X / Twitter
Sam Altman, OpenAI CEO
Sam announced an aggressive Codex promotion: over the next 30 days, any company wishing to switch can use Codex free for two months, backing the claim that "Codex is the best AI coding product." The post received 17,061 likes, making it the day's most widely circulated AI industry news. He also discussed the price–speed trade-off, arguing that speed deserves optimization alongside the price–intelligence trade-off, potentially enabling more aggressive pricing for scenarios where speed does not matter.
https://x.com/sama/status/2054626219858293128
https://x.com/sama/status/2054627102922797323
Alex Albert, Anthropic Research
Alex announced that from June 15, all paid Claude plans will include a monthly allocation of Claude Agent SDK credits. These cover users' own scripts and agents, `claude -p` command-line calls, and SDK-based third-party apps including OpenClaw and Conductor. This is an important step toward Anthropic's developer ecosystem, letting developers and power users run agent workflows without additional payment.
https://x.com/alexalbert__/status/2054613082589298899
Guillermo Rauch, Vercel CEO
Guillermo shared insights from actual production data on Vercel AI Gateway: Google leads by a wide margin in production scale, Anthropic dominates coding tasks and spending, and OpenAI has grown rapidly since GPT-5.4's release. This is among the few public datasets drawn from actual agent usage and offers a valuable reference.
https://x.com/rauchg/status/2054671803264757957
Peter Steinberger, OpenClaw Founder
Peter demonstrated a futuristic remote-control setup: streaming an Android phone's screen through Tailscale to a Mac in a data center, then having an agent, or claw, operate the phone through Appium. His agent can now book an Uber for him; the post received 1,562 likes. He also released Crabbox 0.13.0, adding Modal sandbox execution, full working-directory resynchronization, native Windows script support, and clearer SSH and synchronization failure messages. He now uses this workflow for almost every PR.
https://x.com/steipete/status/2054647734418756012
https://x.com/steipete/status/2054690836613324997
Nikunj Kothari, FPV Ventures Partner
Nikunj said VC has truly become "consensus capital," illustrating it with a personal experience in New York. A friend who had worked at a startup lamented that the company never reached its potential; after their conversation, Nikunj realized how wasteful it is not to provide capital to the right people at the right time. His reflection on VC's current sameness carries a clear element of self-criticism.
https://x.com/nikunj/status/2054599845214650442
https://x.com/nikunj/status/2054612777785045209
Builder Zara Zhang
Zara is developing a skill that transforms "ugly PowerPoints" into polished HTML slides in one pass. She is publicly requesting ugly PPT examples for testing, seeking enough real cases to cover different forms of "ugly."
https://x.com/zarazhangrui/status/2054625264286749138
Swyx, Host of Every CEO's AI & I Podcast (via Latent Space)
Commenting on prompt injection and API-key security, Swyx said the issue has multiple layers and should not simply be reduced to "OpenClaw is insecure." AI security risks need to be understood more systematically.
https://x.com/swyx/status/2054570997865718220
Podcasts
Training Data — Suno's Mikey Shulman: Everyone Can Make Music Now
Key takeaway: Suno's central breakthrough is modeling music purely as sound waves, rather than combinations of notes and instruments. This approach of "throwing away all musical knowledge" instead frees the model's creative boundaries.
Mikey Shulman is Suno's co-founder and CEO, a Harvard physics PhD specializing in quantum computing, and a former Kensho co-founder. Suno is one of the largest AI music-creation platforms today.
Suno's origins were full of surprises. Early back-of-the-envelope calculations suggested high-quality music generation was "two orders of magnitude short" given the available compute and model scale, so they initially used the same technology to understand audio rather than generate it. Following technical breakthroughs, they discovered highly efficient audio compression and pivoted to generation. Early Suno could produce only 12.5-second clips and sometimes ignored the lyrics. But they followed Midjourney's path by launching a Discord bot and discovered that many users genuinely enjoyed creating this way.
Suno's most important technical decision was to model music as continuous sound waves—float32 values sampled 48,000 times per second—rather than telling the model "there are 12 pitches" or "there are 200 instruments." He explained: "If you tell the model there are 12 pitches, it will only produce those 12 pitches, and you'll never create the next Skrillex with Suno." A consequence of imposing no musical-knowledge constraints was spontaneous new stylistic combinations, such as trap with sitar, country with 808 bass, and even microtonal music. Suno works best in country and pop, with uneven quality in less mainstream styles, but "people willing to keep trying always find surprises."
He offered a striking user-behavior figure: on any given day, 90% of Suno users create content rather than merely consume it. He sees this as the true paradigm shift: from only a tiny fraction of the world's 8 billion people making music to creation itself becoming a source of entertainment and satisfaction. The pleasure is not using the finished output elsewhere, but the act of creating.
Suno V5 was a qualitative leap, but he acknowledged that such leaps are hard to predict because research investment and improvements in user experience are not linearly related. An internal "10% preference improvement" does not necessarily translate into user growth. They already know the directions for V6 and V7 and deliberately maintain a regular release cadence instead of waiting for a "perfect version" to launch all at once.
https://www.youtube.com/watch?v=Jq3BIGz4vXQ
AI & I by Every — Claude Code Can Be Your Second Brain
Key takeaway: Building a "second brain" with Claude Code on an Obsidian vault lets knowledge workers complete the entire research, thinking, and writing process on their phones, beyond code generation alone.
Noah Breyer was an early Every interviewee, founded AI strategy consultancy Alethic, previously co-founded Percolate and Variance, and is a dedicated tool enthusiast.
His core setup is a server in his basement hosting his Obsidian vault—Markdown files synchronized through Git—with Claude Code running on top, allowing direct access to and operation of the entire knowledge base from his phone. He believes this breaks the assumption that "phones are unsuitable for deep work," enabling real research and writing in any setting.
He organizes Obsidian according to PARA—Projects, Areas, Resources, Archives—with everything in Markdown files Claude Code can directly retrieve and manipulate. Claude Code asks him questions, searches relevant material across the vault and web, and tracks his recent thoughts and important insights.
For voice, he strongly supports Grok's voice mode, considering it substantially better than ChatGPT's and Gemini's in tool use and research. During a five-hour drive, he used it to deeply study Transformer self-attention, calling it the best explanation he had heard. On another drive, he studied Walter Benjamin's "The Work of Art in the Age of Mechanical Reproduction," reflecting on how each new technology prompts elite anxiety that "technology lets more people do what only we could previously do," a framework he uses to understand criticisms of AI.
Overall, his framework treats Claude Code as a thinking partner integrated with an entire knowledge system, not just a coding assistant. The key is organizing the vault in a format Claude Code can directly operate.