That much learning and motion. Learning is a higher order, like a higher-level process. We have adaptive practice, and you can see we have some very sophisticated models, patient training methodologies, and causal analysis. It’s a very popular model used in various applications, particularly for efficiency. And then this example discusses the model back in 1958, where people started to shift towards a model of interconnected systems, the precursor to what Microsoft has been developing.

Oh, can we make some adjustments? It’s not as clear as it was earlier. I’m trying different settings…yes, I know. Oh, that’s…that’s much better. Yes, all right. So, starting from 2013, we have seen rapid growth in artificial intelligence (AI) models. For example, BERT was introduced in 2016, and then Google’s Transformer model architecture became very popular.

The first version of GPT was actually based on this Transformer architecture. Then, starting from 2020, we saw the emergence of what is now referred to as LLMs, or large language models. In 2023, GPT-3 was released, and it marked a significant evolution in AI capability. Currently, we’re using GPT-4, which is already considered outdated as the newer versions are being developed, focusing more on automated data processing and optimization.

Later versions, like GPT-4.5, are considered upgrades in efficiency. We’re now using GPT-5.0, which is being developed by OpenAI’s collaborative research initiatives.

Yes, I just wanted to ask, is the Transformer more like a neural network? Yes, it is, but it has evolved to handle increasingly complex data structures. For instance, the model now uses a hierarchical approach, which allows it to generalize to various classes.

In the initial GPT model, for example, we saw about 1.7 billion parameters. Now, with GPT-3, we’re at 175 billion parameters, a thousand-fold increase, and with the latest developments, it’s only going to grow. The challenge is managing the energy consumption needed to train such large models. GPT-3, for instance, consumed about 1,287 megawatt-hours of electricity, equating to approximately 5.2 tons of CO2 emissions, which has significant environmental implications.

Microsoft and Amazon are also developing their own foundation models to support efficient data processing. In 2014, data centers in the U.S. alone consumed around 70 billion kilowatt-hours of

electricity, accounting for roughly 1.8% of the nation’s energy use. This number is expected to double, which raises concerns about sustainability.

Another type of model, the Geo-Foundation model, is geared toward tasks related to spatial data, such as geographic and satellite imagery analysis. Since the release of GPT models in 2023, we’ve seen Geo-Foundation models trained specifically for tasks like counting objects in images or estimating distances. These are much more specialized but equally capable models.

To give you an example, back in the 1960s, during the Cold War, the U.S. used aerial photography to monitor Soviet missile sites, which required manual image interpretation. Today, we have thousands of satellites providing real-time data, but we still face challenges in processing this vast amount of information effectively. Geo-Foundation models help automate and improve the accuracy of spatial analysis tasks.

In terms of energy efficiency, a lot of research is focusing on model compression and optimization. Methods like distillation and pruning are becoming popular to reduce model size and energy consumption. Looking forward, I expect that AI models will increasingly use energy-efficient hardware and renewable energy sources to mitigate their environmental impact.

Thank you very much.

All right, any questions?

Yes, is it just satellite images, or can it work with aerial footage as well? Most Geo-Foundation models are designed to work with satellite imagery, but they can be adapted for other sources.

Yes, do you foresee environmental regulations impacting this development? Given that models are only getting larger, I anticipate some future regulations around the energy consumption of large-scale models.

I like your question about linear growth. The scale of these models doesn’t increase linearly—it’s exponential. While hardware improvements are relatively linear, the models themselves are growing at a much faster rate, which could lead to a significant imbalance.

Thank you, great question.

Okay, moving on. Our next speaker is Professor Vijay Manasetti, a cybersecurity and privacy expert from Georgia Tech, who will discuss blockchain and AI.

Thank you for the introduction. How many of you understand the basics of blockchain? I’ll give a quick overview for those who might not be as familiar. Blockchain, as part of Web 3.0, integrates IoT, blockchain, and advanced analytics. Web 3.0, unlike today’s Web 2.0, enables large-scale decentralized data processing.

Initially, blockchain was simply a distributed database. However, as it evolved, it became more than that, supporting complex applications beyond cryptocurrency. With blockchain, we can exchange value directly without intermediaries like banks or governments, creating a trusted, immutable database.

Tokenization is a key feature of blockchain, allowing for frictionless exchanges of value. Unlike traditional models, where transactions are managed by banks, blockchain offers a decentralized approach.

Blockchain transactions are organized into chains of blocks, with each block containing a “hash” of transactions from a given time period. This makes it almost impossible to alter a block without changing all subsequent blocks, providing security and trust in the system.

Most cryptocurrencies rely on proof-of-work (PoW) consensus, which is computationally intensive and consumes significant energy. Due to this, new consensus mechanisms like proof-of-stake (PoS) and proof-of-authority (PoA) are gaining traction. These alternatives are less energy-demanding and are already being implemented by second- and third-generation blockchain platforms.

Proof-of-stake, for instance, relies on stakeholders to validate transactions, reducing the need for energy-intensive mining operations. Bitcoin, however, still uses PoW, and many data centers in Georgia and North Carolina support Bitcoin mining, which raises energy consumption concerns.

Georgia Power and other utilities have prioritized data center growth, but the environmental impact of such decisions is significant. While options like nuclear and geothermal energy are being considered, the infrastructure shift is slow and expensive. For example, nuclear plants take years to construct and are extremely costly.

In Georgia, crypto mines and data centers are often given energy discounts, but the general public ultimately bears the cost. This is why we need to consider both energy efficiency and regulatory frameworks moving forward.

So, what can we do about the energy-intensive nature of crypto mining? We could potentially incentivize more sustainable practices, such as charging higher energy rates for PoW-based operations compared to PoS. But there needs to be pressure on companies to adopt more efficient models.

There’s also been misinformation about electric vehicle (EV) grids using coal, but grids across the country are diversifying their energy sources and getting cleaner every day. It’s important to be mindful of energy sources and their environmental impact.

Finally, let’s discuss how AI and blockchain technologies, like large language models and agents, might continue evolving. Agents powered by LLMs can handle increasingly complex tasks, coordinating information and decisions on behalf of users, potentially transforming entire industries. These developments, however, raise ethical concerns as automation could displace jobs traditionally done by humans.

Thank you for listening, and I’d be happy to answer any additional questions.