Matt's Homepage

I’ve written about Google’s custom silicon TPUs before (Google’s Tensor Processing Units – v1).

One of the big reasons for Google and others web services to develop their own custom chips is that general purpose CPUs are flexible but typically need a lot of power. That power costs a lot of money in electricity bills and cooling costs in huge data centers. So, why buy chips with lots of stuff you don’t need when you can build your own – and save millions of dollars a year in a data center with lower cooling and power costs?

In just a 6 years, Google has managed to design and build 4 ever increasingly capable AI data center chips. They had somewhat humble beginnings – but they are becoming increasingly powerful. Now they have just published information about TPU version 4.

What is this new chip capable of?

• a nearly 10x leap forward in scaling ML system performance over TPU v3 
• boosting energy efficiency ~2-3x compared to contemporary ML DSAs, and 
• reducing CO2e as much as ~20x over these DSAs in typical on-premise data centers

Even crazier, it’s the first system to use purely optical switching.

TPU v4 is the first supercomputer to deploy a reconfigurable OCS (optical circuit switching). OCSes dynamically reconfigure their interconnect topology and are much cheaper, lower power, and faster than Infiniband.  The figure below shows how an OCS works, using two MEMs arrays. No optical to electrical to optical conversion or power-hungry network packet switches are required, saving power.

Add to this, the newest version claims to be 1.2-1.7x faster and 1.9x more efficient than nVidia A100 chips.

Worth a read.

Links:

• TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings: https://arxiv.org/ftp/arxiv/papers/2304/2304.01433.pdf
• https://cloud.google.com/blog/topics/systems/tpu-v4-enables-performance-energy-and-co2e-efficiency-gains

The question of copyright, lawsuits, and AI is going to very quickly come to a head.

Creatives from artists to comedians are filing lawsuits, staging online ‘protests’, and suing various AI-based companies for copyright infringement. In 2022, ArtStation members staged a online campaign against AI generated artwork by posting ‘No AI art’ images in their portfolios.

China entered the fray by recently announced their interim measure to govern AI generated text, pictures, audio, video, and other content [Update: Wow – already redacted, check here or here]. It covers generating AI content in PRC, but may be unclear about what foreign companies can import into China.

But it doesn’t stop there. Now we can add game developers to the fray.

Recently Steam devs were seeing their games with AI generated content blocked from Steam. Valve responded that it was not able to “ship games for which the developer does not have all the necessary rights” or for “utilizing AI tech.”

In a statement to IGN, Valve spokesperson Kaci Aitchison Boyle clarified the position. While developers can use these AI technologies in their work with appropriate commercial licenses, they can not infringe on existing copyrights.

Aitchison Boyle emphasized that Valve is not attempting to discourage the use of AI but the confusion arose due to Valve’s ongoing efforts to incorporate AI technology into its existing review process while ensuring compliance with copyright laws.

Have you heard of the Dead Internet Theory? In 2021, IlluminatiPirate wrote the theory in which they claimed that already, or very soon, the majority of the internet will be just bots and autogenerated content. They claimed this was done by a few illuminati type mustache twirlers bent on controlling opinions, grooming political/society’s opinions, as well as generating customers for particular products.

While the latter part of the theory is pretty implausible and tinfoil hat, ColdFusion does some interesting investigation into the claims by looking at the actual data – and shows how quickly we are actually approaching many of these Black Mirror like claims. Many of which I have written about before.

He investigates current bot usage compared to actual users. While many people rolled their eyes when Elon Musk made these claims about Twitter, we’re increasingly seeing social media companies failing during fiduciary scrutiny when they go to sell themselves. It turns out many have massive amounts of fake users. A major unicorn app was outed for having 95% bots just this week. Facebook took down 5.4 billion fake accounts in 2019 alone – more than twice the number of REAL accounts.

He follows an experiment in which a experimenter uses off the shelf tools to create a fake influencer who posts AI generated social media posts, AI generated pictures of a completely AI generated photorealistic person, and starts picking up follows – many of which were bots themselves. Caren Marjorie created an GPT4 AI version of herself that would be your girlfriend.

The Atlantic did a research project on tweets that all contained repeated text by countless accounts while similar profile pictures with huge engagement levels above their normal levels of those account types. By 2025-2030, 90-95% of the content on the internet may be generated by bots if we continue at the current rates.

Jubilee put 6 humans and 1 AI into a chat room and they had to pick out the bot just by the answers to questions given to them all to answer. It took a lot of rounds before they all guess the AI correctly.

Maybe you’d like to try the famous Turing Test yourself and see if you can spot the bot? Google has a bot that has successfully passed the Turning test, and ChatGPT was the second.

Want to see how fast AI is progressing? Did you know chatGPT 4 is able to get 90th percentile on the Bar exam, solve complex logic problems, build complex apps and games, write books, or make money by founding and running a company for you? He doesn’t even capture all the things AI has been doing.

This is worth a watch if you want to see a smattering of what AI is already doing in 2023

We started with the AI based show about nothing, then AI Spongebob. Now we have a live streaming AI Jesus. The video, audio, and what he says is all generated by AI. What’s surprising is that it accepts a lot of different questions – and often answers them with a higher degree of accuracy than I would have thought (though I would certainly NOT take any of your religious formation from the AI version).

I think it’s more revealing the kinds of questions people ask. While some are clearly humor others are quite serious and reveal the depth of things people are struggling with.

I guess it’s only slightly better than when some Lutherans let chatGPT run an entire service with a sermon.