The AI Rundown by Lightscape Partners - 2/10/25

Good morning and welcome back to another edition of The AI Rundown by Lightscape Partners.

• Apple is preparing to push the boundaries of AI hardware with its upcoming M5 chips, featuring advanced stacking technology for better heat dissipation and efficiency. With a stronger focus on AI performance, including a next-gen Neural Engine, these chips signal Apple’s commitment to AI-driven computing as competition heats up in the semiconductor space.

• Meanwhile, Google is expanding its Gemini 2.0 lineup, rolling out new models optimized for coding, reasoning, and cost efficiency. The introduction of a budget-friendly Flash-Lite model and a 2-million-token context window in Gemini 2.0 Pro reflects Google’s push to capture more of the enterprise AI market while improving accessibility for developers.

• On the cybersecurity front, French startup Riot secured $30M to combat AI-driven cyber threats, using real-time phishing detection and employee training to bolster digital defenses. As AI-powered cyberattacks become more sophisticated, the race is on to develop security solutions that can match the speed and scale of emerging threats.

Stay tuned as we explore these stories and their implications for the future of AI, technology, and innovation.

If you haven’t yet, please support the newsletter by subscribing!

Hardware

Apple’s M5 chips introduce advanced stacking for AI and efficiency gains. Link.

• Apple's upcoming M5 chips, built on TSMC’s 3nm N3P process, promise a 5% performance boost and up to 10% better power efficiency.

• A stronger focus on AI is expected, with a significantly upgraded Neural Engine surpassing the M4’s 38 TOPS.

• Some M5 chips will use SoIC-mH stacking, a new bonding method improving heat dissipation and performance.

• The first M5 chips are expected in iPad Pros later this year, while the Ultra variant likely won’t arrive until 2026.

Machine learning is revolutionizing hardware simulation speeds. Link.

• Traditional engineering simulations rely on outdated numerical methods, making design iterations slow and computationally expensive.

• AI-based physics simulations can speed up processes by orders of magnitude, reducing simulation times from days to seconds.

• Startups and platforms like NVIDIA Modulus are emerging to challenge legacy players like Ansys and Siemens.

• Adoption remains a challenge, as engineers must trust AI-driven results and overcome barriers in training and evaluation.

Models

Google expands Gemini 2.0 lineup with new models for developers. Link.

• Google DeepMind has made Gemini 2.0 Flash generally available via the Gemini API, offering improved speed and reasoning.

• The experimental Gemini 2.0 Pro, optimized for coding and complex prompts, features a 2 million token context window and tool integration.

• A new budget-friendly model, Gemini 2.0 Flash-Lite, delivers better quality than its predecessor at the same cost and speed.

• Google is enhancing AI safety with reinforcement learning and automated security testing to mitigate prompt injection risks.

Consumer + Enterprise Applications

AI at the edge: Making generative AI affordable and offline-ready. Link.

• Running generative AI on edge devices, rather than the cloud, reduces latency, improves privacy, and lowers infrastructure costs.

• Leaner AI models, with fewer parameters, are enabling localized processing on laptops, smartphones, and even cars.

• Dedicated neural processing units (NPUs) like Hailo’s AI accelerators offer cost-effective, power-efficient alternatives to GPUs.

• Edge AI could eliminate reliance on premium hardware and cloud subscriptions, making generative AI universally accessible.

OpenAI's o3-mini launches in Azure with improved reasoning and efficiency. Link.

• o3-mini enhances AI reasoning with a new ‘reasoning effort’ parameter, allowing users to control response complexity and latency.

• The model introduces structured outputs via JSON Schema and supports functions and external tools for seamless automation.

• It improves cost efficiency over o1-mini while maintaining strong coding, math, and scientific reasoning capabilities.

• Now available in Azure OpenAI Service, o3-mini is optimized for enterprise AI applications with enhanced precision and scalability.

Funding & Valuation

Riot raises $30M to fight AI-driven cyber threats with employee-focused security. Link.

• French cybersecurity startup Riot secured $30M in Series B funding, led by Left Lane Capital, bringing its valuation to $170M.

• The company plans to expand internationally, double its workforce, and enhance AI-driven phishing simulations and employee training.

• Riot’s AI-powered chatbot, Albert, detects and mitigates phishing attempts in real time, strengthening cyber resilience.

• Protecting 1M employees across 1,500 companies, Riot aims to scale to 10M employees by 2027 as AI-driven attacks grow more sophisticated.

SuperOps raises $25M to expand AI-powered IT service automation. Link.

• The Indian startup secured a $25M Series C at a $200M valuation to enhance AI-driven tools for managed service providers (MSPs) and IT teams.

• Its AI assistant, Monica, will soon predict IT issues and recommend solutions, aiming to automate troubleshooting.

• SuperOps offers lower-cost alternatives to competitors like NinjaOne, charging $1.5 per endpoint versus $4.

• With 1,300 customers across 104 countries, the company plans to expand to the U.S., Latin America, and Europe, including a London office.

Safety + Ethics

Google drops AI ethics pledge, opens door to military use. Link.

• Alphabet removed its commitment not to develop AI for weapons or tools that could cause harm, citing the need to support national security.

• AI chief Demis Hassabis framed the shift as aligning AI development with democratic values like freedom and human rights.

• The move comes as AI governance debates intensify, with experts warning about the dangers of autonomous weapon systems.

• Alphabet’s announcement coincided with an 8% stock drop following weaker-than-expected earnings and slowing cloud growth.

Data Centers + Energy

DeepSeek’s AI model forces Japan to rethink energy forecasts. Link.

• Japan recently adjusted its energy strategy to account for rising AI-driven electricity demand, but DeepSeek’s efficient AI model is challenging those projections.

• DeepSeek’s AI assistant, which uses less power and data, quickly overtook ChatGPT in App Store downloads, triggering a global market response.

• Some analysts argue that cheaper AI models could accelerate adoption, ultimately increasing power consumption rather than reducing it.

• Japan’s policymakers, wary of past miscalculations, are monitoring developments closely as debates over nuclear versus renewable energy intensify.

OpenAI

OpenAI’s ‘Deep Research’ takes on DeepSeek’s AI breakthrough. Link.

• Days after DeepSeek’s AI assistant surged in popularity, OpenAI launched ‘Deep Research,’ a new feature for AI-driven analysis and synthesis of complex data.

• The tool aims to bridge reasoning gaps and adapt to technical domains but struggles with real-world contextual challenges.

• Critics question whether AI synthesis can truly replace scientific creativity and challenge existing knowledge structures.

• While AI accelerates research, it has yet to demonstrate the ability to generate genuinely novel methodologies or breakthroughs.

OpenAI files trademarks for AI hardware, robotics, and chips. Link.

• OpenAI’s new trademark application covers AI-powered consumer hardware, including smartwatches, AR/VR headsets, and smart jewelry.

• The filing also mentions robotics and AI chips, aligning with OpenAI’s recent hiring for a new robotics team.

• OpenAI debuted a research tool powered by its upcoming o3 model, designed for multistep online research and data synthesis.

• The company is reportedly in talks for a $40 billion funding round, partly to support its Stargate AI infrastructure joint venture with SoftBank.

Thank you for reading the AI Rundown by Lightscape Partners. Please send any questions, comments, or suggestions to [email protected].