Replit CEO Amjad Masad shared an early preview of their newest agent tool, noting they’re not even a month into building it — showing Replit’s continued push into AI-powered coding assistants.
Box CEO Aaron Levie made a sharp observation about cognitive limits in AI oversight: “There is a limit on human cognition. Even if you’re not reviewing everything they’re doing, how much you can hold in your head at one time.” He pointed out that at a certain scale, teams of people become necessary not just for execution, but for managing the complexity of what AI agents can do. https://x.com/levie/status/2040264454982480138
Cursor Lead Designer Ryo Lu announced major updates to Composer 2, highlighting a new interface with a new icon set — noting they’re “made with care and character.” He also shared that the new interface offers 2x usage efficiency. https://x.com/ryolu_/status/2040272055673999730 https://x.com/ryolu_/status/2040270917637648749
Y Combinator President Garry Tan announced that “Garry’s List” now scores 95 on page speed (a subtle flex at codegen skeptics). He also released a new developer experience tool called /plan-devex-review to help create better DX frameworks.
https://x.com/garrytan/status/2040228442130850008
https://x.com/garrytan/status/2040211569775395042
Swyx demonstrated what agentic self-improvement looks like in practice: he copy-pasted blog posts and tweets into Devin and it one-shotted the complete implementation. He also shared videos of recorded feature testing with annotations showing exactly what’s being tested. https://x.com/swyx/status/2040181076237443299 https://x.com/swyx/status/2040245613997219902
Anthropic’s Claude announced Microsoft 365 connectors are now available on every Claude plan — connecting Outlook, OneDrive, and SharePoint to bring email, docs, and files into the conversation. https://x.com/claudeai/status/2040086268562842097
Polyagent founder Peter Steinberger shared criticism of Anthropic’s recent API changes, though he credited them for adding some fixes for better cache usage. He also noted he’s been hitting GitHub API quota limits repeatedly — pointing out the API wasn’t designed with agents in mind. https://x.com/steipete/status/2040298884787032103 https://x.com/steipete/status/2040067429242675523
Linear Head of Product Nan Yu argued that combining design and product management creates poor incentives, as PMM is fundamentally a product concern. He also shared a practical observation about engineers: they may roll their eyes at process tools, but eventually they’ll finish a project with 40 incomplete issues and want to group them by use case. https://x.com/thenanyu/status/2040206778751750255 https://x.com/thenanyu/status/2040177368187212157
Builder Zara Zhang weighed in on a popular product debate, arguing that product management and product marketing should be one job — “they’re two sides of the same coin.” She also noted a new trend: people distilling colleagues, influencers, and even their exes into agent skills. https://x.com/zarazhangrui/status/2040319135507480921 https://x.com/zarazhangrui/status/2040211603875074512
AI tutorial creator Peter Yang shared practical tips for getting better results from OpenClaw, including advice on creating a distinct voice (“Warm, sharp, dry humor. Not a chatbot — a person”) and avoiding filler phrases like “Great question!” and “I’d be happy to help!” https://x.com/petergyang/status/2040269162719887610
FPV Ventures partner Nikunj Kothari quipped that “Claude Code eyes are going to be a thing” after token anxiety — a nod to the growing awareness of AI agent behavior. https://x.com/nikunj/status/2040209921690804611
Latent Space — “Marc Andreessen introspects on The Death of the Browser, Pi + OpenClaw, and Why ‘This Time Is Different’“
The Takeaway: Andreessen believes we’re witnessing an “eighty year overnight success” — AI’s recent breakthroughs are the culmination of decades of research finally paying off, and this time is fundamentally different because the technology actually works.
Andreessen shared his sweeping vision of the AI revolution in a wide-ranging conversation with Latent Space. He argued that what makes this moment unique is that previous AI winters were driven by the technology failing to deliver — but now, for the first time, the four key breakthroughs are all working simultaneously: large language models, reasoning systems, agents, and recursive self-improvement.
“If Linus Torvalds is saying that AI coding is not better than he is, like that’s that’s never happened before. That’s the benchmark count.”
He addressed the dot-com crash comparison directly, acknowledging the telecom overbuild that caused $2 trillion in losses, but drew key differences: the companies investing in AI infrastructure are “the bluest chip companies” (Microsoft, Amazon, Google, Nvidia) rather than speculative startups, and unlike fiber optic bandwidth, every GPU deployed today is already generating revenue.
On open source, Andreessen called DeepSeek “a gift to the world,” noting that the open release of R1 taught the entire world how to build reasoning systems in three months. He expressed optimism about the current US administration’s support for open source AI, contrasting it with the previous administration that “wanted to kill it.”
The conversation also touched on edge inference and supply constraints — Andreessen predicted chronic GPU and memory shortages for years to come, which will keep inference costs high until massive new fab capacity comes online. Meanwhile, even three-year-old NVIDIA chips are becoming more valuable due to rapid software improvements — the opposite of typical hardware depreciation.