A New Open-Source Model: The Competition Heats Up!
Hello! In this blog, we’ll discuss the latest developments in the world of artificial intelligence (AI) and language models. It seems like just yesterday we were marveling at the release of GPT-4o or Claude 3.5 Sonnet, but now new contenders are emerging, ready to challenge the established players.
DeepSeek v3: What’s It All About?
Perhaps the biggest news of recent days is the open-source model from the Chinese company DeepSeek. According to benchmarks, it surpasses even prominent names like GPT-4o and Claude 3.5 Sonnet, particularly in math and coding tasks.
Why Is This Important?
Performance: DeepSeek v3 can generate 60 tokens per second, three times faster than its predecessor.
Open Source: All weights and information are publicly available, allowing researchers and developers to study the model in depth and adapt it to their needs.
Training Costs: Reports claim that training the model with 671 billion parameters cost only $5.5 million and 2.8 million GPU hours (compared to 31 million GPU hours for similar models).
While the results look very promising, further tests are planned to see how well DeepSeek v3 performs in real-world applications—because theoretical metrics are only part of the story.
OpenAI and Microsoft: Betting on AGI
Meanwhile, OpenAI continues to generate buzz. On one hand, the company is exploring new paths for development (including transitioning to a Public Benefit Corporation structure). On the other, it’s negotiating with Microsoft to define the criteria for achieving AGI (Artificial General Intelligence).
$100 Billion: A New Benchmark?
Some reports suggest that if a system can generate $100 billion in revenue, it might be considered AGI. The details remain unclear, but the logic is straightforward: if OpenAI reaches this goal, Microsoft could exit the project without losses, and OpenAI could renegotiate the terms of their collaboration.
Transitioning to a Public Benefit Corporation could help OpenAI attract investment under more transparent terms and allow greater freedom for nonprofit initiatives in areas like healthcare and education.
ChatGPT Search and “Hidden Pitfalls”
Not everything is smooth sailing, however. The Guardian reported that some language models (e.g., ChatGPT Search) can be tricked using hidden text on websites. This could lead the model to ignore negative reviews or even insert malicious code.
What’s the Issue?
The machine can “see” code that is invisible to the typical reader. This allows attackers to “re-prompt” system instructions, potentially manipulating the neural network’s responses.
The Solution?
Developers need to focus on security measures such as filtering hidden data, enhancing contextual checks, and more.
Altman Asks: “What Would You Fix?”
Every year, OpenAI CEO Sam Altman asks users: “What should we improve in our services?” Some of the most requested features this year include:
Voice Mode: Users complain that ChatGPT’s voice mode interrupts them if they pause to think, and they want the interaction to feel more natural.
Family Accounts: Many request profiles for children with parental controls and possibly more affordable pricing.
Memory and Context: Users want a larger input-output window so ChatGPT can “remember” more information during conversations.
Altman has promised to consider these ideas, so we can expect new updates in the future.
Aurora by Elon Musk and Google’s Efforts
“Grok” and Aurora
Elon Musk, through his Grok project, has unveiled the Aurora model, focusing on image generation. Aurora reportedly excels at both creating and editing images (though the editing function isn’t active yet). However, its performance in Russian is weaker, often translating queries into English. Improvements are planned, possibly following the example of DALL·E (editing specific parts of images).
Google and Gemini
Google is also staying competitive. The company is actively developing the Gemini family of models and has released an updated cookbook—a set of tools and guides for developers, available on GitHub.
Scaling Up: Sundar Pichai has set a 2025 goal to take Google’s consumer AI segment to new heights. Gemini is expected to become a breakthrough product for the company.
Competition: While the AI race was intense in 2024, 2025 promises to bring even fiercer competition.
What’s Next?
AGI: Who will achieve it first (or at least generate $100 billion)? OpenAI, Anthropic, Google, or someone else—it remains to be seen.
New Models: DeepSeek v3 shows great promise, but real-world applicability still needs to be tested.
Security: Issues with hidden code manipulation and filter bypasses highlight the need for increased focus on ethics and data protection.
We’ll continue to monitor developments and experiment with various models—whether it’s DeepSeek v3 or familiar names like GPT-4o, Claude 3.5 Sonnet, and others.
Happy New Year to all!
May 2025 (and the journey to it) be productive, successful, and full of new opportunities for you. And yes, we hope that interactions with “non-carbon-based life forms” (i.e., AI) remain entirely positive!
See you in future posts!
Top comments (0)