Good morning, it’s Friday. China’s DeepSeek has created one of the most powerful “open” AI models. With 685 billion parameters and a $5.5 million price tag, DeepSeek V3 is outdoing rivals like GPT-4o and Llama 3.
Meanwhile, ChatGPT and Sora experience second major outage this month, AI may have shaped elections, and big tech turns to geothermal for soaring energy demands.
Also, be sure to catch the latest from the Forward Future 50: 2024’s top AI stories.
🤔 FRIDAY FACTS
Can AI be tricked by optical illusions? If so, what does that reveal about how machines "see" the world?
Stick around for the answer! 👇️
🗞️ YOUR DAILY ROLLUP
Top Stories of the Day
🗳️ AI in Elections: Quiet Influence, Big Impact
AI subtly shaped the 2024 elections, not through dramatic deepfakes but via behind-the-scenes tools like AI-driven speechwriting, translations, and strategy planning. Smaller campaigns leveraged AI for accessibility and outreach, especially in countries like India and Indonesia. While its influence was innovative, concerns about weak detection systems and misinformation persist. This election likely marks just the start of AI’s growing role in politics.
👨💻 Microsoft Copilot Hits 365—Costs and Complaints Rise
Microsoft has added Copilot to Microsoft 365 subscriptions in Australia and Southeast Asia—along with mandatory price hikes for all users. The move, aimed at monetizing Microsoft’s AI investments, has drawn backlash over intrusive pop-ups and unwanted features, reminiscent of the infamous Clippy assistant. Built with OpenAI technology, Copilot is central to CEO Satya Nadella’s strategy to drive AI adoption, but critics argue it prioritizes profits over user choice.
🎓 UTSA Launching AI and Cyber College in 2025
The University of Texas at San Antonio (UTSA) is set to open a College of AI, Cyber, and Computing in fall 2025, enrolling over 5,000 students. Focused on AI, cybersecurity, and data science, the new college will prepare graduates for booming tech careers and enhance UTSA’s research reputation. It will integrate specialized departments and leverage research centers like the AI Consortium for Human Well-Being to address societal challenges and drive innovation.
💸 AGI? Microsoft, OpenAI Tie It to $100B Profit
For Microsoft and OpenAI, AGI success may hinge on $100 billion in profits, not technical breakthroughs. OpenAI is expected to stay unprofitable until 2029, delaying any AGI declaration. The deal also secures Microsoft’s access to OpenAI’s tech until AGI is achieved, easing concerns of premature claims. Meanwhile, OpenAI’s o3 model impresses but faces cost challenges, making the profit-driven AGI goal a distant target.
🌋 Big Tech Taps Geothermal to Power AI Growth
With AI’s energy use doubling every 100 days, Big Tech is turning to geothermal energy to curb emissions and meet sustainability goals. Google’s 50% emissions spike since 2019 threatens its 2030 net-zero pledge, pushing companies like Meta and Alphabet to invest in geothermal power. Despite its clean energy potential, high costs and competition from natural gas pose challenges, leaving geothermal’s role in AI’s future unclear.
🦾 DEEPSEEK DOMINATES
DeepSeek V3 Emerges as a Powerful Open-Source AI Challenger
The Recap: DeepSeek, a Chinese AI firm, has unveiled DeepSeek V3, a 685-billion-parameter model that rivals leading AI systems in tasks like coding and translation. Released under a permissive license, it allows developers to freely download, modify, and utilize it for various applications, including commercial use.
DeepSeek V3 surpasses both open-source and closed AI models in internal benchmarks, excelling in coding competitions on platforms like Codeforces.
The model outperforms competitors such as Meta’s Llama 3.1 405B, OpenAI’s GPT-4o, and Alibaba’s Qwen 2.5 72B in specific coding tasks.
Trained on a dataset of 14.8 trillion tokens, DeepSeek V3 demonstrates significant advancements in handling diverse text-based workloads.
Despite its large size, DeepSeek V3 was trained in approximately two months using a data center equipped with NVIDIA H800 GPUs, costing around $5.5 million—considerably less than models like OpenAI’s GPT-4.
The model exhibits content filtering aligned with Chinese regulatory standards, avoiding responses to politically sensitive topics.
DeepSeek's open-source approach has prompted competitors like ByteDance, Baidu, and Alibaba to reduce usage prices for their models, fostering a more competitive AI landscape.
Forward Future Takeaways:
DeepSeek V3's release marks a significant milestone in the AI industry, showcasing that high-performance models can be developed efficiently and cost-effectively. Its open-source nature is poised to accelerate innovation and competition, potentially leading to more accessible and advanced AI applications across various sectors. However, the model's compliance with Chinese regulatory standards highlights the ongoing challenges of balancing technological advancement with governmental policies. → Read the full article here.
👾 FORWARD FUTURE ORIGINAL
Interview with Steven Smith, VP of Engineering, Protocol, at Tools for Humanity
In an age where artificial intelligence is reshaping the world, Tools for Humanity is at the forefront of ensuring technology remains human-centric. In a recent interview, Steven Smith, Head of Protocol at Tools for Humanity, shared insights on the evolving challenges and opportunities in safeguarding human identity in the digital age. Through innovative protocols and a mission-driven approach, the company, which is a core contributor to World Network, seeks to address some of the most pressing issues facing society today.
The Significance of Human Identity
Human identity is foundational to digital interactions, yet it’s increasingly difficult to determine in the AI era. Steven began by painting a stark picture of the challenges involved. “It’s about much more than just preventing identity theft,” he explained. “It’s about protecting the integrity of what it means to be human when AI systems can mimic our voices, images, and even our decisions.”
Steven emphasized the urgency: “AI will reach the levels—and arguably is already there—of superhuman influence already before it ever reaches superhuman intelligence.” This influence isn’t just theoretical—it has real-world consequences in shaping opinions, spreading misinformation, and disrupting trust.
Steven’s passion for this field is deeply personal, tracing back to his career in decentralized technologies like blockchain. “The realization that billions lack basic access to identity and financial systems was a pivotal moment for me,” he reflected. “I saw the potential for technology to bridge these gaps but also the dangers if it isn’t developed ethically.” → Continue reading here.
🚧 AI SLOWDOWN
AI's Data Dilemma: Is the Tech Industry Approaching a Slowdown?
The Recap: Tech companies like OpenAI and Google are facing a shortage of internet data to train AI systems, potentially slowing recent rapid advancements. Experts are exploring new methods to sustain progress in artificial intelligence.
Google’s Demis Hassabis warns AI may hit a wall as internet data runs out, pushing firms to explore synthetic data.
Ilya Sutskever, former chief scientist at OpenAI, states, "We’ve achieved peak data, and there’ll be no more. We have to deal with the data that we have. There’s only one internet."
Researchers are developing AI systems that learn from their own trial and error, generating "synthetic data" to overcome data shortages.
OpenAI's recent system, OpenAI o1, utilizes self-generated data, though this approach is currently effective mainly in fields with clear right and wrong answers, like math and programming.
Despite potential slowdowns, companies continue significant investments in AI development, with Databricks nearing a $10 billion funding round, the largest-ever for a start-up.
Forward Future Takeaways:
As the tech industry confronts the limits of available internet data for AI training, the focus is shifting toward innovative techniques like synthetic data generation. This transition is crucial to maintain the momentum of AI advancements and could redefine how AI systems learn and evolve, ensuring sustained progress despite data constraints. → Read the full article here.
🏆 2024 HIGHLIGHTS
Forward Future 50: The Top AI Stories of 2024
#34 Neuralink’s Brain-Computer Interface: Neuralink achieved a significant milestone by implanting its brain-computer interface in a human. Initial results are promising, aiming to enable thought control of devices.
#31 NVIDIA’s Climate Platform: NVIDIA’s Earth-2, using AI and Omniverse technology, predicted weather patterns precisely, enhancing climate change understanding and disaster preparedness globally.
#30 Google Introduces Med-Gemini: Google's Med-Gemini family of AI models, built on Gemini 1.0 and 1.5, surpasses GPT-4 in medical benchmarks, enhancing clinical reasoning and multimodal understanding.
#29 Genmo's Mochi-1 AI Video Model: Genmo released Mochi-1, a free, open-source AI video generation model under the Apache 2.0 license, enabling users to create high-quality videos from text prompts.
#27 Google Partners with Kairos Power: Google entered into a pioneering agreement with Kairos Power to deploy advanced small modular reactors, aiming to supply up to 500 MW of carbon-free electricity to its data centers by 2035.
#26 OpenAI's Realtime API: OpenAI’s Realtime API advanced AI communication with near-instantaneous speech-to-speech interactions, enabling seamless, low-latency voice applications.
📻️ Tune in Monday for the next batch of top stories from 2024.
🔬 RESEARCH PAPERS
Wonderland: Turning Single Images into Expansive 3D Worlds with Diffusion Models
A team of researchers introduces "Wonderland," a breakthrough method for generating high-quality, wide-scope 3D scenes from a single image. Unlike prior approaches, which rely on multi-view data and lengthy optimizations, Wonderland employs a video diffusion model to create compressed latents that preserve multi-view consistency.
These latents drive a 3D Gaussian Splatting reconstruction pipeline, enabling fast and visually rich 3D outputs. Extensive tests show Wonderland excels at handling out-of-domain images, marking a significant leap in single-view 3D generation by leveraging diffusion-based latent spaces. → Read the full paper here.
🤔 FRIDAY FACTS
AI Falls for Optical Illusions—Just Like Us!
Surprisingly, yes—AI can be deceived by optical illusions! Studies show that neural networks, especially those used in computer vision, interpret patterns much like the human brain. For example, the café wall illusion (where straight lines appear slanted due to alternating patterns of black and white squares) tricks AI systems into perceiving the lines as crooked, just as it does for us.
This uncanny similarity suggests AI doesn't simply "see" images—it processes visual data in ways that mimic human perception. While this raises intriguing questions about AI’s cognitive parallels to human brains, it also highlights the limits of artificial vision, reminding us that even advanced systems can be fooled by a clever arrangement of pixels.
So the next time you’re staring at an illusion and questioning reality, rest assured—you’re not alone. The AI might be just as confused!
Reply