X / Twitter
Claude Code engineer Cat Wu
Anthropic announced custom environments for Claude Code, allowing users to run it remotely through claude.ai, Claude desktop, and mobile. This substantially expands deployment flexibility across platforms.
https://x.com/_catwu/status/2033705960959840649
Replit CEO Amjad Masad
Masad observed that vibe coding is empowering citizen journalism and OSINT, or open-source intelligence, practitioners. Replit’s code-review agent is designed to be adversarial toward its coding agent and can therefore be overly aggressive. He expressed excitement about users earning their first money from apps, saying it feels as thrilling as his own first earnings. Overall, his posts this week continue to focus on Replit’s product capabilities and users’ entrepreneurial stories.
https://x.com/amasad/status/2033777173757366501
https://x.com/amasad/status/2033777451894231208
https://x.com/amasad/status/2033789817331585178
https://x.com/amasad/status/2033949556828934349
Box CEO Aaron Levie
Levie argued that one of universities’ hardest problems today is deciding what to teach when the whole world says every job will be transformed. Box’s official CLI has launched, letting users treat Box as a full cloud file system through Claude Code, Codex, Perplexity Computer, OpenClaw, and other major AI tools, including on free accounts. This is an important platform move for Box within the AI agent ecosystem.
https://x.com/levie/status/2033749314997280850
https://x.com/levie/status/2034023235654980049
Y Combinator CEO Garry Tan
Tan shared his recent heavy use of GStack for coding, while acknowledging stress from personal difficulties the previous week, including being stranded in Austin by weather and his mother’s hospitalization. He recommended San Jose Mayor Matt Mahan’s interview on Sam Harris’s podcast, calling Mahan an advocate for common-sense reform in California and encouraging more people to learn about him.
https://x.com/garrytan/status/2033729112117018821
https://x.com/garrytan/status/2033736358863507949
Google Labs VP Josh Woodward
Woodward announced Gemini’s Personal Intelligence feature, building on last week’s worldwide memory rollout. US users can now enable it across web, Android, and iOS, marking Gemini’s shift from one-size-fits-all answers toward personalized AI.
https://x.com/joshwoodward/status/2033983926117773517
Claude Code engineer Thariq
Thariq published an in-depth article on using Claude Code Skills effectively, joking that “using Skills is itself a skill issue.” He previewed an open-source iMessage skill example this week and a Skills livestream on Thursday. He also described using Claude to assist his writing, not by ghostwriting but by helping him express ideas more clearly. Together, the posts reflect his consistent view of AI as enhancing rather than replacing people.
https://x.com/trq212/status/2033958799615398346
https://x.com/trq212/status/2033958801259544857
https://x.com/trq212/status/2034043795193663536
Vercel CEO Guillermo Rauch
Rauch rejected the common view that chat is only a transitional form for AI, predicting more work and life will happen through chat interfaces, not less. He also summarized Vercel’s build and deployment optimizations over six months, including a latest 15% speed improvement in the deployment step.
https://x.com/rauchg/status/2033881564824932578
https://x.com/rauchg/status/2033943845613027615
Cursor designer Ryo Lu
Lu reflected on the value of struggle: when things come too easily, neither the mind nor emotions get exercised. Growth is often cyclical, and ease can look like progress while skipping the crucial repetitions that truly change us. The post received more than 1500 likes and resonated with many technology workers.
https://x.com/ryolu_/status/2033937719685222598
Builder Zara Zhang
Zhang shared her practical Skills-building process: first brute-force the task with Claude Code, iterating until satisfied, then have Claude read Anthropic team blogs and documentation she downloaded beforehand and capture the full workflow as a skill. She also emphasized spending an hour every day playing and exploring with models to keep improving in AI—not always seeking immediate utility, but testing limits and enjoying building.
https://x.com/zarazhangrui/status/2033978001126564186
https://x.com/zarazhangrui/status/2033982697899757848
FPV Ventures Partner Nikunj Kothari
Kothari launched Lenny’s Roundtable: he organized all Lenny Rachitsky’s content, videos and newsletters, into a repository and built a website that retrieves relevant material and simulates different guests’ debating perspectives in response to questions. Built with Conductor Build and Railway, it is powered by Gemini and OpenAI models.
https://x.com/nikunj/status/2034055256318026155
Sam Altman (@OpenAI)
Altman posted a heartfelt tribute to predecessors who manually wrote complex software character by character, reflecting on the difficulty of that era and thanking them for bringing humanity to the start of the AI age. The post received nearly 28000 likes.
https://x.com/sama/status/2033935276079510011
Podcasts
Latent Space — Why Anthropic Thinks AI Should Have Its Own Computer — Felix Rieseberg of Claude Cowork/Code
This Latent Space episode features Anthropic’s Felix Rieseberg, exploring the product philosophy and engineering decisions behind Claude Cowork, the cloud collaboration workspace.
Felix defines Claude Cowork as “Claude Code for nondevelopers.” It gives knowledge workers unfamiliar with terminals access to Claude Code’s powerful agent capabilities by running Claude Code in a VM with a friendlier UI and safer defaults. He emphasized that Cowork is not a stripped-down version but a superset, like VS Code relative to Visual Studio: apparently aimed at beginners, yet its extensibility attracts advanced uses.
Architecturally, Felix considers the most important decision to be running the agent on the user’s local computer rather than moving it to the cloud. A truly useful AI assistant needs access to the same tools as its user or it will be constrained. Asking a developer to work without a computer, exchanging code only through email, might have been feasible 50 years ago but is unacceptable today; the same applies to AI.
On iteration speed, Felix described an internal Anthropic shift from a waterfall process of memos, designs, and execution to “quickly build all candidate approaches, then use small-group testing to select the best.” Lower execution costs make parallel prototype validation routine.
Skills are Cowork’s core extensibility mechanism. Felix explained that they are essentially Markdown files: describe in natural language how Claude should perform a task, and it understands and executes. Host Swyx demonstrated packaging recurring tasks such as YouTube uploads and calendar-conflict checks as Skills for daily automation in Cowork. Skill portability also came up: the current solution is file-based GitHub repositories, while the future may require separation between personal and public components.
On AI’s future social impact, Felix candidly said Anthropic is deeply concerned about effects on entry-level employment. Repetitive work previously assigned to junior employees is exactly what AI handles best. He believes society, economists, and governments need to discuss this seriously, and current public conversation is inadequate.
https://youtube.com/watch?v=ZpZ7lFoWaT8
Training Data — Greetings, Earthlings: Philip Johnston of Starcloud on Data Centers in Space
This Training Data episode interviews Starcloud founder and CEO Philip Johnston about an idea that sounds like science fiction but is becoming real: moving data centers into space.
Johnston explained why this is an economic inevitability rather than a technical gimmick. The marginal cost of new energy projects on Earth is soaring: permitting alone takes 5 to 10 years in North America, and suitable sites are dwindling. Space follows the opposite logic: as SpaceX Starship manufacturing and launch costs fall, each additional batch of satellites becomes cheaper at the margin. He believes space data centers become cheaper than terrestrial ones once launch costs fall below $500 per kilogram, because space needs neither land permits nor battery storage under continuous sunlight, and each square meter of solar panel generates 8 times as much power as on Earth. He predicts close to a trillion dollars in annual capital expenditure deployed in space within 10 years, calling it the largest market opportunity ever.
The main engineering challenge is not that space is too cold, but the opposite: vacuum prevents convective cooling, leaving infrared radiation. Starcloud spends 70% of engineering time on heat dissipation, using heat pipes and heat pumps to raise chip-cooling fluid from 60 degrees Celsius to a radiator temperature of 100 degrees Celsius, then dissipating heat through huge deployable radiators. The second challenge is chip reliability under cosmic radiation. Tests at a cyclotron proton-beam accelerator in Knoxville and a heavy-ion accelerator at Brookhaven National Laboratory simulate 5 years of radiation in 24 hours.
Surprisingly, GPU workloads have natural tolerance to cosmic rays: AI inference is stochastic, so even bit flips barely affect output quality. Two ChatGPT poems might differ in wording yet retain the same quality. Starcloud’s first satellite carries 5 Nvidia GPUs, including 1 H100, and has not yet suffered a chip-related failure requiring a restart.
On scaling, Johnston said Starcloud 3 is designed for 8 kilowatts, with the next generation substantially more powerful. The ultimate goal is gigawatt-scale orbital data centers, and he believes launch costs will make space the default choice for global data center construction by then.