I Tested Claude & ChatGPT's New Knowledge Connectors—Here's Your TLDR + Pros & Cons

OpenAI’s Atlas browser is shipping as a public MVP, and the biggest differentiator isn’t just faster iteration—it’s personalization built on ChatGPT memory. Atlas is rolling out without yet matching the quality of established AI-powered browsers, but the OpenAI team is already collecting feedback aggressively and prioritizing improvements, including heavy use of codecs. The practical edge is how OpenAI plans to reuse “ChatGPT memories” inside the browser experience—then carry that same personalization into other AI products. That “app-in-app” style approach is already showing up in ChatGPT-style in-app experiences that can launch applications while preserving memory context, signaling a broader strategy: make memory a persistent layer across the product surface.

Anthropic’s parallel push is “agent skills,” and adoption is accelerating quickly. Skills are spreading faster than earlier ecosystem primitives like MCP, with GitHub stars rising sharply as developers remix and reuse them. A key architectural shift sits behind the hype: instead of relying only on a prompt plus a context window, Anthropic introduces a third layer—reusable skill patterns—sitting between the prompt and the remaining context. That structure makes skills portable across Claude surfaces, and it can even be “jury-rigged” to work in other environments like ChatGPT. The implication is that skills could become a new default prompting architecture, prompting other major model makers to respond with competing implementations or to adopt skills-like patterns as standard.

Hardware and operating-system integration are also moving into focus. Apple’s M5 laptop went on sale with peak GPU compute performance aimed at AI workloads, reinforcing a trend of embedding AI capability directly into consumer hardware. That matters because OpenAI acquired Sky, described as the team best positioned to connect natural-language queries to macOS. Two plausible futures follow: sooner, OpenAI could ship an agentic capability tightly integrated with macOS and optimized for M5-class hardware; later, the same learning could support a more ambitious “native AI OS” direction—speculative, but increasingly aligned with how LLMs are learning to interact with real environments.

Security is the pressure point across all these AI-native browsers. Researchers keep finding critical vulnerabilities, and the current approach—essentially relying on users to watch what the AI does—doesn’t qualify as a real safeguard. Prompt injection remains the clearest example: a malicious webpage can embed instructions that fit inside a model’s context window, tricking the browser into leaking sensitive data or performing harmful actions (like extracting Gmail or bank credentials). The risk extends beyond browsers too; uploaded documents or files with malicious prompts can steer LLMs in cloud and chat interfaces. For now, there’s no robust hedge, so the “front lines” are expected to be browser-side, even as solutions are still emerging.

On the business side, AI productivity claims are getting sharper. Cityrg CEO Jane Fraser said AI deployment can free up 100,000 developer hours per week—equivalent to adding 50 full-time developers annually for every week saved—while also warning that ROI depends on getting deployments right. The recurring pattern: companies that invest in the right “agentic architecture” and restructure teams and tech stacks see faster ROI; those that don’t often blame models instead.

Finally, Meta’s AI division layoffs—about 600 roles—signal a potential strategy problem rather than a talent shortage. The concern is whether Meta can catch up if shipping cycles lag behind frontier competitors like Gemini, Anthropic, and OpenAI, since delays mean months of lost ground. A short bonus thread ties everything together: both Anthropic and OpenAI are pushing memory and company-knowledge features, currently limited and recency-focused, but expected to expand quietly in the background with deeper data connections.