Photo by panumas nikhomkhai

Introductory

You may have heard that the energy requirements of GenAI solutions are very high.

Why is this?

If we ask chatGPT a question, then it searches the Internet: it compiles the relevant answer to your question from the information there and from the huge database in its Large Language Model (LLM). In order to answer your question not in minutes, but almost immediately, we need high computing power, which has significant energy consumption.

How much energy do Google searches and GenAI requests consume?

The energy requirement of a Google search is about 0.3 Wh, with which we could operate a 60 W light bulb for 17 seconds. Of course, the number of daily searches is very high, but we can say that Google’s energy consumption is still modest compared to AI.

In contrast, a single question to chatGPT or another chatbot consumes roughly 2.9 Wh. (source) This is equivalent to about 10 Google searches.

ChatGPT’s annual energy consumption is roughly 227 GWh.

This amount of energy would be sufficient:

• to charge 95% of US electric cars
• cover the energy supply of more than 26,000 American households

The energy requirement for training BERT was estimated to be that of a trans-American flight in 2019 (source). The training of GPT-3 cost approximately 1.287 MWh of energy (source).

The challenge of AI in the future

Based on the forecasts, the energy demand of AI data centers will increase by 44.7% until 2027. This is the compound annual growth rate (CAGR)), so it can reach 146.2 TWh (source).

Another challenging area is the water demand of AI solutions. This is also a very significant factor in terms of environmental effects. I will not cover this topic in this article.

So it doesn’t hurt to keep these facts in mind and start using AI more consciously.

How can we do this? This is what the next chapter is about.

Possible solutions

Now you understand how serving the ever-increasing GenAI needs affects energy consumption.

What can we do to reduce AI’s insatiable appetite?

Global level

The solution at the corporate level could be if AI companies can turn to renewable energy sources (e.g. solar and wind energy). Other possible solution is to optimize their hardware, algorithms and business model.

By the way, big companies like Google and OpenAI are increasingly trying to use more energy-efficient data centers and renewable energy sources in order to reduce the environmental load.

There are also attempts to use nuclear energy, which I personally do not consider an ideal solution. There is some risk of a nuclear power plant disaster (Chernobyl, Fukushima) and nuclear waste is also generated during the process.

Wind and solar energy can therefore be a better solution. Although the production of solar panels is not an environmentally friendly solution, it can still be a better alternative than the use of fossil fuels in the long run. (source)

In addition, companies offering AI solutions can limit access to services, for example, by making them paid or by providing a better service to subscribers.

State-level regulations can also help keep energy consumption under control.

What can we do as a community and as individuals?

You might think, “One person doesn’t matter, I can’t do anything.”

In fact, it is precisely the case that if we form good habits together, we can greatly support a good cause (similar to how we started to use GenAI more and more intensively).

It is a fact that even large companies are increasingly pushing for the use of GenAI, as they would lose their competitiveness if they did not do this.

On this blog, I would like to regularly give tips on how to optimally use AI solutions. In some cases, how to replace AI with classic non-AI tools.

My plan is that there will be regular noAI challenges. For a few days we will not use GenAI tools, but only “traditional” digital solutions to draw or edit the images.

I know that this is not feasible in all cases, since AI is now present as an integrated service in many devices and software (smartphones, office software, operating systems, etc). Nevertheless, I hope that this article was of concern to you and that you too can find an area where you can moderate the use of AI.

What I, as a blog owner, can do for reduce energy consumption of AI:

• I write articles that help you make optimal use of AI, such as:
  • efficient prompt writing, so you can achieve the desired result with fewer attempts
• writing most of the articles without using AI (currently I write 60-70% of articles without AI)
• illustrating articles with photos or self-made graphics instead of AI generated images
• generating as few AI videos as possible
• organization of noAI and wiseAI challenges
• instead of weekly AI Art Gallery, a maximum of 1-2 galleries per month
• providing image editing tips

Summary

Generative AI solutions are developing rapidly and are now available to everyone. This explosive development has resulted in a massive increase in the energy demand of AI solutions on a global scale.

In order to manage this change, we need conscious and reasonable solutions. The huge appetite of AI systems cannot be sustained in the long term.

We may have to wait a little while longer, but maybe we will be able to calm down the AI ​​hype a little and reopen the traditional digital and analog solutions in the field of arts as well. Remember that many masterpieces in the fields of graphics, music, film production and text writing have already been created without AI.

P.s.: I wrote this article about AI without AI.

The post Energy consumption of AI and possible solutions appeared first on AI Art Blog.

Who Can Try the o1 Model?

At the moment, the o1 model is available to ChatGPT Plus and Team users, with Enterprise and Educational users getting access soon after. If you’re subscribed to ChatGPT Plus, you can already experiment with two versions of the model: o1-preview and o1-mini. The first one is the flagship model focused on solving particularly challenging tasks, while o1-mini is a more affordable and faster alternative, especially useful for coding-related tasks.

There are some limitations, though, such as message caps:

• 30 per week for o1-preview and
• 50 per week for o1-mini

but these are likely to change as OpenAI refines the model further.

When Will It Be Available for Everyone?

For now, the public can’t try o1 unless they’re paying for the Plus or Enterprise tiers. However, OpenAI has mentioned plans to eventually roll out access to ChatGPT Free users, especially the o1-mini version, which is both faster and cheaper. This might happen in the coming months, but there’s no fixed date yet. It’s something worth keeping an eye on it, if you would like to try it, but don’t want to pay for subscription just yet.

Videos Worth Watching

On OpenAI’s official o1 page, there are some videos that demonstrate the power and flexibility of this new model. Here are three that caught my attention:

HTML Snake Game
In this video, o1 helps create a classic snake game in HTML. It shows how the model can handle multi-step coding tasks efficiently. This is particularly interesting because it demonstrates o1’s ability to generate and debug code for developers quickly. Even if you’re not into game development, it’s worth watching to see how the model works through the logic step by step.

Solving a Complex Logic Puzzle
This video showcases o1’s reasoning abilities by tackling a complex logic puzzle. What’s fascinating here is how o1 constructs a detailed “chain of thought” before delivering its final answer. It gives you a peek into how the model breaks down intricate problems, something that GPT-4o wasn’t as strong in.

Quantum Physics Challenge
Ever wondered how an AI would handle quantum physics? In this video, o1 goes through the paces of generating complex mathematical formulas needed for quantum optics. This is where o1 really shines—its ability to deal with the sort of difficult, niche problems that you wouldn’t normally throw at an AI. Watching it process these advanced concepts makes it clear why OpenAI is touting this as a breakthrough model for science and research.

I’m curious about how this new model handles basic mathematical tasks, like counting letters in a word. The previous models were not too good at calculations, but if I ask GPT-4o to write a program for this task, it generates a well-functioning script.
The o1 seems to be able to do simple math operations that the GPT-4o couldn’t.

Why Try Out OpenAI o1? Is It Better Than GPT-4o?

So, what makes o1 worth trying? For one, the model is designed to spend a little more time thinking before it answers. This extra computational effort makes a huge difference in fields like math, physics, and even coding, where the ability to reason and understand the problem before generating an answer is crucial.
For example, o1 outperformed GPT-4o in solving math problems at the International Mathematics Olympiad, solving 83% of the problems compared to GPT-4o’s 13% success rate.

Moreover, o1 is highly specialized for tasks that require multi-step reasoning, like generating scientific hypotheses or debugging complex code. While GPT-4o is still useful for broader, more generalized tasks, o1’s ability to “think” deeply makes it an excellent choice for more technical challenges.

Despite these advantages, there are limitations:

• As an early model, o1 lacks some features that make GPT-4o practical for everyday use, such as browsing the web or handling files and images. So, depending on your needs, GPT-4o might still be the better choice for now if you’re aiming to use these capabilities.
• The operational costs of using the o1 model are significantly higher than those of GPT-4o, with input costs three times greater and output costs four times higher. Additionally, processing speed can be slower for complex queries.

Summary

While I haven’t personally tested OpenAI o1 yet, the early demonstrations and use cases suggest it’s a game-changer for anyone working in fields that require complex problem-solving. Whether you’re a researcher tackling quantum mechanics or a developer debugging code, o1 seems to have the potential to be much more than just a language model—it’s a tool for reasoning.

Right now, only those on ChatGPT Plus or higher tiers can test it out, but with plans to make it accessible to Free users, it’s something I’d recommend keeping on your radar.

The post OpenAI o1: Better Reasoning in complex tasks appeared first on AI Art Blog.

In this article, we’ll take a look at what OpenAI’s latest GPT-4o model can do. If you’re curious about its capabilities, be sure to read on.

On Monday, OpenAI announced its newest, fantastic model named GPT-4o. They practically wrote an article and there are plenty of videos on their YouTube channel in which they demonstrate its capabilities.

We will go through them in broad outline, so that by the end of the article you will have a comprehensive picture of what makes this new large language model (LLM) so groundbreaking.

It is noteworthy that GPT-4o will also be available for free.

News about GPT4-o

Its most notable ability is that you can practically have a conversation with it. The reaction time for answering has been reduced to a minimum with this AI model, and you can also interrupt his speech.

The previously available Voice service of ChatGPT needed 4-5 seconds for the answer, now that has completely changed.

Watch this example video to see what’s new.

The GPT4-o model

In the description, you can read that this model is actually an omni model. This means, that it can not only interpret your voice (Voice mode) and text, but can practically handle both.

The interpretation of images and videos is made possible by the Vision mode.

The Memory mode remembers our previous conversations, which ChatGPT can use later. With the Analyze mode, complex data can be analyzed quickly and efficiently. Examples: Excel spreadsheets or even a chart that you show as an image to chatGPT.

Practical examples

If you look at the introduction page of GPT-4o, you can see many examples of people using their phone’s camera with the ChatGPT app and asking ChatGPT what it sees in the camera, and it then answers them “verbally”.

So the same thing can now be done in real time with your phone’s camera. The big difference between the current GPT-4o and the previous Voice service is that a completely new model was actually trained here with voice, text, videos and images.

It is also worth watching the video where they show that one person speaks Italian, the other person speaks English. The phone interprets it in real time. This means that as soon as the Italian text is spoken, ChatGPT immediately translates it into English and vice versa.

You’ve may seen conversations, where politicians spoke in their own mother tongue and still understood each other. It was thanks to the dubbing service in the background.

In the future, this will probably be completely replaceable with such an AI solution.

Performance of GPT-4o

There is a comparison chart on the hello-gpt-4o page in the Model evaluations section, with which we compare the GPT-4o model through various benchmarks, for example with GPT4 turbo, or even with the paid Claude 3 Opus and Gemini Pro 1.5.

The speed difference according to the graph will become visible in practice. All models are already very fast and are therefore capable of responding in real time.

You can feel the difference between a free ChatGPT, i.e. GPT 3.5 and GPT 4, but the difference between the above big models is not really noticeable anymore.

Security and limitations

The good news is that they also put a lot of emphasis on security. It’s highly likely that you won’t be able to say obscene words with it and you won’t be able to ask ChatGPT for other silly things either.

This model is subjected to serious, in-depth testing, which means that a team tries to hack these systems and extract answers from them. The name of the testing method is the so-called red teaming.

The future

These services are constantly being improved. Maybe by the time you read this article, the entire interface will look completely different.

Thanks to the Memory mode, ChatGPT will also remember previous conversations. You can turn this off if you don’t want it to store information about you.

The Analyze mode is good for uploading documents and requesting all kinds of analytics from it. So there will be improvements here in the free GPT as well.

Summary

This new model contains huge innovations. It almost feels like talking to a live person. It can also be used to facilitate scientific tasks, such as the analysis of complex data, graphs, and images.

If this is also combined with image and video generation, or perhaps with the generation of sound and music, the limit to what we can create with it is the imagination.

Of course, he won’t cook lunch for you, but he can help you in creative and scientific tasks.

Thank you for reading this article!

Image by Franz Bachinger, Pixabay

The post OpenAIs GPT-4o model is free to use! appeared first on AI Art Blog.