Good morning, it’s Wednesday! Tetsuwan Scientific is building AI-powered lab bots that can run experiments and fix mistakes—no lab coats required.

In other news, Google benchmarks its Gemini model against Anthropic’s Claude, OpenAI might venture into humanoid robotics, and Chinese AI companies innovate around U.S. chip restrictions.

Plus, we’re starting our countdown of 2024’s top 50 AI stories. Let’s dive in!

🆚 Google Benchmarks Gemini AI Using Anthropic’s Claude
Google contractors are benchmarking Gemini AI against Anthropic’s Claude, sparking concerns over permissions and competitive practices. While Google denies training Gemini on Claude’s data, it admits using rival AI for comparisons. Contractors noted Claude’s stricter safety measures, contrasting with Gemini’s occasional safety lapses. Both companies declined to clarify whether permissions were granted, adding to ongoing concerns about Gemini’s reliability.

🤖 OpenAI Explores Humanoid Robots
OpenAI is exploring humanoid robot development as part of its renewed robotics focus, following investments in startups like Figure and Physical Intelligence. These robots aim to navigate human environments more effectively with human-like designs. However, the initiative is still in early discussions, as OpenAI prioritizes advancing reasoning models and AI agents for software automation, keeping robotics as a secondary focus for now.

🎯 OpenAI’s o3 Sets New AI Benchmarks—At a Price
OpenAI’s o3 model sets new performance standard, scoring 88% on the ARC-AGI test but at a staggering $10,000 per task. Using “test-time scaling,” it boosts accuracy by applying more compute during inference. While its breakthroughs suggest human-like adaptability, high costs and occasional errors limit practicality for everyday use. Instead, o3 targets high-stakes applications in finance, academia, and industry.

📈 China’s AI Advances Despite U.S. Chip Limits
Chinese AI startups like DeepSeek and Moonshot AI are closing the gap with U.S. models, leveraging techniques like reinforcement learning and “mixture of experts” to optimize performance with fewer resources. DeepSeek’s models rival OpenAI’s o1 in math tasks, and Tencent’s AI matches Meta’s Llama 3.1 despite limited access to high-end Nvidia chips. However, U.S. hardware advancements and pricing challenges raise questions about China’s long-term AI competitiveness.

Artificial intelligence is rapidly reshaping industries, reimagining urban systems, and redefining human interactions. David Ly, CEO of Iveda, offers a unique perspective on how AI can address critical challenges while driving innovation and efficiency. In our recent interview, Ly outlined the trends shaping smart cities, ethical AI adoption, and the transformative potential of technology in 2025 and beyond.

Stop-and-Go AI: A Cautious Path Forward

As AI adoption accelerates, David Ly predicts a "stop-and-go" approach where cities strategically pause to address ethical and operational challenges before advancing. "Municipalities want to ensure the technology integrates smoothly into existing frameworks," Ly explains. By carefully managing these rollouts, cities can mitigate risks, improve transparency, and build trust with their communities.

This approach reflects broader societal apprehensions about AI's implications. Ly highlights the role of AI in improving efficiency, but he also acknowledges the infrastructure gaps that still need to be addressed. "We’re getting data a lot sooner and quicker," Ly says, "but there’s still the challenge of figuring out how to respond to and manage all this." For cities, this means embracing innovation while taking the time to implement solutions thoughtfully and sustainably. → Continue reading here.

#50 Meta’s MTIA Chip Boosts AI Speed by 3x: Meta unveiled its custom AI chip, the MTIA, designed to enhance ranking and recommendation models. It delivered 3x faster performance and supported advanced AI workloads.

#49 NASA’s Rover Hunts for Life on Mars: NASA’s Perseverance rover used AI to autonomously analyze Martian rocks and identifying minerals. This innovation accelerated data collection for potential signs of ancient life.

#48 AI Doubles Accuracy in Rare Genetic Diagnoses: Researchers developed AI-MARRVEL, an AI system that doubled diagnostic accuracy for rare genetic disorders, reducing diagnosis times and improving identification of past undiagnosed cases.

#47 AI Decodes Whale ‘Language’ for Conservation: Scientists used AI to uncover whale communication patterns, including a "phonetic alphabet," advancing conservation efforts and supporting calls to grant whales legal personhood.

#46 AI and Humans Team Up for Better Stock Predictions: AI outperformed most analysts in stock predictions, but combining AI’s data processing with human insights reduced errors and improved accuracy.

#45 AI Supercharges Chip Production by 60x: NVIDIA’s cuLitho, integrated by TSMC and Synopsys, accelerated semiconductor manufacturing by 40–60x using AI and GPUs, enabling faster, cost-efficient production for advanced chips.

📻 Tune in tomorrow for the next batch of top stories from 2024.

The Recap: Tetsuwan Scientific, co-founded by Cristian Ponce and Théo Schäfer, is building AI-powered robots that can run scientific experiments autonomously, combining machine learning with physical automation. Inspired by recent AI advances, their goal is to create “robotic scientists” that can evaluate results and modify experiments without human intervention.

Highlights:

Forward Future Takeaways:
Tetsuwan Scientific’s approach could fundamentally change research by accelerating experiments and minimizing human error. As AI gains physical agency, it raises both opportunities and ethical questions about the automation of discovery. With competitors like FutureHouse also exploring this frontier, the field of autonomous science is progressing quickly. → Read the full article here.

SepLLM is a novel framework designed to accelerate LLMs by compressing information between special separator tokens, reducing redundancy without sacrificing accuracy. Researchers found that separators disproportionately influence attention scores, enabling them to store segment details more efficiently.

SepLLM achieves over 50% KV cache reduction on benchmarks like GSM8K-CoT while preserving performance, even handling sequences exceeding 4 million tokens in streaming tasks. Compatible with pre-trained and custom models, SepLLM combines inference speedups with training optimizations, offering a scalable solution to LLM computational bottlenecks. → Read the full paper here.

Matt dives into OpenAI's unveiling of o3, which has left the AI industry reeling with its groundbreaking performance in complex math and coding challenges. Boasting 25% accuracy in Frontier Math and unmatched adaptability, o3 marks a significant leap in AI capabilities—fueling debates on its economic viability and AGI potential. Get the full scoop! 👇

Looka | Logo & Branding: Looka helps businesses design logos and marketing materials quickly with AI, no design expertise needed.

Steer | Writing Assistant: Steer improves grammar, tone, and clarity across apps, enabling fast, professional, and multilingual communication.

Bolt | Web Development Platform: Bolt enables full-stack web app creation and deployment in-browser, streamlining development with AI and WebContainers.

Reply to this email if you have specific feedback to share. We’d love to hear from you.

Follow us on X for quick daily updates and bite-sized content.
Subscribe to our YouTube channel for in-depth technical analysis.

Prefer using an RSS feed? Add Forward Future to your feed here.

Thanks for reading today’s newsletter. See you next time!

The Forward Future Team
🧑‍🚀 🧑‍🚀 🧑‍🚀 🧑‍🚀