- Pongo Points
- Posts
- Pongo Points 12/7/23
Pongo Points 12/7/23
Google broke ChatGPT | US goes nuclear | Bitcoin evaporates swimming pools | Swiss research on CBDCs | Google releases Gemini AI
1. Google AI Researchers Break ChatGPT
Read it on Github here: Extracting Training Data from ChatGPT
Why it’s interesting: Artificial intelligence models are trained on large data sets, but are generally believed to synthesize mostly novel answers to user prompts - a notion that is now in question after a recent report from some Google DeepMind employees’ testing on ChatGPT.
What stands out: The researchers were able to determine that some outputs were, in fact, training data and not unique answers by comparing them to 10 terabytes of data downloaded from the internet that were generated before ChatGPT’s creation.
What’s next: The researchers’ methods to extract data that exploited ChatGPT have since been fixed, but the vulnerability likely remains unaddressed and could be exploited again in different ways - leading to concerns that the underlying issue might have more complex repercussions in the future.
2. John Kerry Leads Global Pledge to Increase Nuclear Energy Capacity
Read it on the US Department of State website here: Declaration to Triple Nuclear Energy
Why it’s interesting: This is the most concrete declaration by the US in favor of the development and rehabilitation of nuclear energy production facilities.
What stands out: The pledge makes numerous references to the 1.5°C limit on temperature rise and admits that nuclear is the only feasible solution to achieving a sustainable energy future - a claim that many climate activists are resistant to acknowledging in favor of solar and wind power.
What’s next: This declaration is a positive step in de-stigmatizing nuclear energy after decades of public misunderstanding and media denigration.
3. How Much Water Does a Bitcoin Transaction Use?
Read it on Cell Press here: Bitcoin’s growing water footprint
Why it’s interesting: Alexander de Vries released a new perspective piece that proclaims a single Bitcoin transaction consumes enough water to fill a (very small) swimming pool, assuming Bitcoin mining facilities have the same water usage and electricity mix as data centers in similar geographical regions.
What stands out: Despite public information that provides more accurate data from which to draw conclusions (such as public crypto-mining company disclosures), de Vries continues his tradition of releasing Bitcoin research that is easily debunked for faulty data and assumptions.
What’s next: The ongoing battle between environmentalists and the cryptocurrency industry rages on, albeit with increasing transparency from crypto miners and fewer valid assumptions made by environmentalist researchers.
4. BIS Explores Private and Secure CBDC
Read it on Bank for International Settlements website here: Project Tourbillon demonstrates cash-like anonymity for retail CBDC
Why it’s interesting: With the help of contributors from IBM and ETH Zurich (no relationship to Ethereum), the BIS is generating research and pilot programs that aim to set the foundation for future CBDC projects - suggesting that central banks around the globe are seriously considering issuing digital money for retail use.
What stands out: Project Tourbillon succeeded in demonstrating the feasibility of a CBDC that has strong privacy and security, but was limited in its scalability - meaning the program is still relatively early and that widespread CBDCs aren’t an immediate possibility.
What’s next: The focus on providing cash-like anonymity is essential to ensuring a fair economy and avoiding a “1984” surveillance state, but there are still other critical considerations before any CBDC can be deployed - such as offline payments and improved transaction speeds.
5. Google Launches Its Most Powerful AI Yet Called Gemini
Read it on Google here: Introducing Gemini: our largest and most capable AI model
Why it’s interesting: Google’s Gemini project is the most capable AI model released to date, with the company showcasing its ability to observe video, image, and audio cues in order to interact with a users in an impressive hands-on video.
What stands out: Gemini is natively multimodal, meaning that it can seamlessly receive any combination of information (images, video, audio, etc.) to derive an output that incorporates that unique blend of inputs, rather than stitching together an answer that is derived from each the cues independently.
What’s next: Gemini is being released in three different “sizes” to accommodate hardware constraints, suggesting that Google aims to make its AI models more accessible for use in smaller devices like smartphones and tablets.