X / Twitter
Anthropic Claude Code Team Member Thariq
Thariq announced compensation on Anthropic's behalf: users affected by earlier changes to Claude's usage limits would receive one month of credits free, with a link to claim them. The announcement received more than 1,700 likes, reflecting Anthropic's attention to user experience.
https://x.com/trq212/status/2040215427931156595
OpenClaw Founder Peter Steinberger
Peter described a significant developer frustration: repeatedly hitting GitHub API quotas, saying the API was “not designed for agents at all.” The post received 1,832 likes, clearly resonating with developers. Another post, with 3,220 likes, revealed that he and Dave Morin had jointly approached Anthropic, but “the best we could get was...” The vague account attracted intense community interest.
https://x.com/steipete/status/2040067429242675523
https://x.com/steipete/status/2040209434019082522
Claude (@claudeai)
Anthropic officially announced that the Microsoft 365 connector is available on all Claude plans. Users can connect Outlook, OneDrive, and SharePoint, bringing work documents and email into Claude's context. This marks a significant step toward deeper integration with enterprise work. The post received more than 14,000 likes.
https://x.com/claudeai/status/2040086268562842097
Box CEO Aaron Levie
Aaron shared his most important architectural lesson from building AI agents: be “brutally unsentimental” about architecture. Agent systems evolve rapidly with requirements, and emotional attachment to a particular architecture impedes that evolution. Continued progress requires rational choices based on actual results.
https://x.com/levie/status/2039931799414194621
Cursor Designer Ryo Lu
Ryo explained Cursor's core design philosophy: “glass,” rather than a “black box,” so users can see and control everything. He considers terminals a classic example of transparent design and believes AI tools should follow that tradition. Another image post, “Simple beats complex,” received 411 likes, succinctly expressing his consistent position on product design.
https://x.com/ryolu_/status/2039895634313187619
https://x.com/ryolu_/status/2040061299053519099
Y Combinator CEO Garry Tan
Garry called Perplexity Computer “quite special.” He also released `/plan-devex-review` in GStack to help users build excellent developer experiences. His overall position is that tools in today's AI era should empower developers and reduce friction.
https://x.com/garrytan/status/2039943351278190840
https://x.com/garrytan/status/2040211569775395042
Latent Space Podcast Host Swyx
Swyx showed a practical example of “agentic self-improvement”: simply pasting blog posts and tweets into Devin AI let it complete the task in one shot, without repeated iteration. He also criticized AI Twitter's signal-to-noise ratio, saying this is why nobody takes it seriously.
https://x.com/swyx/status/2040181076237443299
Builder Zara Zhang
Zara observed a growing trend: people are “distilling” the styles of colleagues, influencers, and even former partners into personalized, reusable agent skills. This suggests a new direction for personalized AI agents.
https://x.com/zarazhangrui/status/2040211603875074512
Linear Head of Product Nan Yu
Nan strongly supported treating product marketing (PMM) as a responsibility of the product function itself. He also criticized structures that combine design and product management, arguing that designers ultimately report to product managers, undermining design's independence and influence.
https://x.com/thenanyu/status/2040177368187212157
South Park Commons Partner Aditya Agarwal
Aditya made a pointed observation: when a company worth $10 billion sues because your app “looks and feels like its product,” imitation is clearly a real threat. His advice is to focus on building a genuinely differentiated competitive advantage instead of worrying about being copied.
https://x.com/adityaag/status/2040107083841053145
Podcasts
Latent Space — Marc Andreessen introspects on The Death of the Browser, Pi + OpenClaw, and Why "This Time Is Different"
Key takeaway: Marc Andreessen calls today's AI era an “eighty-year overnight success.” Deep learning's victory unlocks decades of foundational research; it did not emerge from nowhere.
In this in-depth Latent Space interview, Marc Andreessen traces his interest in AI back to the 1980s, explains why this wave differs from earlier ones, and discusses browsers, agents, and future infrastructure.
He describes AI as an “eighty-year overnight success”: the original neural-network paper dates to 1943, followed by decades of alternating summers and winters. AlexNet in 2013 was a turning point, followed by the Transformer in 2017, then four quiet years from 2017 to 2021 before ChatGPT sparked mass adoption. He recalled that ordinary users could initially access GPT-3 only through AI Dungeon: “You had to pretend to play Dungeons & Dragons when you really just wanted to talk to GPT.”
He offers a clear distinction for why “this time is different.” From ChatGPT's arrival through 2025, an informed, good-faith skeptic could still dismiss it as “pattern completion,” unsuitable for coding, medicine, or law. The reasoning breakthroughs of o1 and DeepSeek R1 changed that. The coding breakthrough came over the holiday period: “If even Linus Torvalds says AI coding has surpassed him, that's unprecedented.” His conclusion: once AI can do this in coding, it can do it everywhere else.
He has a distinctive view of AI browsers and infrastructure. Traditional browsers as the interface between people and the internet face a fundamental challenge when agents act autonomously, reshaping the logic of existing interfaces. He mentioned Pi, their conversational AI project, and OpenClaw as different explorations of agent-native interaction.
From an investment perspective, he acknowledged that “this time is different” is the most dangerous four-word phrase, but the difference now is that it actually works. AI researchers spent forty years on it, some never seeing results in their lifetimes; Hinton is among the few still alive to see it succeed. His advice is direct: “If I were 18, I'd spend all my time on this.”
https://www.youtube.com/@LatentSpacePod
No Priors — AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus
Key takeaway: Liam Fedus believes AI must connect to the physical world to truly affect science. Language and reasoning alone are insufficient: experiments and experimental data are needed to ground its understanding of reality.
Liam Fedus, a co-creator of ChatGPT and former OpenAI VP of post-training, founded Periodic Labs after leaving OpenAI, focusing on AI for materials engineering and the world of atoms.
He began in physics, researching dark matter as an undergraduate, before spending some of deep learning's richest years at Google Brain in 2016–2017. He worked on distributed training strategies, mixture-of-experts models, and early Transformer development. He described a “Cambrian explosion” when small teams with a few GPUs could advance the frontier.
At OpenAI, his mission was to turn the pretrained GPT-4 into a product. Ideas included writing bots, coding bots, and even a meeting bot taking notes in Google Meet. Ultimately, John Truman overcame objections to choose the most general approach: a chatbot, which became ChatGPT.
Periodic Labs' central premise is that language models make powerful orchestration layers, but materials science requires a closed loop with experiments. They use language models as “directors” while developing specialized neural networks for atomic systems, with symmetry awareness and low latency. Large models invoke these networks as tools or reward functions. He said, “We spend zero effort on coding models. Codex and Claude Code have already accelerated the company enormously.”
Data is materials AI's central challenge. Reported values for the same material property extracted from the literature can differ by orders of magnitude. Models trained on such data can model its distribution without approaching the truth. The real breakthrough combines experimental data with a closed-loop system: use results to check the literature, identify anomalies, and drive the next experiments, enabling active iteration rather than passive accumulation. The business aims to provide an “intelligence layer” for manufacturers and materials companies, first validating internally with Periodic as customer zero before expanding externally.