Also, humanity has managed built Python API wrappers around this alien API with 30k Github stars and nobody really understands what this wrapper is useful for.

NLP in 2022/2023: Aliens left a black box technology that humans may call via a REST API (with API limits of course). Humanity is struggling to figure out what this thing can do, what it cannot do, how to tell it what to do, why it works, and whether it's dangerous.

Join millions who have switched to Grok.

I believe that trying to keep models private and "curating" them to prevent spam generation is the wrong approach. Let the all spam happen and then let the free market figure out the best solution to deal with it.

Good-enough AI generated content will finally push us over the edge. The spam problem will become so obvious that it can't be ignored. This will force us to invent novel curation/verification/incentive techniques that lead to a better web.

The web is already full of spam. Nearly all App Store and Amazon reviews are fake. Mainstream news media is shallow clickbait. Google's results are full of SEO spam. So far we've gotten away with ignoring the issue and basic curation. But browsing the web has become painful.

Controversial opinion: Spamming the whole internet with low quality AI-generated content will turn out to be a great thing in the long-run.

ChatGPT is better than I am at almost everything. Life sucks. Time to have a drink. .

Of course the causal/symbolic reasoning ability is still missing, and that won't change because of the substrate. But I no longer think it's that important. The vast majority of applications don't need it.

ChatGPT is the most impressive tech demo I've ever seen. With the previous models I was still skeptical about the broader impact, but now that's impossible. Humbled by how close this get to human-like understanding. We're not so special after all.

Also, why haven't we gotten the whale bus yet

In 1900, this is what artists thought school will look like in the 21st century. Feed in training data books and turn the crank to run backprop. The poor guy is responsible for adjusting the learning rate over time.

RT @cgst: 🚨 Now hiring a backend engineer to build simulation environments for DeFi at 🚨. If you're itching to tran….

RT @alive_eth: Crypto and AI are intersecting to create a design space that we've barely begun to explore. @cgst and @dennybritz at Runloo….

Our goal is to build something akin to OpenAI gym for the full financial ecosystem and use it to train profitable agents. If this sounds exciting to you, we just closed a seed round from @a16z @alive_eth and @DCGco and are hiring a founding engineer:

All code is public. This means you can build high-fidelity simulation environments without needing to make a lot of assumptions about hidden mechanics. The model of the world is given to you for free.

All data is openly available and free. It's a level playing field where anyone can build agents on top of the same data. It doesn't favor large corporations with big internal data moats or expensive infrastructure.

The past ~8 months I've been working on a new company called Runloop together with @cgst. We're building simulation environments and RL agents for DeFi (Decentralized Finance). Really cool opportunity because:.

RT @neilhoulsby: New paper from Brain Zurich and Berlin!. We try a conv and attention free vision architecture: MLP-Mixer (.

Never give up on your learning curves

What are the things you keep looking up and but never seem to memorize? How to. - Write a for loop in bash.- Generate a TLS certificate from LetsEncrypt.- Name the AWS credential env variables.- Start Jupyter with extra modules in sys.path.- Soft-reset a git repo.- . ?.