AI development is progressing along two distinct paths. One path involves companies raising significant capital at extremely high valuations. The other involves more focused entities taking a different approach. Both strategies have their advantages in different situations. The funding landscape continues to evolve rapidly. SSI quadrupled its valuation to $20B by February 2025 and later secured $2B at a $32B valuation in April 2025, led by Greenoaks with Alphabet and Nvidia participation. Meanwhile, Mira Murati aims for a $2B seed round at a $10B valuation. Companies such as OpenAI and Grok are pursuing scale-driven strategies, while others, like Mistral and Hugging Face, are exploring a different approach. They are focusing on open-source models, developer communities, and specific niches instead of chasing the AGI moonshot, which requires billions in capital. Meta occupies an interesting middle ground with its Llama family of models. Despite being a tech giant with vast resources, Meta has embraced an open-source approach by releasing increasingly powerful models, including Llama 3 and, most recently, Llama 4. Llama 4 features innovative multimodal capabilities and a mixture-of-experts architecture. Mistral is targeting underexplored areas like local language models such as Arabic, and value-added niches like OCR and small models that can run without extensive infrastructure. Similarly, Hugging Face has built the "GitHub for machine learning" - a collaborative platform hosting over 500,000 AI models, most of which are freely available and open-source.

I've been using Gemini 2.5 Pro for just a short while now, but it easily feels like months. That's the funny thing about habits: the truly transformative ones integrate so quickly, you forget what life was like before. That's how it feels having left the landscape of previous LLMs after years inside what now feels like a more constrained environment. It wasn't a difficult switch, technically, but adjusting my workflow took a moment. Since then, it's consolidated into one massive, ongoing chat containing thousands of lines of code, ideas, and reflections – my primary interface for creative and technical problem-solving. It taps into that fundamental human capacity to adapt rapidly to powerful new capabilities. I find myself relying on its unique ability to tackle problems that stumped other models, its knack for creativity, and its skill in connecting disparate concepts. The raw numbers back up my subjective experience too – Gemini 2.5 Pro leads benchmarks like GPQA (science questions) and AIME (math problems) without needing specialized test-time techniques. It's not just marketing – there's a qualitative leap in how it approaches complex tasks. The benchmarks are impressive enough to mention specifically. Gemini 2.5 Pro has claimed the top spot on LMArena by a significant margin, outperforming both Claude 3.7 and GPT models. On technical benchmarks, it's reached 63.8% accuracy on SWE bench for coding tasks, and an impressive 18.8% on Humanity's Last Exam – a challenging benchmark created by hundreds of subject matter experts to test the limits of AI reasoning. Maybe it takes more conscious effort to switch tools as you gain experience, resisting the comfort of the familiar. But the reward here is rediscovering the sheer joy of leveraging a tool that expands your own potential – the speed, the intelligence, the massive 1M token context window allowing for unprecedented depth. And that context window is no gimmick – I've watched it analyze entire codebases (100,000+ lines) in a single session, maintaining coherence throughout. Where other models fragment understanding across multiple chats, Gemini can hold the entire conversation history plus documentation, making problem-solving feel genuinely cumulative rather than disjointed.

I've been using Devin from Cognition for one month now, but it easily feels like a whole quarter of development work. That's the funny thing about habits: they can compress or expand time depending on how much they change your workflow. That's how it feels having left traditional engineering practices after years inside the closed garden of manual coding. It was initially challenging to trust an AI with my codebase, but since then, I've found a balanced approach where Devin handles the predictable tasks while I focus on architecture. Humans naturally resist change until the benefits become undeniable. I've seen this play out in my own skepticism that gradually transformed into practical appreciation. The older you get in the tech industry, the more you've seen innovations come and go, making adaptation sometimes feel exhausting. But working with Devin has rekindled a curiosity about what's possible when you automate the predictable parts of engineering. I believe in Devin's capabilities for certain tasks while recognizing it's not a silver bullet. There are legitimate reasons why complex, cross-component work still requires human engineers with holistic understanding and creative problem-solving abilities.

Last weekend, my girlfriend and I were immersed in a Roblox gaming session when she presented an intriguing challenge: "Wouldn't it be cool if we could create something that follows us around in the game?" This question sparked my curiosity and presented the perfect opportunity to dive into Lua programming and Roblox scripting. I'm also currently watching Mythic Quest new season which is a sitcom about a group of people who work at a gaming company. Also you may be familiar that vibecoding games with Three.js is very popular these days on X. In this tutorial, I'll walk you through how we created a companion cube that loyally follows your character throughout any Roblox world and serves as a digital guardian. The History and Purpose of Lua Before we dive into the code, let's understand what makes Lua special. Lua was created in 1993 at PUC-Rio (Pontifical Catholic University of Rio de Janeiro) in Brazil by Roberto Ierusalimschy, Luiz Henrique de Figueiredo, and Waldemar Celes. It was developed as a solution for a partnership between the university and Petrobras, the Brazilian oil company, which needed portable and customizable tools. A Scripting Language by Design Lua was never intended to replace system programming languages like C or C++. Instead, it was designed as a scripting language to extend applications. Think of it this way: the core of your system (the "kernel") is written in a compiled language for speed and efficiency, while Lua provides scripting capabilities on top of that system.

Special thanks to Zac, Josh, Cooper and Bruno from the Aztec team for their valuable feedback and reviews. What is Noir? Noir is a domain-specific language (DSL) for zero-knowledge (ZK) programming, released in October 2022. It is a generalized ZK programming language that exposes a high-level programming language with syntax similar to Rust. The language compiles down to ZK circuits but is deliberately platform agnostic. Noir is not specific to Aztec or PLONK, and it is an open-source project that aims to be used by as many people as possible. Aztec Labs and the community wanted to create a generalized ZK programming language that anyone could use, regardless of the platform they were using. The goal is to make Noir an open-source public good, where anyone can use it and build their backends and front ends, even if it doesn't touch any other Aztec system. A Layer 2 (L2) that recreates the smart contract ecosystem that Ethereum has, but as a L2 where the private state is a first-class default primitive. This means that things become private by default, and as a developer, you can easily include private data in your program logic. However, this requires a lot of disparate components to work, such as an exceptionally fast zk proving system. This architecture enables all this and a language to program these contracts. Noir now has an official reference repository called Noir-Starter, which is a full-stack project using Next.js and Hardhat and serves as the perfect starting point for anyone who wants to dive into Noir development.

web3 with a16z crypto https://twitter.com/a16zcrypto/status/1690522827391422464 Speakers Sonal Chokshi, Ali Yahya, Dan Boneh Overview Sonal Chokshi discusses the intersection of AI and crypto with his guests, Dan Boneh and Ali Yahya. They explore the potential benefits and challenges of combining these two technologies, including the decentralization of AI, the use of zero-knowledge proofs in machine learning, and the verification of training models. They discuss how these technologies can mutually benefit each other and also touch on the tension between centralization and decentralization in AI and crypto.

Special thanks to Andrew Miller for his valuable feedback and reviews. :::info :bulb: Check out the awesome-sgx-blockchain repo on GitHub. ::: :::info 🪧 Check out the announcement tweet. ::: Introduction