Fleeting Loyalties – A Love Story in Ten-Day Chapters

Ten days ago, a video titled “Smartest Open-Source AI Ever Built” drew me to Z.AI, a masterpiece of agentic code that breaks tasks into subtasks with uncanny precision. I’m gobsmacked by its reasoning, hooked on answers delivered with a clarity that feels almost unfair, like cheating at trivia night.

Out goes Kimi, unfortunately, but not unusually. This is how it’s been. A parade of short-lived loyalties. Back then, around the end of the previous month, I was singing songs about its reports polished enough to impress seasoned consultants. Manus’s work appeared pale in comparison. And, it was Manus who changed the industry a few weeks prior by exhibiting a type of Deep Research that made everyone else’s claimed deep research seem shallow, like a kiddie pool.

It took me a while to recall the name of Minimax’s maker, even though I was certain just a few months ago that its handling of long contexts would become a permanent part of my workflow. Those were the days when Qwen, juggling 119 languages, was the first go-to outside the three core Western models for me. Amid the new entrants, my browser history groans under taps to unearth last uses of Perplexity’s revolutionary real-time search or DeepSeek’s R1, trained for $6 million yet scoring 92% on complex problems versus GPT-4’s 78%.

Amid this blur, space must be made for fresher stars. Last week’s Kimi K2 and xAI’s Grok 4, with its PhD-level smarts across fields, yield to whispers of OpenAI’s GPT-5, rumored for August 2025. Perhaps NotebookLM could track it all, but I’ve forgotten its podcast-like summaries since Gemini’s 1-million-token window does the same without container fuss. Time on staples like ChatGPT, Gemini, and Grok dwindles, though nothing’s permanent in their usage decline as well.

We are living in an era of unprecedented technological cannibalism. The rate of innovation is so swift, so brutal, that product loyalty is a myth. While some companies announce consistent revenue growth, the underlying shifts in usage and the risks of sudden obsolescence are a silent alarm. The question, then, is not about who has the best model today, but who can navigate this chaos tomorrow. This fleetingness, this lack of loyalty, points to something much bigger: the fact that even the most fundamental business philosophies of the past are being completely rewritten.

The Open-Close Reversals

The most visible casualty of this AI carousel is consistency. Take Meta, the self-proclaimed evangelist of open-source. For years, Mark Zuckerberg positioned the transparency of his company's Llama models as a key differentiator. But recent internal discussions, revealed in tech journals and news reports, show a noticeable shift. While Meta will continue to release open-weight models, it has also begun to hold its most advanced work behind closed doors, citing a need for greater control and monetization. 

Yet, in a bizarre mirroring, OpenAI, the long-standing fortress of closed-source, just released its first open-weight models since GPT-2. The new "gpt-oss" models are available for download under a permissive license, a move that CEO Sam Altman himself called an acknowledgment that his company may have been on "the wrong side of history" regarding open sourcing its technology. This peculiar flip-flop by both industry giants isn't just an amusing anecdote. It is another proof that nothing is given about these technologies and their paths ahead.

The Meta Blind Spot

Meta’s Llama 3.1 was heralded as technically excellent, but that hasn’t translated into everyday mindshare. Despite being placed prominently at the top of WhatsApp and Instagram chats, usage is tepid. A recent Sensor Tower survey revealed that fewer than 3% of WhatsApp users engaged with the Meta AI feature in June 2025, despite it being just one tap away and free. 

Meta’s struggles show that reach alone doesn’t build usage. Even free, one-tap convenience hasn’t been enough to change behavior. There has to be something lacking in its models and it is in Llama’s inability to master the reasoning trends. And Meta is not alone. Mistral and Cohere, despite their engineering quality, have similarly trailed in reasoning development. It was only in Q2 2025 that Meta caught up to chain-of-thought reasoning techniques that the closed-source giants and multiple Chinese models adopted last winter. 

Zuckerberg’s recent talent raids are an urgent attempt to buy a ticket back onto a train Meta somehow missed.

The Man vs. Machine Conundrum

If Meta is betting on human talent, that is not what others feel is the main differentiating factor. Until now, the prevailing assumption is that raw compute power guarantees dominance. In other words, one who builds the biggest data center is the one who should win. Musk’s xAI, unlike Zuckerberg’s Llama, is clearly in that camp with the announcement of a Memphis supercluster of 100,000 Nvidia H100 GPUs. It is, perhaps, a fittingly stark contrast between two titans who will likely never agree. 

Lest anyone thought the choice was a simple matter of picking a direction for one’s billions, the truth is far more complex. None of the half a dozen or so Chinese models or LG’s EXAONE or Hon Hai’s Foxbrain have attributed their success to any superstar programmers or massive compute structures. Their boasts tend to be more technical in efficiency-generating innovations through a tight-knit R&D unit focused on optimizing processes. 

So, should one say it is a choice between men or machines or methods?

AI into Apps or Apps for AI?

The most entrenched players in the industry are dealing with the AI revolution by doing what they've always done: bolting the future onto the past. We see this with Microsoft's Edge browser, which now has a Copilot sidebar that can summarize an entire webpage or PDF, and in Google's Search, which offers AI-powered overviews. The integration of Gemini with Workspace, allowing users to ask natural language questions about files in their Drive, is a powerful example of this. But these companies are also learning that simply adding an AI helper isn't enough. They are being forced to massively re-engineer their products from the ground up to make best use of the new paradigms, a process that is often clumsy and slow.

The contrast, however, is being made by a new guard, which operates on a different philosophy entirely: don't bolt AI onto an old app; design a new app for AI. Perplexity's browser Comet, for instance, isn't a traditional browser with AI added on. It is a search-first, answer-centric application that fundamentally changes the user experience. Rumors persist of OpenAI developing a productivity suite, with reports suggesting it's building spreadsheet and slide alternatives that aren't AI-enhanced copies of Excel or PowerPoint, but reimagined containers for agentic reasoning. In code, the battle is visible. Cursor, Sweep, and Devika are not plug-ins to VS Code; they are full AI-native coding environments. Startups like Attio are creating AI-first CRMs, not just adding a chatbot to a legacy system.

The choice appears binary: retrofit the past or plant a flag on fresh ground. That said, some are trying to have it both ways as well. For example, Adobe has been integrating AI into its existing suite of products through efforts like Firefly, and it has previewed “Project Turntable”, an AI-native image editor that exists only in the browser and speaks plain English layers.

The Agent Grab: Everyone Wants to Wake Up Before You Do

Everyone now wants the right to wake up your device. Microsoft’s Recall, though delayed after privacy outcry, will still launch this summer on Copilot+ PCs, taking screenshots every few seconds so that Copilot can answer things like “what was that chart I saw yesterday?” Google’s Project Mariner will pilot a Gemini agent that clicks, scrolls, and books flights across any site you allow. OpenAI’s desktop app needs screen recording access to “improve suggestions,” as it edges toward becoming an always-on digital butler. Meanwhile, Apple’s iOS 19 promises a new “Personal Context” layer that allows Siri to act across apps and notifications using on-device data.

These assistants aren’t just tools. They aim to see everything, predict tasks, and launch apps on users’ behalf. Even relatively modest-seeming players aren’t immune. Perplexity’s mobile app asks for microphone, camera, and file access; Claude’s app, newly released, wants access to notifications and calendars. These models, collectively, are becoming a new layer in our workflow, residing somewhere between the operating systems and applications.

Curiously, the only companies not asking for such control are the Chinese models. They live mostly in the browser, respond only when prompted, and ask for nothing beyond a text input. Whether this reflects architectural limitation, regulatory caution, or a different philosophy of engagement isn’t yet clear. 

APIs in the Shadows vs. Front-End Real Estate

One last quirk in AI’s hall of mirrors: some models lurk in the backend like stagehands, powering apps without stealing the spotlight. Anthropic’s Claude shines via Amazon Bedrock, invoked in enterprise tools but rarely named. Its Opus 4.1 handles coding tasks invisibly, per AWS’s August 2025 rollout. Mistral zips through hosted endpoints on Together.ai, Fireworks.ai, and OctoAI, fueling apps at warp speed without a logo in sight; Cohere joins the shadows on AWS Marketplace, embedding in custom workflows. These API-kind silent engines scale quietly and miss the fanfare.

Others crave the limelight, building brands that scream “use me.” ChatGPT’s app boasts 2 billion visits monthly in 2025 stats. Perplexity’s Comet browser weaves AI into every tab, and Grok chats front-and-center on X. Phind pivots to UI-first with visual answers for devs, while You.com crafts a search app around its model. Visibility forges loyalty, yet invites scrutiny as every glitch becomes a headline. 

Smarter LLMs, Same Price Tag?

On one axis, it’s exhilarating to observe how every major model now solves problems once reserved for specialists. This spans fields as diverse as complex reasoning, coding, image or video creation, creative flow, and memory management advice. Whether it’s Grok, Gemini, or DeepSeek, each appears to edge closer to handling most of our mental load. Yet, another trend is unfolding quietly: the price for “all-you-can-eat” access has settled into a surprisingly narrow band. Despite brand pedigree, infrastructure, or unique smarts, the first pricing tier - normally called “premium” - is priced around USD20 per month. The super-tier with more usage limits is about ten times more expensive at around USD200 per month.

In a Field Where Nothing Settles

Two years ago, Sam Altman, during a visit to India, famously articulated a sentiment many shared: that it was "totally hopeless to compete with us on training foundation models." It was an understandable position then, given the staggering costs and colossal computing power required. That exchange, now just a footnote in the archives of AI history, has aged poorly. Since then, not one or two, but dozens of new players have emerged, many of them approaching the frontier of capability with startling speed. Today, models with names barely known six months ago produce work indistinguishable from those trained with orders of magnitude more capital.

This isn't a simple story of progress; it is an epic of disruption. For those looking at the top players’ revenues, it is still common to assume that the race is over as they race ahead to award them with valuations dozens of times their revenues. Of course, the likes of Meta and xAI would not agree and are looking for new momentum through extraordinary resource allocations. 

Every observation in this essay — from the fluidity of usage patterns to the reversal of open and closed strategies, from the erosion of app boundaries to the fight over devices — points to the same conclusion: nothing here is settled. A model with lower training cost can outperform a giant. A product without a user interface can quietly dominate behind APIs. A firm with no premium plan can still reshape usage behavior. It would not be irrational to expect that one of today’s middle-tier firms could sit atop revenue or engagement charts within two years.

As unsettling as this may appear, this is not a lament, but for observation. And for investors, it is a clear directive: to navigate this tempestuous sea of innovation, one must operate with an open mind, for the rules of yesterday have no bearing on the game of tomorrow.