New EU AI unicorn Lovable💛, ChatGPT agent 🤖, Windsurf Wave 11 🧑💻, Reflections on OpenAI 🤔, Huma AI personality cloning voice assistant🎤 and more
AI Connections #60 - a weekly newsletter about interesting blog posts, articles, videos, and podcast episodes about AI
TOP 3 NEWS IN AI THIS WEEK💎
“Lovable becomes a unicorn with $200M Series A just 8 months after launch” - article by TechCrunch: READ
This article is about Lovable, a fast-growing Swedish AI coding startup, which has become Europe’s newest unicorn just eight months after launch, raising $200M at a $1.8B valuation by helping over 2.3 million users—mostly non-technical—build apps and websites using natural language, while reaching $75M ARR and attracting both enterprise clients and high-profile investors
“Perplexity sees India as a shortcut in its race against OpenAI” - article by TechCrunch: READ
This article is about Perplexity’s aggressive expansion into India, where it’s partnering with telecom giant Airtel to offer free subscriptions to 360 million users, aiming to scale rapidly in the world’s second-largest internet market as a strategic move to rival OpenAI’s dominance in the West and grow its user base despite monetization challenges.
“Top AI Companies Have ‘Unacceptable’ Risk Management, Studies Say” - article by Times: READ
This article is about two new studies by SaferAI and the Future of Life Institute, which reveal that top AI companies—including OpenAI, Anthropic, Meta, and Google DeepMind—have “unacceptable” risk management practices and weak safety commitments, with none scoring above 35% in responsible AI scaling assessments.
READING LIST 📚
“Reflections on OpenAI” - blog post by Calvin French-Owen: READ
This blog post is about one engineer’s firsthand reflections on working at OpenAI from May 2024 to July 2025, offering a detailed, insider view of the company’s fast-growing, bottoms-up culture, intense Codex product launch, research ethos, infrastructure quirks, and deep ambition to build AGI—while highlighting the tradeoffs, lessons, and human intensity behind one of the world’s most influential AI labs.
“MirageLSD: The First Live-Stream Diffusion AI Video Model” - blog post by Decart: READ
This blog post is about MirageLSD, the first real-time, zero-latency AI video model capable of generating infinite, interactive video streams by solving long-standing challenges in temporal coherence, frame-by-frame responsiveness, and error accumulation using custom diffusion, GPU, and training techniques.
“How to Build a Brain (without losing yours): Intro to Neuromorphic computing”- blog post by Arya: READ
This blog post is about building a brain-inspired spiking neural network (SNN) that mimics how biological neurons process time-based information, demonstrating how to train it on MNIST and showing it can be up to 8× more energy efficient than traditional ANNs — all while exploring the fundamentals of neuromorphic computing and its hardware challenges.
“All AI Models Might Be The Same” - blog post by Jack Morris: READ
This blog post is about the Platonic Representation Hypothesis, which argues that as AI models scale, they converge on shared, universal representations, offering a possible path to cross-model embedding inversion, decoding ancient languages like Linear A, and even understanding whale speech by leveraging the structural similarities between different models’ internal representations.
“How I Use Claude Code to Ship Like a Team of Five” - blog post by Kieran Klaassen: READ
This blog post is about how Claude Code transforms software development by acting as a tireless AI teammate, allowing one developer to ship like a team of five by offloading implementation, debugging, reviews, and GitHub workflows—redefining coding as outcome-driven delegation rather than manual typing.
“Asymmetry of verification and verifier’s law” - blog post by Jason Wei: READ
This blog post is about the concept of asymmetry of verification and the "Verifier’s Law", which argue that tasks easy to verify but hard to solve—like coding, puzzles, and scientific optimization—are especially suited for AI and reinforcement learning, and that as long as a task is objective, fast, and scalable to verify, AI will eventually master it, leading to a future where AI rapidly solves any measurable, verifiable challenge.
NEW RELEASES 🚀
“OpenAI introduced the ChatGPT agent”—a unified agentic system combining Operator’s action-taking remote browser, deep research’s web synthesis, and ChatGPT’s conversational strengths: TRY
“Huma AI released personality cloning voice assistant”: TRY
“Windsurf and Cognition released Wave 11” - Wave 11 introduces major upgrades to Windsurf’s agentic IDE, including voice input, smarter browser integration, named conversation checkpoints, default planning mode, and expanded JetBrains support—making AI-powered software development faster, more intuitive, and deeply integrated across tools: TRY
RESEARCH PAPERS 📚
“Apple Intelligence Foundation Language Models” - research paper by Apple: READ
This research paper is about Apple’s new multilingual, multimodal foundation models, a ~3B-parameter on-device model optimized for Apple silicon, and a scalable server model using a novel Parallel-Track Mixture-of-Experts (PT-MoE) architecture, designed to power Apple Intelligence features with strong performance, low latency, privacy safeguards, and developer-accessible tools for generation, tool use, and fine-tuning.
“TransEvalnia: Reasoning-based Evaluation and Ranking of Translations” - research paper by Sakana AI: READ
This research paper is about TransEvalnia, a prompting-based translation evaluation system that uses reasoning to provide fine-grained, multidimensional scores and rankings of translations, outperforming or matching state-of-the-art MT-Ranker on multiple language pairs and showing high correlation with human judgments, while also addressing position bias in evaluation order.
“How Many Instructions Can LLMs Follow at Once?” - research paper by Distill AI: READ
This research paper is about IFScale, a new benchmark that tests how well large language models follow up to 500 simultaneous instructions in business report writing. It reveals that even top models struggle, maxing out at 68% accuracy, and uncovers key degradation patterns, biases, and tradeoffs that impact real-world prompt design.
“Agentic-R1: Distilled Dual-Strategy Reasoning”- research paper by Carnegie Mellon University: READ
This research paper is about DualDistill, a fine-tuning framework that merges diverse reasoning strategies from multiple teacher models to train Agentic-R1, a student model that adaptively chooses between tool use and natural language reasoning, achieving improved accuracy and robustness across both complex logical and computation-heavy tasks.
“Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation” - research paper by KAIST AI: READ
This research paper is about Mixture-of-Recursions (MoR), a unified framework that combines parameter sharing and adaptive computation within Recursive Transformers to achieve large-model performance at lower cost, enabling dynamic token-level depth selection, memory-efficient attention, and faster inference while outperforming existing baselines across multiple scales.
“Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety” - research paper by Apollo Research: READ
This research paper is about Chain-of-Thought (CoT) monitoring as a safety method, arguing that since AI systems often reason in natural language, inspecting their thought processes can help detect harmful intent—though imperfect, CoT monitoring shows promise and should be further studied and integrated with other AI safety techniques, with careful attention to preserving its reliability during model development.
“Scaling Laws for Optimal Data Mixtures” - research paper by Apple: READ
This research paper is about a principled method for optimizing data mixtures in large model pretraining, using scaling laws to predict model performance across domain combinations, enabling accurate estimation of optimal domain weights for any target domain—validated across LLMs, multimodal, and vision models—without relying on expensive trial-and-error.
VIDEO 🎥
OTHER 💎
"Brainrot Education"