AI Builders
Digest

The Takeaway: OpenAI is on track for research-intern level AI capabilities by September, with continual learning as the core path they're scaling toward rather than a hyped side quest.

OpenAI Chief Scientist Ako Paioki is literally one of the most important people shaping the future of AI. He explains that coding tools have exploded at OpenAI, handling the majority of actual coding, and math and physics progress provides measurable signals for closing intelligence gaps. For tougher domains like medicine or law, the challenge is longer-horizon tasks where models must self-assess partial progress, coinciding with RL scaling.

On alignment, chain-of-thought offers a powerful inspection tool because it's not directly supervised, revealing true motivations in natural language. 'I think there's something very exciting here about just not having the training signal fight against us.' Paioki stresses urgency in using model advances to accelerate research itself, while roles evolve to higher complexity with AI - humans will set the vision for bigger problems.

Vercel CEO Guillermo Rauch gave a transparent rundown of a sophisticated security breach. An employee's Google Workspace account was compromised through a breach at an AI platform customer, escalating to access Vercel environments via non-sensitive environment variables.

The attacker was highly sophisticated and likely AI-accelerated, moving with deep knowledge of Vercel. Customer impact is believed to be limited, with priority notifications sent. Rauch recommends secret rotation, monitoring, and proper sensitive env var usage. New dashboard features for env var overview and management have been rolled out.

The team is collaborating with top cybersecurity firms, peers, and law enforcement to strengthen defenses.

Box CEO Aaron Levie shares a contrarian take: AI productivity gains won't replace roles but will make them more sophisticated as tools enable tackling bigger, harder problems.

As everyone accesses the same augmenting technology, job definitions evolve—the AI-augmented engineer builds beyond lightweight apps, requiring deeper expertise. Basic tasks like contract review or financial research will no longer define roles. Those who understand their field and grow skills will stay differentiated by addressing elevated challenges.

FPV Ventures partner Nikunj Kothari warns that cybersecurity companies are going to be worth a lot as AI model capabilities improve, accelerating the pace of attacks.

Humans remain the primary vector, and we haven't seen anything yet. He offers 'Hugops' to all infrastructure providers facing these challenges.

Roblox product Peter Yang spent over an hour trying to get OpenClaw working with GPT for a simple weekly stats recap email task that Opus handled easily, but GPT repeatedly messed up the template and follow-through.

He notes frustration with the model's agentic capabilities for cron jobs or simple tasks, suggesting it may be a skill issue or current limitation. Yang is hopeful Spud / GPT 5.5 will solve this, while also sharing how AI reduced his open terminals to just two apps.

Y Combinator President & CEO Garry Tan released GStack v1.4 open source, adding a new /make-pdf skill that works great as a tool inside OpenClaw/Hermes and Claude Code.

He builds tools by prompting Claude Code for his needs and open-sources them. Tan advises telling OpenClaw to implement cron and subagent replacements to the extent possible until better plugin APIs arrive.

Builder Zara Zhang argues that as AI gets more capable, product teams should spend more time on external communication—talking to users and customers—than internal coordination.

Figuring out 'what' to build matters more than 'how' since AI handles implementation well; inspiration comes from diverse external perspectives. Teams using agents see fewer internal meetings, and indie entrepreneurs dump customer recordings to agents for execution.

FirstMark VC Matt Turck quipped that when asking Claude to evaluate a potential investment, it replied 'cute, just not as good as me'.

He also noted that as software went serverless and now headless, there's not much left for VCs to diligence.

Linear head of product Nan Yu reflected that if people think you’re bad at PR, they tend to believe you’re not lying—which ironically makes you good at PR.

She questioned which major lab's press release would have the shortest distance between what they said and what they believe. Nan also shared a quote tweet critiquing how people frame decisions, like asking red vs green Lambo instead of whether to buy one at all.