• AI KATANA
  • Posts
  • Transistor Mimics Brain Functions in AI Breakthrough

Transistor Mimics Brain Functions in AI Breakthrough

Also: OpenAI's New Models Can "Think with Images" and Understand Sketches

Author

AI KATANA
April 17, 2025

Hello!

Today’s edition captures a high-powered moment in the AI world, where the boundary between biology and silicon continues to blur. Researchers at the National University of Singapore have transformed standard transistors into brain-like components—laying the groundwork for a new class of energy-efficient neuromorphic hardware. Meanwhile, AI is outpacing economists in predicting central bank moves, and OpenAI is making headlines again—this time for both expanding its model capabilities to reason through sketches and whiteboards and inviting activist Dolores Huerta to advise on its controversial nonprofit-to-for-profit transition. Also in the spotlight: the U.S. mulls a crackdown on China’s DeepSeek, diffusion models are poised to reshape the music industry, and new tools from OpenAI and Microsoft signal a shift toward smarter, more visual agents. Catch up on these major developments that are redefining creativity, economics, and geopolitics in the AI era.

Sliced just for you:

  • 🧠 Transistor Mimics Brain Functions in AI Breakthrough

  • 📈 Economists show AI beats humans at guessing ECB rate moves

  • 🤝 OpenAI taps Dolores Huerta to steer its philanthropy during for-profit pivot

  • 💭 OpenAI’s New Models Can “Think with Images” and Understand Sketches

  • 🚫 Trump Administration Considers US DeepSeek Ban

  • 🎵 AI Diffusion Models Revolutionize Music Creation

Researchers at the National University of Singapore (NUS) have achieved a significant breakthrough in neuromorphic computing by demonstrating that a standard silicon transistor can mimic the functions of both biological neurons and synapses. By operating the transistor in a non-traditional way and adjusting the resistance of its bulk terminal, the team could control physical phenomena (punchthrough impact ionization and charge trapping) to replicate neural firing and synaptic weight changes (plasticity). They developed a two-transistor cell, dubbed "Neuro-Synaptic Random Access Memory" (NS-RAM), which exhibits low power consumption, stable performance, and consistent behavior. This innovation paves the way for developing scalable, energy-efficient AI hardware using existing, widely available silicon chip technology, potentially leading to more compact and responsive AI processors.

📈 Economists show AI beats humans at guessing ECB rate moves

Germany’s DIW Institute reports that a transformer model trained on speeches, inflation prints and bond spreads predicted European Central Bank decisions with 78 percent accuracy—nine points better than the median economist. Researchers say textual sentiment analysis around “persistent inflation” weighed heavily, suggesting policymakers telegraph more than they realise. The study adds to growing evidence that AI can crunch forward‑guidance subtleties faster than human analysts, potentially reshaping how banks, funds and even the ECB itself model monetary policy.

OpenAI has appointed legendary labor activist Dolores Huerta and three other philanthropic leaders to a 90‑day advisory board tasked with ensuring its charitable arm remains true to public‑interest goals as the company seeks regulatory approval to complete its controversial shift from nonprofit to for‑profit status. Huerta’s brief is to weigh benefits against risks of advanced AI and recommend guard‑rails, but critics say the gesture is “window dressing” while OpenAI chases a $40 billion raise at a prospective $300 billion valuation and fends off an Elon Musk lawsuit over governance.

OpenAI has released its latest AI models, o3 and o4-mini, which can "think with images" by understanding and analyzing user-uploaded sketches, whiteboards, and diagrams, even if they're low quality. This capability means the models don't just see images but can integrate visual information directly into their reasoning chains. The o3 model is specifically tuned for math, coding, science, and image understanding, while o4-mini operates faster and at a lower cost. Both models can independently use all ChatGPT tools including web browsing, Python, image understanding, and image generation, helping them solve complex, multi-step problems more effectively. Available to ChatGPT Plus, Pro, and Team customers, these models follow OpenAI's September release of o1, which focused on solving complex problems through multi-step deliberation. The company has faced recent criticism for changes to its safety precautions, including no longer requiring safety tests for certain fine-tuned models and avoiding releasing a "model card" for its GPT-4.1 model.

The Trump administration is considering new restrictions on Chinese AI lab DeepSeek that would limit it from buying Nvidia's AI chips and potentially bar Americans from accessing its AI services, according to The New York Times. These potential restrictions are part of the administration's broader effort to compete with China in the AI sector, following Tuesday's move to strengthen rules restricting Nvidia's AI chip sales to China. DeepSeek has gained significant popularity among US AI developers in recent months, with its competitive pricing forcing Silicon Valley companies to offer frontier AI models at lower costs. However, questions remain about whether DeepSeek engaged in intellectual property theft to create some of its more competitive models, with OpenAI alleging that the Chinese lab distilled its models in violation of OpenAI's terms of use. This development comes amid increasing tensions between the US and China over technology leadership, particularly in artificial intelligence.

Diffusion models, a type of AI that transforms random noise into coherent patterns, are now advancing into the music industry with unprecedented impact. These models can generate new songs across various genres—from orchestral performances to heavy metal—that are increasingly indistinguishable from human-created music. The article explores how AI-generated music is likely to blend seamlessly into streaming platforms, playlists, and soundtracks, raising profound questions about creativity, authorship, and originality. Major record labels are currently suing top AI music generators, alleging that these models replicate human art without compensating artists, while model makers argue their tools are designed to assist human creation. The article delves into the cognitive science of human creativity, contrasting it with how diffusion models work, and examines what it means to be emotionally moved by something without a human creator. This development represents a significant milestone in AI's evolution, as music—deeply embedded in our experiences, memories, and social lives—becomes the latest creative domain to face AI-driven disruption.

🛠️ AI tools updates

OpenAI debuted two reasoning models: flagship o3 and lightweight o4‑mini. Both can embed sketches or whiteboards directly into their chain‑of‑thought, zooming or rotating visuals while autonomously invoking ChatGPT tools such as Python or image generation. Plus, Plus/Pro/Team subscribers get them today; o3‑pro follows in weeks, while older o‑series models are retired. The launch lands days after GPT‑4.1, underscoring OpenAI’s rapid cadence.

Users running the latest Edge can now opt‑in to “Vision,” letting Copilot interpret on‑screen content, highlight key elements and walk you through recipes, résumés or even Minecraft without clicking. Mustafa Suleyman says no page data leaves the device, and broader OS‑level vision remains a Copilot Pro perk. Early testers report some lag and missing controls, but the feature hints at a browser‑centric future for multimodal agents.

💵 Venture Capital updates

Bloomberg‑sourced reports say OpenAI is negotiating to acquire Windsurf—the “agentic IDE” spun out of Codeium—in what would be its biggest purchase yet. Windsurf claims its autonomous coding companions already run on GPT‑4.1 and could slot directly into ChatGPT’s toolchain. The deal would deepen OpenAI’s enterprise moat against GitHub Copilot Enterprise and Replit, while giving Windsurf backers like Kleiner Perkins a rapid exit. Talks are fluid and could still collapse.

🫡 Meme of the day

⭐️ Generative AI image of the day