@primateriacom, EE @stanford and berkeley, train software @ycombinator s19
Palo Alto
Joined February 2018
- Tweets313
- Following525
- Followers6.2K
- Likes2.2K
Pinned Tweet
How we built lazy updates in our fully onchain game @primodiumgame:
A fully onchain game is gas-intensive, so we need ways to not have players sign as many transactions. Enter lazy updates where we calculate how much game resources are needed only when players perform actions.
In the first version of Primodium, players had to place mines next to nodes before placing conveyor belts to transport resources to the main base.
When the player claims their resources from their main base, I had to figure out the following:
1. How to keep track of each player's mine
2. Allow mines to connect to factories
My first implementation consisted of running depth-first search starting from a player's main base when resources are claimed. Players could place up to four mines next to each adjacent tile.
When we introduced factories that take in resources from mines, players rushed to build complex designs for optimal resource production. For instance, player @CincauHangus created the following factory design:
After joining our team, @AlirezaN95 redesigned our resource components, built with MUD by @latticexyz. In v0.6.0, we gave each player a global resource production rate that increased when each mine was built on a resource tile.
This meant that updates were now atomic, and when a player performed an action, we could calculate their resource count via their production rate in constant time. Factories would then simply subtract from the global rate as they consumed resources to produce goods.
In subsequent versions, we observed an engagement drop-off right after building a mine, since the path-building mechanic was confusing for some players.
A global resource production rate meant that we could remove paths and nodes altogether, enabling the complex gameplay loops that we have in @primodiumgame today.
For our latest release, we re-engineered how we store resource production rates within our smart contracts, which paves the way for an exciting upcoming feature.
123 looks more predictable than 12 because three digits updates the bayesian posterior confidence that I'm typing sequential numbers instead of just twelve
Emerson Hsieh retweeted
Introducing edith-1, a decision model for online trace monitoring. We trained it to catch agent failures, reward hacking, and problems with tasks, graders and environments.
On our eval, it beat Sonnet-5.5 in Claude Code on balanced accuracy, at an estimated 1/274 of the cost.
Luofei is a brilliant and mission-driven engineer. Building a personal intelligence machine has been his dream since I met him more than ten years ago, and I’m glad we’re at the right time for his vision to come true. Congrats on the launch! 😊
Say hello to Oki Home, your Personal AI Computer.
We're building a future of limitless, personalized intelligence for every consumer, while keeping all your data in your hands.
Reserve our first batch, shipping mid-december! okihome.ai
Emerson Hsieh retweeted
Introducing pgrust v0.3: the bug-smashing release.
We evaluated pgrust v0.2 and v0.3 against the same held-out corpus: 164,000 test cases from 50 open-source projects.
pgrust v0.2 diverged from Postgres on 1,454 cases. pgrust v0.3 diverged on 11.
pgrust is 300x faster than Postgres for analytical workloads, and we're a drop-in replacement.
Held-out corpus included: @djangoproject, SQLAlchemy, Hibernate, @rails, and more.
@pgrustdb is now much more Postgres-compatible and reliable than before. @JasonSeibel and I don't recommend running pgrust in production (yet), but we are successfully running it ourselves internally.
- @AntithesisHQ and @workersdotio for simulation testing
- Detail.dev by @danlovesproofs for LLM-assisted code analysis
- Multiple home-built differential fuzzers
- Kani for formal verification
Our goal is for pgrust to be truly 100% compatible with Postgres.
There are still bugs. You should still run Postgres for any data you can't afford to lose.
But we want people trying pgrust on real workloads in non-critical environments. If you need Clickhouse performance queries in Postgres, are frustrated by the many gotchas in Postgres, or just geek out about databases and Rust, feel free to send me a DM as we would love to chat.
Emerson Hsieh retweeted
Primateria is led by two of the most ambitious, mission-driven founders in the industry.
Throughout the Primodium -> Primateria journey, they've developed leading-edge research from first principles while holding themselves to an exceptionally high bar: building a commercial enterprise capable of the largest possible positive impact.
I'm excited for what comes next!
Announcing Primateria, an AI-native scientific lab.
We’re starting with one of the hardest problems in quantum mechanics: the electronic Schrödinger equation.
Our first model tackles it in transition metals, a notoriously difficult regime.
Try it: app.primateria.com/
Emerson Hsieh retweeted
In 1929 Dirac said we already knew the equations for all of chemistry - they were just impossible to actually solve. Almost 100 years later, and we are getting close!
Announcing Primateria, an AI-native scientific lab.
We’re starting with one of the hardest problems in quantum mechanics: the electronic Schrödinger equation.
Our first model tackles it in transition metals, a notoriously difficult regime.
Try it: app.primateria.com/
Emerson Hsieh retweeted
a cool launch from the coolest brothers I know
Announcing Primateria, an AI-native scientific lab.
We’re starting with one of the hardest problems in quantum mechanics: the electronic Schrödinger equation.
Our first model tackles it in transition metals, a notoriously difficult regime.
Try it: app.primateria.com/
Emerson Hsieh retweeted
Announcing Primateria, an AI-native scientific lab.
We’re starting with one of the hardest problems in quantum mechanics: the electronic Schrödinger equation.
Our first model tackles it in transition metals, a notoriously difficult regime.
Try it: app.primateria.com/
I built a Lean prover farm with Leanstral on two RTX Pro 6000s. On a small benchmark, it generated and compiler-checked ~2,500 proof attempts per hour.
Then Fable 5 formalized the disproof of Erdős #608 overnight without needing it.
The mathematics is by Füredi and Maleki, described by Grzesik–Hu–Volec. The new contribution is the Lean formalization, now linked from the Erdős Problems database.
Repo: github.com/primateria/erdos6…
Merged PR: github.com/teorth/erdosprobl…
Emerson Hsieh retweeted
Working on something very difficult in hardware, but if works, should have very little market risk because of the tremendous value it provides.
Does anyone know of any hard tech founders or operators that are happy to give some advice on what the pitfalls may be for us? Thanks!
Emerson Hsieh retweeted
I’ve got a great designer who worked on the @primodiumgame client, very passionate about games, but also picked up iOS dev (Swift) for @tubdotapp.
if you liked the Primodium interface, you should work with her. DM for intro!
Emerson Hsieh retweeted
A few months ago, I wrote this article internally on how we would consider launching a token at @primodiumgame.
You can also read the full article or this 🧵for a summary: xitter.cf/tax_cuts/status/189153…
Thank you @primodiumgame team, it was great working with you all!
Read about everything we've worked on together: developer.primodium.com/
Today, I am announcing that we are open sourcing our fully onchain game and that @intemerson and I will be helping the team transition to new opportunities.
A brief 🧵 on what happened:
