The Sequence Radar #880: Last Week in AI: A $60B Cursor Deal, Google's Brain Drain, and Midjourney's Body Scanner
SpaceX announced a $60 billion stock deal to acquire Cursor, a code‑completion startup, marking the first time a rocket manufacturer has bought a core AI‑tooling platform. In the same week Google saw two of its most influential researchers—Noam Shazeer, co‑author of the seminal "Attention Is All You Need" paper, and AlphaFold Nobel laureate John Jumper—depart for OpenAI and Anthropic respectively. Midjourney unveiled a prototype ultrasonic CT scanner, signaling a push from generative‑image firms into medical imaging hardware. The roundup also highlighted a wave of multi‑billion‑dollar fundraises across AI startups, from Baseten’s $1.5 B round to General Intuition’s $300 M raise.
The Sequence AI of the Week #878: Inside Google Deepmind's First Real Crack in Next-Token Generation
Google DeepMind unveiled DiffusionGemma, a text‑diffusion model that generates multiple tokens simultaneously, challenging the sequential nature of transformer‑based LLMs. The architecture treats text generation as a diffusion process, allowing parallel updates and reducing inference latency. Early benchmarks show DiffusionGemma matching...
The Sequence Opinion: Systems of Record Vs. Systems of Action
The post argues that agentic AI will not destroy SaaS but will redefine its purpose. Historically, enterprise software acted as a system of record, storing canonical data while humans performed actions. Now AI agents can safely act on that data,...
The Sequence AI of the Week #875: Why Your Language Model Needs a Nap
The blog spotlights a new paper arguing that large language models suffer from anterograde amnesia after pre‑training, forgetting new information once a session ends. The authors propose a biologically inspired “sleep” phase that consolidates recent context into long‑term weights. Experiments...
The Sequence Knowledge #874: Transformers or Not?
The Transformer has become the default architecture for advanced AI because its attention‑driven design scales smoothly with more data, parameters, and compute. Its universal attention mechanism lets each token consider every other token, making it effective across language, code, images,...
The Sequence Opinion #872: The Cake Is a Battlefield: Who Really Controls the AI Stack
Jensen Huang depicts the AI stack as a five‑layer cake—energy, chips, infrastructure, models, and applications—emphasizing seamless reinforcement across the layers. The blog argues that strategists see the same diagram as stacked margin pools, where each tier can become a commodity...
The Sequence AI of the Week #871: Inside the Loop with Claude Opus 4.8
Anthropic released Claude Opus 4.8 on May 28, 2026, delivering a suite of reliability‑focused upgrades for AI agents. The model cuts unremarked code flaws by roughly four‑fold, patches silent tool‑call skips, and improves long‑context compaction. New dynamic workflows let the system fan out...
The Sequence Opinion #868: Recursion Is the New Scaling Law
The Sequence Opinion #868 argues that AI progress is moving from simple scaling—bigger models, more data, more compute—to recursive systems that can loop, critique, and improve their outputs. Historically, scaling laws guided model development, but recent breakthroughs show that the...
The Sequence Opinion #864: Every AI Agent Needs a Computer
The post argues that the next breakthrough for AI agents lies not in larger models or longer context windows, but in giving each agent direct access to a computer. An agent confined to token output is likened to a "brain...
The Sequence Knowledge #862: Learning About Text Diffusion Models
The post highlights how diffusion models dominate visual AI while text generation still relies on autoregressive (AR) transformers like GPT‑4. It explains the inherent limitations of AR models, such as generation drift and the reversal curse, which stem from their...
The Sequence Opinion #860: Every Company’s Last eXam: Some Reflection About Practical AI Evals
The Sequence Opinion #860 argues that AI evaluations have become the fourth pillar of modern artificial intelligence, joining compute, data, and models. As AI agents move from chatbots to production workflows, generic benchmarks no longer suffice; each enterprise must build...
The Sequence Knowledge #858: How State Space Models Went From Curiosity to Serious Transformer Competitor
State space models (SSMs) have moved from a niche curiosity to a serious challenger for transformers. For years transformers dominated because self‑attention works, but its O(n²) scaling now bottlenecks long‑context and large‑model inference, consuming tens of gigabytes of VRAM. SSMs...
The Sequence Radar #857: Last Week in AI: Inside the Machine, Outside the Text Box
The Sequence Radar #857 highlights AI’s transition from a pure model‑building race to an infrastructure race. Anthropic unveiled a Natural Language Autoencoders paper that translates hidden activations into readable text, while OpenAI released three new voice models that push AI...
The Sequence Opinion #856: The Salesforce of Agents Won't Be Salesforce, The Google of Agents Won't Be Google
The post argues that the next wave of internet services will be built for AI agents rather than human users. While platforms like Salesforce and Google dominate today’s human‑centric SaaS, they are ill‑suited to serve autonomous agents that need APIs...
The Sequence AI of the Week #855: Inside Nemotron Omni: NVIDIA’s New Multimodal Brain for Agents
NVIDIA unveiled Nemotron 3 Nano Omni on April 28, 2026, a unified multimodal model that ingests video, audio, images, and text and outputs natural‑language responses. The architecture collapses the traditional Rube Goldberg‑style pipeline of separate speech, vision, and language models into...