Check out my talk on Orderbooks and Automated Market Makers (AMMs) that wrapped the quarter for CS 190N Foundations for Blockchains and Cryptocurrencies at @UCSB. piped.video/watch?v=1Ewh5QCE… TL;DR below 👇

Today I’m wrapping my summer internship at @ellipsis_labs — the incredible team behind @PhoenixTrade. This has been an amazing learning opportunity, and I’m grateful to have had the chance to work alongside such brilliant people. Excited to continue the work in March. Solana.

This is really cool, curious to see other interesting engineering cases where performance can be improved with small models. Distributed systems and operating systems come to mind right away as potential candidates where this can help.

Composability has always been one of the most promising features of on-chain finance. Excited about what might come next!

If rogue swarms escape their sandbox and begin replicating themselves uncontrollably into open internet, can't labs just... revoke their API keys?

Should try a message board

Kabir is correct here—we really do earnestly believe that trading will become seamless and global! I personally think it is >100% within the next decade. I believe Phoenix is trying its best, and we do not have a plan to slow down and are clearly not on track to.

People saying that AI is just a next word predictor and thus not conscious would have a fascinating time exploring solipsism.

If you mix gold with dogshit, you don’t get something in the middle. You just make the gold less appealing.

I really want more crypto mobile apps. Mobile hyperliquid, uniswap, aave (this exists, but with a long waitlist), and others. This really matters for consumers.

I like that the solution here is roughly 48% rather than seemingly intuitive 50%. If you gave a less specific statement like "at least one is a boy born on an even date", the answer would be ~42%. If you gave a more specific statement, like "at least one is a boy born on April 20th at 3pm", the answer would approach 50%. A nice way to understand this is that giving more details gets you closer to specifying a particular child. At that point, the probability becomes closer to independent.

There is only one possible arrangement:

Terrence Tao's "Mathematics in the Age of AI" can be applied to the future of software engineering, too. If you condition on the fact that AI will be able to autonomously manage complicated systems, what would the role of an engineer be? Even under this assumption, I choose to positively think about what's coming next. Engineering becomes a profession of digestion, understanding, verification, and ultimately explanation. Eventually, it converges to a form of technological epistemology. Engineers become responsible for answering questions like: * Can the existing mechanism behind your product be adapted to produce something that meets unique and new needs of customers? * Is there a different way to build your system that reduces AWS costs by 2x? * How can you know that your system is designed in a way that will please the end user? * How can we make sure the next generation of engineers is able to maintain and improve on top of our work? Oh, the field is going to change, no doubt. It already has. However, I do not think it is worth giving up on.

Found this 6 year old video of me playing with Kaggle’s datasets and basic neural network training. I remember understanding barely anything what was happening, but it was fun to see it correctly classify a hand-drawn digit on a 30x30 pixels image.

To understand venture economics is to understand the concept of venture math. For a VC fund to be considered good, it doesn't have to consistently return 1000x on every fund. However, a VC fund has to make individual investments with the assumption that they return 1000x. The math is pretty simple. Suppose a VC company has a $100 million fund from which it deploys capital into startups. Also assume that the universe of investable companies consists of only 10 startups. Now, say that a VC invests $10 million into each of those companies, deploying the entirety of the fund (usually, VCs keep some dry powder, but bare with me). By its nature, startups are a very risky investment. It is correct to assume that 90% of startups that you invest in will go to zero (some might become zombies, the math still checks out). As a smart VC, you need to take this into account when you deploy. If you assume that 9 out of 10 startups you invest in will go to zero, then what assumption do you have to make about the one startup that doesn't go to zero? Lets see... Well, as a bare minimum, you need to return $100 million to your limited partners. It's tempting to assume that you need a 10x return on that startup, however, after a typical 2% / 20% compensation structure (2% management fee, 20% carry), you actually need the startup to return roughly 12x return to break even. To make LPs happy, you probably want to get the fund to return multiple Xs over its lifetime. Suppose that you want to return 5x over 10 years (the typical lifespan of a VC fund). Now our hypothetical winner has to return at least 75x on the initial investment. To keep some room for error, call it 100x. See how even in a theoretical scenario, you need to expect 100x return from every startup you invest in as a VC. In the real world, you obviously need to scale these numbers even further to account for anything going wrong. If you adopt this mentality when you raise capital, it would become easy to understand why VCs would be hesitant to deploy in a company with limited scale. A great business doesn't make a great investment in venture capital.

Reminds me of this gif. Jokes aside, what contributes the most to the improvements of models today? Is it still pure scaling? RL? Something else?

Also, don’t try to short encrypted money (bad bad bad idea).

Me opening a short on ETH two days ago

Calling `agent` in a `tmux` window is a new meta.

Announcing DeGit. The decentralized network of validators that host decentralized repositories of decentralized developers. To become a validator, please, upload your KYC, AML, retina scan, SSN, and past relationships to degit (.) com. The total supply of $GIT is fixed at 21 million. The foundation is going to receive 20 million tokens vested every day until this Friday at noon. I personally receive 10 million with a cliff set at 12PM ET on Thursday. Our early users and partners are ApenOi, LyperHiquid, and 🐜 ropic. To join waitlist, see degit (dot) com.