GPT-5.4 Unveiled, Saudi Arabia Scraps The Line, Perplexity's Jarvis

Welcome to The Aigency Works Dispatch, your backstage pass to what's fresh, fascinating, and flying off the innovation shelves in the world of AI. From breakthrough tools to bold new use cases, we're serving up bite-sized updates to keep you (and your Aigents) ahead of the curve. Let's dive into what made waves this week

Crossing the Line

Saudi Arabia’s futuristic mega-project, The Line, has been one of the most ambitious urban experiments ever proposed. The idea was a 170-kilometre-long mirrored city stretching through the desert, designed to house millions of people in a hyper-dense vertical environment powered by renewable energy. But reports are now suggesting the kingdom may be shifting focus away from parts of the project, redirecting resources towards large-scale AI infrastructure and data centres instead. With global demand for compute exploding, the logic is clear. Owning the physical infrastructure that powers AI could prove far more strategically valuable than building a city that still exists mostly as a concept render.

In some ways this pivot feels like a reality check. The Line was always an architectural spectacle, but questions around feasibility, cost, and actual demand never fully went away. Data centres, on the other hand, are becoming the oil fields of the AI era. Countries that control energy and compute infrastructure are positioning themselves at the centre of the next technological economy. But it also highlights something about the current AI boom. Enormous resources are now flowing into infrastructure before we even fully understand what the long-term economic model of AI will look like. Betting billions on compute might prove visionary - or it might become the next speculative bubble.

GPT 5.4 unveiled

OpenAI’s latest model update, GPT-5.4, continues the pattern we’ve seen across the AI industry recently: steady incremental improvements rather than dramatic leaps. Each version tends to get slightly better at reasoning, slightly better at reliability, slightly faster, and slightly cheaper to run. These updates rarely arrive with the kind of theatrical launch events we saw in the early days of generative AI. Instead, they quietly appear inside APIs and products, gradually improving the tools developers and companies rely on. In practice, these small changes compound quickly. A model that is just a bit more accurate or just a bit faster can unlock entirely new workflows when it’s deployed across millions of interactions.

But there’s also a growing sense of “model fatigue” setting in across the industry. Every few weeks there’s a new version number, a new benchmark score, or a new claim about reasoning ability. For most people actually using these tools day to day, the differences are harder to feel than they are to measure. The real breakthroughs are increasingly coming from how models are integrated into products rather than the models themselves. The AI race may slowly be shifting away from raw model capability and towards experience design, distribution, and ecosystems. In other words, the question is becoming less about who has the smartest model and more about who builds the most useful systems around it.

Perplexity x Jarvis

Perplexity has quietly been pushing further into voice, and it’s starting to feel a bit like the early sketches of something we’ve all imagined for years: a Jarvis-style assistant. The latest updates to Perplexity Voice allow you to speak naturally, ask follow-up questions, and receive conversational answers that feel more fluid than typical voice assistants. Instead of the rigid command-response style we’ve become used to with Siri or Alexa, the interaction feels more like talking through a problem with someone who can actually reason through it with you. Combine that with Perplexity’s real-time search capabilities and it becomes something slightly different from a traditional chatbot. It isn’t just answering questions - it’s actively pulling information from the web and responding in a way that feels context-aware and up to date.

That said, the “Jarvis moment” narrative always needs a bit of grounding. Voice assistants have promised this kind of natural interaction for over a decade, and they’ve mostly delivered frustration. The difference now is that the intelligence layer has caught up. Large language models can actually maintain context and handle messy human questions. But there are still big practical hurdles before this becomes a true everyday interface. Voice in public spaces is awkward, accuracy drops in noisy environments, and people still default to typing when speed matters. The real test will be whether voice becomes something we choose to use rather than something we occasionally try because it feels futuristic.