. so sad.
Incredible overview with a clear eye on the market. Other comments agree but this is an absolute goldmine of insights for anyone building a business in this space
As someone who is building a project that integrates LLMs deeply, I was really very worried about someone stealing my idea until I started working on it, and then I realized how hard it actually was hahaha
Great video!
what is your idea, lol
@@randomperson619 Nice try, PERSON WHO IS TRYING TO STEAL MY IDEA! ;-)
🙂this
@@MattSimmonsSysAdminlol 😂
Now this is what i love deep dive with founders themselves.
As a founder building a HR tech SaaS infused with a LLM I really enjoyed this. However, I don't think the opportunities to build $B businesses like this are endless.
If you get 100% accuracy with an LLM based system, it means you haven't tested enough.
100% accuracy is not possible. It's a probability test and over fitting will give you results that are too good to be true. You need to test on new data to get more accurate results.
Theoretically, using algo's you can get so close near 100% but costs are the bottleneck. And also it's never truly 100% like you state. But you can get close. 99.9%~
Our products have reached 98% accuracy for healthcare applications
@@mohammednisham3211 fair enough. That means out a 1000 cases you get 20 failures. That's not 100%. Also, how do measure accuracy?
@@pawsjaws on a test set, perhaps. In reality, unlikely
14:00 They had access well before GPT-4 became public. This is why the average Joe cannot reproduce these kind of success stories. You must be at the right place at the right time with the right friends...
Some people look for a reason why they couldn't have done it and some look for a way to do it.
I totally agree with you, but don’t hate the player. Hate the game !
You had me at “injecting domain expertise” - kudos and thanks for sharing the lessons learned
Super episode. Love these deep dives with real founders.
Thanks for all that it took to put this together.
As a former attorney who is now in YC, this resonates. Hard to justify the $1k+ hourly rate when you can get 90%+ of the way there in pennies. Cant wait to do this for real estate agents.
why hasn't this been done already with real estate agents?
Imagine how good their sales team is...trying to sell and cut deal with lawyers, their company should be the hyperbolic time chamber for sales pros.
Great video! One note: evals are great but I would be careful assuming that if it passes 100 evals you'll get to 100% accuracy, particularly in use cases where there's a potentially adversarial user element.
6:00 bruh if you're working on software on someone else's company time on their computer making software, they own it. Bro was playing a CRAZY game with a law firm 😭
Vertical is about speed.
Expand your SaaS as fast as possible, board many users on, and dominate your niche.
Quite interesting.
Yc bringing another banger.
Thanks!
00:04 First-hand experience with AI revolutionizing tasks
02:24 Successful transition to AI technology.
06:18 Starting a company requires time to find the right solution.
08:13 Recommendation algorithms power music platforms like Pandora and Spotify.
12:02 Achieving real product-market fit
13:53 Extended NDA allowed for pre-launch testing and positive feedback
17:27 Customer feedback was crucial in convincing people and driving quick change.
19:05 Accuracy and improvement in AI models
22:31 Legal research process involves diligent search and analysis
24:09 Importance of test-driven development in prompting
27:30 Bridging the gap between technical and non-technical users
29:13 Importance of thorough testing and avoiding 'raw dogging' in engineering
32:37 AI's ability to understand nuanced details
34:24 Investigating prompting AI to think strategically
Great video and insights. It is interesting that they are able to solve Vertical LLM integration into the domain with just Prompt Engineering. Integrating domain specific data into your LLM is still the biggest challenge that needs to be solved for
REPLIT is deploying an even more powerful auto agent with q* based reasoning capabilities. this will handle your concern. stuff will get very very crazy once Q* based reasoning AI agents come online by q1 2025.
@@paulmuriithi9195
Nothing came up with on google when I tried looking up for "replit q*"
Excellent Episode - lot of value. Prompting from a surface level may look simple but that is just tip of the iceberg - specially when it comes to Vertical / domain specific use cases. It is not about just asking questions but it is more of a Art & Science - right context, right question etc..This is very much like skills of Leadership - listening more, asking right question etc....New World, New Skills !!
Love Garry Tan's comment against the fine-tuning discussions at the end 🤣
when it's all about speed and you have early preferential access to the latest model... i wonder if there's a case-law llm to run this by ;)
Yes, of course, there’s LawLLM. Additionally, the most notable fine tuned is Harveyai, which I’ve heard is extremely impressive but I couldn’t get my hands on it to test. It’s based on ChatGPT 4 and could probably be the answer to the country’s lack of resources allocated to civil/family division.l have tried Casetext, but only had a week., and understandably, their tools are not feasible unless you have an established practice. Their tools are indeed impressive. I’m not an attorney, which makes me an even better test case, coming from the perspective of a victims of growing epidemic of pro se litigants in the US. In my case I had 3 attorneys then went pro se and now I need a bankruptcy attorney too
My big takeaway here is SPEED and CONVICTION. If you have a hunch that your edge is time bound, your job is to convince and inspire your team to act now and execute quickly.
The other takeaway is leading from the front. When the founder builds something great, it inspires the team.
Yeah, agreed. The fact that they had early access to GPT4 under NDA also provided such an advantage. Seeing the technology and what is was capable of would have put a lot of pressure to create value from it before others.
Luck. He said it right in the beginning.
Outstandingly real and useful.
The prompt engineering tips were super helpful. Gonna apply the principles to my genAI app!
I think this is how they trained o1.
Take a question let model think and reach an answer. Now compare it's answer with correct answer(ground truth). The reasoning which reach correct answer Finetune model with those answers.
It's based on paper called star: bootstrapping reasoning with reasoning.
What is really important about AI, like it was important with the internet 30 years ago is to dress in a cool simple fashion, ditch the tie, the suit, the chanel dress and start wearing a black or gray $200 t-shirt.
This dude has so much zen
As a German and EU citizen, one key information is missing: What about data privacy? Is identifying info scrubbed before using it to train or fine-tune LLMs? How is this feasable at such a large scale?
Do you guys know that your feed is out of date? on Spotify and on RSS it only shows episodes up until April 22.
Very helpful for new founders 👌
I am not sure If anyone noticed, Jake's voice is very similar to Andrew Ng.
Well I guess the main factor here is the access to GPT4 before it’s out because you’re “close to the folks at OpenAI” 😅
Thanks for answering my questions on my startup)
If lawyers are being upgraded, what would stop them from launching legal offensives on more companies and individuals? Could it become economically viable to sue for minor things?
Man I don't know but I feel like this comment is causing me pain (points to back) here and suffering (points with other hand to back) here 🎉💰 🤑
Lightcone ! ❤ what a name!
"I do not want nothing to change" - it's funny how expressions like this mean the complete opposite of what is intended.
His experience with O1 is the exactly what I saw. The underlying question is with o1, you might need chain of thought prompt engineering anymore. If chain of thought prompt engineering is a big piece of Casetext IP, where does the new technology from the frontier models leave the AI products like Cast text. Case text, along with some others has early access to open AI models, which give them some early mover advantage. yet it still requires someone like Jake to recognize the opportunity, jump on on and even smarter, sell it within 12 months. Smart move!
what is the difference between condos for sale and condos unsold?
Good one, YC!
Great stream
thank you I watch all your episodes
what is a Vertical LLM Agent?
Going deep into solving a single problem = going vertical
any chance can provide us the website of vertical ai agent, so that we can learn more about it. thank you
“At the time it was not called AI, just NLP and Machine Learning” 😂
100% accuaracy for legal sector is a huge feat.
They are not at 100%, but yes it is a big feat to strive for
There is no such thing as 100% accuracy with LLM based systems, that's not mathematically possible. Also in legal, 100% accuracy is not part of the field, it's very much opinion based.
What do you say about the displaced human souls accumulating on the streets?
At what point do we return to the idea…
… corporations must have some public benefit beyond profit for a few?
How do you think they get revenue if not by creating value for millions of other people?
…please clarify, who is the they to which you refer?
It's a pleasure to watch.
Thanks ❤🙏
Great video thank you for this
This is really great
Good stuff. Were any live cases won using AI ppowered Casetext Co-counsel ?
Thank you! 🙏
Very interesting!
Superbly！
The best - thanks!
the url slays
Love this
i'm a designer spending all of my energy to learn, build, and invest in this space.
Thank you!
So it is entirely possible to build a accurate LLM-powered system! If it is good enough to pass in the legal domain... then it's pretty good...
Inspiring interview! I'm developing an AI system that creates custom CRM prototypes. Both the agent data and the final CRM are embeded in a well known environment which has the most flexible organized data and on-the-fly code capabilities. The process involves interviewing business owners. Like Casetext, I'm translating domain expertise into AI capabilities. Anyone working on similar AI-driven business tools? I'd love to connect and exchange ideas. Any recommended communities for discussing such projects?
Find anything? In the space as well.
Always wondered if an LLM could be trained on symbolic logic rules to identify logical fallacies in legal arguments.
great Q. can it not already do that?
Phenomenal
So all you need to build a business that's $650M is 10 years grinding, being a part of alpha programs with cool tech companies, be right in the middle of innovation and pivot right before a shakedown is about to go down. Oh and probably have had a career specialising in something niche would be necessary.
This is a good one.
I‘m a lawyer myself building a legal tech product. It was especicially valueable to hear that it was a hard 10 year ride to get to the breakthroughs. You feel lonely on both sides. The legal colelleagues don‘t understand the Tech the Techies don‘t understand how important precision is in legal work. Good things usually don‘t come easy. Thanks for sharing!
Are LLMs and AI the only interesting things to talk about now?
Yes, because nothing else you're doing right now is getting the traction and value they're getting...Learn!!!
AI is the last thing we will ever need, so yeah.
@@ronilevarez901 idk maybe. The skeptical part of me knows that when there’s a profit motive, as exists with AI, there will be a lot of grift and hype; See Web 3, blockchain, Apple Vision Pro…
Considering language models are the most significant information technology in the history of the world, yes they are sort of interesting and far more interesting than practically anything else in the world of technology and commerce.
Ai for now till we can talk robots and are market ready
The best teammate you didn’t know you needed: AI 🤩
Hi guys please can someone tell me how this works
I need to watch this every day,
Folks please like my comment so I can get notified every month to check this video :D
The video is good, but the title doesn’t correspond the content.
Notice that he is a lawyer by trade and build a law-related project. Too many AI projects are being done by people who have no domain knowledge.
Why even build this when you could build the framework that creates this for any vertical instead and let the customer train it on their data?
The specifics of actual use cases makes building a general case difficult or potentially impossible right now. The models conceivably can get better but it’s no guarantee
@@ycombinator Right now, sure. But mark my words -- this kind of business will become more and more commoditized. The ones who enable the framework will be the real sustainable business creators.
@@mastershredder2002 How would you approach building this framework?
Me & my boy Sam were raw doggin' some sick prompts with our bros the other day, and he confided in me how o1 works. But he made me swear not to tell you guys. Sorry.
Cool story bro
Which gpt model r u?
You better tell me.
" George, is getting upset 😡 "
"$1 billion?" Hope that's a typo... more like $100 billion
Did he say, "...back then we didn't call it AI." Was that some time before 1956?
Obviously the term AI has existed for a long time. But even 5 years ago it wasn’t common in the industry to refer to an ML model as an AI. Generally models were specialized enough that they were referred to by the task they performed: sentence encoder model, ham/spam filter model, recommender model, etc. it’s just that LLMs and multimodal models have become so general purpose that describing them as AI has really caught on.
Read about the history of AI. All the AI winters, where AI was a forbidden word and people made fun of you when you mentioned it. Even the word neural networks was forbidden to mention, so they came up with the word deep learning. Machine learning was the term generally used.
ICP?
Ideal customer profile. This is an avatar of the customer who makes you the most money, who is the best buyer compared to the other customer profiles you have.
Me too
16:00
Wait till the IRS gets ahold of this tech :"(
I’ve lived through several technology revolutions in my lifetime. The AI revolution is the just the next one. Time 26 to 27 had one of the best insights that a founder needs.
What lessons have you learned from going through the previous tech revolutions?
Por favor traducir al español
Busca en Settings > Captions > Spanish (auto translate). No es 100% precisa pero te ayudará.
Fine tuning is still necessary - whoever think that's a waste of time is just stupid.
💛
$650 million in cash 😮 why cash ?
Meta is making superpowerful AI available for FREE and so how are these valuations realistic?
Nothing is free. Compute costs money. Smart prompts are hard to get right.
You're going to need very powerful machines to run llama 400, probably costing in the order of several dozens if not hundreds of dollars PER HOUR, PER MACHINE. You'll want at least three in different availability zones.
That alone already makes llama 400 out of reach for at least 90% of use cases, certainly +98% of startups.
Even if spining up such an expensive compute ring is viable to a given project, you'll need specialized engineers which happen to be among the most in-demand and scarse right now - meaning upwards of USD300,000 / year.
On the other hand, OpenAI or Anthropic offers you more capable LLMs, faster time to market, very high scalibility, less financial risks (pay as you go), for very little money in comparison.
tell me how you run llama 3.1 400b for free in your computer?
the most valuable resource in the world, surpassing oil back in ~2016ish, is data. Think about it.
@@ycombinator While this is indeed true, the cost of compute is reducing @ an accelerating rate, and there are a growing number of studies showing promising ways of reducing the load on compute, thus reducing the need for compute. The margins in the current game are attracting interest from capital allocators like sharks rushing to a chum ball. So the status quo is at risk, but these are still very early days.
🎉
😂 the new bubble. Let's burn investor's money
"I will defeat Elon Musk before anyone could"
Elon is Elon and himanshu is himanshu...two different people with different strengths ....find yours 🎉
@@dpdeepnan yep, everyone want to be the forehead of AI, I want to be the forehead at cosmic level
@@0xhimanshuoh brother
@@nullvoid12 himanshu entered the chat
😅😅😂🤣🤣
😂🎉🎉
Ahm, error propagation.
bye.
Amen !! God bless AMERICA!! Harris Vance, Trump Walzes !!!
Hey Y combinator , awesome video! Your unique style really stands out. I'm a video editing specialist focused on TH-cam growth, and I can help enhance your content with high-quality edits and engaging thumbnails. If you're ever open to exploring new ideas or strategies that can boost your channel further, I'd love to connect and see how we can collaborate. Let me know your thoughts!