Mark Zuckerberg believes that 2025 will be the year of artificial intelligence, and in a recent Facebook post he writes that Meta intends to invest $60-65 billion in AI in 2025. This is a significant increase from 2024, when the budget was about $38-40 billion. This is not the only such large investment, and in the process a new rival is emerging.
A large part of the aforementioned investment will be used to build data centers. These are key to providing the computing power that the Meta needs to develop AI products. Zuckerberg said the initiative is not just about increasing infrastructure, but also about driving innovation and maintaining technology leadership in the US. As part of its strategy, Meta plans to acquire more than 1.3 million GPUs by the end of this year.
This is not the only such major investment this year. Recall that Microsoft, OpenAI, Oracle and SoftBank are beginning to co-create the Stargate Project. This is the largest-ever investment in AI, or more precisely, in AGI (strong artificial intelligence or superintelligence), which is expected to amount to $500 billion (over the next 4 years).
At the same time, a new Chinese rival to the US giants, DeepSeek, has emerged. The Chinese company has released the inexpensive DeepSeek-V3 and DeepSeek-R1 models, which beat several other leading OpenAI and Meta models in benchmarks. DeepSeek-V3 was trained using just 2048 GPUs with more than 2.78 million GPU hours at a cost of about $6 million. This is a fraction of what other leading models typically require. By comparison, the Llama models developed by Meta, including the Llama 3.1, cost more than $60 million and required 30.8 million GPU hours to train.
These models have been released under an open source license, like Llama, which means anyone can run them on their hardware. The price for using DeepSeek-R1’s inference/understanding API is also significantly lower than rivals such as OpenAI. DeepSeek charges $0.14 per million tokens for input, compared to $7.5 for OpenAI.
Why is DeepSeek so cheap? The reason is its architecture. The model uses the Mixture-of-Experts (MoE) framework, which allows it to activate only some of the parameters during processing. The manufacturer claims that this method guarantees higher performance and lower computational requirements compared to traditional models. In addition, unlike OpenAI o1, which uses supervised fine-tuning (SFT), DeepSeek uses pure reinforcement learning (RL), which allows it to autonomously develop advanced reasoning capabilities.
There is no doubt that strong artificial intelligence requires a considerable amount of money. However, ordinary, generative artificial intelligence can be much cheaper to develop, as DeepSeek just exemplified. Interestingly, its app has just dethroned ChatGPT in the Apple App Store, so Americans should be on the lookout!
OpenAI o1 available for free from Microsoft Copilot
OpenAI o1 is currently the best artificial intelligence model from this manufacturer available to the public. Among other things, it is superior to GPT in solving complex tasks. Until now available in premium services, it has now become available for free to everyone in Microsoft Copilot. What ChatGPT doesn’t have!
Microsoft has introduced the OpenAI o1 inference/understanding model to Copilot for all free users, which means you no longer need to subscribe to the paid Copilot Pro or ChatGPT Plus plans to use it. In other words, Microsoft is once again making the new AI functionality available for free first, long before OpenAI does in its ChatGPT. Think Deeper, as Microsoft calls this o1 integration, allows Copilot to answer more complex questions. Using this feature, the AI assistant says it “considers your question from all angles and perspectives,” which takes about 30 seconds.
Today we are making Think Deeper available for free to all Copilot users. This now gives everyone access to Copilita’s world-class OpenAI o1 inference model – everywhere and at no charge. I encourage you to try it out. It is truly magical. Think Deeper will help you:
- Get in-depth advice on how to manage a career change with a detailed breakdown of milestones and educational options, sources, where to look for positions, job entry strategies and industry trends you absolutely need to know about.
- Plan an epic project. Dump everything from your brain into Think Deeper and watch it process it all and spit out a step-by-step guide on how to do it. I’ve tried it out for a few things (fitness routine, upcoming big launch) and it’s really very helpful.
- Take a deep dive into any topic. Want to learn something about something? This is a shot in the arm! Not a substitute for human teaching, but an amazing extension. When I’m fascinated by something, lately I reach for Think Deeper. Most recently: the history of India, what happens to ocean currents.
The practical applications are simply endless. I’m sincerely excited that our tens of millions of users have this capability. We have so much more on the schedule that I can’t wait to tell you about it. But in the meantime, life is full of chaos and complex problems. Let Think Deeper take on some of that mental burden! – Mustafa Suleyman, CEO of Microsoft AI
OpenAI o3-mini available on Microsoft Azure and GitHub Copilot
OpenAI has released the o3-mini. This is the latest version of an affordable artificial intelligence model that matches the performance of the OpenAI o1 in math, coding and science (STEM) with the low cost and reduced latency of o1-mini. From day one, the model has been available in ChatGPT (paid subscriptions), via APIs, and in Microsoft Copilot Chat, GitHub Copilot and GitHub Models. Let’s check out the details.
OpenAI o3-mini is the company’s first inference model that supports developer-awaited features such as function calls, Structured Outputs and developer messages. Similar to o1-mini i o1-preview, o3-mini will support streaming. Developers can choose between 3 inference options – low, medium and high – to optimize the model for different use cases. This flexibility allows o3-mini to “think harder” on complex challenges, or to prioritize speed when latency is an issue. The publisher notes that o3-mini does not have image processing (computer vision) skills, so developers should rather use o1 for these tasks.
The new model is also available in Microsoft products. As we read in the blog AI + machine learning:
We are pleased to announce that OpenAI o3-mini is now available in the Microsoft Azure OpenAI Service. o3-mini adds significant cost savings over o1-mini along with enhanced inference, new features such as inference effort control and tools, while providing comparable or better responsiveness.
o3-mini’s advanced capabilities combined with its performance gains make it a powerful tool for developers and enterprises looking to optimize their AI applications.
With faster performance and lower latency, o3-mini is designed to handle complex inference workloads while maintaining efficiency.
– Yina Arenas, Vice President of Product, Core AI, Microsoft
GPT-4o better at understanding images and STEM tasks
OpenAI has announced improvements to GPT-4o, its AI model that powers ChatGPT, among others. It has gained more up-to-date knowledge, deeper understanding and ability to analyze images, better performance in STEM, as well as. increased use of emoji. Details below.
More up-to-date knowledge
ChatGPT-4o is no longer limited in terms of knowledge to events prior to November 2023. Now the boundary is June 2024, allowing the model to offer more relevant, timely and contextually correct answers, especially to questions about cultural and social trends. Fresher training data also makes it easier for the model to frame its answers from the Internet.
Deeper understanding and analysis of uploaded images
GPT-4o is now better at understanding and answering questions about images (entered as input), including better performance in multimodal MMMU and MathVista benchmarks. The model is better at interpreting spatial relationships in images, analyzing complex graphs or diagrams, and combining visual data with textual content. Responses to uploaded images will provide richer insights and more accurate guidance in areas such as spatial planning and layout design, as well as solving mathematical or technical problems based on visualization.
Greater STEM skills
ChatGPT now performs better on math, science and programming problems. The model has achieved better scores on academic benchmarks such as GPQA and MATH, as well as MMLU, a comprehensive benchmark that tests language skills, knowledge span and inference. This means it can now handle more complex problems of these domains.
Increased use of emoji
GPT-4o is now more enthusiastic about using emoji and will use them more often!
