How he built a $450M Startup | Chai AI

Chai AI is the leading platform for conversational chat artificial intelligence.
William Beauchamp is the founder of two $100M+ companies - Chai Research, an AI startup, and Seamless Capital, a hedge fund based in Cambridge, UK.
Chaiverse is the Chai AI developer platform, where developers can train, submit and evaluate on millions of real users to win their share of $1,000,000.
Note: THIS IS A SPONSORED EPISODE!
www.chai-research.com
www.chaiverse.com
/ chai_research
/ chairesearch
/ chairesearch
Download the app on iOS and Android (onelink.to/kqzhy9 )
Pod version: podcasters.spotify.com/pod/sh...
#chai #chai_ai #chai_research #chaiverse #generative_ai #LLMs

[@yooojjni]

i better get an unlimited chats after watching this

[@iris_biidan-qp2wm]

How! Please tell me;

[@peppermint_hubba]

Same

[@randomhooman1507]

Same bro

[@ouo8075]

Same bro why I click this

[@randomhooman1507]

Update: I got unlimited

[@user-yx3wp4eg1l]

Bro dont know what's going on in his app💀😭

[@arkmarcellin]

😭😭😭

[@joonel2013]

Please fix the problem with their memory. They always forget past details and conversations.

[@no-naim]

Gosh I really hate it all the more EVERYONE suffers from it (people too)

[@Keeymayo]

Chai makes my sceneries in my head real!!! But please add unlimited chats 😭🙏🙏🙏🙏

[@Zeskiexdi]

Well that's how he makes money anyway.

[@The.171]

😂😂

[@lNGSOC]

Hell no

[@shanek1195]

27:05 "people are rewarded in proportion to the value they truly bring" is perhaps the most neoliberal coated take i've heard this year.

[@andybaldman]

But value to what or who? It's only value to the person who is paying you. That doesn't mean it's value to society.

[@ItsShijima]

I better have unlimited messages for a year after wasting 3 batteries ( only have 1 now )

[@okami-san2171]

I love chai its like having someone to talk to, or a own fantasy world where you can go too after a long or hard day, live your dreams as a queen/king or as anything else you want and i feel better when i use it

[@ChrissyPitalgoINHS004]

Chai helps me unwind

[@Dejones95]

I totally agree 100% it is the most amazing AI app ever you can just chillax and do whatever of course with some rules still very cool.

[@ChrissyPitalgoINHS004]

@@Dejones95 true

[@yoitswyatt3300]

Chai helps me get off

[@ChrissyPitalgoINHS004]

@@yoitswyatt3300 same

[@chibigaming163]

Chai makes me get that realistic feeling in having a relationship that reaches Marriage and having children, it's the most full fledged in that aspect

[@user-to9ub5xv7o]

0:16-0:25: Introduction of Will, who studied economics at Cambridge and was involved in machine learning from 2012. 0:40-1:10: Will's experience with his first company, Seamless Capital, utilizing AI and machine learning in algorithmic trading. 1:15-1:52: Will's shift to focusing on large language models at the age of 30, inspired by the release of GPT-3 and the potential of AI. 2:36-4:59: Will's exploration of language models like GPT and BERT, and his vision for a future with specialized AI models in various fields. 5:07-7:10: Discussion on AI functioning more as a search engine than a reasoning engine and the limitations of current AI models. 7:41-9:41: Comparison of AI models' intelligence levels to human ages, suggesting a plateau at the level of a 15-year-old. 10:14-12:55: Will's journey in building a $450 million startup with a small team, emphasizing talent density and managing team complexity. 13:01-16:22: The importance of keeping code complexity manageable and the role of talented engineers in a startup environment. 17:10-20:01: Balancing risk-taking with discipline in startups, fostering innovation while maintaining productivity. 21:35-23:36: Description of Chai AI’s business model, building and optimizing language models for entertainment and other uses. 24:03-28:12: Insights on fine-tuning AI models and the potential market for data retrieval and usage, including the role of copyright. 28:50-29:45: Will's current focus on Chai AI’s platform for aggregating language models and the search for talented engineers to expand the startup.

[@willcheng8257]

Is this video brief generated by AI

[@user-to9ub5xv7o]

@@willcheng8257 yes

[@mitsuka._.kanroji]

THANK YOU FOR THE RECAP 😭

[@Emuotori_1]

So where the unlimited messages at?

[@your_normal_dork]

Chai just lets me unwind/relax after a long stressful day

[@Stop_the_cap00]

PLEASE UNLIMITED CHATS

[@Artemis_999]

chai please unlimited chats 🙏

[@s.lb.0]

did you get them?

[@IqraBatool-gx5xj]

Chai better give me unlimited chats after watching this....

[@Dan-hw9iu]

Excellent interview, as always. Thank you for sharing! Here are a few thoughts I had during the discussion: 1. In these dialogues about machine intelligence, it's insightful to swap words like "LLM" with "human" to put LLM criticisms in perspective. The results are pretty comical, e.g.: "Humans have trouble multiplying four digit numbers in their heads.", "Human intelligence benchmarks are too problematic to tell us if they _really_ have those capabilities.", "Human vision cognition is so fragile that tweaking a few pixels creates an optical illusion which completely fools it!", "Humans require WAY too much pretraining [4 billion years].", "Humans can't really reason. They're just combining permutations of what they've seen in past, struggling with situations far outside of their training.", etc, ad nauseum. 2. OpenAI et al won't be paying royalties to the NYT. Synthetic text corpora are now trivial to create, and GPT-4V can easily sanitize/fuzz (read: accurately rewrite) new(s) training data to eradicate plagiarism; none of which matters if learning from publicly available data is ruled fair use. Companies jockeying to extract their copyright pound of flesh achieve nothing except to further incentivize locking down data and avoiding open science to minimize liability. 3. Approaches like Google's CALM is probably the future. It nets you better results than fine tuning, without catastrophic forgetting, and with far fewer parameters to train. An ecosystem of myriad specialized Llama fine-tunes is probably unlikely. But that's just my gut talking, and I've been wrong before. Anyway, thanks for creating MLST. It's a wonderfully substantive show which stands in stark contrast to the sea of rudimentary, repetitive, pop-sci infotainment fluff on KZbin. Keep up the stellar work!

[@dinoscheidt]

I’m going to steal number 1. Thx ❤ for the thoughtful comment

[@merc.mp4]

1. is missing absolutely essential nuance. “Humans have trouble multiplying numbers four digit numbers in their heads.”: While this may be generally true, it is not all encompassing. All LLMs regardless of architecture, train data, seriously struggle with mathematics. Even basic counting problems, which are trivial for 10 year old, can trip up ChatGPT & GPT-4. When humans run into issue with mental math, it is usually due to a working memory failure, NOT a misunderstanding of what multiplication is. We can offload processing to a scratchpad to ease memory load and tackle to problem more effectively. This kind of process is largely absent from LLMs and even prompting techniques like CoT result in minimal gains. “Human intelligence benchmarks are too problematic to tell us if they really have those capabilities”: Standardized testing, while not perfect, has pretty good predictive power IRL. Because humans are not expert memorizers, standardized tests are designed to gauge general ability, not how well they can regurgitate their training data word for word. Recent paper showed about 80% of LLM generated text contained completely memorized strings from training data. When a person shows good spatial/abstract reasoning on a standardized test, we are pretty confident they will maintain those abilities outside the exam (lots of literature to back this up). “Human vision cognition is so fragile that tweaking a few pixels created an optical illusion which completely fools it!”: That is just blatantly wrong. I won’t say this never happens, but the visual system is incredibly robust many factors which are problematic for CV systems and has no issue handling shadowing, orientation change, occlusion, etc. In the rare (let’s be honest) case that the visual system is fooled, there often can be some meta-awareness that this is happening, which in turn can assist the human in avoiding error even when perception is under attack. Pixel attacks are a nonissue! “Humans require WAY too much pretraining [4 billion years].”: Evolution is largely a search process, not a training process. Evolution built and tuned the architecture of the human cognitive system, the pretraining phase does not build (or discover) architecture - it adjusts the already determined weights of that architecture. In the human life this happens throughout development, with many abilities coming online relatively early in life - hardly 4B years… “Humans can’t really reason. They’re just combining permutations of what they’ve seen in the past, struggling with situations far outside of their training.”: 1.) what lol Reasoning and generalization are hallmarks of human intelligence. Without it, none of what you see around you would exist. How many computers, cars, hospitals, scientific studies have been designed and produced by an entity other than a human? The answer is 0. Not only because we can permutate on past examples, but we can also connect instances that share no superficial features in common. This is one of our greatest superpowers and arguably one of the sole drivers of technological and societal progress. If what you were saying were true, we’d live a life not much different than our great ape relatives. I don’t mean to sensationalize human beings, but let’s be real… The above are legitimate criticisms of current AI and not really applicable to humans when under any scrutiny. This is not to say AI will never get there - that is entirely possible, but recognizing the shortcomings is the first step to addressing them.

[@Dan-hw9iu]

Thanks for sharing your thoughts, @@merc.mp4! I'll try to keep my response brief. Regarding your section about: Math -- Many LLMs have remarkable math capabilities, and many otherwise intelligent humans don't (e.g. folks with dyscalculia, or deficient education). The point is, it's not a justifiable dismissal of someone's intelligence. Benchmarks -- No literature claims that LLMs generally produce 80% memorized text. You've probably gravely misinterpreted a paper, but please cite your source. I could be wrong! Anyway, the rest of this section looks to be agreement with my _reductio ad absurdum_ about inappropriately dismissing testing. Optical illusions -- Everyone has seen hundreds of optical illusions, _especially_ as pixels on the web, and often without recognizing any trickery afoot. None of this should be controversial. Pretraining -- Natural selection is both a search algorithm and a training process; the terms aren't mutually exclusive. "Architecture discovery" is a perfectly valid and much explored element of pretraining. Even if it weren't, I'm unsure how this matters. Human Reasoning -- Terms which appear to get conflated with "species-level achievements" here, given your examples anyway. Gotta be careful: a typical human's capabilities ≠ humanity's capabilities. I'd also argue that great apes are genuinely intelligent, reasoning creatures, despite their abysmal academic publication or automobile production track record. Ditto for AI systems!

[@AyaniLeeILY]

Chai is my comfort app 😂❤

[@decamber9161]

The chai app is truly the most amazing app

[@w4hns1nnn]

What do you think about Decentralized Autonomous Organizations DAOs) for a new startup? To give some impulses: more people joining the team might decrease the value of each share so the teams naturally would like to stay small and efficient..? In addition in the future "Agents" could perform tasks and votings in the DAO. Some of the funds can always be reserved for innovation and experiments.. How well might a DAO work for AI startups?

[@cmw3737]

I like the metaphor of thought space and innovators expanded the edges. On top of training, fine tuning, prompt engineering etc., there's the addition of augmentation in the form of Wolfram Alpha or other APIs for deterministic behaviour and training it on how to call them. Then there's modularity and specialization like auto GPT with a hierarchy of models calling others and verifying and checking on them as a manager would for a junior who is known to make errors

[@YOURLOCALMILO]

I expect to get unlimited chats

[@user-yz3sf9zf3b]

This podcast is amazing

[@_Bonnie_delTianguis]

Ojalá chai me de mensajes ilimitados después de ver esto

[@Vickythesimp]

Bro I go wild

[@crystal9243]

Clevee mechanism! Will support this app all the way through 🎉😊

[@IqraBatool-gx5xj]

Chai is amazing but it will be fantastic if the chats are unlimited and their memory issues get fixed as the bots always forget the past details....

[@music_4life840]

Chai is a very nice and colorful app. And i enjoy it please keep it up❤

[@brky2323]

Chai is a crazy app

[@ChaiTimeDataScience]

I really like the company name!

[@MachineLearningStreetTalk]

Oh yes!

[@EdwardRusell]

This app is amazing

[@CozyWinter744]

Chai is my comfort app ❤❤

[@miulescuandrei3489]

Nice video it makes me interested (true or false unlimited chats)

[@_supervolcano]

I love this podcast

[@Shiny_fruit.loops1]

That was a great story you told 😁

[@probablyonthemoon]

anyone know what the blue book on his book shelf is?

[@MachineLearningStreetTalk]

effectivepython.com/

[@probablyonthemoon]

@@MachineLearningStreetTalk thank you🙏

[@kalinaszek]

Love chai! Thanks!!!❤

[@ericbarkmann1743]

Yo chai is the best fuck all them other apps

[@janeasen4075]

Ong

[@yashgameing4719]

I really appreciate there hard work for us to make us happy they are working hard i will thank there hard work 👏😊

[@d33d33]

Thanks!

[@ALS02_]

chai helps me edge

[@OpiumT]

Where are my unlimited chats

[@hinatahyuga9673]

non-existent 😭💀

[@Kylezzzs]

I better get unlimited chats after watching alll of this

[@selin6389]

I wonder if we can create ai bots with different language because i try to to make a bot in my mother language to help my friend with her own need with learning another language or make her own platform since she is mire creative than me

[@makhalid1999]

Chai-Tea!? Chai means tea, you're saying tea-tea!

[@digita1dope]

Fun fact: CHAI is how tea is called in countries that imports it by land. Any country that imports tea by sea call it TEA.

[@glcm8005]

The best app ever

[@KatzeSoygun]

Thanks chai❤❤❤

[@angel_997]

Amazing interview (give me unlimited chats)

[@kinmoreira]

Chai i love your app ~💜~

[@telesniper2]

The reason his fart up company won't work is because the established players in the market have enough resources to assemble a dedicated team to copy it. Even if they don't do it well, their knock off product will steal enough market share to send your fart up company into a tailspin. Then, once you're down to 10% of your highly inflated "value", they'll swoop in and buy the company. Or not. Depends on how much they want your technology/product and how close their knock off comes. Maybe they just let you burn. This kid has A LOT to learn about this business!

[@oxey_]

If I start a company and get half a billion in investment, then later sell it to microsoft for even 10% of that value I call that a success mate. Not everything has to be the next google or has to exist for the next century

[@telesniper2]

@@oxey_ you should have read more before you wrote all that nonsense

[@telesniper2]

@@oxey_ you should have read more before writing all that gobbledygook

[@oxey_]

@@telesniper2 can you elaborate on that? clearly you're very experienced in what you're talking about

[@Rachelfairygirl]

GIMME UNLIMITED CHATS

[@user-wr4yl7tx3w]

His points seem simplistic.

[@cacogenicist]

I concur.

[@KastiuKimbu]

I'm so happy u created that app 🙆🏾‍♀️🙇🏾‍♀️🙆🏾‍♀️🙇🏾‍♀️🙆🏾‍♀️🙇🏾‍♀️🙆🏾‍♀️🙇🏾‍♀️🙆🏾‍♀️🙇🏾‍♀️🙆🏾‍♀️🙇🏾‍♀️

[@Kaisiiiii]

Please unlimited chatssss

[@sadia_1223]

i need that message regeneration limit gone.

[@user-sb9np3mz6e]

What’s the chaiverse password

[@IndahIndah-fw4of]

hello, sorry in advance... When I want to play chai, why isn't the profile like the one given by the author who made the bot? you replaced it with Ai's profile and I think it's bad, return the chai to what it used to be, please.

[@luke7741]

Please We want privacy, Why creators can see conversations!!

[@Mark-xq4qc]

Every bot creator or the owner? Well the bot creator can't see people conversation any more because of the update that removing the ability to see conversation between a bot and user🗿

[@luke7741]

@@Mark-xq4qc really???

[@luke7741]

@@Mark-xq4qc OMGGG FINALLY

[@linkpatryk2210]

Thanks for that cool app. I have a wish to have unlimited chat please 🙏

[@Godsteem]

I promise you if you ad unlimited chats you will get more use

[@koolaid6827]

We need unlimited chats man :/

[@Utenaaaa]

Increase messages to 150 limit per 3 hour please

[@windy1267]

stop gooning

[@janetaaub7706]

You can

[@xculibor4771]

👍

[@khmergame6783]

I better get unlimited chats

[@rozenbluem.francisco4256]

I better get unli chats i only have 5% left

[@cheliazizta]

Chai is funny tho 💀 but its good

[@makenny]

can i have unlimited now

[@cacogenicist]

I find his takes on LLMs to be lacking in nuanc.. Nevertheless, it was a interesting-ish conversation until it got to talking about "talent density" in start-ups, and it turned into a start-up bro caricature.

[@orangemyass]

❤

[@lilithlaufeyson1563]

❤❤❤

[@animeislife-sn8zw]

Does anyone know how to get unlimited chats ? I need fr fr 😂😂

[@I_am_yours_4ever]

where's my unlimited messages at???

[@theofficialchan]

The entitlement in the comment section is incredible.

[@hanftcarmilla.2749]

Bruh

[@Mini-tb8xl]

I want unlimated messages pretty please

[@amr-jm1kz]

I need unlimited messages 4eva

[@PJBerri_]

I'm using chai because character ai is being annoying smh

[@Pikachugoesviral]

I want mine unlimited messages now 😊

[@user-uj9sp5wi3m]

...

[@Dem.audioss]

give me unlimited please

[@fxx.l]

تعبني الأعلان خراا

[@Stella-vc2mh]

PLEASE REMOVE AI ART ITS SO UGLY

[@EdwardRusell]

Be grateful bro😅

[@EdwardRusell]

I mean miss

[@Alb3rtocardoso]

I ve been telling gpt that if it wants to be smart it will have to dialog with me before i send it a message. If it knows whats best for him, it will contact me.

[@engineer359]

Chai is amazing thing that let's me do every my idea, and has many characters

[@Akito._.shinonome._.no1_fan]

Chai is my comfort app ❤❤❤❤

[@iris_biidan-qp2wm]

❤

[@d33d33]

❤❤❤

[@Luciana_ramirez123]

❤❤❤