So excited to team up with @MaxJunestrand and @WeAreLegora!

If this works, it’ll be the biggest business model transition, possibly ever. On the scale of Apple’s transition from niche PC maker to iPhone maker and app platform

Every week a new model comes out that “beats” the previous SOTA on benchmarks. But benchmarks don’t matter - models need to be evaluated on YOUR data. Introducing Superpipe, a framework to build, evaluate, and optimize LLM pipelines on your data. Built w/ @benscharfstein

The future is starting to look like the future

Prediction: web3 will flip the social media stack upside-down. A thread

Thrilled to announce Stelo v2 and our developer API @stelolabs on our mission to make web3 safe and easy for everyone! Couldn't have asked for better partners than @AriannaSimpson, @janehk, and the whole @a16zcrypto team of superstars on this journey

The training set may incidentally contain information about needle in a haystack tests. Or it may include text about “testing if models pay attention”

18 months ago I was in a big-tech job, unable to leave and pursue my dreams because of US immigration laws. Tomorrow, I launch my first real venture-backed startup. I'm so grateful for the freedom I have now and believe every immigrant deserves a shot at their dreams.

Side note: Prof G's about to hit 420.69k followers

The Twitter game, in 3 steps: 1. Do something unique, challenging & useful at a high level (99th-ish percentile) 2. Reflect and break down the process you use and share it to create an audience 3. Create a course, newsletter, videos etc. to monetize (optional)

Let’s not ignore the role that FB played in getting IG to where it is today. IG - hired great people from FB - built their ad stack on FB’s - Used FB as a growth channel - Used FB’s deep pockets to acquire users Unclear if it’d be as successful without FB.

Introducing SuperOpenAI, a lightweight logging & caching python framework for LLM development ✅ all LLM inputs and requests logged locally ✅ includes token usage, cost and latency ✅ responses cached in-memory, never wait when re-running code built w/ @benscharfstein 🧵👇

Financial freedom is a dubious goal. I learned this when I hit my financial freedom number multiple times, but each time the goalpost moved, and I never felt "free". Here's what you should strive for instead. #ship30for30 (10/30)

We've made the difficult decision of sunsetting all of @stelolabs web3 products. I'm proud of our small but mighty team for building 4 best-in-class products in less than a year. Unfortunately, our assumptions around the timing of mass crypto consumer adoption turned out to be false. We're continuing to build something new™ - will share more soon! In the meantime, make sure to install @PocketUniverseZ , @_joinfire or @wallet_guard to replace Stelo.

Made it to ETH Denver, who's around?

“Culture is a better onboarding to crypto than interest rates” - @ricburton at @launch_house

This makes sense when you realize the person hiring for the position is a mediocre mid-level manager wanting to hire cogs that fit seamlessly into the machine, not independent thinkers who push the machine beyond its existing limits.

When you pivot from consumer social to B2B SaaS

An under-appreciated driver of crypto/web3 is the $13 Trillion $ that has been printed since Mar 2020. This has led to bubbles everywhere: SPACs, growth stocks, and maybe - jpegs and early-stage web3 startup valuations. When the tide goes out, we'll see who's swimming naked

Web3 has been critiqued because Alchemy, OpenSea and Coinbase are "centralized". What matters is decentralization of power, NOT decentralization of technology/infrastructure. These cos are less powerful than web2 cos because of the user's ability to exit. Can't do that with fb

I spent years working on fake accounts at @Meta & payments fraud at @Square and learned a thing or two about fighting fraud at scale. Today @benscharfstein and I are launching @stelolabs to address web3 fraud & phishing. 👇

The interfaces we use to interact with crypto are still stuck in the terminal era. That's why we're building Stelo Explore - the GUI for every smart contract. It is to web3 what the command line was to the early internet. I'd love to hear what you think! stelo.com

Yesterday, news broke that Mistral is reportedly raising $600M at a $6B post-money valuation. ​ A few days ago Arthur Mensch (CEO) did an amazing 20VC episode with Harry Stebbings. ​ The episode is a masterclass on the current AI landscape. My top insights from the episode: ​ 1. Balance scale and efficiency. ​ Their Mistral-7B model filled the gap in the efficiency to performance trade-off space that was sorely missing. ​At 7B parameters, it could run efficiently on a MacBook or smartphone while still being sufficiently smart for real applications. Its popularity showed that there was a lot of interest for efficiency rather than scale and hence Mistral continued creating more efficient models. ☝️ ​ 2. The biggest bottleneck for model quality. ​ Data quality. ​ He thinks compute is no longer the bottleneck (for companies that have it), it's refining the data and feeding it to the model to drive continual improvement. 3. Robust evaluation methods. ​ Often models fail on niche tasks that are out of domain compared to the data you have. ​ The challenge then becomes mapping where they're failing and figuring out ways of improving it through tailored evaluation and iteration ​ Whether it's enhancing math reasoning abilities or demonstrating theorems versus medical diagnosis - which require very different approaches. ​ 4. Verticalization edges out generalization. ​ For mission-critical low-latency apps (trading, medical assistance), compact verticalized models may beat massive general ones for UX. ​ 5. Key factors that AI developers really care about: ​ • Cost and customization are obvious. ​ • Deployment flexibility: Devs want the freedom to deploy anywhere - cloud, on-premises, edge devices. ​ • Data Control: For enterprises, data privacy and securing proprietary data is a major consideration. That's why Mistral, OpenAI, and other big players use Azure and AWS for enterprise-grade security. 6. Brand reputation and trust are critical. ​ Arthur: "brand is important because trust is important in this domain, and open source brings trust." Mistral's approach of open-sourcing and distributing models has contributed to building its brand. ​ 7. Capital = Compute ~ Quality ​ Arthur thinks capital equals compute. And compute is correlated to quality, it's not completely dependent on it. ​ He thinks there's a strong opportunity for providing models that are the best of their class (for example, Mistral 7B is the best model at those parameters) because they might be sufficient to solve certain use cases. ​ The funding news makes it clear: they need more compute to be at the cutting edge of research since OpenAI, Anthropic and the like have raised a lot already. ​ Mistral only has ~1,500 H100s. In contrast, Meta will accumulate 350,000 H100s by the end of 2024 which is insane! 8. On future efficiency. ​ Arthur acknowledges it's an open question whether we can extract much more performance at lower compute costs, or if we're hitting diminishing returns. ​ But he believes "we can make models that are much better for a certain size" through experimenting more. ​ 9. How quickly does the cost of compute go down? ​ Compute costs reduce by around 30% every 2 years. But the real gains come from algorithmic efficiency. ​ There's been a 100x improvement in training efficiency over just the last 3 years through algorithmic advancements. ​ So Mistral's bet is going all-in on extracting massive remaining efficiency gains through hardware and algo advancements. ​ 10. What is the end state for the model landscape? ​ Arthur sees the end state as having more developed platforms that allow customization and making low-latency models for specific purposes. ​ Models themselves are just "a tiny part" - the differentiation comes from the data, user feedback, and deep understanding of the application domain, which is not commoditized. ​ --- ​ great episode @arthurmensch @HarryStebbings 🙌

Foundational models are going to get commoditized. Meta is spending $40B this year on chip designs, data centers, and their foundational models. Llama-3 is free while GPT-4, Claude 3, and other models are not. As @bgurley said, the "$20 concept is gone" - charging for consumer access to LLMs seems unsustainable if giants like Meta offer them for free.

Web3 needs to listen to its critics and address valid criticisms instead of getting stuck in a wagmi echo-chamber It's important to seek out and engage opinions/evidence that contradict your beliefs, not confirm them. Makes us all better

(1/11) An upcoming crypto bear market is both inevitable and healthy for the long-term success of web3. A massive amount of money ($13 Trillion) was printed by the fed over the last 2 years. This has caused unsustainable overvaluation among early-stage startups, protocols & NFTs.

Your book inspired me to write mine amandhesi.net/writing/my-gui…

Introducing Superpipe Studio: a free and open-source tool to curate datasets, run evals, conduct experiments and optimize your LLM pipelines for accuracy, speed and cost. @benscharfstein and I have spent the last year building with LLMs and working closely with a number of companies and encountered the same problems repeatedly: - evaluating LLM software is hard, because... - collecting high-quality, representative, correcty-labeled data is hard... - which makes it hard to objectively compare different techniques, models and parameters... - which prevents continuous iteration and improvement of LLM software If you want to build a winning AI product - it all comes down to continuous evaluation and optimization. Unlike traditional software, you can't build it once, write some unit tests and sleep well at night. You need a virtuous cycle where product usage generates data that goes into your eval system, which is used to evaluate new models/techniques or fine-tune your own model. @HamelHusain captured it best in this diagram -- We tried existing tools in the market and found none of them really worked well to implement this virtuous cycle, so we built a lightweight tool for internal use: Superpipe Studio. It helps you do 3 things: 1. curate and manage datasets 2. run experiments and compare them on accuracy/speed/cost 3. monitor your production pipelines (and expand your datasets with production data) We think everyone should have these abilities, so we're making Studio completely free and open-source with no restrictions. Studio is a work in progress and rough around the edges but still robust. We encourage you to try it and modify it to suit your needs. Building a winning AI product is more about process than about any one feature or idea — and it all starts with high-quality labeled data and a robust evaluation and optimization system.

in web3 twitter is the acquisition channel, discord is the retention channel

Couldn't agree more. Watch Avi Wigderson (one of the fathers of zero knowledge) explain the simplest ZKP out there piped.video/watch?v=5ovdoxnf…

This is what the web3 social stack might look like Reputation Content App | App | App | App Social graph protocol Identity protocol You will "own" your identity, graph, content and reputation and "bring" them to the application.

🔥🔥 session on Twitter for founders by @thatguyBG at the @launch_house web3 residency this evening

As a startup founder most days it feels like the ground is shifting beneath your feet. And then one day the ground LITERALLY shifts beneath your feet and falls out… scenes from our NYC office

David Perell @david_perell & Write of Passage transformed my relationship to writing. A master of metaphor, David uses them to great effect to teach his ideas on writing. Here are 3 of his most powerful metaphors to improve your writing. #ship30for30

For the average person the security challenges in the first 10 years of crypto were around secure custody. The security challenges of the next 10 years will be around transaction security.

gm. Realized this weekend that DeFi staking is hard to understand because it has come to mean many different things So I mapped out the 3 broad types of staking and their risks & rewards (spoiler: includes $OHM) Here's a 🧵

fr fr

Meditation = staking for the mind

Elon’s tweets reaching 120m people are the most effective ad campaign on earth

Reporting from the future 🫡 AI x Apple Vision Pro at the @southpkcommons @OpenAI hackathon w/ @ryanorban @benscharfstein @andrewcyu

Public and private keys to prove ownership is a feature. Users needing to think about them is a bug.

The year is '89. Everytime you buy coffee with your credit card there's a 1% chance you lose ALL your money. So you have 3 cards - one for coffee, one for tele-shopping, and one you keep locked in your parents' basement. This is the state of crypto in '23. @stelolabs fixes this

I'm assembling a team

I'm now onto my third open-source project and man, writing good documentation is hard work and a labor of ♥️ So much respect for doc writers who sweat the details. Particularly impressed by @jxnlco and the instructor docs

You know that feeling. With sweaty palms and nervous excitement, you hit the publish button on that essay you've been working on all morning. And then.. crickets. It sucks. Fellow writers, here are 5 ways to make peace with the crickets of indifference #ship30for30 (11/30)

I came to the US in 2010. Got my green card in late 2020, and freed myself from the shackles of the US immigration system. It took 10 years. And I'm one of the lucky ones. Ever wondered why there are no immigrant actors, comedians, performers, writers? Here's why #ship30for30

Creating material abundance is the long-term arc of all technological progress. In developing countries people embrace this because they see their lives improve in their lifetime. In developed countries many people take their quality of life for granted

Is data truly a moat for AI companies? You can't go to a coffee shop in SF without hearing some version of "data is the new oil". But Marc Andreessen makes a provocative claim... According to Marc, "the amount of data available on the internet and generally in the environment is a million times greater... it swamps the value of your data. Your proprietary data will be a little bit useful on the margin but it's not actually going to move the needle" He cites this as proof — that there is no large, sophisticated marketplace for data. If the data were valuable, it would have a market price. Ben agrees that the idea of data as an artifact for sale is massively overrated, since most data is either widely available or not general-purpose. But he adds — "Almost every company can increase its competitiveness through the use of its own data... every company has data that if fed into an intelligent system would help their business and almost nobody has data they could go sell" My take: I think Ben has it right. That said, there are clearly some companies that have data worth selling. Google and OpenAI have signed deals to use Reddit data for model training, rumored to be for $60m/year. But this data is only valuable if you're building a cutting-edge model for a general purpose consumer product. It's not valuable if you're building a small domain-specific model. There are at least 3 types of data within companies that is valuable to others and has a market: ⋅ User-generated content — social platforms like Reddit/X have data that's optimized to be engaging to humans ⋅ Video — video is expensive to create and store, so most high-quality video is locked inside walled gardens like YouTube and Netflix ⋅ Real-time data — similar to UGC and video, high-quality real-time content like news media is often locked behind paywals But for most companies that don't have this type of data, the questions they should be asking are: 1. Do I have a valuable proprietary data asset? 2. How do I use the data to create a competitive advantage? 3. Do I train my own model on my proprietary data or trust OpenAI/Google enough to feed it to them?

"Revenge anything" is a toxic mindset. It's great to start a company and take your destiny into your own hands. But do it because you believe in yourself and want to build something you love, not to get revenge

when it comes to tech week, i try to find the most unusual non-mainstream events, like @stephsmithio's internet pipes event yesterday really enjoyed the conversation about the intersection of creators, niche internet media and niche software w/ @shl @walthickey @danshipper

Sam Altman said taking a sabbatical was the best decision of his career. I had a similar experience and I think more people should take them. Here's why... In 2021, after 7 years working in tech I left to take a creative sabbatical I wanted to disconnect from work but at first, I just couldn’t do it. I threw myself into figuring out “what’s next” like it was a math problem Slowly, I became aware of the subtle anxiety and subconscious beliefs that were driving me. I recognized the degree to which my self-worth as a human was tied to my work For years this affected my relationship with work and caused me to work from a place of fear, insecurity, and scarcity But getting off the ‘success treadmill’ or the guilt of constantly producing something to be a valuable member of society was liberating It gave me the much-needed time to focus on my ‘why’. And the experience changed me forever I’m more connected to my intuition. I have a better understanding of what motivates me and what I’m genuinely interested in And most importantly, I learned to focus on expressing myself rather than on impressing others, on doing work that is spiritually aligned rather than what's deemed admirable by others

Reminder that it's all a game

Decided to be a contrarian and not go to NFT NYC ama

After getting my green card and leaving @DoorDash in March, I took a sabbatical to recharge, explore, and decide what's next. I'm so glad I did because I discovered the magic of Web 3.0 which has now become an obsession.

1/ Towards decentralized social media: token incentives are a start but not enough 🧵

🎄 Nifkey is now open to everyone. Merry xmas friends!

Tokens are like grease in a car's engine. They can make the car go faster and smoother. But it's the car doing the moving, not the grease. Value comes from creating products and services people want, not from dropping a token.

Alan Turing's 1950 paper "Computing Machinery and Intelligence" was the first paper on AI. It's one of those papers that everyone in the field has read, or at least pretends to have read. But what makes it so groundbreaking?🧵

Them: you should touch grass everyday, it's good for mental health Me:

"This is why we need crypto" - crypto twitter, when literally anything happens

Prediction: DAOs that start with concrete goals will succeed over more generally-aimed DAOs. eg. 📈 a DAO to buy the constitution at an upcoming sotheby's auction 📉 a DAO to buy valuable collectibles

$BTC 69000 $ETH 4200

1/ Reputation transfer is an important, unsolved problem in crypto. Especially as we the pseudonymous economy grows. Quick thread

. @danboneh has a couple of lecture slides on it too. Dan, wen video lectures? :) cs251.stanford.edu/lectures/…

After spending a week at @launch_house LA it's clear that the web3 energy just hits different. So many smart, driven, young people pouring themselves into reinventing the web. web2 doesn't stand a chance. Some people on a 🚀 trajectory you need to follow 👇🏼

What is success? Every society, culture, family, and individual will tell you it's something different. So how do you define it for yourself? Here's a #ship30for30 atomic essay on my definition of success 👇🏼

That’s incredible. So grateful that @davidmarcus @kittukolluri @KamakshiSivara1 and a few others signed mine back in 2017. Helping someone get the freedom to pursue their dreams is a priceless gift

Prediction: major universities will have a blockchain/crypto/web3 major within the next 3 years. It will be an inter-disciplinary collab between computer science, economics, math, political science and policy departments.

How do you think about the perverse incentive this creates for users to game usage metrics to try and extract ownership out of the network without contributing value? And the resulting inability of builders to accurately measure product-market fit.

2 days ago Evolutionary Scale launched ESM3, a protein language model that can generate novel proteins and a lot more ​ The numbers: • 98 billion parameters • 2.78 billion natural protein sequences in the training set • 771 billion unique tokens of training data ​ What's the hype about? ​ [1] It speaks fluent protein ​ ESM3 doesn't just understand protein sequences - it grasps structure and function too. This multimodal (sequence, structure, and function) approach is a first. ​ [2] Architecture custom-built for protein data It's a masked language model specifically designed to handle the complexity of protein data: ​ • Multi-track input/output: Processes protein sequence, structure, and function as separate token tracks. ​ • Geometric attention: The first transformer block includes a special geometric attention layer to handle 3D atomic coordinates. ​ • Masked prediction: During training, random masks are applied across all tracks, and the model learns to predict the masked tokens. ​ [3] 500 million years of evolution in one model ​ ESM3 generated a novel green fluorescent protein (GFP) that's so different from known proteins, that it's comparable to what we'd expect to see after 500 million years of natural evolution. ​ [4] Data quality is still king ​ The team didn't just rely on natural proteins. They augmented their dataset with synthetic examples to give ESM3 an edge. [5] Chain of thought isn't just for language anymore ​ They used chain-of-thought prompting to guide ESM3 in protein design. NLP techniques are proving their worth even in hard sciences. The team is also releasing ESM3-open, a 1.4B parameter model, for academic research. ESM3 is the best example of the ideas behind language models being applied beyond language and code. ​ Kudos to @alexrives and team! 🙌

Yesterday @kevinrose, tech entrepreneur and founder of @proof_xyz got hacked and lost $2m+ of cryptoassets. THIS is why we're building @stelolabs. We just had our highest-ever user growth day today. And we're hiring founding engineers (DM me)!

yeah. the pay package being voided would set a truly awful precedent for the future of american capitalism. america attracts the best in the world precisely because you can get rewarded for doing hard things. that's at stake

Congratulations @dwr and @varunsrin! The crypto ecosystem is so much better off for your vision, clarity and execution

Alright y’all I did it. Flying to ETH DENVER right meow

The two most important days in your life are the day you are born and the day you deploy your first ethereum smart contract

Chamath talking about AI

gm. woke up like dis..

I'm 100% sold on mkdocs + mkdocs-material. They strike a beatiful balance between structure and freedom. Amazing job @squidfunk and Oleh (who is not on X afaict)

gm to everyone building the new web

IMO no one wants to "own their data". People want financial upside, convenience, security, delight etc. Anyone building web3 products to help people own their data needs to ask how they help people get what they really want. Human needs and problems don't change much.

@hasanminhaj oh wait that already happened

Excited to be a founding fellow for @beondeck product management and help @andrewcyu and the ondeck team build the product management school of the future.

the lesson I've learned over and over again is to not chase success at the expense of personal alignment as Letterman said to Seinfeld: make sure you fail doing exactly what you want to do. that you can live with or as @david_perell said: surrender to your nature

The Stelo developer API is now live. Give it any transaction or signature and it'll give you: - a rich human-readable explanation, including simulation results (assets going in & out) - a risk score and explanation of risk factors Human-readability and risk analysis is going to be tablestakes for wallets. Docs: docs.stelolabs.com

the question is not whether we’re gmi. the question is what we’re going to do after we’ve made it. figure out your values. hold them dearly. live them daily.

One simple way to visualize this is to imagine this: If you wanted to check out @naval's twitter feed you'd go to naval.com/twitter not nitter.app/naval

alien gang wya @chain_runners

gm to all my new followers. i have a feeling you're all gmi

Experiencing this opened my eyes to the potential for NFTs + IRL experiences. Last night an NFT (ticket) unlocked the IRL experience (gallery) which unlocked an NFT (merch). Now extend this to concerts, festivals, etc and the possibilities are endless

I used to think knowing a bit about a lot of things was great. Starting to question this. The world rewarded this 20 yrs ago (hedge fund managers, consultants). But maybe today's world rewards people who know a lot about one domain and aren't distracted by anything else

lmao JCal just called me mid because I called out his encouragement of "revenge startups" remember, you too could be an investor in Uber if you were a little more toxic

I'm constantly looking for the latest AI developments since the space is developing at warp speed. Here are the 12 most fascinating developments that piqued my interest in the last 4 months. 1. Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data

Tribe AI is putting project-based ML on the map. As an early member I'm continually impressed by the quality of the community that @wis3foo1 and @jacsrice have built @tribe_ai

yesterday Trump appeared on the all-in pod and made promises to increase high-skilled immigration by giving all college grads a green card but I don't believe him. quick story... in 2017 after he came into office I was almost forced to leave the country because of changes he made to immigration rules in late 2016 I had just left my job at facebook as a PM and was trying to build a fintech startup. I was on an H1B work visa and applying for an O-1 extraordinary ability visa which would let me be a startup founder and then Trump came into office and removed "premium processing" for H-1B visas. this meant you'd have to wait 9-12 months to get your H-1B approved so if you ever got fired or left your job on an H-1B, you might have to leave the country for 9-12 months unless you immediately found another job this was a needless change. there was no justification for removing premium processing — applicants had to pay extra to get it anyway I had to scramble to find a job because I didn't want to take the risk of having to leave for that long. my whole life, my friends and my career were in the US and i'm just one example — i've heard many stories of friends who were impacted much more because of this change so forgive me for not believing Trump's promises (even if he had the power to enact these changes, which he doesn't)

No moat other than brand. Low switching costs for hosts and guests. No business diversification. They've been working on experiences for 3+ years and haven't nailed it.

Still the best writing advice you can in read in <5 mins by @morganhousel h/t @david_perell

Confirmed that the @NotionHQ office looks like you literally walked into a Notion page Also confirmed that AI events are ⚡️

Just experienced my first rug-pull. It could have been a lot worse than it was. Part rite of passage, part reminder to DYOR and be careful out there!

So @MidsummerDAO should focus on getting 1 specific film funded and produced @Climate_DAO should focus on buying a stake in one specific company @goodkarmadao should focus on signing one specific artist Motto: dream big, start small

Web3 social media will look more like the early FB/twitter dev platform. Users will own their identity and social graph through permissionless, decentralized, cryptographically verifiable protocols. Social media apps will be built on these identity and graph protocols.

Our bodies contain the wisdom of centuries of evolution passed down through our ancestors. And yet in the Brain centric knowledge economy we rarely pause to listen to the body. Here's a story of how listening to my body got me to the best physical shape of my life #ship30for30