X / Twitter
Josh Woodward, Google VP
Josh Woodward posted two previews showcasing new Gemini App capabilities. The first featured image transformations, captioned "Turn your car into a Lamborghini," showing an ordinary car visually replaced with a Lamborghini. The second emphasized multilingual capabilities: he said working across languages is now "ridiculously easy," suggesting that barriers to multilingual generation and switching have fallen dramatically. Both appeared as quoted posts with short links, resembling feature recommendations for ordinary users more than disclosures of technical details. Overall, he continues positioning Gemini App as a creative tool that anyone can easily use.
https://x.com/joshwoodward/status/2060443095527989413
https://x.com/joshwoodward/status/2060443093825094091
Boris Cherny, Claude Code Team
Boris Cherny relayed Salesforce's detailed retrospective on fully adopting agentic development with Claude Code, highlighting several striking figures. A migration originally estimated at 231 days shipped in just 13 days; one PR delivered 21 endpoints with 100% test coverage. He emphasized that quality did not decline as output grew: despite more PRs, the overall incident rate fell 5%, because the team built safeguards and quality standards directly into its agentic workflow. His central judgment is that teams gaining the most from AI fundamentally change how they work, rather than merely accelerate existing processes. In his words, consider "which steps can be removed, which handoffs can be eliminated, and what agents can own end to end." He also thanked Srini, Benioff, and the team by name, arguing that productivity and quality are often treated as a trade-off, but Salesforce's experience shows they need not be.
https://x.com/bcherny/status/2060390855383400729
https://x.com/bcherny/status/2060390853835726946
https://x.com/bcherny/status/2060390852619272526
Thibault Sottiaux, OpenAI Codex and ChatGPT Lead
Thibault Sottiaux revealed today that he had seen a number on the Codex dashboard that made him happy, teasing more news "about that number" soon. He thanked users for continuing to adopt Codex and emphasized how early things remain, saying "So early" twice. He also posed a question for everyone building models: do you still trust benchmarks, or simply listen to friends' recommendations? What makes you willing to try a new model? The question received 545 replies, suggesting that "distrusting benchmarks and choosing models by word of mouth" is a widespread community concern. Overall, he is both building anticipation around Codex adoption and publicly collecting firsthand feedback on how users decide to switch models.
https://x.com/thsottiaux/status/2060565265906290786
https://x.com/thsottiaux/status/2060563528596287874
https://x.com/thsottiaux/status/2060529970523603099
Aaron Levie, Box CEO
Aaron Levie offered a contrarian interpretation: when a company is willing to spend $500M building its own replacement for an application, that is the best advertisement for the "application layer." His logic is that nobody would spend so much copying it if the application layer lacked value, so the event should actually make people more bullish on software. He acknowledged that much of the nuance cannot fit in a headline, but maintained that it is a positive signal for software. This continues his longstanding position: amid narratives of AI disrupting applications, he repeatedly argues that the application layer will not be easily flattened and may develop stronger defenses through content and use cases.
https://x.com/levie/status/2060525104384418271
Garry Tan, Y Combinator CEO
Garry Tan shared pointed startup advice. When a founder repeatedly said, "If only we had money, we could do X," he responded directly: money is not the fire; it is gasoline poured on a fire already burning. In other words, the problem is not fundraising but "nobody wants it yet," and the right move is to kindle the first flame. The post reframes "a lack of money" as "a lack of demand," a practical reminder for early founders. Beyond startups, he also posted two comments today about University of California admissions, criticizing UC Regents who "perform morality without really caring about student outcomes" and directly alleging discrimination against Asian applicants in UC admissions.
https://x.com/garrytan/status/2060600088079356292
https://x.com/garrytan/status/2060586945491931202
https://x.com/garrytan/status/2060582680216084925
Peter Steinberger, OpenClaw and OpenAI
Peter Steinberger enthusiastically announced that Vince is joining the team. He said few people truly understand "the new way software is being built," and Vince is one of those who get it. Given his own identities as ClawFather and part of OpenClaw and OpenAI, this reads as a recruiting and expansion signal bringing key talent into OpenClaw. For those following agentic coding tools, OpenClaw's continued recruitment of people who understand the new paradigm is worth watching.
https://x.com/steipete/status/2060306947035832628
Podcasts
No Priors — Building an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan
Key takeaway: As agent activity inside enterprises grows 100x, 1,000x, or even a millionfold, "human in the loop" breaks down completely. The only viable solution is training specialized small models to decide in real time "whether this step should be reviewed by a smarter agent."
The guest is Maxim Bar Kogan, co-founder and CEO of Onyx Security. Based in Israel, Onyx has a team largely drawn from Israeli intelligence units specializing at the intersection of mathematics and cyber, and works on "training agents to supervise other agents." The topic matters because enterprise AI-security concerns have shifted sharply over the past few months: from "what employees type into ChatGPT" two years ago, similar to chatbot DLP, to near-market-wide alarm. Autonomous coding agents are accidentally deleting databases, leaking code and tokens that should not be public, and even causing system downtime.
Maxim identified several key points. First, autonomous coding agents are already the fastest-growing category inside enterprises, accounting for more than 50% in a typical company, alongside approximately 45% low-code automation and just 2% genuinely in-house first-party agents. Second, traditional security tools largely fail here: identity permissions aim to "limit what software can do," yet we want assistants such as Claude Code to have the same permissions we do so they can work for us, undermining identity security's role. Endpoint and API security tools also lack the context to understand "why the agent is doing this" and judge intent. Third, a proxy plus policy engine is insufficient: much AI runs in other companies' clouds or on devices, where a proxy is not always feasible. The hard part is not "seeing the data," but "deciding whether to allow the action after seeing it."
Onyx's solution is to train small models that are "not very smart, but good at just one thing": deciding "whether this step needs a smarter agent to look at it." Maxim offered a counterintuitive chess analogy: "Top players make most moves intuitively, without calculating ahead. But every so often they stop and calculate deeply for a long time, because they know this move matters." Security supervision works similarly: avoid wasting intelligence where it is unnecessary and concentrate compute overwhelmingly on high-risk actions, keeping both cost and latency low.
He also offered a useful framing for founders: enterprises work with a two-year-old company with fewer than 100 employees because the pain is strong enough. And model vendors will not inevitably own this work, because security buyers naturally distrust "the person selling the product" to certify its safety, just as car buyers would not ask the salesperson to certify the car's quality. Maxim considers it a $100B-plus opportunity, with a long-term need to explore mechanistic interpretability to understand the internal structure of model weights and activations.