Can we trust ai agents to stay on track without human oversight? 🎯
I'm excited to announce AgentMonitor, a context dependent framework designed to keep agents aligned with their objectives. Created within
@auto_gpt
, it was recently accepted to the NeurIPS SoLaR workshop!
As
i'm addicted
i wake up thinking about it, i go to sleep thinking about it, every action and second throughout the day is directed towards it
is my baby
Latest demo of
@stackwiseai
. Just explain what you want a function to do, and AI builds it. 😲🦾
Skip the back and forth with ChatGPT. No more hunting for documentation to integrate with APIs. Just pure functionality! 🔧
Check out the GitHub repo below to try for yourself :)
Joining
@_buildspace
for nights & weekends s4 🔥🔥
Currently building an agnostic way to evaluate agents with AutoGPT and writing a paper about it
Maybe will ship some other agent related stuff :)
🎊 Exciting Announcement! 🐙
We've raised $12M to take AutoGPT to the next level!
Our aim? Making it the largest open-source project in history, and democratizing access to a brighter future of work.
A future where everyone is an ai engineer, whether they know it or not 🌄
thinking is so underrated
seriously, try sitting down for just 30 min and do nothing but write
pen to paper
no distractions, no music
just you and your thoughts
you’d be surprised how many problems you can solve
or how many problems weren’t problems at all
few.
sometimes we cancel ourselves out without even testing reality
there is 0 downside to respectfully asking
it’s led to
- jobs
- deeper relationships
- access to high signal places I had no business being
break the thin veneer of unknowing
and seek base truth
just ask
I can now generate 10000 AI generated emails by clicking a button 🥶
Doing it manually takes literal days
Clay is too expensive and requires setting up a Lambda function
I wrote a Google Sheets script that can do it free with your laptop closed 🤯
Code and example sheet 👇
As a founder, there is absolutely no greater feeling then a potential customer getting excited and asking to invest during a demo
It's moments like these that make it all worth it.
Why am I starting a company?
> nothing else comes close to stimulating my mind on so many different ways on a daily basis
When I look back at the journey so far I wonder how I'm still sane
It's because even through the highs and lows there's nothing else I'd rather be doing
This might sound crazy - it is not crazy.
Lee Kuan Yew talks at great length throughout his memoirs on building Singapore about trees specifically, trees *as such*, being harbingers and indicators and inspirers of great things:
In Boston ppl are in bio/chem-tech bec they have a high degree of interest or passion. The environment is much more electrifying when everyone’s goal is to solve a generational problem. And not mostly motivated by money, ease of work, or shortcuts. In sf it’s more mixed
One of the things I’ll miss most about Cambridge/Boston is the sheer concentration of incredible people building valuable things in hardtech. Every lab and moonshot company is just a bike ride away :)
UI is UX
this is the 'aesthetic-usability effect'
visually pleasing products are perceived as being easier to use
@endflowai
has a ways to go
but Notion is a master at this, and we try to learn from the best
software should be delightful to use. micro interactions matter
these decisions are made too frivolously
it's a decision akin to getting married
a choice that should be experimented with and personalized to yourself should not be made on a whim
if you have the grit you will succeed regardless
Bezos, Elon, Reed, all did
you can too
solo founders are underrated.
i used to think i needed a cofounder until i realized that i can just do everything myself.
some solo founders i respect that helped me realize this
@0interestrates
@pk_iv
@AviSchiffmann
an 'agent' is any entity that can perceive it's environment and act upon it
before LLM agents existed, reinforcement learning agents were all the rage
RL agents interact with dynamic environments, adapting their behavior to achieve their goals like playing tag such as in
Roadman AI: Chat with a real-life roadman using AI 🧢🔥
Ever wanted to ask a roadman a question or translate your speech to roadman lingo?
No? Built it anyways using Whisper, OpenAI, and Eleven.
Nothing like a fusion North-Indian/British wedding. Vibes are incredible, music is banging, Indian food is marvelous, and emotions are flowing
I think Bollywood is my new hype music of choice
developing an ai agent is a monumental task, made even slower + harder due to a lacking feedback loop.
the solution is a diverse standardized agent benchmark!
would love to hear your thoughts :)
cc:
@_buildspace
@_nightsweekends
you'll read this in <15 sec, but it took me 15 minutes to write
i experimented with different permutations of format and verbiage to capture nuance
to get across exactly what i intend. no more, no less
writing is the crystallization of thought
How close are we to achieving recursively self-improving AI? 🔁
It seems to be within reach given the tools at our disposal. Q* + the tweet below made it click.
Some backstory - the 'OpenAI Five' was an AI that beat Dota pros consistently. Its training method? Self-play,
Why does this work? Because judging an output is typically a lot easier than generating it. For example, you can check the correctness of a program by unit tests or compiler errors, but it’s way more cognitive burden to write the code.
RLHF, for example, is a type of synthetic
pain becomes discomfort when you're comfortable with pain
any unsettling thing you seek is discomfort
anything you don't is pain
the less pain you seek the more things are painful to you
push the boundaries of discomfort
any ephemeral pain is just discomfort hidden behind a
to Rockefeller, he had a God given mandate to make boatloads of money
our brains are powerful
why not rationalize positive things?
im so serious
sit down for 30 minutes
your only goal is to convince yourself something negative being positive
just write
and rationalize.
The analogy here is:
- context window = RAM
- toks/sec = GhZ
- param count = gate size
Agents are no longer hyped as they were, but with the rate of advancements being made in LLM OS the trough of disillusionment may not last very long 🚀
Imagine the possibilities with a model that has:
- 1m context like Gemini pro 1.5
- Instant and cheap inference like Groq
- Gpt5 like reasoning
We’ll be building insane things 🤯
developing an ai agent is a monumental task, made even slower + harder due to a lacking feedback loop.
the solution is a diverse standardized agent benchmark!
would love to hear your thoughts :)
cc:
@_buildspace
@_nightsweekends
bits are cool, but atoms are cooler
when I work with manufacturing companies there's a tangible sense of reality to it
with
@endflowai
i feel super lucky to get to see the factories that I'm making more efficient
the smells of raw materials 🪨
the sound of a circular saw 🪚
Next Friday (29th),
@AlexReibman
,
@hackgoofer
, and I are co-hosting a discussion based agent event.
Focused on problems in the agent world and their potential solutions. In-person and live streamed
Rsvp to come!
First with DallE, it was artists. Now with Devin, coders. We never thought it would happen, but it was obvious in hindsight.
It reminds me of The Bitter Lesson by Richard Sutton, where he makes the case that betting on the scaling of computational power is the winning formula.
spent the weekend at the
@FoundersPodcast
retreat
the method is clear: knowledge + love + focus over time
i just need to keep building unfair advantages
We live in a world of loss functions.
Whether it's error messages to fix bugs in a project, mse for an ai model, or the visual difference in your painting and a tutorial.
@merwanehamadi
and I are creating the loss function for ai agents at
@Auto_GPT
‘action creates direction’
because
1. action creates data for reflection
2. there is no objectively correct path
any action with a two way door should be made without hesitation
wait too long -> diminishing returns vs taking *any action*
also why doing more compounds
the
Thanks to everyone who came out, was a blast!
No better place to be then Sf, so much energy here right now :)
Huge shoutout to co-hosts
@AlexReibman
@hackgoofer
and all speakers
We invited 130+ AI engineers to discuss challenges, strategies, and achieving state of the art with AI agents.
The space is growing fast, and there’s a new breakthrough almost every week.
Here’s a what we saw at the SF Agent SOTA forum (🧵):
Impressive, Spotify can match languages based on the phonetic sound regardless of language.
They probably have a proprietary version of Neural Search (Algolia) behind the scenes
Generate subtitles, a thumbnail, and a summary for any video 🎥🖼️
Brilliant contribution by
@focusedVK
using
@replicate
hosted models :)
It's open source - use it free:
living your life without reflecting is like doomscrolling but in real life
i used take calls, code, and sell without reflecting and days would blur before I realized that iteration is the name of the game.
for fitness, for my company, and for mental health.
it's most effective
The US has figured out how to do this vaccine thing. I booked late last night (took me 5 minutes) and just got my shot (10 minutes of waiting). Canada needs to learn
advice is not a source of truth
but it is valuable as data points
maximize advice by maximizing empathy
1️⃣ what has their path looked like regarding the object of advice? (empirical/second hand, successes/failures, etc)
2️⃣ based on their knowledge, how do they perceive you in
Bring GPT into your spreadsheet
Extensions -> Apps Script -> paste the below code -> add OpenAI key and save. Super simple
=GPT() to use
Useful if you're looking to save money on Clay and on the ridiculously overpriced Sheets extensions
(Copy pastable below)
action requires direction
-> direction requires iteration
-> iteration requires information
-> information requires action 🔄
compounding is the open secret to all greatness. take action and don’t stop. reflect in between
the only thing that can stop you is yourself
Landed in Sf greeted by the gentle fogs of the bay, the glowing lights of the golden gate, and the feeling of latent energy this city has.
Glad to be back 🫡
I've found this to be the case for every subject I did not enjoy learning in school including biology, chemistry, physics, math, etc. I'm glad they never tried to teach me AI in high school
I do wonder how many more passionately curious people an optimal school can create
@fchollet
@Josh_Ebner
Have you heard of autonomous agents? With an llm as a reasoning machine, we can perform actions in the real world. It’s early but already we’ve seen sparks
I'd be rich if I had a dollar for every time I heard directly contradictory advice from smart people about startups and life
There is no perfect path.
All we have is a patchwork of guesses and optimizations that may lead us to fulfillment and happiness now and in the future 🛣️
every exponential is flat
until it isn’t
every ounce of effort is in vain
until it converts
we underestimate how long it takes
and how hard it is.
Warren Buffet made most of his money after 80
if he died at 75 we would have no idea who he is.
we weren’t built to
I was introduced to the concept of '0th order principles' listening to a pod with
@bryan_johnson
0th order principles are completely new ways to think about things.
For example, the number 0 itself was a 0th order invention.
It's a 0th order invention because because it
The goal of a startup is to create immense value out of thin air.
You do this by pulling together the strands of reality to lower the entropy around one idea, cultivating a collective pursuit of a potential future.
It ain't easy, it's not meant to be.
I love capitalism.
survivorship bias is an excuse
Dyson worked for 12 years on a bagless vacuum before getting it right
today he’s worth billions
survivorship bias or having conviction and being right?
maybe the real lesson is not about where the bullets hit the plane, but the actual pilot
@paulg
Agreed, it’s not necessary. But doing great things isn’t easy, and often times great people do have that extra motivator - the ‘chip on the shoulder’. It doesn’t have to stem from an unhappy childhood.
Being “great” requires sacrifices only few are motivated enough to make
A role model of mine is James Dyson. An extremely underrated founder and human being.
But everyone has heard of his vacuums.
Vacuums used to require bags, which made them lose efficiency quickly and cost money to upkeep.
And, after seeing a large industrial vacuum that did
before Alexander Hamilton was anyone, he wrote that he hoped for a war
status quo shakeups are the best way for ambitious people to find gaps and create value quickly
the revolution - Napoleon
computing - Steve Job
internet - Sergey & Larry
AI - ??
these are exciting times.
delusion is a massive indicator of how successful a founder will be
it's necessary
"the ones who are crazy enough to think they can change the world are the ones who do"
New feature for Auto-GPT-Benchmarks 📊
Now includes a minimal implementation to run challenges from the front-end
This makes it easier for you to iterate and objectively improve your agent!
Clone and follow instructions in :
Stay
Over the past couple years there's been a lot of talk of AI and AI agents changing the world in a radical way.
But truthfully, end customers don't care about AI.
The equation is simple - does it make someones life easier?
✔️ / ❌
Since the industrial revolution we have been
🐙🤝⚡️ We're thrilled to share the penultimate judge for our hackathon 🤖
@silennai
- Head of R&D at
@Auto_GPT
will be joining us on 18th August - 9 AM.
More details about the collaboration are on the way 🚀, stay tuned.
Register here:
Everyone should be familiar with the concept of 'lot size one' - even outside of the manufacturing industry.
It refers to the concept of keeping production flexible to demand, bringing costs down and increasing personalization.
The goal is to only produce an item when someone
Iceland is wonderful. Feels like I’m in alignment with the Earth with the earthquakes, volcanoes, and geothermal activity
The culture too. Lots of trust, nice people, good food, even the tourists aren’t as obnoxious
Unveiling Isomorphic: Visualize and interact with your Pinecone embeddings in an understandable 3d vector graph 🚀📊
Discover individual data points, visualize similarities in latent space, and explore your embeddings!
Check it out 👇
@paulg
There are a lot of duplicate ‘gpt wrapper’ startups, but there will be a winner in each space as most provide real value. Same can’t be said for web3.
'sales' is a misnomer ❌
and 'making money' is a fake goal 🙅♂️
the best salesmen understand that sales is just effectively communicating how you're going to solve their problems
which means
1. understand the problems
2. communicate how you solve them
3. provide sufficient trust
Made aitemplates: a one-stop Python package for working with
@OpenAI
API 🐍🚀
Easy to create prompt engineering templates from the cli. Offers Python typing support, error checking, and an API cost tracker.
Check out the GitHub repo for more 👇
the ladder every startup has to climb to justify its existence
1. emotion, identifying a real pain
2. sales, justifying ROI to decision makers -> offer market fit
3. product, execution aligned with offer -> PMF
in consumer all 3 are 1 person
Sometimes I get swept up in the daily grind, feeling like I'm not moving fast enough and not doing enough to get ahead.
Ironically, this leads to me feeling like I have less time, not more. It's not a good feeling.
When I notice, I always try to take a step back and cultivate
I love it when I do something/make a suboptimal decision and then cringe about it.
At least I'm aware of what to change and can reflect on what went wrong.
It makes me wonder how often I do something suboptimal but just don't realize it.
It's why I like failing :)
There are 3 types of 'AI companies'
1. AI as a feature
2. AI company
3. AI native
90% of companies claiming to be AI companies fall into the first bucket.
If you are calling an API like
@harvey__ai
, you are not an AI company.
If you are using ML like
@Spotify
you are not an
@gokulr
It's true. A lot of people don't understand what it means to wake up excited about something.
If they were truly motivated about something even for a day their perspective would change.
my hot take is that tech wise it won't take long for most digital jobs to be replaced by AI.
distribution and trust are the bigger problems
if that's solved, why would a company hire a high maintenance human when they can hire a tireless AI?
it's
@merwanehamadi
30th birthday today 🎉🎉
thanks for everything man, meeting you was one of the best things that happened this year. Wish you all the best today and into the future :)
can we get some more well wishes going his way??
I do my best to live with a Kobe work ethic and a Kanye level of self belief.
I'm not always there, but it gives me a good baseline
Haven't been able to get that line out of my head since
@FoundersPodcast
dropped it at FoundersOnly a couple weeks ago
Success is a game of n*p.
n (num hours) represents the effort you put in. Think about this as the speed to achieving your goal.
p (probability) represents focus/direction in which you put work into. This is analogous to creating a plan.
100x gains come from p, not n. 1/3
@Auto_GPT
The AgentMonitor sits at the action layer, preventing the agent from executing misaligned actions.
It starts with a regex filter to ensure that the agent's 'thought' is an action, not monologue. Then, context is given to gpt-3.5-turbo, which assigns the action a score 0-100.
the Egyptian old kingdom built the pyramids 5000 years ago and we still don’t fully understand how
i wonder what that apparatus could accomplish in 20 years if it existed today
space elevators? warp drives? infinite energy? immortality? ASI?
SAAS - software as a service is a thing of the past
SAAS - service as a software is the present
The optimal user experience for most ‘AI’ tasks is often an outsourced human.
For now
Years ago QuickBooks let users upload receipts and then "AI" would identify the expense amount and category automatically. Sometimes it worked fast and sometimes it took hours.
It was actually just workers in the Philippines, and it took awhile sometimes because they were asleep