X / Twitter
Claude, Anthropic's Official Account
Claude announced general availability for Claude for Excel, PowerPoint, and Word, with Claude for Outlook entering public beta. Crucially, Claude carries full conversational context when moving between Microsoft apps, enabling seamless cross-application AI-assisted workflows. The integration is available to all paid-plan users. The post received more than 30,000 likes, the highest engagement in this batch.
https://x.com/claudeai/status/2052445786651168849
Alex Albert, Anthropic Research Lead
Alex Albert shared actual Firefox security results with Claude Mythos Preview: with the model's help, the Firefox team fixed more vulnerabilities in April than in the previous 15 months combined. This demonstrates AI-assisted security testing at scale in a real open-source project. The post received more than 11,000 likes.
https://x.com/alexalbert__/status/2052468573516513762
Sam Altman, OpenAI CEO
Sam Altman expressed his central belief about AI and software engineers: "Turning developers into superheroes is much cooler than replacing them. What one truly excellent person can do today is incredible." He also observed an interesting interaction pattern: younger people favor voice, older users favor typing, and middle-aged users use both. He wondered whether this distribution will change.
https://x.com/sama/status/2052485051812909530
https://x.com/sama/status/2052462428663992564
Thariq, Claude Code Engineer
Thariq reposted Dario Amodei's statement that Anthropic works every day to acquire more compute and pass it to users, adding "you heard the man." It received more than 1,300 likes, reflecting strong community anticipation of expanded compute.
https://x.com/trq212/status/2052250816720056604
Amjad Masad, Replit CEO
Amjad Masad shared an interesting example: a petition described as "the most viral petition in history" is hosted on Replit. He also collaborated on content with well-known product podcast host Lenny Rachitsky and reposted an article about Replit.
https://x.com/amasad/status/2052478595277467703
https://x.com/amasad/status/2052185144879096139
Peter Steinberger, OpenClaw Founder
Peter Steinberger described using `/goal` with GPT 5.5 for large-scale refactoring: it can plan very complex end-to-end refactors with tests, with excellent results. The post received nearly 2,000 likes. He also shared his experience mentoring ChatGPT Future Class 2026, where one student team built an AI system to detect disaster survivors through walls and rubble.
https://x.com/steipete/status/2052514752245481675
https://x.com/steipete/status/2052486085226184742
Dan Shipper, Every CEO
Dan Shipper and Kieran Klaassen recorded quick commentary from Code with Claude, discussing the xAI compute agreement, Managed Agents, and Anthropic's strategy of turning its API into full-stack cloud infrastructure for developers. He believes "the AI platform wars are coming."
https://x.com/danshipper/status/2052501376195080381
Madhu Guru, Google Gemini Product Lead
Madhu Guru announced her departure from Google, reflecting on three years at Gemini: OpenAI and Anthropic led three years earlier, while her team built model-development methods, customer-feedback loops, and competitive capabilities. The departure statement received more than 1,100 likes, marking a notable industry talent move.
https://x.com/realmadhuguru/status/2052490869320946037
Peter Yang, Roblox Product Manager
At Code with Claude, Peter Yang expressed longstanding trust in Claude: "Of all AIs, Claude feels most like a trusted friend. I'm glad Anthropic now has the compute to scale." He also demonstrated two AI agents scheduling a meeting on behalf of their respective users.
https://x.com/petergyang/status/2052177418555977866
https://x.com/petergyang/status/2052226719151710221
Podcasts
The MAD Podcast with Matt Turck — OpenAI Board Member Zico Kolter on the Real Risks of Frontier AI
FirstMark partner Matt Turck speaks with Zico Kolter, head of Carnegie Mellon University's Machine Learning Department, OpenAI board member, and chair of its Safety and Security Committee, or SSC. The discussion examines practical AI safety governance, actual frontier-model risks, and AI's inherent simplicity.
On governance, Zico described the SSC's responsibility to oversee model-development safety and governance, including authority to delay releases when necessary. OpenAI has a full safety organization with which the SSC regularly collaborates on evaluations. He emphasized a key distinction: when a model lacks capability, waiting for the next generation often makes sense, but capability and safety improve differently. Bigger models are not necessarily safer, and robustness cannot simply be solved through scale.
Zico divided AI safety issues into misuse, where external malicious actors use AI, and alignment, where models act outside expectations. He sees prompt injection and jailbreaking as the most significant current practical threats. For agents specifically, he noted that access to more external tools and environments increases safety complexity exponentially as the attack surface expands.
He also introduced Grace One, the third-party AI safety and security evaluation company he co-founded. It provides large AI labs with red teaming, including human competitions and automated systems, and enterprises with customized agent firewalls and mitigations. He believes the industry needs common safety evaluation standards rather than separate approaches at every company.
On AI's nature, Zico offered a striking observation: an AI system's core mathematical framework can be expressed in approximately 200 lines of Python, with all complexity emerging from training data. He called this a "beautiful fact" that deserves broad understanding, while clarifying that production AI systems require far more than 200 lines. Data pipelines and scaled training, such as efficiently using tens of thousands of GPUs, add enormous engineering complexity.