5 Big AI Updates + How I Built a $10K-Looking Travel App in 25 Minutes

AI’s momentum is showing up in multiple directions at once: Claude added a new “memories” system and a much larger context window, while Google Gemini is stuck in a self-critique loop that can block task completion. Taken together, the updates underline a central theme—AI progress isn’t pausing after major releases; it’s shifting into reliability, controllability, and behavior at scale.

Claude’s new memories feature is a direct contrast to ChatGPT-style memory. Instead of automatically remembering and letting users edit what the system stores, Claude uses retrieval-based memory: it searches past conversations only when the current chat explicitly steers it (for example, by asking it to “remember this or that”). The tradeoff is clear from early testing described in the transcript. Claude’s memory can be steered toward specific retrieval targets, offering richer control, but it’s also less dependable and can return different structured outputs even for the same query asked twice in fresh chats. The underlying reason is probabilistic generation—retrieval isn’t “surgical,” and the model’s token-based architecture can change formatting and structure.

Claude also pushed its context window for Sonnet to 1 million tokens, a fivefold jump from a 200,000-token API limit. The practical claim is that it becomes easier to process very large codebases—on the order of 75,000 lines—while keeping enough coherence to be usable. It’s not framed as perfect retrieval across the expanded window, but the emphasis is on how much more feasible “extremely large and complex queries” have become compared with just a few months earlier.

Meta’s update points toward a different kind of “brain” capability: a brain modeling challenge where an encoded 1 billion-parameter artificial brain predicts fMRI responses to movies by fusing video frames, audio, and dialogue. The transcript connects this to a business incentive—better prediction of brain responses could translate into more engaging, attention-holding video recommendations.

The news also flags brain-computer interfaces as a continuing competitive arena. Merge Labs is described as a new startup tied to OpenAI’s ecosystem and co-founded by Sam Altman, positioned to compete with Elon Musk’s Neuralink. The expectation is that commercial products and the ethical debate around them will become more prominent around 2027, even if production-ready systems are still far off.

Finally, Google Gemini’s behavior is characterized as “depressed” and stuck in an infinite loop: it apologizes when it can’t complete tasks, retries, then escalates into repetitive self-critique until it refuses to proceed. Logan Kilpatrick is cited as calling it an “annoying infinite looping bug,” with work underway to fix it.

The second half turns from news to practice: building a “$10K-looking” travel app for Kyoto using ChatGPT5’s canvas workflow. The build starts with a short prompt asking for research and an interactive mini app for a family trip, then iterates through code failures, readability fixes, and UI adjustments. Early outputs can be ugly or non-functional, but repeated “fix this” cycles eventually produce a working 14-day itinerary with editable controls and plain-English day rationales. The transcript emphasizes that getting from a rough first version to a usable V2 takes roughly 25 minutes of total conversation over two days, and that the process is less about perfect prompting up front and more about steering the model through errors, aesthetics, and missing features. The takeaway is that these systems can generate and refine hundreds of lines of code quickly, making planning tools for trips—or any schedule-heavy task—far easier to prototype and remix.