AI research agents 3.0 - I built a team of AI agents via
@pyautogen
,
They can read/write airtable data, generate high quality research & verify each other's work
I built an AI sales that do Cold Call & Whatsapp followup - powered by
@GroqInc
,
@Vapi_AI
&
@RelevanceAI_
This AI Sales can:
1. Outreach people on whatsapp
2. Make phone call to follow up
3. Take actions in CRM based on call transcript
Thread about how i built it 🧵
@jasonzhou1993
, a.k.a YouTube tech wizard AI Jason has just dropped a mind-blowing video showcasing his latest creation with Relevance,
@Vapi_AI
and
@GroqInc
: a real-time, multi channel AI voice sales agent.
🗣️ This isn't your run-of-the-mill chatbot. Jason has ingeniously fused
Sam altman revealed key aspects of GPT5
At
@ycombinator
W24 kickoff,
@sama
suggested startups build w/ the mindset GPT-5 and AGI will be achieved "relatively soon";
But what does GPT5 actually gonna look like?
In an interview with
@BillGates
, Sam revealed some key milestones
Can you imagine how powerful an AI Agent with long term memory will be?
What if AI Agents can keep learning and improving its workflow and skills through external feedback? And they can retrieve the information whenever needed from its database.
We can train agents to develop
Referral/affiliate marketing is a hugely underutilized growth channel in web3
It generated:
- 30% of volume for Lido
- $11B trading volume for GMX
I analyzed 6 web3 projects to try to understand how BIG this could be.
Here's the data I collected 🧵
Build a real time AI Conversation Co-pilot run on your iPhone with 0-latency?
1 year ago, building an AI listens to conversation & generate real time response is almost impossible
But the model speed has shifted dramatically past few month...
👇🎥
I'm building a web scraper agent where you can:
- Give it a list of websites + data points you want to extract from each website, and it will do the work for you
- e.g. Give it list of B2B Sass companies, and extract all case studies on their website, OR find specific info like
Airdrop incoming!
Any cow that is staked has a chance to get an airdrop. To increase your chances pick up some more cows from
@MagicEden
and then head over to our staking site and stake your cows. Only 1,000 airdrops available
Excited to join
@Safaryclub
batch5 🦁🔥
Super impressed by the quality of the community, and can't wait to jam with web3 growth leaders from dYdX, Zerion, Arbitrum and more!
@OpenAI
unveiled more details about Sora in an interview:
- Sora will be released within 2024
- Instead of hours, Mira mentioned it takes a few mins to generate a 20 sec 720p video
- The biggest challenge of AI generated video is consistency between frames, and that's what
@jasonzhou1993
Just released an awesome video on Local Agentic RAG w/ llama3 🦙
It covers how you can build a reliable and accurate RAG system with Firecrawl 🔥,
@llama_index
and
@LangChainAI
Go watch it 👇
Looking for a developer to build a research agent with me, ideally someone who is:
- full stack
- move fast
- passion about AI
DM me if you have good recommendation!
There is a hugely underutilized growth strategy in web3
When I was working at Safetyculture, it led to millions of organic traffic every month
I saw the same thing adopted by web3 projects like Dune analytics.
Here is the unpack 🧵
How to reduce LLM cost by 78%+?
For AI builders, LLM is a variable cost that you can’t ignore.
Sometimes, it can even decide if a product can be successful.
I tried to launch an AI product with free trail, and it almost destroyed my balance sheet, here are my hard learnings…
I’m looking for a talent to help edit videos & grow AI Jason channel with me together, who are:
1. A killer video editor, producing fast-paced no bullshit videos.
2. Super passionate about AI
Think this is you? Or know someone perfect for this?
Please DM me!
Many people didn't realise that
@GroqInc
LPU is not only good for LLM inference, but also img processing
This is a demo where they processed 8 img in 0.18 seconds
"Would a referral program work for my web3 project?"
Many web3 founders thought about this, but don't have a clear answer.
Analyzed 7 web3 projects' referral program & their results, read this til the end... 🧵
ReAct Agent framework is outdated, welcome LLMComplier?
ReAct has been the most popular Agent planning framework, where agent follows Thought-Action-Observation structure; However, the limitation is that the agent can only action on the very next step, which is inefficient &
SEO is outdated; it's time for GEO now.
As LLM-powered search engines like Perplexity, Bard, and Bing gain more adoption, users are relying more on direct generated answers from LLMs instead of search result rankings.
This poses new challenges & opportunities – how should you
Don’t try to build a product that will be used by one million users at the beginning,
Focus on serving 10 users really really well until they can’t live without your product
Q: How do you design an amazing user experience?
In the clip below, Airbnb co-founder and CEO Brian Chesky explains that one route to a great UX and word-of-mouth growth is designing the perfect experience for one person:
“How do you make something for a million people? I don’t
Speed comparison between chatGPT vs Groq
🤯 Groq is so freaking fast, 453 token/s
This speed can unlock many real time use case, can't wait to try out the API
GPT5 unlocks LLM System2 Thinking?
For human, problems often occur when we try to solve System 2 level complex problem, with System 1 fast, intuitive answer
However this is exact the state of LLM now, there is no difference between answer 1+1 VS Complex math equation; All it
The "weaknesses" section of Sora gave me lots of fun;
Didn't realise Sora is not just video generation, but simulation of the physics, this itself has lots of implications
Here is "Basketball through hoop then explodes."
Introducing Sora, our text-to-video model.
Sora can create videos of up to 60 seconds featuring highly detailed scenes, complex camera motion, and multiple characters with vibrant emotions.
Prompt: “Beautiful, snowy
I cant believe it took me 1 hour to figure out how to call external API from
@salesforce
How could developers at salesforce sleep at night shipping such unnecessarily complex system? 😤
One step closer to the AI Agent-cy. This workflow is a working mod and implementation of
#ChatDev
incorporating
@jasonzhou1993
's Research Agent 2.0 on the frontend—so the process starts with live web search, and is not confined to a training cutoff.
@Vapi_AI
is the best Voice AI platform with lowest latency I’ve tried so far, and it works with
@GroqInc
If you want to build real time conversation AI, definitely recommend, I got a tutorial here:
We’re the founders of Vapi. We're building Voice AI Infrastructure for the Internet. Build, test and deploy voicebots in minutes rather than months.
Here's our story 🧵👇
Why is so hard to build web & mobile agent?
As device like Rabbit R1 became popular, mobile & web agent is under spotlight;
They are type of agent that can directly view & operate your computer & mobile phone by simulating human interactions, powered by multimodal models
Model
What made a new agent framework popular?
1. Langchain - enable easy retrieval use case
2. Autogpt - enable fully autonomous agent
3. Autogen - enable multi agent collab
None of those framework are the first to introduce the concept, but they made development work easier
6 reasons why I believe web3 affiliate marketing will be bigger than web2:
1/ Higher margin: Digital goods (NFT) has a much higher margin than physical goods (Clothes) to incentivise affiliates
2/ Web3 projects have a community of believers who skin in the game.
👇
One of the biggest unconventional learnings I got past month:
"Majority of No code platform users are actually people who knows how to code"
Because the real hurdle for non technical people is not about coding syntax but the way of problem solving in machine
Just had a chat with
@ZooniesXYZ
about
@qwestive
, the first
@solana
NFT project as
@opensea
launchpad partner. Very talented team with clear articulation of their thoughts & vision.
Very bullish on them.
Any AI Agent Builders in Singapore?! 🇸🇬 🤖
I will be in Singapore next week for a meetup, where I gonna share first hand real world AI Agents use case & learnings!
RSVP open for the next 48 hours:
🤩We’re excited to announce that we are joining
@BinanceLabs
Incubation Program! Qwestive, the Web3 CRM (Community Relationship Management) platform, is coming soon. We help token communities retain and grow their user base.
More details:
#BNB
#Qwestive
Question for AI builders:
Assuming GPT5
- Achieved superior AGI reasoning ability
- No hallucination
- Connect to external data source easily
- Can be customised & personalised
What are the core value AI startups should be focusing on for building AI agents?
Sam altman revealed key aspects of GPT5
At
@ycombinator
W24 kickoff,
@sama
suggested startups build w/ the mindset GPT-5 and AGI will be achieved "relatively soon";
But what does GPT5 actually gonna look like?
In an interview with
@BillGates
, Sam revealed some key milestones
2/
@RangoExchange
has a simple referral program launched in Aug 2021:
- Rango received $1+ Billion trading volume from this referral program during the bull run
- It was launched together with a trading competition, which strength the result
My mom called me 3 times past 12 month:
May 2022 for Luna crash
Nov 2022 for FTX
Mar 2023 for SVB
Every time she thought it will be the end for crypto and suggested I go find a normal job.
But boy, I’m more bullish than ever now about Defi & Crypto.
chatGPT helps me stay objective & resolve conflicts;
Now everytime when I have an argument with others, I put the discussion in chatGPT and let it help me see both sides;
Having an objective 3rd party always available is so useful
OpenVoice: Instant Zero-shot voice cloning with accurate tone voice
It requires only a short audio clip from the reference speaker to replicate their voice & generate speech in multiple languages;
1. Accurate Tone Color Cloning. OpenVoice can accurately clone the reference tone
In the age of LLM, You can create a valuable Sass if you have:
1. Curated niche dataset - doesn’t even need to be proprietary data, even handpick high quality public data provides value
2. A series of good prompt that articulate your domain workflow
This is a good example
SEC Insights AI is live on Producthunt 🥳🥳
Currently trending at
#1
Powered by
@llama_index
by
@jerryjliu0
Analyze complex financial documents in seconds
Link to launch post below
Just joined
@withBackdrop
and made my first post, it is super valuable to have a curated space where you can get valuable advice & connections with web3 professionals.
Gemini 1.5 Pro - A highly capable multimodal model with a 10M token context length
Today we are releasing the first demonstrations of the capabilities of the Gemini 1.5 series, with the Gemini 1.5 Pro model. One of the key differentiators of this model is its incredibly long
Two big problems I faced building AI agents, how to get agents to:
1. Knows its limit and ask me for input
2. Follow my tool usage instruction
Anyone have good solution?
Back at
@juiceboxETH
town hall meeting again, and I'm mind blown by how much good stuff happening in just 1 week.
12 months ago, I worked in "DAO" for the first time as a product designer.
🧵 I learned so much, and here was my notes from working in Juicebox:
Will 2024 has chatGPT moment of Robotic AI?
OpenAI invested $ millions in Robots, and Industry leaders
@DrJimFan
@adcock_brett
all shorten the timeline prediction of when Human Robots will take off to next 12~36 month
One of the biggest evolution that pushed us towards the
The programmatic display ad is a $400 Billion industry to be disrupted by web3
Why?
Web3-enabled ads can provide a much better experience for both advertisers & consumers
Here is how... 🧵
Is it ever worth training your own domain LLM? GPT4 is beating domain specific model like Bloomberg GPT in financial text analysis
Last year, Bloomberg spent $3m & introduced its own 50-billion parameter LLM, specifically trained on a wide range of financial data to support a
Lots of discussion around how Sora is not just a text-to-video model but a physic engine or simulation of the real world;
@runwayml
actually brought up similar concept Dec 2023, with a video explaining their General World Model concept, which is an appealing future;
By feeding
Small model running directly on mobile device is gonna be one of the biggest trend 2024
Tried running whisperKit from
@argmaxinc
on my phone was mind blowing, the ability to process real time data with low latency just provides such better UX
This means 2 things:
1. Mobile
The UX of co-pilot agent seems to converge:
- Chat interface to give feedback & update
- Task list to align on priority
- Specialised tools to operate mission critical systems (e.g. Code, Browser, File editor, etc.)
Devin for Legal is coming
This is the first sneak peak of Spellbook's agent
It reviews a VC term sheet and starts editing a set of financing documents
It's not perfect yet. But it's improving quickly. Agents are about to leap from their GPT2 era to their ChatGPT era.
More about GPT5 reasoning ability:
"At minimum, We need some sort of adaptive compute, right now we spend the same amount of compute on each token for both simple/dump question and figuring out complex math question;"
Likely GPT5 will develop system 2 thinking, and auto switch
“Action produces information” - Coinbase CEO Amsterdam
When decisions have multiple options with no clear winner & mixed pro and cons,
Just flip a coin and move on.
Insights generated from actions are more value than insights from analysis
Bytedance’s new Multi-modal model can capture more media details than GPT4V?
Existing multi-modal models are good at getting big pictures while handling pictures, sound, text etc., but sometimes overlooking important tiny details, which will limit their capabilities in
But most important thing is what kind of use cases AI developer can unlock;
I've been thinking about this a lot, and this is where i came up with the idea of fully autonomous AI sales that can do multi-channel communication (Including phone call!)
Full tutorial in this video:
Could programmatic advertising be the real next use case for web3?
Wallet data provides advertisers with hyper-segmentation,
also enable new UX for end users
Let's dive in... 🧵
This wearable AI is sick
The fact that it shows a gallery of what I "see" through out the day is quite amazing
Wondering what is the cost per day at the moment?
If you are wondering what I've been working on after selling my company:
Meet Jetson - the world’s first wearable assistant that remembers everything you saw
🤹♂️ Does everything that other AI-wearables do
📷 10x more useful because records not just audio, but also video while
The current challenge in LLMs training? it's constrained by the limited size and quality of human feedback. Standard methods like Reinforcement Learning from Human Feedback (RLHF) rely on this feedback to create a fixed 'reward model' that guides LLM learning. However, this
#web1
: give consumer access to information that was exclusive to few brands
#web2
: allow consumers to yell at brands & public if I don’t like you
#web3
: enable consumers to influence brands’ decision & co-create, co-own the IP
👇
One big challenge I had in startup sales:
I got customers excited and commit to use the product when demoing them initially,
but 2 months later, after the product is ready, their priorities shift.
Sales is way ahead of the product. Anyone else has the same issue?
Have any web3 projects tried "retargeting"?
Met an ad network that drove $4+ Billion volume to CEXs via displaying ads to people who land on CEX website but didn't complete signup
The same technique should work for top Defi protocol too, I'm gonna try it out.
There are 6.7M unique wallet addresses made Defi transactions,
but how many REAL Defi USERS do you reckon are out there in total?
I heard some founders think there are only 500K.
I saw
@RangoExchange
built a cool in-app royalty reward system
It drove $500M trading volume with 7k+ people participated
This made me wonder what other web3 projects adopted such royalty systems
Found 4 great examples if you are looking for design reference 🧵
We've started adding loaders to help load data in a format such that you can finetune (usually for a custom tone)
So far we've got Facebook Messenger, Slack, Telegram, WhatsApp, Twitter (via
@apify
), Discord
What else should we add?
What types of tasks are:
1. Long tail,
2. Need to be conducted constantly
3. Have lower quality bar
?
I think those are tasks that will be automated by AI agent first
Has anyone tried to use chapGPT to build cold emailing automation?
1/ Give it a list of VC & clients
2/ Auto scrab their recent podcast, blogs & make a summary
3/ Send personalized email
Don’t be afraid if you have a different background/path than others in your industry
Being different is good,
It is what gives you a unique perspective and value.
@balajis
Smart doctor, lawyers and artists should be start working on product that enables this AI disruption now.
Better disrupt yourself rather than others do it
Those are the key methods, to do any of those optimisation, the ability to log & observe all LLM calls is critical;
Here is a quick example of how I used langsmith to monitor & optimise cost for one of the agent & reduce 78% cost
Below is my full video:
AI tutoring will be big:
70 medical students did a clinical trial of surgical skills; Ones with immediate feedback provided by AI tutor is delivering significant better results
One mistake I made was trying lots of growth tactics while the product is still immature
Do not waste time "hacking" the growth before you figure out early product market fit
Otherwise, it speeds up death instead of growth
"How should web3 projects do user acquisition during a bear market? 🐻"
Has been discussing this topic with multiple founders & marketers past few weeks,
Here are the top trends 🧠
Business expect at least 99% accuracy when adopting new tech, while consumers AI Assistant use case have much lower bar & tolerence;
What's the Implications on AI agent platforms? 👇
"What should I aim for the Beta test?"
Here is the guide from
@amyjokim
that was used for Netflix, The Sims:
MVP - Build the core habit-building UX
Beta - Nail a killer onboarding experience
The Launch - Establish user acquisition channels
Method 2: LLM Cascade
- Create a cascade chain of LLMs to use, from cheap to expensive
- When user input the question, using the cheap and small model to answer the question first, if the confidence score is high, it will accept the answer. Otherwise, moving to the next model.
-
"Potential marketing traffic for web3 apps"
Web3 native traffic - Dapps, alpha communities & web3 publishers
Web2.5 traffic - CEX, e.g. that's why Coinbase Learn charges $ Millions
Web2 traffic - e.g. people visit Bloomberg with metamask installed, this is the biggest.
In LLM era, UI is becoming dynamic & adaptive to scenario & user, instead of the other way around;
I'm pretty keen to see what if we get an agent generating 5 versions of UI based on a brief, w/ a manager agent critique for improvement, and repeat again & again;
Will it able to
Introducing Alvea!
An adaptive AI assistant creating unique user interfaces personalized to your tasks just in time. Your evolving tool, designed to enhance productivity and user experience 🚀
Won 🥇
@agihouse_org
w/
@sockcymbal
@narphorium
@vhurgenes
“Company’s culture & value should be controversial.”
It should almost scare away people who we don’t want, and make specific group feels belong
Wise words from a wise man
@userlastname
that is worth tweeting