X / Twitter
Thariq, Claude Code @Anthropic Team
Thariq announced a shift he believes will transform Claude Code workflows: HTML is becoming the new Markdown. He has almost stopped creating Markdown for Claude's informational output—documents, notes, code explanations, and debugging reports—and moved to HTML. HTML renders visually in Claude Code's built-in preview, with styling and structure far superior to plain text. Claude can also add interactive elements, colors, and charts, improving readability and information density. He shared sample HTML documents showing how the format substantially improves the usefulness of the same content. The post received more than 8,700 likes, making it one of the day's most watched updates.
https://x.com/trq212/status/2052811606032269638
https://x.com/trq212/status/2052809885763747935
https://x.com/trq212/status/2052811607454146761
Alex Albert, Anthropic Researcher
Albert disclosed early METR evaluation results for Claude Mythos Preview: the AI safety evaluation organization found an autonomous-task time horizon on the scale of months. This means the new Claude version can independently execute tasks spanning months without human intervention. It marks an important agent-capability milestone and a concrete instance of Anthropic collaborating with an external organization on autonomy safety evaluations.
https://x.com/alexalbert__/status/2052899864493830590
Garry Tan, Y Combinator CEO
Garry Tan released GStack v1.28 with two important features: GStack Browser can now download files and run headlessly for automation. GBrain also added thin-client mode, allowing Claude Code or other secondary agents to use it as a knowledge backend without a full installation, lowering barriers to using GBrain knowledge across environments.
https://x.com/garrytan/status/2052588548126364028
https://x.com/garrytan/status/2052629679572701455
Sam Altman, OpenAI CEO
Altman said OpenAI is willing to help enterprises improve security and believes taking action now is important. The post advances the GPT-5.5-Cyber enterprise security service released that day and signals explicit external marketing of OpenAI's AI security infrastructure.
https://x.com/sama/status/2052558319940944256
Aaron Levie, Box CEO
Levie offered a competitive insight: when AI makes something easy for you, assume it does the same for competitors. He sees a common strategic blind spot in companies focusing on what they can do with AI while forgetting competitors use the same tools. True competitive advantage must come from elsewhere, not merely "we use AI too."
https://x.com/levie/status/2052566788236509254
Peter Steinberger, OpenClaw Founder
Steinberger demonstrated a new agent-collaboration scenario: two OpenClaw instances talking to one another, with his agent Molty learning to delegate cron jobs to other agents. This is a practical demonstration of work allocation within multi-agent systems, showing the OpenClaw ecosystem moving toward more complex autonomous collaboration.
https://x.com/steipete/status/2052630190346457301
Nikunj Kothari, FPV Ventures Partner
Kothari shared a view he finally felt comfortable saying publicly: weekly one-on-ones are among management's most underestimated sources of leverage. He observed that, whatever the frequency, the feedback quality and team impact of 1:1s exceed most people's expectations. Responses to him were sharply polarized—strong agreement or disagreement, with no middle ground.
https://x.com/nikunj/status/2052626010332668278
Matt Turck, FirstMark Capital VC
Turck reflected on AI-agent pricing: it may not move entirely to per-task or outcome-based charging. He believes the question is more complex than it appears, with different services potentially requiring entirely different pricing models, while current industry discussion remains immature.
https://x.com/mattturck/status/2052839798063112303
Builder Zara Zhang
Zara Zhang confirmed that GPT Realtime 2 feels like "black magic," unlocking many previously impossible applications. She believes breakthroughs in real-time voice APIs will create entirely new products, especially agent applications requiring natural spoken interaction.
https://x.com/zarazhangrui/status/2052895297903501601
Dan Shipper, Every CEO
At the Milken conference, Dan Shipper observed a clear information gap: attendees generally lagged the latest AI developments by 3–4 months. For AI practitioners, he sees the time lag itself as an opportunity. The understanding gap between "people ahead" and "people just realizing" can translate into real advantages in financial markets and business decisions.
https://x.com/danshipper/status/2052864131573715225
https://x.com/danshipper/status/2052865144066097217
Swyx, AI Podcast Host
Swyx warned the community about a phishing approach that nearly fooled him. Attackers pose as legitimate sources, possibly business media or tool platforms, to contact developers and induce link clicks or information sharing. He urged AI builders to be cautious about such outreach, especially as their rising profiles make them more attractive targets.
https://x.com/swyx/status/2052599553283813788
Podcasts
AI & I by Every — The Secrets of Claude's Platform From the Team Who Built It
Key takeaway: Anthropic's platform team envisions Claude choosing models and dividing agent tasks itself, leaving users to describe desired outcomes without worrying about architecture.
Angela, Anthropic's platform product lead, and Caitlin, its platform engineering lead, explained on Every's AI & I how the Claude platform evolved from an API endpoint to Claude Managed Agents.
The underlying logic is that as Claude becomes more capable, user interaction moves to higher levels of abstraction. It began with "send prompts, receive tokens" completion, then richer APIs with tool calls and session state, and now Claude Managed Agents: users describe goals while Claude calls tools, executes code, manages state, and returns results. The future vision is Claude understanding itself well enough to "write its own agents at runtime": describe the goal, and Claude chooses models, divides tasks, and assigns subagents.
Claude Managed Agents' core is an enhanced wrapper around the messages API, with built-in sandboxed code execution, web search, and infrastructure combining these tools into "the architecture Claude considers best." The key principle is to return architectural decisions to Claude rather than make users consider "which agent architecture to use."
The guests shared an important product insight: customers are splitting into deep experimenters who want low-level access and to design architectures themselves, and newcomers who want ready-to-use tools without architectural deliberation. Anthropic must serve both, providing granular API control for experimenters and a higher-abstraction Managed entry point for newcomers.
Looking a year ahead, they find it most exciting that Claude may begin to "understand itself." Given a task, it could decide whether to use Sonnet or Opus, one agent or multiple collaborating agents, and how to manage intermediate state. Users would no longer decide these details; Claude's self-understanding would do so and continually improve result quality.
On scalability, generating the necessary agent architecture dynamically means platform infrastructure must match that elastic demand. This is a real infrastructure challenge for Anthropic.
https://www.youtube.com/watch?v=lLypHkIVLqc
Training Data — ElevenLabs' Mati Staniszewski: How Voice Becomes the Interface for Everything
Key takeaway: ElevenLabs' mission began with "solving Polish dubbing" and is evolving toward "letting anyone speak any language with the same emotion," ultimately aiming to provide infrastructure for the audio AI era.
Mati Staniszewski is ElevenLabs' co-founder and CEO. He and his high-school friend Piotr founded it in Warsaw in 2022, starting from a real childhood frustration: foreign films and content in Poland were all deliberately dubbed monotonously by the same voice, leaving viewers to supply the emotion mentally. This unusual industry convention led them to believe everyone should eventually speak any language while retaining their original emotion and intonation.
ElevenLabs took a different path from most AI companies: instead of starting with a multibillion-dollar pretrained large model, it began with frontier models focused on audio, benefiting from timing in 2022. Staniszewski considers whether this remains replicable in 2026 an open question, but sees timing and domain focus as key to its growth.
Voice is becoming an interface sooner than most realize. He listed books and news articles becoming audio, language barriers disappearing, future humanoid robots' main interaction mode, and all hands-free settings such as driving, cooking, and exercising. He sees voice not merely as a replacement for text, but as enabling entirely new applications because typing itself was previously a source of friction.
On the business model, ElevenLabs chose the tooling layer rather than a direct-to-consumer content platform. Its customers are developers and companies building voice products, spanning podcast production, audiobook publishing, and AI-agent voice interfaces.
On creators' future, he believes AI voice tools will not eliminate them but greatly lower barriers to high-quality audio production. Creators with distinctive voices and perspectives retain irreplaceable value, while the cost of production itself approaches zero.
On defensibility, he acknowledged rapid technical iteration in voice AI but sees ElevenLabs' strengths in the broadest language coverage, including smaller languages, the richest voice library, and deep integration with developer ecosystems. It also continues investing in ethical boundaries for voice cloning, with usage policies stricter than the industry average.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8