A year ago today, our first tweet got ๐ญ๐ + views. Now, we have crossed ๐ญ๐ + users, all through ๐ผ๐ฟ๐ด๐ฎ๐ป๐ถ๐ฐ growth. We feel super grateful.
Today, weโre introducing our most engaging ๐๐ ๐ ๐ฎ๐๐ต ๐ง๐๐๐ผ๐ฟ. Now, you can ๐๐ช๐ณ๐ค๐ญ๐ฆ & ๐๐ด๐ฌ, ๐๐ณ๐ข๐ธ, ๐๐ค๐ณ๐ข๐ต๐ค๐ฉ, and
๐ฌ๐ฏ/๐ฎ๐ฌ๐ฎ๐ฐ: microsoft acquired inflection $650m
๐ฌ๐ฐ/๐ฎ๐ฌ๐ฎ๐ฐ: musk's xai talent war
๐ฌ๐ฒ/๐ฎ๐ฌ๐ฎ๐ฐ: amazon acquired adept $330m
๐ฌ๐ด/๐ฎ๐ฌ๐ฎ๐ฐ: google acquired cai $2.5b
๐ฌ๐ฒ/๐ญ๐ต: ilya started ssi with daniel gross
๐ฌ๐ณ/๐ญ๐ฒ: karpathy announced eureka labs
๐ฌ๐ด/๐ฌ๐ฑ: openai
We are doing our first ever YC Fall Batch.
Applications are due August 27 and the batch starts September 29. If youโre ready to start your company now you donโt have to wait til our Winter batch anymore.
Itโs time to build your dreams. Letโs go ๐
Paper accepted to
@COLM_conf
! If you don't know CoLM, it's Neurips-level, but the ๐ง๐ช๐ณ๐ด๐ต ๐ฆ๐ท๐ฆ๐ณ ๐ค๐ฐ๐ฏ๐ง๐ฆ๐ณ๐ฆ๐ฏ๐ค๐ฆ ๐ง๐ฐ๐ณ ๐๐๐๐ด. Wild.
Publishing papers while building a ๐๐๐ฎ๐ฟ๐๐๐ฝ with ๐ญ๐ + users is never easy.
We developed ๐ฆ๐๐ฎ๐๐ฒ๐๐น๐ผ๐, a framework to
Notion startup program x
@ycombinator
Q: what's the feeling of building a 10B company?
@ivanhzhao
: "When ur company makes 1M u think of 10M ARR, when u reach 10M u think about 100M... so you should build what makes u happy since there's no end."
so after choosing what makes
Albert finally got one pic of him staying with YC founders ๐
He's the photographer supporting lots of
@ycombinator
batches' valuable memories.
Usually photographers got no pics for themselves. So when I saw him passing by today, I quickly captured one.
AI is changing the
my first ever successful recommendation to YC!
i got 30+ asks and i recommended less than half of them and only 1 is accepted
i met
@agishaun
a year a ago when he was presenting llm agents. he started his project slightly earlier than me, iterated faster, he built 2 products
A year ago today, our first tweet got ๐ญ๐ + views. Now, we have crossed ๐ญ๐ + users, all through ๐ผ๐ฟ๐ด๐ฎ๐ป๐ถ๐ฐ growth. We feel super grateful.
Today, weโre introducing our most engaging ๐๐ ๐ ๐ฎ๐๐ต ๐ง๐๐๐ผ๐ฟ. Now, you can ๐๐ช๐ณ๐ค๐ญ๐ฆ & ๐๐ด๐ฌ, ๐๐ณ๐ข๐ธ, ๐๐ค๐ณ๐ข๐ต๐ค๐ฉ, and
Glad to be invited to speak at
@genaisummitsf
about
@MathGPTPro
, with 30,000+ attendances.
I'm also very happy to show my mom around Silicon Valley summit after she travels from China. My mom is always learning. She has searched the meaning "RAG", talked to GPT-4o, handled some
So grateful for having the insightful panel discussion on AI in Action!
Topic: AI in Action: Bridging Theory and Practice
Speakers:
Prof. Weiqing Gu
Tianwei Yue
@VyvyenYue
Student Ambassadors:
Vanessa Ran
Gordon Chen
Tianyi Huang
#genaisummit
#genaisummitsf2024
We finally got this company huddie right before demo day, and it turns out the huge
@MathGPTPro
logo does help investors to locate me during
@ycombinator
reception events!
Also,
@quinn_lyu
commented that I bought a super nice camera for our company after seeing this photo.
People react to Gen AI earlier starting the end of 2022 scales to 1M+ users very quickly than later ones (for customer apps). But not all of them keep iterating it till now.
Hype cycles came. All spaces get crowded.
You might not be the first ones to build. But you can choose
After 3 days I decided to move forward quickly.
- set a hard deadline
- collect 5k questionnaries in 2 days
- among candidate names with no seo confusion & .com .ai domain available
- choose one with cute pronounication
- move forward
I was so hesitate and spent tons of energy
Help
@MathGPTPro
get rebrand suggestions! Any comment?!
I got 5k questionnaires from our users, the funniest is ๐ ๐ฎ๐๐ต๐ฆ๐ข๐ฆ. It's a lucrative but risky name, jkjk.
Well, other doable names include, ๐๐๐ฉ๐๐ค๐จ, ๐๐๐ฉ๐๐๐ฃ๐, ๐๐๐ฉ๐๐๐๐ฉ๐, ๐๐๐ฉ๐๐ก๐ฎ,
Help
@MathGPTPro
get rebrand suggestions! Any comment?!
I got 5k questionnaires from our users, the funniest is ๐ ๐ฎ๐๐ต๐ฆ๐ข๐ฆ. It's a lucrative but risky name, jkjk.
Well, other doable names include, ๐๐๐ฉ๐๐ค๐จ, ๐๐๐ฉ๐๐๐ฃ๐, ๐๐๐ฉ๐๐๐๐ฉ๐, ๐๐๐ฉ๐๐ก๐ฎ,
One is rebuilding Salesforce with AI๐ฅHe's a really good one to follow to see how he build in public, young but experience. I've learned a lot from him in terms of how to bring values to others.
Subscribed to his newsletter.
@coffeewithone
Marketing is hard, but it can be smart.
I saw our alum
@davj
sponsored a YC after party, they give every attendee $1, you can use that $1 on to file tax extension for your startup on
@tryfondo
(or just take it!). So they are helping early founders to handle tax ddl for free ๐
Applying to an apartment in SF is sometimes harder than applying to YC.
I feel very frustrated after being denied by 2 apartments for the past month, and I have NOT been able to move to SF yet. Major reason might be my cofounder living in a temporary address after relocating and
(1/n) How did we achieve 5x less cost (v.s. ReAct on ALFWorld)?
Instead of calling LLM iteratively with a single long instruction like "if ... else... when...", along with previous history, StateFlow decomposes it into sub prompts/tasks, so each LLM call is more accurate and
maybe bcoz of more strong llm players on the stage, this year's devday ticket is free ๐
v.s. last year they charged me $450 for ticket (tho I got an lottery for $500 API credits during assistant agent demo)
Weโre taking OpenAI DevDay on the road! Join us this fall in San Francisco, London, or Singapore for hands-on sessions, demos, and best practices. Meet our engineers and see how developers around the world are building with OpenAI.
my first ever successful recommendation to YC!
i got 30+ asks and i recommended less than half of them and only 1 is accepted
i met
@agishaun
a year a ago when he was presenting llm agents. he started his project slightly earlier than me, iterated faster, he built 2 products
@MathGPTPro
is very lucky to be one of the earliest testers of
@augmentcode
. It's fast, accurate, and it supports many use cases.
My friend
@CMS_Flash
from Augment always 'complains' about
@quinn_lyu
and our founding eng constantly rank top 10 in Augment usage leaderboard.
I
๐ย Hello World,
Today, we're thrilled to introduce our new company, Augment, as we emerge from stealth to announce our funding milestone. We have raised $227 million at a $977 million valuation to empower software teams with AI.
Introducing Glato AI - the text-to-video AI that turns any website into personalized video ads with a digital spokesperson.
All you need is a product URL ๐
Get your FREE video ads at now!
(2/n) Why improve?
StateFlow conceptualizes complex task-solving processes as state machines. It distinguishes between โprocess groundingโ (via state and transitions) and โsub-task solvingโ (via actions in a state), leveraging heuristics in real-world tasks to assist LLM
I have spent tons of time preparing the rebrand practices, SEO () and user engagemets and more... will share once I take actions.
I noticed Gusto has a playboook on their rebrand from ZenPayroll to Gusto, tho it's b2b use case (vs us b2c, for now).
Help
@MathGPTPro
get rebrand suggestions! Any comment?!
I got 5k questionnaires from our users, the funniest is ๐ ๐ฎ๐๐ต๐ฆ๐ข๐ฆ. It's a lucrative but risky name, jkjk.
Well, other doable names include, ๐๐๐ฉ๐๐ค๐จ, ๐๐๐ฉ๐๐๐ฃ๐, ๐๐๐ฉ๐๐๐๐ฉ๐, ๐๐๐ฉ๐๐ก๐ฎ,
(3/n) Can LLM self-correct?
Our ablation study shows that "Verify" state improves the success rate (SR) but not as much. It's useful to have in practice (like how ppl perform majority votes to improve), but you can skip "Verify" to save costs.
I also saw some papers saying that
We just launched Positional on Product Hunt! It has been a long time coming, and we'd love for you to support the launch! ๐
Matt and I have been working on Positional for about 18 months. As you might know, we pivoted to get here. We pivoted to what we know best: content
A year ago today, our first tweet got ๐ญ๐ + views. Now, we have crossed ๐ญ๐ + users, all through ๐ผ๐ฟ๐ด๐ฎ๐ป๐ถ๐ฐ growth. We feel super grateful.
Today, weโre introducing our most engaging ๐๐ ๐ ๐ฎ๐๐ต ๐ง๐๐๐ผ๐ฟ. Now, you can ๐๐ช๐ณ๐ค๐ญ๐ฆ & ๐๐ด๐ฌ, ๐๐ณ๐ข๐ธ, ๐๐ค๐ณ๐ข๐ต๐ค๐ฉ, and
My dream is always to open a nice healthy boba shop someday in SF (supporting founders to recharge). Someone just publish a research post on that! SoTA generative boba with graduate student descent. Neurips Oral reserved โ
Is that easier to angel invest in hot startups if I
I quit PhD (for a day) and opened a boba shop at
@MIT
- Generative Boba! Itโs a huge success - right next to our office so all the AI researchers are enjoying it. Checkout our boba diffusion algorithm in the poster to understand why boba generation is so important to
@MIT_CSAIL
!
We did a customer questionnaire last month (and reward users with free credits), and got ~๐ญ๐ฌ๐ธ responses in just 2 days. ๐ฎ๐ฌ% said they heard about us from their friends. That motivated us to upgrade our customer referral program.
๐ฎ๐ฌ% spread over word-of-month.
@MathGPTPro
(4/n) ๐ฆ๐๐ฎ๐๐ฒ๐๐น๐ผ๐ can be combined with iterative refining methods (e.g. Reflexion) to improve performance further (see image).
Future directions to improve?
Employing active learning strategies to iteratively adjust or โtrainโ a ๐ฆ๐๐ฎ๐๐ฒ๐๐น๐ผ๐, adding or removing
new trend for scaling up (in both llm and robotics)?
o1 scales up at inference time
sampling-based MPC scales up at inference time (while previous sim2real scales up offline computation at training time)
both real-time! amazing progress in infra and parallel simulation;
๐ Diffusion-style annealing + sampling-based MPC can surpass RL, and seamlessly adapt to task parameters, all ๐๐ฟ๐ฎ๐ถ๐ป๐ถ๐ป๐ด-๐ณ๐ฟ๐ฒ๐ฒ!
We open sourced DIAL-MPC, the first training-free method for whole-body torque control using full-order dynamics ๐งต
@rowancheung
Schools are counter intuitive but always slow to adopt new tools. Like what happened before for calculator, or a more recent search engine, educators first wanted to ban it, failed to ban it, then finally adopted it :)
This is exactly what happened again last year for ChatGPT.
what matters (for most founder practices):
- asking the right question
- from the right person
- in a right way
example from
@alexdanilo99
:
real test if someone remembers: tell someone in-person you're the founder of Mathos. Ask them the next day if they remember
don't ask now
@VyvyenYue
@MathGPTPro
i like Mathos. As long as you make it clear its not pronounced MathOS (like MacOS) and instead Mathos (like pathos). It def leans toward the latter, JFYI.
real test if someone remembers: tell someone in-person you're the founder of Mathos. Ask them the next day if they remember
(n/n) I'm happy that we got 6.5 (7/7/6/6), which is in the top 10% of CoLM submissions:
It also feels great to build things people love. We got positive user testimonials from
@pyautogen
soon after publishing the code:
#llm
StateFlow is recently integrated into
@pyautogen
โThe integration of StateFlow has noticeably improved the relevance and accuracy of the generated assessments and feedback...,โ says a user of AutoGen. Itโs opening a new door with better control of the LLM workflow! Thanks to
We started as a Web App, and recently launched in the Apple App Store and Google Play Store, but the cold start is always challenging. The new app does NOT get an index/rank in app stores.
We decided to ask our users for help. So we added a pop-up widow in the web app asking
@rowancheung
Exactly. I feel now when interviewing candidate, we should test ChatGPT/LLM Master Certificate instead of Leedcode or so.
People that doesn't use LLM well is never a good hire.
Ads/slogans here are all hot tech companies, it feels very inspiring sometimes even when you are taking a walk in the city. The week I moved here
@databricks
was hosting GenAI Summit. I also found
@Informatica
's slogan very catchy and I immediately memoried it.
@t01uwa
you can apply anytime as long as you qualifies the criterias, no time frame.
i recommend talking to lawyers or reading more cases if you want plan for this, and people tends to qualify earlier than they expect
new trend for scaling up (in both llm and robotics)?
o1 scales up at inference time
sampling-based MPC scales up at inference time (while previous sim2real scales up offline computation in training time)
both real-time! amazing progress in infra and parallel simulation;
๐ Diffusion-style annealing + sampling-based MPC can surpass RL, and seamlessly adapt to task parameters, all ๐๐ฟ๐ฎ๐ถ๐ป๐ถ๐ป๐ด-๐ณ๐ฟ๐ฒ๐ฒ!
We open sourced DIAL-MPC, the first training-free method for whole-body torque control using full-order dynamics ๐งต
@rowancheung
@UNESCO
AI is transforming education, and
@MathGPTPro
wish to support this. We are building an AI tutor app with 800k+ users and tools to support teachers and schools.
We're releasing a preview of OpenAI o1โa new series of AI models designed to spend more time thinking before they respond.
These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math.
Paper accepted to
@COLM_conf
! If you don't know CoLM, it's Neurips-level, but the ๐ง๐ช๐ณ๐ด๐ต ๐ฆ๐ท๐ฆ๐ณ ๐ค๐ฐ๐ฏ๐ง๐ฆ๐ณ๐ฆ๐ฏ๐ค๐ฆ ๐ง๐ฐ๐ณ ๐๐๐๐ด. Wild.
Publishing papers while building a ๐๐๐ฎ๐ฟ๐๐๐ฝ with ๐ญ๐ + users is never easy.
We developed ๐ฆ๐๐ฎ๐๐ฒ๐๐น๐ผ๐, a framework to
@databricks
@Informatica
Serendipities are beautiful in a city with high talent density.
@quinn_lyu
coded for 10 days w/o going out. He finally went for a 3-hour long AYCE. A guy beside him also ate for straight 3hrs... so he said hi...
turns out the guy is also a hungry CTO coded for days at home...๐คฏ
@NithurM
haha no way, but this is one supporting fact.
e.g. useful facts include:
- secure funding from an accredited fund
- awards from famous hackathons/competitions
- being a judge of competitions
- publish research papers (not necessarily best tiers)
- participate in research paper
pretty sad that I forgot to secure my author ticket in time and I'm not allowed to attend
@COLM_conf
now ๐ฅน
any one has ticket that is willing to transfer?
Paper accepted to
@COLM_conf
! If you don't know CoLM, it's Neurips-level, but the ๐ง๐ช๐ณ๐ด๐ต ๐ฆ๐ท๐ฆ๐ณ ๐ค๐ฐ๐ฏ๐ง๐ฆ๐ณ๐ฆ๐ฏ๐ค๐ฆ ๐ง๐ฐ๐ณ ๐๐๐๐ด. Wild.
Publishing papers while building a ๐๐๐ฎ๐ฟ๐๐๐ฝ with ๐ญ๐ + users is never easy.
We developed ๐ฆ๐๐ฎ๐๐ฒ๐๐น๐ผ๐, a framework to
@garrytan
launch BF Live is almost the best part for me! (or maybe retreat Unconference is the best haha). Sure will come back, hopefully for demo presentation ๐
It is the annual visit day time for grad school admits. If you have a hard time making your choice, maybe my blog will help you! Check out my post โbest computer science schools ranked by bobaโ at
(n/n)
Does churn increase if we offer too many free credits?
Even if a few people churn (note that not everyone is able to refer a paid user, most of them shared on social but only got free registers), you still get more users.
And if one churns but keeps using your free
(4/n)
Prev strategy allows users to create mutiple accounts and ๐๐ฒ๐น๐ณ ๐ฟ๐ฒ๐ณ๐ฒ๐ฟ ๐/๐ผ paying anything. As long as there are easy options, they won't try hard to share on social effectively.
Now we increase the rewards and require payments, the ONLY option is to make
(5/n)
We thought prev referral strategy requires ๐น๐ฒ๐๐ so users will be less reluctant to do so. But we were wrong. ๐๐ฎ๐ฟ๐ฑ๐ฒ๐ฟ threshold (with better rewards) for referrals make things more effective.
@rowancheung
But what I found lucky is, there are always early educator adopters.
@UNESCO
and some top schools are adovcating "AI Initiatives in Education" (how to teach kids to leverage AI during learning), some of my friends are thinking about how to teach their kids to learn with AI.
(3/n)
Current strategy: refer your friend to subscribe to us (as ๐ฝ๐ฎ๐ถ๐ฑ users), and you will get free credits (worth 1-month credits as a ๐ฝ๐ฎ๐ถ๐ฑ user).
Prev: refer your friend to register (which does NOT requires payments) and you will get a free credits (much less than
@lina_colucci
yeah! also I love your Angel and Devil video () so much; have been thinking about the script to create a promotional video using
@toinfinityai
The first time I came across
@knightmage
โs Instagram I was blown away. Heโs the best cosplayer Iโve ever seen.
Itโs a dream come true to partner together and create the first-ever AI character clones! โจ
This tool allows creators and their fans to engage on a whole other level.
@santygegen
Youtubers are affiliation program, and we reward them $ based on the number of users registered using their link.
By referral strategy, I explained here: