TL;DR
Claude pushed into Word with tracked edits, and Claude Code moved planning to the web with auto mode approvals. Garry Tan called agents the Altair BASIC era, while Aaron Levie warned software without a real API gets left behind.
Claude for Word is now in beta, letting people draft, edit, and revise docs from the sidebar while preserving formatting and showing tracked changes. Anthropic is also tying Word, Excel, and PowerPoint together so Claude can carry context across open documents in one conversation.
He says Chinese AI work culture is built around 11am-to-11pm days, young teams, and heavy use of US tools like Claude Code through VPNs. The bigger picture: government backing is strong, Beijing is the main hub, and even one-person startups are being pushed with subsidies as youth unemployment stays ugly.
Claude Code’s new /ultraplan mode pushes implementation planning into the browser: Claude drafts the plan on the web, you can edit it there, then run it either on the web or back in terminal. The pitch is simple — planning is mostly about reading code and intent, so it doesn’t need a local interactive loop, and it uses about the same tokens and rate limits as plan mode.
He says enterprise CIOs and AI leaders are converging on one thing: vendors without a solid headless/API mode are at risk in the next 3-5 years. The bigger shift, from the Box CEO’s view, is that software has to be useful to agents as much as to humans — which could force new business models, but also open up way more workflows and revenue.
He boils early-stage investing down to 11 blunt lessons: get on the flight, don’t trust weekend momentum, and remember that conviction doesn’t come from the data room. His bigger point is that the best founders can argue both sides cleanly, and the best VCs are the ones who keep asking hard questions without pretending they’re smarter than the market.
He just says Skillshare would be a killer domain to own now — more flex than insight. As Linear’s head of product, it reads like a quick hit of startup-brain, not something actionable.
He says the “rationalist” AI-doomer mindset he warned about two years ago is now showing up in real-world violence, pointing to the alleged Sam Altman Molotov attacker as evidence. He also throws in a sharper geopolitical jab: if American enterprise needs saving, maybe China’s open models and Europe’s platform regulation end up doing the job.
She argues the fastest way to understand AI is to become a builder: use coding tools to learn, not just to ship faster. Her other take is more practical than flashy — stop asking models to merely summarize long content and instead have them remix it into formats that surface better insights, like magazine articles or Socratic dialogues. She also sketches the new default workflow as Markdown, CSV/JSON, and HTML replacing the old Word/Excel/PowerPoint stack.
He says the current setup for getting OpenClaw, GBrain, and an LLM knowledge wiki talking to your phone is still annoyingly rough — basically the Altair BASIC phase of agents. But he’s also pointing to a future where a strong PM-style devex review can be automated, so founders can “just do things now” instead of wrestling tooling.
He says the wild part of "free" software is how fast you can change it: hate the UI, push a new one; performance sucks, refactor the data layer and let automation optimize it. That’s a classic builder take from an ex-Facebook, ex-Dropbox CTO — software stops being a fixed thing and starts behaving more like clay.
He says Claude Cowork is Anthropic’s answer to a simple gap: Claude Code was powerful, but too technical for most people. The conversation digs into why the product was built fast, how it uses VMs, tools, memory, and local files, and why the real bottleneck in an AI-agent world may shift from execution to taste and trust.
Anthropic: agentic coding scores shift with infrastructure setup
Lead: Anthropic found that infrastructure choices can move agentic coding benchmark scores by as much as the leaderboard gaps people use to rank models, and argues that eval resource settings should be treated as a first-class variable.
Numbers:
So What: Resource limits don’t just stabilize evals; past a threshold, they change what the benchmark measures by enabling heavier tools, bigger installs, and more memory-intensive strategies. Anthropic recommends specifying both guaranteed allocation and hard kill threshold per task, and calibrating the band so floor and ceiling scores stay within noise. As the post puts it, “a 2-point lead on a leaderboard might reflect a genuine capability difference, or it might reflect that one eval ran on beefier hardware.” For builders and buyers, the takeaway is blunt: don’t trust small leaderboard deltas unless the eval configuration is documented and matched.
Anthropic adds auto mode to Claude Code with classifier-based approvals
Lead: Anthropic launched auto mode for Claude Code, a new permission mode that skips most manual approval prompts by using model-based classifiers and prompt-injection checks to block dangerous actions while letting routine work run unattended.
Numbers:
So What: Auto mode is aimed at teams that want more autonomy than manual approvals but more protection than `--dangerously-skip-permissions`. It uses two defenses: a server-side prompt-injection probe on tool outputs, and a transcript classifier that gates risky actions like force-pushes, credential hunting, external data sharing, and shared-infra changes. Anthropic says the classifier is intentionally conservative: “The classifier is deliberately conservative,” and the system is designed so “the agent shouldn’t be able to hide a dangerous operation behind a benign-looking wrapper.” For builders, the practical takeaway is that routine in-repo edits can flow without friction, while higher-risk actions still get blocked or escalated. Auto mode is not a replacement for careful human review on high-stakes infrastructure, but it is a safer middle ground for everyday coding agents.
Anthropic: Build less around Claude as it gets smarter
Lead: Anthropic says Claude is “grown more than it is built,” and argues agent harnesses should evolve with the model by using what Claude already knows, asking what can be removed, and setting tighter boundaries only where needed.
Numbers:
So What: Builders should shift orchestration from the harness to Claude where possible: use general tools like bash, let Claude manage filtering and context, and rely on skills, compaction, subagents, and memory folders for long-horizon work. Anthropic’s core advice is to keep pruning old guardrails as capability improves: “what can I stop doing?” For UX, security, and observability, promote only the actions that truly need dedicated tools or confirmation gates. The practical takeaway: re-test your assumptions every model step-change, or your harness will become dead weight.
AI’s real bottleneck is trust, not intelligence
The Takeaway: The next software leap won’t come from smarter models alone, but from making them safe, local, and easy to trust.
Rieseberg, who leads engineering for Claude Cowork at Anthropic after stints at Slack, Stripe, and Notion, comes at AI like a product engineer obsessed with how real people actually work. His point is that models are now good enough to handle long, messy, multi-step tasks — the hard part is turning that power into something humans will actually use without babysitting it.
That’s why he keeps coming back to local-first design. “I have a strong belief that the data that is relevant for your work probably lives in two different places,” he says: on your computer and in the cloud. For him, asking users to upload everything to a remote system is both a trust problem and a practical mess, especially when banks, logins, and security checks get involved.
Cowork reflects that philosophy. It gives Claude its own sandboxed computer, lets users define skills in plain markdown, and stores memory as text files. The result is less sci-fi than it sounds — and more useful. As Rieseberg puts it, “most of the buttons you add and most of the product services you build are probably more for the human than they are for the model.”
Peter Steinberger called AI code review the killer dev workflow, while Thariq said Claude Code turned files into a workbench. Cursor bet on specialized models plus brutal RL infra, and Anthropic detailed how it contained Claude across products.
Steinberger said AI agents need lean skills, not bloated prose; Levie argued AI won’t kill jobs, it’ll raise the bar. Yang said Codex tested itself while Claude still owned frontend, and Garry Tan pushed hard evals as the way agents got better fast.
Levie warned AI demos fool CEOs; the hard work starts after. Tan said high agency plus taste beat everything, while Yang urged builders to build the system that builds the MVP first. LeCun kept the pressure on: LLMs help, but they’re not the road to real intelligence.
Levie said AI kills tasks, not jobs, so roles just expand. Masad said Replit’s AI stack can ship MVPs in a weekend, while Tan showed Qwen fine-tunes in hours and personal AI got real. No Priors said Cerebras won by betting on speed before the market cared.
Levie said AI will surface more bugs, making humans the bottleneck. Steinberger argued GitHub’s native PR limits kill bot policing, while Tan pushed the “ship the 60%” rule and Swyx said AI coding should harden codebases, not just move faster.