The AI Vision Board for 2025

DeepSeek Launches V3: A Milestone in Open-Source AI

Hey there, 

Wishing you all a very Happy New Year! To recap, let’s take a trip back to the beginning of 2024:

OpenAI was the big player at the start of the year. They had their flagship model, GPT-4 turbo, and they even announced Sora in February. It seemed like OpenAI was unstoppable - they had the team, and the money, and everyone was backing them. But then, things started to go downhill!

By the end of the year, GPT-4 was facing tough competition from Claude, Gemini, and Llama. Sora felt outdated compared to Veo 2. Qwen was the best model for coding. GitHub Copilot is now free for everyone. Elon Musk’s xAI raised a ton of money and is growing faster than anyone else.

Don’t get me wrong - OpenAI still seems to be in the lead. o1 is better than anything you’d see for reasoning, and they’ve already announced o3. But the lead doesn’t feel as big as it did at the start of the year. This is even though OpenAI had 12 days of events in mid-to-late December!

Let’s quickly take a look at the current landscape and see what the leaders of these companies have been saying about 2025!

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection

Table of Contents

Technology's most influential voices are painting a fascinating picture of AI's evolution. Here's how these visionaries see the next chapter of our technological future unfolding:

Sam Altman's Predictions

  • Advancements in AGI: Altman believes that Artificial General Intelligence (AGI) is closer than previously anticipated, with predictions that superintelligence could accelerate AI breakthroughs by tenfold each year, fundamentally reshaping society and the economy (source)

  • User-Centric Features: In a recent tweet, Altman highlighted user requests for OpenAI in 2025, which include enhancements such as better memory, longer context handling, and the introduction of a "grown-up mode" for more mature interactions. (source)

  • Integration of AI Agents: He anticipates a future where AI agents will be more proactive, capable of anticipating user needs and managing tasks without explicit instructions. (source)

Elon Musk's Views

  • Concerns About AGI: Musk has expressed caution regarding the rapid development of AGI. He predicts that while AGI may not be fully realized until 2026, its implications will require careful consideration due to potential existential risks associated with advanced AI systems. (source)

  • Impact on Workforce: He shares concerns similar to Altman about AI's impact on jobs and the economy, emphasizing the need for ethical considerations in AI deployment. (source)

    Jensen Huang's Vision

    • Year of AI Agents: Huang foresees 2025 as the "year of AI Agents," where digital workers capable of understanding tasks and executing actions will emerge. These agents could engage in customer service, execute marketing campaigns, and optimize supply chains. (source)

    • Collaboration Among Agents: Huang emphasizes that these agents will work collaboratively to mechanize workflows and improve enterprise productivity, potentially handling up to 50% of tasks for employees. (source)

Satya Nadella's Insights

  • AI as a Core Component: Microsoft’s CEO Nadella reflects on 2023 as a pivotal year for AI and expects 2025 to further solidify AI's role in enhancing productivity across various sectors. He notes that organizations focusing on integrating AI will see significant performance improvements compared to those that hesitate. (source)

  • Continued Evolution of AI Tools: Nadella predicts ongoing advancements in AI tools that will redefine digital experiences and enhance user interaction through better integration with existing technologies like 5G and IoT. (source)

Marc Benioff's Perspective

  • AI-Driven Business Transformation: Salesforce CEO Benioff predicts that AI will fundamentally transform business operations, enhancing customer engagement through personalized experiences.

  • Focus on Ethical AI: He stresses the importance of ethical frameworks guiding AI development to ensure it serves humanity positively.

DeepSeek Launches V3: A Milestone in Open-Source AI

DeepSeek has unveiled its V3 model, a groundbreaking step forward in open-source AI. This latest release features an innovative Mixture of Experts (MoE) architecture, delivering three times the speed of its predecessor—processing 60 tokens per second—all while maintaining full API compatibility.

Key Technical Advancements

  • 671 billion total parameters, with 37 billion activated for efficiency.

  • Trained on 14.8 trillion high-quality tokens.

  • MoE architecture ensures optimal balance between power and resource efficiency.

This release demonstrates the narrowing performance gap between open and proprietary AI systems, while maintaining transparency and accessibility. DeepSeek's roadmap includes multimodal capabilities and ecosystem expansion, indicating continued investment in open-source AI development.  (source)

What's the Story?

  • The Achievement: OpenAI's o3 system has reached an 85% score on the ARC-AGI benchmark, matching average human performance and setting a new AI record. The system demonstrates remarkable "sample efficiency" in pattern recognition with minimal examples.

  • The Context: The achievement marks a step forward in AI's adaptability to novel situations, particularly in complex pattern recognition tasks.

Key Considerations

  • Testing Methodology: Pre-training conditions significantly differed from human testing scenarios, raising questions about direct performance comparisons. Notably, top human performers continue to outperform the model.

  • Limited Scrutiny: External scientific validation remains restricted to select institutions, prompting calls for broader examination and peer review.

  • Expert Clarification: François Chollet (test creator) and OpenAI's leadership have explicitly stated this achievement does not constitute AGI, tempering public excitement.

Why It Matters?

  • Communication Impact: The announcement's framing has sparked industry-wide discussions about responsible communication in AI achievements.

  • Evaluation Standards: This milestone highlights the need for standardized frameworks in assessing and communicating AI progress. (source)

OpenAI is evaluating a plan to transform its for-profit arm into a Delaware Public Benefit Corporation (PBC) to balance advancing AGI and securing the capital needed for its mission. Here’s what this means:

Why The Change?

  • Rising Costs: Building AGI requires billions in funding, far beyond the reach of traditional donations.

  • Mission Alignment: A PBC structure ensures decisions prioritize public benefit alongside shareholder returns.

What’s Changing?

  • If implemented, the PBC would raise conventional equity while keeping OpenAI’s mission front and center.

  • The non-profit retains significant ownership, funding initiatives in healthcare, education, and science.

This proposed restructuring could ensure access to substantial funding through conventional equity, empowering the organization to accelerate AGI research while supporting charitable initiatives in healthcare, education, and science. (source)

Microsoft and OpenAI are reportedly evaluating a specific, internal definition of artificial general intelligence (AGI) — one tied not to technical milestones but to financial performance, according to a new report from The Information. If implemented, this definition would mark AGI as achieved only when OpenAI’s AI systems generate at least $100 billion in profits.

What’s the Story?

  • A Profit-Based Definition: OpenAI’s definition of AGI, as outlined in the reported agreement, diverges from traditional technical or philosophical benchmarks, focusing instead on financial success.

  • A Long Road Ahead: OpenAI projects significant losses this year and estimates it won’t turn a profit until 2029, indicating AGI under this definition remains many years away.

Why It Matters?

  • Microsoft’s Access: Microsoft’s agreement with OpenAI includes a clause that revokes its access to OpenAI’s technology once AGI is reached. This profit-based definition could secure Microsoft’s access to OpenAI models for a decade or more.

  • Compute Costs: Recent debates around OpenAI’s o3 model—despite its improved performance—highlight the high compute costs involved, complicating the path to profitability under this framework.

This evaluation reveals a shift in how AGI could be defined, blending financial benchmarks with technological progress. The implications for OpenAI’s timeline and Microsoft’s partnership remain significant.  (source)

Elon Musk’s xAI is reportedly exploring how to leverage its recent $6 billion Series C funding round to scale its AI ambitions. Here’s what’s happening:

The Headlines:

  • Major Investment: xAI’s valuation has doubled to $45 billion, bringing total funding to $12 billion.

  • Strategic Backers: The round includes tech giants like Nvidia and AMD, traditional finance leaders like BlackRock, and venture capitalists from Andreessen Horowitz and Sequoia.

What’s Next?

  • xAI plans to use the funds to train its most powerful AI model yet, expand its product lineup, and boost R&D infrastructure.

  • Sovereign wealth funds like Kingdom Holdings are taking significant stakes, signaling a global interest in AI development. (source)

Inside Google’s Use of Claude AI to Improve Gemini

Google is reportedly using Anthropic's Claude AI to evaluate and enhance its own Gemini AI. Here’s a closer look at the situation:

What Happened?

  • Contractors assessing Gemini’s performance noticed references to Claude in the evaluation process, raising questions about Google’s methods and whether Anthropic approved this use.

  • Google has denied using Claude to train Gemini, stating that comparisons are standard in industry evaluations.

Key Concerns

  • Terms of Use: Anthropic explicitly prohibits the unauthorized use of Claude to develop competing AI systems.

  • Safety and Reliability: Contractors noted that Claude prioritizes safety, often declining unsafe prompts. In contrast, Gemini was criticized for safety violations, including generating inappropriate content. This raises concerns about Gemini's reliability in sensitive applications like healthcare. (source)

  • Tool: ReadPartner

    ReadPartner is a tool designed to effortlessly summarize content from text or YouTube videos, making it an indispensable companion for quick insights and simplified learning.

    How to Access

    • Sign up for ReadPartner.

    • Paste a YouTube video link or text into the provided box.

    • Click summarize, and voilà! You're good to go.

I tried using the tool to summarize a YouTube video. It was quite helpful - you can request different formats like bullet points, short executive summaries, or detailed breakdowns, and adjust the length to what works best for you.

  • For those interested in Generative AI, the course ‘Reimagining GenAI: Common Mistakes and Best Practices for Success’ by Shahebaz Mohammad provides an in-depth exploration of the challenges in adopting AI technologies. It covers why high-fidelity demos often fail in real-world applications and offers practical strategies to ensure scalable, ethical, and ROI-focused implementation of AI solutions.

What are you looking forward to in 2025? What is your prediction for the year? If there was one bet you would take, what would it be?

Looking forward to hearing from all of you.

Regards,

Kunal

Reply

or to participate.