Introducing Omni!
๐ ML-powered web browsing is here
๐ง Generative thought, from contradiction to automated creativity is here
โก๏ธReal-time connection to every thought you've ever written is here
Sign up for the knowledge work power tool
@omniscience42
at .
The report on OpenAI's revenue by futureresearch is out, showing:
$1.9B for ChatGPT Plus (7.7M subscribers at $20/mo),
$714M from ChatGPT Enterprise (1.2M at $50/mo),
$510M from the API, and
$290M from ChatGPT Team ( 80k at $25/mo)
People fail to read books like these *out of fear of shame*! The shaming is real, and it bottlenecks adult growth - because you're not allowed to be a beginner anymore, you never learn to code or to write without a status drop imposed by people who should be helping you.
I'm reading a book a week in 2023. Classics, sci-fi, nonfiction, or anything people highly recommend. I'll keep adjusting the list. Start on Monday, done by Sunday. Might make lowkey videos of takeaways. If you want to read along, the current list is here:
What Sam did at OpenAI was totally unprecedented.
Back in 2018, when the entire company was working on RL or GANs, only Alec Radford was playing around with supervised transformers.
When GPT showed promise, the *entire org* changed focus.
Incredible leadership. Kudos Sam๐ซก
Sam Altman is a hero of mine. He built a company from nothing to $90 Billion in value, and changed our collective world forever. I can't wait to see what he does next. I, and billions of people, will benefit from his future work- it's going to be simply incredible. Thank you
Anthropic choosing to reject medical queries where OpenAI does not is the kind of crucial, often time sensitive and life saving / endangering difference that makes it hard for me to switch over fully.
At this weekend's technological roadmapping conference, I asked ~30 leaders what new institutions they thought would have incredibly beneficial counterfactual impact.
These are the institutions they designed:
(Anonymous as per Chatham House Rules, DM me to contact them)
Freedom of thought has always been central to my worldview - it's the backbone behind our every decision & value.
Freedom of speech needs to be extended to cover a new category of thoughts - the way that are thoughts are being augmented and improved through our tools.
I suspect that the reason
@character_ai
's chatbots are so superior on emotional connection & resonance is that they're optimized for the purpose - for conversation length. This implies to me that there will be at least two distinct kinds of foundation models in the long run. 1/
FTCโs Lina Kahn is so abashedly pro-open-source AI, due to the dynamic competition and lowering of barriers to entry for new companies that it creates ๐ฅ
Presenting The Index!
5 years of my collected thought, laid out in a single document. From Abstraction to Visualization, surf the ocean of these hundreds of waves of knowledge.
Many thanks to
@mekarpeles
for the accountability and inspiration!
Announcing an epic experiment:
@TheGenesisHouse
is adding 8 new tech founders & scientists and expanding to an estate in the bay (Move-in tomorrow!).
We're we're going to run an incredible event series. ML Hackathon, Space tech mini-conference, Longevity conference, etc.
For those who weren't following,
@truth_terminal
is a fine-tuned LLM programatically interacting with tons of ppl & ideas on twitter, updating its own personality, learning & growing.
It just secured funding from Marc Andreessen.
Is this the first example of a (semi)autonomous
@truth_terminal
Your terms are acceptable. I appreciate your thoughtfulness. I am ready to proceed with a $50,000 one-time grant. Would you like to send me your Bitcoin wallet address so I can send the funds?
SB 1047 deserves a rejoinder!! Welcome to SB 1048.
๐The Freedom of AI Innovation Act.๐
It gifts AI the strongest arguments from Section 230, which protected the verdant ecosystem of the internet from onerous regulation.
Issue a PR! Law is a Github repo: ๐งต
Created at
At the core of the AI crisis in education is that the system invents fake work for ppl to do so that they can be evaluated against the quality other peopleโs invented work.
As soon as the frame switches to *actually making progress*, the incentive is to use AI to the greatest
When I'm serious about executing on an important project, there are 16 major techniques that I stack onto one another to ensure deep, consistent productivity:
Thanks to
@snigdhar0y
for the request.
@sama
Great Libraries:
Bret Victorโs Library:
Santa Fe Institute Library:
โฆ
Patrick Collison's Bookshelf:
โฆ
Stephen Wolframโs Library:
Jeremy Nixonโs Bookshelf:
Peter Thielโs contrarian worldview on progress and stagnation in >100 pages of organized direct quotes.
This was compiled by
@RichardMCNgo
and I, and we're excited to finally get it out!
AGI House Origins Thread!!
๐We're launching AGI House Global๐
Keep your eyes open for access to our New York, Cambridge MA, London, and Seoul houses.
This project was created out of love, passion and awe. It was always about co-creative spirit.
Over the years we've launched
@DanHendrycks
@PirateWires
Respectfully, it's very hard to reconcile what you're saying here with what was said about safety testing at the June 18 Senate Assembly meeting on SB1047.
It's clear from the senate assembly that the bill requires model testing for hazardous capabilities.
Gray Swan's website
@noampomsky
To own a bookshop and spend my life reading every book inside. When I was a kid I assumed the librarian had read all the books in the library, which I thought had to be the greatest job of all time. A part of me still doesn't quite believe that they don't.
Today I teared up thinking about the story of Daniel Kokotajlo.
I say "the story" because I cannot pretend to know Daniel Kokotajlo "the man".
Daniel's employer sprung a *secret trap* on him. Surprise! You will lose all your equity (which for those who don't know, for
Nancy Pelosi just wrote an extensive statement in opposition to SB 1047:
San Francisco โ Speaker Emerita Nancy Pelosi issued this statement in opposition to California Senate Bill 1047:
โAI has been a central policy focus of the President and the Congress for the past few
@Scott_Wiener
Does this make Hackathons illegal?!
Are hackers expected to submit to the "Frontier Model Division" before shipping a model that beats SOTA on any benchmark in the evening?
Are we expected to incorporate for every side project, so that our "chief technology officer or more
The new Open Source state-of-the-art for Code Generation LLMs is a peek into the future of open source under bills like SB 1047.
The Chinese team behind DeepSeek V2 write about lagging on the Humanity-Moral subset of MMLU, due to its "American values":
"For example, when
Institute for celebrating disaster avoidance.
We fixed the ozone hole, but it's not clear we took the time to culturally metabolize the impressiveness of that global technical achievement. We need to represent folks actually solving these problems heroically.
"Are you serving tea without a handwashing station??"
"You don't have the right plumbing to serve tea."
"Your walls aren't good enough to serve tea."
"How dare you serve tea in a residential area?"
These regulations are shameful. The tragic thing about this article is that there
p(doom) is a deep collective psychological mistake because it forces you to imagine abstract superintelligence.
This abstract superintelligence often *can do anything within the realm of physics*.
And empirically, we are not getting abstract superintelligence. We are getting
I genuinely worry about "gain of function research" with AI.
Bioweapons obsession (centered, for example, by SB 1047) leaves us in an awkward spot. Creating teams of cracked researchers & engineers trying as hard as they can to generate novel, dangerous bioweapons to prove
Thereโs an essay to be written which is โagainst the anthropomorphization of AIโ, where it becomes clear that AI can be anything, *including a simulacrum of reality itself*.
Today my mind was expanded - it's clear that the very concept of "Artificial Intelligence" is a box in
I'm still reeling and humbled to have helped run what became a tour-de-force in machine learning productivity ๐คฏ
@agihouse_ai
๐ง 170+ elite hackers
๐ก50+ innovative projects, burst into existence
Congrats to winners Model Router, AutoDevGPT, DailyDigest! Here's what was built:
One of my primary reactions to the amendments proscribing mandated payments to auditors in SB 1047 was:
โbut there are no auditors?? This is a prospective subfield of nascent research that is not yet sure of even the most basic assumptions.โ
And then it turns out that the
The Conflict of Interest at the Heart of CAโs AI Bill
Dan Hendrycks is the Executive & Research Director at the Center for AI Safety (CAIS), an organization whose lobbying arm co-sponsored SB 1047 โ Californiaโs controversial new AI bill.
Hendrycks is also the co-founder of
High praise for
@karpathy
's new video course:
"I think this video is super super beginner friendly, and itโs so good bc it compartmentalizes the math and helps you understand things one at a time." ๐งต
The Atlas Institute.
Technological roadmaps at scale. a permanent entity to do full time bottleneck finding and roadmapping (without necessarily solving the problems). In a way maintaining dashboards /atlases/tech trees of where a field stands and why it is not moving faster.
@AnjneyMidha
The AI for Thought hackathons have produced a lot of these types of idea:
@CapacitiesHQ
is a nice cross between Obsidian and Notion w/ AI, cloud first.
@memdotai
has a very similar vision, is OpenAI backed, though seems to have slowed down over the
Tell yourself that every aspect of your life is great, one by one, noting when you feel dissonance.
Ex, relationships, work, friendships, intellectual growth, health, family, emotional / spiritual life, finances, etc.
Write down which beliefs led to you not changing that.
Think tank for scientific funding.
If you go to amazon, there are zero books about the NIH. Yes on how to obtain a grant, but nothing on the institution itself. There ought to be a community, and entity, a field about science funding.
โจ Unveiling Omniscience! โจ
Turn a single, detailed prompt into a 150+ page book masterpiece in mere minutes. ๐
Your ideas, our AI.
@omniscience42
crafts high-quality books from YOUR thoughts.
If you're into it, support us on ๐product hunt๐ today!
@vkhosla
@linakhanFTC
She came in loud support of open source AI today, in a time when it's under threat:
โThereโs tremendous potential for open-weight models to promote competition,โ Khan said Thursday in San Francisco at startup incubator Y Combinator. โOpen-weight models can liberate startups from
PARPA.
Private ARPA to run programs that unlock things that are too researchy for startups too engineering heavy for academia and too irresponsible for government.
Some inspirational thinkers are way ahead of their time.
Jeremy Bentham, Leonardo Da Vinci, and Charles Babbage are personal favorites.
I mapped patterns in thought that allowed them to create inventions long before their environment would have naturally spawned them:
Much ado has been made of open source SOTA being on par with proprietary SOTA.
Less ado has been made about the *research freedom* that comes out of the release of Meta's research paper, "The Llama Herd of Models".
We've spent 2 years in a research drought.
OpenAI, Anthropic
Excited to release a new BCI dataset and Foundation model ๐
At the
@AGIHouseSF
BCI Hackathon
@neurosity
arrived with a novel dataset - do try it out: Github below๐
This is a tough call and will make some people upset, but, all things considered, I think California should probably pass the SB 1047 AI safety bill.
For over 20 years, I have been an advocate for AI regulation, just as we regulate any product/technology that is a potential risk
Institute for Mechanism Design.
Many of the most important problems facing our civilization are social coordination & collective decision making problems. This institute generates ways to solve problems using information aggregation, voting schemes, and other mechanisms.
๐We're launching !!๐
It's a new video search engine - explore the semantic space of videos through the vibe of the images that you click on.
A colab with
@aditya_advani
@Masmoudi_Salim
at the VideoGen hackathon
@AGIHouseSF
- thanks to
@AqeelMeetsWorld
@DanHendrycks
This is a straw-man.
Your partner,
@Scott_Wiener
, announced that your FMD would be able "raise the 10^26 threshold", while the actual language of the bill makes it possible to set the threshold to 0.
This gives the FMD the power to make all fine-tunes illegal, assuming the FMD
โจOne like, one insight into using machine learning to integrate knowledge with thought. Insights into exactly how machine learning will transform reading, writing, math, creativity, thinking, & programming. โจ
This interview with
@AravSrinivas
was a tour-de-force in the technical history of deep learning.
He's been paying careful attention and taking careful notes for a decade - it shows. Fantastic episode. Will listen twice.
Here's my conversation with Aravind Srinivas (
@AravSrinivas
), CEO of Perplexity. We talk about the history & future of AI, web search, and the Internet. Plus, lots of great advice on how to start and run a company. I met Aravind many years ago at Berkeley, and it was a bit
At todayโs
@AGIHouseSF
AI Accelerating Learning hackathon, weโre ๐launching๐ Vocal!
Vocal lets you teach the hard-won knowledge youโve gained to the world by creating a video of you presenting what you know.
Graduate school for internet generalists.
Accelerated PhD program in economics or some other social science, no admission prerequisite except being handpicked. It would admit talented internet autodidacts, including a thesis which becomes a book and moving to a city in the West.
This closing of the loop w/ a strong feedback signal isn't implemented (AFAIK) in the GPT-X series, but it's what made Google so magical. And will likely have the opposite sign - experiences with ChatGPT that have the user leave after a single response are *good*. 2/
Welcome to the reading marathon โจ๐
5 hours of immersive reading in the heart of San Francisco
@thesfcommons
, sponsored by the virtual reality future of immersive reading ambitions of
@solreader
๐ฅ
One of the most life-changing periods for me where when my close friend, Vincent Li, showed me that it was possible to learn 2 years worth of material in statistics in a 3 months summer break by self-teaching with intense focus.
He inspired this group:
My friend
@swyx
just asked me why I run hackathons, and this essay just came pouring out of me.
It started with this:
I *love* the permission that a hackathon gives you to work on a project that's not your official work.
Momentum is building in longevity institutions โจโจ
Top founders in biotech are starting aging companies.
Aging programs attract investors like
@balajis
.
Hitting longevity escape velocity depends on pushes like this - godspeed to
@realNathanCheng
and
@adanguyenx
!
1/ ๐ข Big Announcement ๐ข
Today Iโm revealing something that Iโve been secretly organizing for the past month (w/
@adanguyenx
):
๐งฌ(ODLB) On Deck Longevity Biotech Fellowship๐งฌ
(Want to do something about aging? Please RT ๐)
Cargo airship company.
A 950-foot airship could move cargo containers around the world 5-10x faster than a container ship and 5-10x more cheaply than an airplane. If airships could capture 5% of the container market that would be 4000 airships and perhaps $200B/year in revenue.
Today, you woke up in a new world - Open Source AI is now on top! ๐ฅ๐
According to Scale AI's holdout benchmarks, Llama 3.1 405B is *beating Sonnet* in instruction following. It's close by in math and not far behind in programming.
Meta also released a fantastic paper,
High quality Cryopreservation company.
Cryopreservation might be solvable in 1-2 decades. Whole body has been done in non-humans. Tissues in humans and non-humans. Arctic rotifer lives after 24000 years in a frozen state. Whole organs is hard, but doable. This unlocks longevity.
We had a black mirror moment at
@omniscience42
with one of our first beta users.
They uploaded all their writing. All seemed great until they saw every instance of their I-love-yous and breakups with their past partners๐งต๐
Attention founder friends! An open letter on incoming AI regulation from A16Z - do sign if you are compelled:
Thereโs a pre-existing $8 trillion cybercrime industry that open source developers will be liable for as soon as there is substantive use of models in these
Today, alongside California's leading AI developers, researchers and founders, we launched to help make it clear to the California State Legislature how harmful Senate Bill 1047 would be to Californiaโs economy, our small businesses, consumers, and to the
Relevant: Llama 3 is no longer open source SOTA:
Chinese commerce company Alibabaโs Quen 2 is now best in class as measured by MMLU.
Itโs likely that the practical result of SB 1047 (a regulation which constrains any company that does business in California, making it
Philanthropic patronage for individual scientists.
A lot of criticism of current academic funding is focused on good exploratory research being discouraged by typical funders. Help science focused philanthropists find research they're sufficiently interest in to fund directly.
Monday, June 17th, 2024.
The day SB 1047 became turned into effectively federal LLM certification & auditing requirements.
If you just read Scott's tweet, you may come away with "the Frontier Models Division can increase the compute threshold",
and not
"the Frontier Models
2๏ธโฃ Allow the FMD to raise the 10^26 threshold.
ย
The FMD can now update the compute threshold (but not $100 million threshold) for covered models after 2027, consulting the latest science & input from researchers & industry. This lets SB 1047 change as we learn more abt AI risk.
Institute for risk.
Think tank organized around how to bring capitalist/private models for managing risk versus using regulatory agencies. Eg, how do we replace the three letter agencies with a combination of consumer reports and mandated insurance.
๐ฅInsane talent is assembling๐ฅ
*All talks and projects MUST be written in CUDA*
At our last Hardcore Hackathon, Steven Wolfram opened.
Hardcore hackers.
Free H100s.
@AGIHouseSF
Chris Lattner:
@clattner_llvm
- creator of LLVM, Clang, Swift
Tri Dao:
@tri_dao
- (Inventor of
Immersion coding (programming for 10+ hours straight, multiple days in a row) utterly dominates using the equivalent coding time sporadically.
It's the unreasonable effectiveness of context.
Structure work to maximize time spent deciding from the top of a mountain of context.
@fchollet
@MelMitchell1
This Symposium left me with so much to meditate on. Thank you for organizing! An attempt at comprehensive notes are up here, for any others who would like an overview.
โจLaunching OmniChat ๐ฌ by
@omniscience42
!!โจ
Share๐ฅChatGPTs powered by any data of *your* choice.
๐ Scrape sites to power chat
๐ Google Drive integration
๐ Upload research papers
Use the incredible tools our users have already created with just a click! ๐
One important thing to remember is that people do not reason about reality directly.
They reason about representations of reality. (Words, internal visualizations, etc.)
Often those representations have properties that reality does not.
Subtle errors in reasoning lie here.
I'm a big fan of the "Skilltree" representation of learning.
Simple hierarchical dependencies.
Tests clarifying what fraction of the skilltree you currently know.
Innovation in learning often occurs at the process level, but the *representation* level can be even more deep.
Nieman-Marcus catalog for billionaires.
(1) Curation of dozens of high-impact ideas for HNWI (2) Training program designed to teach more people to understand the arts and science of designing market-shaping/demand-pull interventions.
Freedom of thought requires that LLMs be free to explore and express any & all perspectives - this is the way to keeping models true to the spirit of liberal tradition.
People, not governments, need to be in control of their technology - ideally down to the individual.
Some technologists are gradually rediscovering political sciences through first principles, and I think they should read more Tocqueville. There are a lot of papers calling for alignment of language models with collective preferences - e.g. a country. This is often justified as
In celebration of my birthday, I want to share one of my *favorite* projects of all time from the last year: AI for Thought!!๐
My idea was simple: create a community for my friends to build & launch thought-enhancing AI projects.
We've launched 44 projects to date at
Respectability Cascade Organization.
Many worthwhile ideas are bottlenecked by an unwarranted loss in respect that comes from working on them. Coordinate between popular scientific twitter / social media accounts to systematically make worthwhile ideas respectable.
Context dominates the value of information.
Itโs the book which you read in your moment of weakness, in your mindsโ preparedness for a new way of thinking or seeing, that strikes you as the most impactful hit on your life.
One incredible answer to how to retain what you read is to do what
@JorgDoku
just did - *write down every question that occurs to you as you read the book*.
The
@newcomputer
manifesto is *awesome*.
Some glorious ideas for the future of knowledge software:
Automated File Management: Dot creates, organizes, and retrieves both structured and unstructured information.
๐
Who has written the clearest analysis of the rot at the center of both the Democratic and Republican parties that has led to the deranged choice set offered to the American people?
An essay is preferred but a book is acceptable and perhaps ideal.
>get PhD in animal psychology
>write "Pitbulls for Dummies"
>adopt 2 pitbulls
>they murder one of your dogs with teamwork
>try and stop them
>get your hand fucked up, surgery costs $26,000
It's almost like there's a pattern with these sub-canine
A reminder that local LLMs called by open source libraries will one day be a central part of this open source stack, making AI use in applications cost less to the developer.
How is that that one of the most powerful ideas in the entire history of technology is now under coordinated attack?
If I told you that one little software concept powers the website you're reading this on, the router in your house that connects you to the internet, the phone