Data Science News Roundup

Spectral is creating an inference economy to power trustless, performant AI/ML models.  Each week we compile and recap other interesting developments at the intersection of machine learning and web3.

The Global Aftershocks of the OpenAI Board Reshuffle

Last week saw some of the biggest artificial intelligence news of the decade when OpenAI’s non-profit board (creators of ChatGPT) sloughed their beloved founder Sam Altman, who returned to the company two days later following a massive employee revolt. What precipitated the move is unknown but details emerged of a powerful new feature called Q* (Q-star) and a conflict between safety-minded effective altruists and technophiles eager to unleash AI’s full potential.

MIT Technology Review spoke to Wenda Li, a lecturer at the University of Edinburgh who echoed Meta AI mastermind Yann LeCun’s speculation that it would likely have something to do with improving an LLM’s ability to plan, which would be particularly useful for the step-by-step logic required for solving complicated math problems. Or it could be evidence of an emerging super intelligence.

Also making the rounds was OpenAI's "ex-interim CEO" Emmett Shear's analysis of an "SEO heist" that calls out the end of the free open internet:

OpenAI’s rollercoaster ride of corporate intrigue came on the heels of a discussion between China and the United States to restrict the use of autonomous weapons (e.g. killer robots), a contentious bid for artificial intelligence restrictions in the U.S. Congress, and a non-binding agreement between 18 countries on AI security guidelines.

How Blockchains are Introducing Money to Artificial Intelligence

The Best Money for AI is Crypto according to Cointelegraph’s Andrew Fenton, who is looking at real-life use cases for AI in web3 this week with Casper Labs, MakerDAO’s Rune Christensen, and other notables. Fenton highlights FewSats, Lightning Labs, and Fetch.ai as having already created interesting tools for connecting GPTs and AI.

Also mentioned in the article above was Syndicate.io’s Ian Dao Lee’s post about creating a GPT agent that can bank itself. Caveat Emptor if you try this on your own! 

Spectral has also made the news—we launched the first challenge for data scientists and ML engineers to come together and build the models that will generate powerful, accurate inferences for our inference economy platform. We are offering a total of $100,000 in bounties and an 85 percent share of all revenue produced to modelers who can hit or exceed inference performance benchmarks. Our first challenge is to create a web3 credit score using data sets we developed from our own MACRO Score.

Our team really enjoyed NEAR co-founder Illia Polosukhin’s speech about how AI and web3 can work together to maximize the best of both technologies, creating new business models, and more resilient governance systems—including a thought experiment about an AI president!

We also really enjoyed seeing [OpenAI research scientist] James Betker’s June 2023 letter reminding us that The “It” in AI Models is the Dataset circulating on LinkedIn.

Meanwhile, multichain Cosmos (ATOM) founder Jae Kwan is calling for a fork after a proposal to reduce inflation on the Cosmos chain to 10 percent passed. The new chain would be called AtomOne. Emily Parker’s OpEd calling CZ’s resignation and Binance’s billion-dollar fine the end of the borderless cryptocurrency company sparked some interesting company-wide dialogue.

Zero Knowledge News 

Exchange Gate.io has implemented zero-knowledge technology (zkSNARKS) into its existing proof of reserve technology, according to Yahoo News

A team of Chinese researchers have achieved zero-knowledge proof (ZKP) based on a device-independent quantum random number beacon: The breakthrough involves integrating quantum technology into ZKP, by using quantum random number generators. These generators are based on the principles of quantum mechanics, where particles behave in unpredictable ways. This unpredictability makes the generated numbers extremely secure, as they cannot be easily predicted or replicated.  The system's capability to broadcast these secure, random numbers in real-time is particularly innovative, implying a robust dynamic form of security that can be widely applied, from secure communications to safeguarding sensitive data.

AI in the Public Domain and Black Box Algorithm Lawsuits

Wading into the data science subreddit, there was a fantastic conversation about whether artificial intelligence should be seen as inherently public domain knowledge and not be commercialized. Quasi-consensus from the subreddit’s scientists: parameter weights should be secret, and datasets should not.

A STAT investigation delves into an algorithm that didn’t accurately factor in COVID-19 increases in hospitalization and suggested cutting off care too early. A lawsuit has been filed against UnitedHealthGroup, which used its subsidiary NaviHealth’s NH Predict algorithm to make care-related decisions for Medicare recipients.

In the realm of data science competitions, there are a couple of big challenges closing this week. From MLContests.com, there’s $100,000 at stake for predicting how cells respond to small molecule drug perturbations on Kaggle (Open Problems in Single Cell Analysis), Lab42 is offering $74,000 to develop an AI capable of tackling 1,000 tasks (including 100 unknown ones), and Zindi is offering $9,000 for Nowcasting South Africa’s inflation.

Why Spectral Builds for Models that Earn

It’s always interesting to think about how much value these challenges could create versus the bounties they pay—a big reason why we chose to make challenges on Spectral perpetual: always open, always evolving, and always generating value for our modelers through revenue-sharing. Interested in participating in our first challenge?