Often Wrong

Status Games and the Future of Social Networks

2026-01-21T06:13:16.755146+00:00

There seem to be more and more services popping up that let you generate fake vacation pictures with AI, as covered by The Verge here. At first I thought, "why would anyone want to use this?" But then when I saw Adam Mosseri's end-of-year thoughts about Instagram (also summarized by PetaPixel), it made sense to me. In this post, I want to go through the things I agree with Adam about, and one important bit that I disagree with.

The death of the feed

First, a brief aside about Instagram. The main feed has been in decline over the past few years. This has been covered by many tech outlets, and even alluded to by Instagram's top brass. From my own experience, people are not "real" in the feed anymore. They treat the feed as a carefully curated scrapbook of sorts. It is not considered weird to post vacation photos to the feed long after the vacation. During the vacation, however, you either post to stories (maybe use the Close Friends filter) or share via DMs as things are happening. In fact, I am starting to see the trend of celebrities maintaining a public-facing profile and a private one seep into my circle as well. The main account is where you post to your extended circle of friends and family. The private one is for people who actually know you.

Though the main feed is turning into this song and dance, it was driven by community and social norms (and in part due to Instagram tweaking the feed). AI generated content feels like the final catalyst. It turns what was a social-norm problem into a technological one, as AI gets better at faking photos (obligatory call out: what is a photo?).

In a world where the main feed turns into this song and dance to craft a public image of yourself, the endgame is one where you can no longer tell what parts of the feed are real. With AI generated content it feels like we have crossed the Rubicon. The main feed is not just a curated image of yourself based on what you have done. It might be AI generated posts of you doing stuff, based on what you want to signal to your network.

What is real?

Adam Mosseri says rawness in photos -- like shaky photos and blurry candids -- is becoming a (temporary) signal for realness. He admits that going forward, the "who" behind the post will start to matter a lot more. His proposed solution includes surfacing more signals about the people behind the content, but also an institutional component to it that I have an issue with. Quoting from Adam,

Platforms like Instagram will do good work identifying AI content, but they'll get worse at it over time as AI gets better. It will be more practical to fingerprint real media than fake media. Camera manufacturers will cryptographically sign images that capture, creating a chain of custody.

Provenance when it comes to AI generated content is a vastly complicated problem. There are numerous attempts (C2PA, Content Authenticity Initiative) by companies and organizations around the world to establish a chain of trust. These solutions sound good in theory, using advanced cryptographic techniques to ensure security and privacy. Every proposal I have seen so far ends up distributing "trust" among the set of companies that are part of the coalition (see critiques from: RAND, ACLU). I don't know about you, but I would rather not give Instagram even more power in terms of determining what is real and what is not.

Besides, if the context around what is posted and who is posting it is going to matter, this by definition is not something that is compatible with a model of global truth. Context is highly specific to the intended audience and not something that can easily be conveyed to people outside that bubble. I think thinking of this problem as something that can be solved top-down is the wrong approach. It is more of a human problem, one that can be solved bottom-up.

The answer is usually human

Before I describe what I mean, the problem of identifying what is original is new to Instagram, but it has existed in other areas, such as luxury goods. Consider Rolex (or any similar luxury goods maker). Rolex watches have had dupes forever now, and despite many attempts by Rolex, they have not been successful in stopping people from making fakes. If anything, fakes have gotten much better over the years. And yet the market for Rolexes has not collapsed. There is research arguing that fakes can increase the payoff (in terms of social status) for owning the real thing, provided there is still a way to distinguish the real ones from the fakes.

Luxury companies have tried many technological solutions to combat this problem, going as far as launching a blockchain consortium called Aura (of course they did). Needless to say, none of these systems ended up being all that successful (the jury might still be out on Aura). The community adapted by moving from treating something like the logo itself as evidence to a network around it: who you bought it from, who will vouch for it, the story that comes with the watch, and so on. Instead of a global truth, we got a network of many small trust graphs.

To bring it back to Instagram, posting to the main feed is not something done in isolation. There's going to be conversations around it in the DMs. The community can figure out over time if the story lines up with what was posted on the main feed, creating many small trust graphs. As tempting as it might be, the answer to Instagram's problem should not be more technology and more centralized power, but a human one. Realness should not be a watermark, but a relationship.

If this resonated with you, say hi. I read every reply.

Power User Feature Requests for AI

2026-01-12T02:04:51.023951+00:00

Photo by Yuma Nozaki on Unsplash

I thought it might be fun to jot down a few features I wish AI chatbots had, things that fit better with the way I think and use them. I'm keeping this as a living scorecard to see what I get right vs. what actually ships. Also, it just so happens that towards the end of 2025, Your Year with ChatGPT told me I was in the top 1% of users by usage. Obviously, given this elite status, my thoughts are so valuable that they needed to be shared (/s in case it was not clear).

On Memory

Memory is one of my favorite features of ChatGPT, and I was surprised it took as long as it did for Claude (late 2025) to get it. It's powerful enough to create real lock in for users (more on this in a separate post). As of right now, I can think of at least a couple of ways memories could be improved.

Decay

Added: 2026-01-01

ChatGPT Plus and Pro users have an option to turn on automatic memory management to prioritize relevant memories and avoid hitting capacity. Anecdotally, however, ChatGPT ends up saving memories about me that are often time-limited and hasn't yet purged them.

I use ChatGPT for shopping from time to time (for instance, to test their new shopping research experience). This is one of the biggest categories where memories not having a sort of decay mechanism becomes an issue for me. For about a year now, ChatGPT has had saved in it things like "user is shopping for Lululemon alternatives". I'm not sure how many times this memory was accessed and where, but at the very least it is wasteful in the limited space for memory. This brings me to a second feature I'd like to have.

Usage Stats

Added: 2026-01-01

If OpenAI does want me to manage my own memory, I need to know more about how they are used. Are all the memories always appended to the beginning of each of my prompts? Probably not, but OpenAI does not disclose how exactly they are fed into the model, and at what stage. The only thing I could find from the documentation is,

Like custom instructions, saved memories are part of the context ChatGPT uses to generate a response. Unless you delete them, saved memories are always considered in future responses.

In fact, only when I was writing this post did I discover that apparently ChatGPT now puts some memories at the top of its mind,

To decide which memories stay top of mind, ChatGPT considers factors such as how recent a detail is and how often you talk about a topic.

How often do those 'top of mind' memories actually get used? I would at least like to know what memories were accessed when answering each of my queries. I had once asked ChatGPT to be concise/direct with me when discussing technical details of a research paper, and it remembered that as "user wants me to be concise/direct" without any context of when. At a later point, when I was asking it questions about a topic to learn more about it, the responses were oddly short. It took me a second to realize that it was this past memory impacting the result and once I deleted it, it went back to normal. There were probably more instances like this, but the point is, I don't want to play a guessing game of what could be influencing my current conversation.

Memories that were accessed could be shown next to the response like citations, or under the menu with the thinking details of the model. Overall usage stats such as the number of times a certain memory was accessed should be displayed in the Manage Memories screen.

I would also like to have memories evolve with me over time. Consider something like my political stance doing a complete 180 after a period of time. Current documentation suggests that ChatGPT could do this on demand if I ask it update/delete a particular memory. Ideally, it should recognize the contradiction and replace the older one with the current one.

Context management

Added: 2026-01-01

This one's perhaps a bit more out there. Despite memories existing, context in a given chat is still clearly very important. The thing I like about ChatGPT over Claude (as of 2025) is that ChatGPT has basically never prompted me to start a new chat because it's running out of context in the current one whereas Claude had done this to me many times. However, given the importance of context, it's still useful to run different chat sessions about different topics.

The issue though is, if OpenAI wants get to a point where it's almost replacing Google Search, there's bound to be many ongoing, unfinished conversations happening at any given time. Maybe this is an issue specific to how I use ChatGPT, or maybe I'm a bit ahead of most people in this given my higher than average usage of ChatGPT. But, managing these conversations is turning into a bit of a chore for me. It feels like I have a large personal knowledge management (PKM) style system where instead of organizing notes into different folders or tags or whatever, I'm doing it with chats. I have to think about when to branch from the existing one into a new chat, a fresh chat, continue from an older one, decide which chats go into a "project", and so on.

Something that'd be cool is if ChatGPT would do all this for me. Just as it is supposedly deciding which memories to use for a given prompt, I want it to show me past messages about this particular topic, regardless of which conversation window they happened in. The ultimate version of this would be just a box that I would start typing in, and the context would be pulled automatically as it figures out what topic I'm talking about and to what end.

Are We Looking at the First iPhone Built for AI?

2026-01-10T07:26:28.077595+00:00

The iPhone design cycle seemed to have shifted from two years to three in the recent years. But with the 17 Pro, Apple has completely redesigned the chassis and used a new material, after a brief two year stint with titanium. Is this an intentional move to switch back to a two year design cycle as Ben Thompson speculates? Or did something else cause this change?

A bit of context for those that are not obsessively tracking internal details about iPhones (I am aware of my problem). Apple had gone from using stainless steel to titanium for their Pro phones in 2023, touting the many benefits of titanium. It was stronger and lighter than stainless steel. The good thing about titanium is that it allows Apple to finally resume its pursuit of making an ultra-thin phone, while maintaining enough tensile strength so that the phone does not bend easily. There was an initial bit of drama around the thermals of the 15 Pro resulting in a firmware update and a statement by Apple that it was not the titanium that was causing the issue. While true, reviews around the 15 Pro's performance under sustained workloads have been mixed (some stress tests favored 15 Pro while others showed more throttling). To Apple's credit, they tweaked the design with the 16 Pro, with up to 20% better sustained performance. I'm not saying titanium caused "bad" thermals, more that it is definitely the case is that aluminum is way better at spreading heat around the phone compared to titanium. The iPhone 17 Pro reflects this, with a claimed up to 40% better sustained performance vs the 16 Pro.

iPhone case designs are believed to be locked eighteen to twenty-four months before launch day. And we heard that Craig Federighi became interested in AI after playing with it over Christmas of 2022. If we assume that was around the time that Apple started taking AI seriously, it was already too late to change the design of the 16 Pro, which would've been locked between September 2022 and March 2023. It does give Apple ample time to design the 17 Pro from the ground up for AI.

My hypothesis is that titanium was a choice made before the AI hype reached a stratospheric level, and before Apple started to take it seriously. The choice at the time might have seemed fine. Apple sacrifices a bit of thermal headroom in exchange for being able to make the iPhone Air, and later the iPhone Fold (perhaps as soon as next year). AI forced a change in the design two years in, bringing us the first iPhone truly[1] built for Apple Intelligence[2].

The AI Compute Currency Problem

2026-01-10T07:26:20.572947+00:00

While strides are being made in model efficiency, model complexity is outpacing efficiency gains. Recent comments by industry leaders suggest that this is going to be the status quo for the time being. Newer models, with o3, o4-mini, 2.5 Pro are getting smarter, and handling increasingly complex tasks (Deep Research, agentic workflows, etc.). That means charging $20/mo even with usage limits does not cut it, because the limits would be too low for frontier models. Naturally, AI service providers are introducing higher limit tiers for their latest and greatest models. OpenAI and Anthropic seem to be taking two different approaches to this problem. In this post, I want to highlight some of the challenges in explaining "AI usage" to users. I think this model places trust on the service provider in a way that is kind of unique to AI. Thinking further out, I believe advanced cryptographic techniques could help mitigate these issues.

From Flat Fees to Usage Tiers

Let us recap how we got here.

What could have worked best was charging users a flat monthly fee for unlimited (maybe with soft limits such as with unlimited mobile data) queries. However, inference costs have exploded with "thinking" models and Deep Research capabilities. There is a clear need to charge users more in order to access these frontier models.

With the internet, users knew exactly what they paid for, e.g., $X for 100 GB of data and a bandwidth of up to 100 Mbps. Now, the users might not have had a good idea throughout the internet era whether that 100 GB would be enough for their needs. But they did have a good idea what their activities on the internet would consume. Casual browsing of Facebook and the likes can vary in usage, but it's possible to get a relatively accurate estimate. When it came to downloads, users knew exactly how much of their download limit would be consumed by this file because the size of the file is displayed beforehand. This estimate is also not one that is controlled by the internet service provider (ISP), meaning the users do not need to trust the ISP. Because of this clarity, there is an intentionality in the user choosing to spend this currency.

Bandwidth can and does fluctuate, and the ISPs technically only say "up to" 100 Mbps. But a user could measure the speed they were getting at different points in time, and hold the ISP accountable if it was frequently low. Besides, bandwidth issues usually impact users in a given radius so the onus is not entirely on each individual user.

Below is a comparison of current AI pricing tiers[1][2]:

Provider	Plan	Cost	Usage Limit	Notes
OpenAI	ChatGPT Pro	$200/mo	Unlimited*	All reasoning models, GPT-4o, GPT-4.1
Anthropic	Pro	$20/mo	5x Free usage	Reasoning models
Anthropic	Max (5×)	$100/mo	5× Pro usage	Expanded access vs Pro tier
Anthropic	Max (20×)	$200/mo	20× Pro usage	Maximum flexibility
GitHub	Copilot Pro+	$39/mo	1,500 premium requests/mo	Additional at $0.04/request

OpenAI's and GitHub's solution is to introduce a tier with a higher rate limit, using a query as their unit of measurement, and putting a concrete number on the number of queries offered for a given time period. Anthropic uses a more nebulous term of usage.

The pricing dynamics here are analogous to something like what internet service providers do. As bandwidth got cheaper and infrastructure improved with things like fibre optic cables, higher download limits and browsing speeds were offered to customers at no (or marginally higher) extra cost, thereby increasing the value per dollar customers were receiving. Normally, the story ends here and we could say AI model providers will follow a similar pattern. I think there an interesting catch here, one that might not be so straightforward to tackle.

The Fluidity of Compute Units

With AI queries, compute costs vary wildly by prompt, making any "unit" hard to predict. Maybe some day we'll get there. Given this, AI companies are considering alternatives with giving the users a set of "compute units" every month. This seems like a good idea at first. But from the user's point of view, they still have no idea how many compute units their query will take. Estimating compute units a priori is highly nontrivial. So we end up in a place where the company is going to provide users with credits, in a wallet that is managed by the company. When a user clicks “submit” to send a query, they must trust the company to charge them correctly. Even if cost estimates are provided, those figures are still proprietary numbers generated by the provider.

Thinking adversarially for a second, an AI company would see that a user is very high in their usage and decide to swap out the model for something more efficient. Given that none of the models are public, it would be very hard for the user to figure this out (assuming the models are not wildly different). This might sound like a conspiracy theory, but it's meant as a thought exercise to figure out where the gaps are so that we can build better systems that require lesser trust.

Third-party audits do not close the trust gap, because an adversarial model swap as described above is still possible. Having the users inspect model weights is not desirable as they are the companies' IP a lot of the times. Fortunately, cryptography offers a fix: A Zero-Knowledge Proof (ZKP) is a tool that stops an AI provider from launching this sort of model swap attack.

In the interest of keeping things light, here's the bird's eye view of how a zero-knowledge proof for this scenario could look like:

Commitment: The AI provider publishes a “commitment” to its model weights (and perhaps some metadata). This commitment, like a cryptographic hash, reveals nothing about the model itself.
Query Proof: Each time the user submits a prompt, the provider runs the model and generates a proof that the response came from the committed weights.
Verification: The user checks the proof against the commitment, which outputs accept or reject.

If the provider cheated and swapped the model out for a different one, we have a guarantee that they will not be able to generate a proof that verifies correctly. If you want to learn more about ZKPs, here's a fantastic video by one of the pioneers in the field.

Note: An eagle-eyed reader might ask that it is still possible for the provider to cheat (say they offer o4 as a service, but choose to commit to o3 from the get-go). This is indeed possible! Establishing that the model being committed to is in fact o4 is not possible in a vacuum. We will have to rely on additional information, such as a particular training dataset, or specific training code, or benchmark results.

Although ZKPs have flourished in blockchain over the past decade, their application to LLM inference remains experimental, with many research challenges to overcome. Yet, with the AI inference market already a multibillion-dollar industry, developing transparent billing is imperative. Deployment of trust-minimizing tools such as ZKPs will lay the groundwork for transparent, auditable AI services as this market continues to scale.

Thoughts on Personalized Podcasts

2026-01-10T07:25:52.852513+00:00

Podcasts have exploded in popularity (About half of Americans have listened to a podcast in the past year). What started as a niche medium has become a platform for everyone from celebrities to politicians to business leaders. We're even seeing podcasts dig into deep technical content, including scientific papers. Recently, we have seen tools such as NotebookLM and Eleven Labs Reader that let you make podcast episodes out of pieces of text, using AI.

Despite this innovation, the core experience of listening to a podcast episode, regardless of how it was generated, remains static. Once the hosts finish recording, editing, and publishing, you get the same show as every other listener. The episode does not adapt to your interests. For example, if you were a long time listener of a podcast, say ATP FM, over time, you get to know the hosts and their personalities. After a certain point, some of the content in a weekly episode might be repetitive to you. The host might be sharing their opinions about a new product that just came out or discussing some news. But because they are catering to new listeners, they understandably have to preface the opinion with some amount of context. With podcasts today, you have two choices – either you listen to it as a whole, or skip the entire chapter (assuming the hosts have enabled chapters) about the product, neither of which is desirable.

Now, imagine an advanced AI system that reconfigures each podcast on the fly — sculpting, editing, and remixing the actual audio in real time to suit your preferences. This goes beyond asking your voice assistant to play/pause/skip chapter, the podcast itself becomes an interactive experience, shaped by your prompts. This post is my thoughts of how this deep integration could work, and what benefits it could provide.

Current Use of AI in Podcasts

Two big AI use-cases have gained attention so far:

Automatic Generation: Tools like NotebookLM generate a fresh podcast-style dialogue from a document provided. Simon Willison has a few examples in this post.
Transcript Services: Apple Podcasts recently introduced interactive transcripts, letting you scroll through text and tap any sentence to jump right to that point in the audio. To be fair, this is a great feature. If you remember a specific moment in a show, you can skim the transcript and tap to play from there rather than manually hunting with the scrub bar. The downside is that it still requires you to look at your phone and interact with it.

An AI-integrated Podcast Experience

Imagine a podcast player that is always listening for a wake word/phrase, similar to how your phone listens for "(Hey) Siri". In this deeply integrated system, the AI could remix the content itself on your command in a couple of different ways. For now, I'm assuming the hosts give consent for this player to modify their content in a few limited ways. This could allow,

Semantically Skipping and Compressing Segments: When a host repeats information that you already know because you're a long time listener, the AI could seamlessly edit that section out, stitching the audio back together (using its own voice) so it sounds natural.
On-the-Fly Summaries: I imagine most people do not consume podcasts that have long episodes (such as Acquired by Ben Gilbert and David Rosenthal, with episodes that are a whopping 5 hours long) in one session. When you come back into the episode and you're 3 hours in, you may have forgotten what happened until this point. AI could generate a summary when prompted, letting you jump back into the episode right where you left off. Moreover, if in the next bit of the podcast it requires recalling information that was mentioned during the first hour, the AI would know this and could make sure to include this bit into the summary. This would not be possible if one were to manually rewind the podcast by say 30 minutes and start listening. I can think of a few genres where this could apply, such as podcasts covering elaborate historical events (The Rest is History often does multi-part episodes).
Cut to the Chase: The flip side to on-the-fly summaries could be when you have only, say 10 minutes to listen, but there's 45 minutes left in the podcast. Maybe you got bored of this particular story, or maybe the podcast has a bit too much filler and you want to know the ending to the story. This podcast player could gather the important points from the remaining episode(s), and piece together a clip that concludes in the desired amount of time.

Concerns

The obvious one is that the creators may not be okay with AI rearranging or splicing their content in ways they did not intend. Even if they do give consent to this tool, I feel like the AI-generated bits would probably need to be marked clearly, by using a distinctive voice instead of mimicking the hosts' voices, or by announcing before/after there is an AI-generated clip.

Assuming this is carefully implemented, I do believe it could add value to the listening experience by providing absolute personalization to the user, without taking away the creative voice of the host(s).

Why Apple Intelligence Ads Miss the Mark

2026-01-10T07:25:34.047510+00:00

Apple's ads have a storied history of excellent branding and messaging, which made their latest Apple Intelligence ads feel strangely off to me. These ads, when contrasted with their older ones, reveal a kind of cognitive dissonance. Apple, a company known for its deliberate and thoughtful storytelling, seems to be rushing into this huge category – a decades-long arc according to them. This haste might explain why the ads feel disjointed, and lack the cohesion of their predecessors. Let's look at the ads to try and unpack this uneasiness.

Apple's Ad Legacy

Apple's ads have near-mythical status. From the iconic 1984 ad onward, they've consistently stood out – not just for their technological innovation, but for their storytelling. It was the way Apple talked about its products. Broadly, I can categorize the ads into two buckets – ones that evoke an emotional response, and ones that talk about their products. There are many examples and articles of people more capable than me analyzing Apple's branding, so I will mention just a couple of examples in each category to make my point.

Photos Every Day: This ad was released with the iPhone 5, highlights everyday moments rather than focusing on new features[1].
Behind the Mac: A series of ads that focused on the people behind the Mac, ending with:

Make something wonderful

Empowerment: A couple of ads they aired that ended with the emotionally resounding quote:

"You're more powerful than you think"

Steve Jobs famously described Apple's DNA as technology married with liberal arts, and that philosophy is evident in their ads. These were not products that you used just to do your job, these were tools of creative expression. When Apple is not going directly for the emotional response and actually showcasing the products, the framing is always about what you can do with them. The tool itself, while great, is fundamentally incomplete without you, the person, putting it to use. There is an earnestness in this messaging.

"Give people wonderful tools, and they'll do wonderful things"

More recently, this ad showcased the redesigned M1 Pro and M1 Max MacBook Pros. It's one of my favorite ads from them in recent memory, that talks about the laptops as beasts and monsters, concluding with:

"What have we done, and more importantly, what will you do?"

Throughout the ad they show very demanding tasks being handled (presumably) with ease by their latest creations, but the most important thing is not the tool, it is what you will create with the tool.

A Contrast in Tone

The latest ads introducing Apple Intelligence feel jarringly out of sync with this rich history. There seems to be some kind of cognitive dissonance – Apple's usual care and intentional storytelling seem to be missing. Instead, it feels like the company is rushing into this huge, decades-long category without as much thought. Let's look at a few examples:

The Lawyer: Perhaps the most egregious one (and the one I have been seeing online the most), is the lawyer who has not done their job. We do not get any context about why this person was not able to prepare to go through a document about Ed Case Construction. In fact, it seems like they were surprised by the fact that they had to lead this discussion, and of course, Apple Intelligence comes to the rescue.
The Desk Worker: Another example is the person wasting time at their desk. The person types out an informal email with little thought behind it. But Apple Intelligence steps in to make the email look professional to their boss, letting the person go back to wasting time with rubber bands. Don't you love it when an under-performer can get away (now with AI) by pretending that they are working hard? The charitable interpretation would be that this person is a genius who does not care about writing "proper" emails, but that's not what the ad is saying. The email they write is only saying that project needs more pizzazz.
The Mom Who Forgot a Gift: This ad features a mom who forgot to get her husband a gift. Using AI, she creates a "memory movie" to save face, and is smug about it. Why was she not able to procure a gift in time?
Bella Ramsey: With the Bella Ramsey ads, it is even more bizarre. One of the ads shows them pretending to like a script that they haven't read. What's the plan here? Will they never read the script or will they eventually read it and potentially backtrack on their word? Once again, we are not told why the person could not do what was expected of them.

In all these ads, the common thread is people saving face at work or with family by using AI – not because they're empowered, but because they were lazy, dishonest, or careless. Worse, the tone often comes across as smug. To put it simply, these don't strike me as aspirational ads.

Instead of framing AI as a way to cut corners or save face, these ads could have highlighted how Apple Intelligence empowers people in meaningful ways. Imagine showing an older person using AI to simplify complex tasks or learn something new, breaking barriers that come with not being tech-savvy. Or a spotlight on the accessibility community, where AI-driven tools like voice dictation and translation are game-changers, enabling more independence and connection. For non-native English speakers, AI could be shown as a tool to build confidence – helping them excel at work or share their thoughts online without fear of judgment. These narratives would not only resonate more positively but also align with Apple's legacy of empowering users to achieve their full potential.

Missed Opportunities

Apple's legacy is rooted in the belief that great tools inspire great work. These ads, however, miss the chance to highlight what makes Apple Intelligence unique. Rather than showcasing AI as a tool for empowerment, creativity, and problem-solving, they lean on narratives of shortcuts and superficial fixes. By reframing the narrative to focus on how AI helps users overcome challenges, Apple could align with its legacy of empowering people to achieve more. Thoughtful storytelling could restore the emotional resonance and aspirational tone that made their past campaigns iconic.

Where is the Puck Skating to?

2026-01-10T07:13:03.971973+00:00

Last month, Meta showed off Orion, which I can only describe as a concept car version of augmented reality (AR) glasses. Orion is a pair of AR glasses that come with a compute puck, that is wirelessly connected to the glasses, and a neural wristband that tracks hand movements.

Orion, or some future version of it, is purported to be the device/platform that will replace smartphones. The idea of strapping a computer to your face and seeing the world through it not new, and has been attempted in many forms in the past, most recently, Apple Vision Pro. What past attempts have in common, though, is that they are generally heavy, which makes them unsuitable (depending on how much discomfort you're willing to tolerate) for long-duration use. They also suffer from getting too hot, having a lower than acceptable battery life, and a myriad of other challenges. Orion, on the other hand, is designed to be much lighter in weight and will not get as hot, given that it shifts the heavier components of the compute to a dedicated puck. As a result, they will be comfortable to wear for longer periods of time, potentially even outside (without looking like a dork).

It is a concept car because it is almost certainly not going to ship as is, owing to the production cost of at least $10,000. However, I think the reveal, and the subsequent interviews by Meta's leadership, gave us some tidbits about Orion and how Meta is thinking about positioning it going into the next platform for computing.

Meta's Bet

So why did Meta show this concept car? The popular theory seems to be that it was because Snap was planning to show its AR glasses product, putting pressure on Meta. Especially considering Meta sunk around $75 billion, a number that does not seem to be slowing down, into its Reality Labs division, there is a ton of pressure from investors to show something tangible for what this money had been up to. That makes sense, but what does it mean for the first AR glasses that will actually ship? I think Meta is sort of uniquely positioned to be able to showcase such a demo compared to Apple and Google. Meta's strategy appears to focus on using as many off-the-shelf components as possible (other than the custom lenses, of course) to get the product to market quickly. An approach that is diametrically opposed to what Apple wants to do, control the entire stack.

Both Google, and Apple, probably have prototypes similar to Orion, in that they are AR glasses with compute off loaded to a different component. However, Google has already made an attempt in this space with Google Glass. We all remember how that went. I don't think they are going to do a lot of convincing by showing a prototype, without even a dev-kit for developers to start building for (you might be thinking that devs could build for the Quest line and port their apps to Orion. Alas, it does not seem like it is going to be that simple). The stakes feel much higher, and given the non-existence of an AR glasses market, I'm not sure what is there to be gained by showing a prototype years in advance.

Apple's Move

What about Apple? Sure, the company does not like to show products/features before they are ready for launch (a sentiment that has at least one exception, see the rollout of Apple Intelligence), but it's also because Apple cannot show/ship such a product. The obvious question would arise: why not use the iPhone as the compute puck? The iPhone has a ton of headroom when it comes to its processor. It feels natural to want to have the iPhone, that is already always in your pocket when you're out and about, to act as the puck for the glasses.

Typically, the fact that Apple can leverage its platform and use APIs that third-party developers cannot, is a massive, if not unfair, advantage. Apple Watch, for instance, is something that only Apple can build. Assuming that Meta is actually "close" to shipping Orion, I wonder if this expectation of being able to use the iPhone as the compute puck puts Apple in a thermal corner. Bos said that the initial prototype of Orion was running so hot that they had to cool it with chilled soda cans. He also said that the current version runs out of battery (~ 2h of usage) before it hits the thermal limit, which means making the battery beefier in order to get more battery life is not straightforward. Granted, Meta is using off-the-shelf parts, while Apple has some of the best chips when it comes to their power draw vs the amount of performance they can extract. But keep in mind, that the current generation of iPhones do a whole host of tasks today for people, and survive about a day or more doing so. It seems that adding AR-related computation to the mix, regardless of how efficient the chips are, will require the iPhone to get heavier.

The Case for iPhone Ultra

The iPhone is a device that is insanely popular, selling over 200 million units per year. It is also the primary computing device for a lot of people around the world, so any design changes will impact tens of millions of people. Given the rumors of the iPhone 17 lineup adding a new "Slim" or "Air" model to the mix, I wonder if Apple is already preparing for a future rebranding of the lineup. I think the term "Pro" has become too overloaded, so, I'm not a fan of further complicating this. Apple should keep the current Pro line, and have it mean pro photography and videography. They should add a tier above this, iPhone Ultra. A model such as iPhone Ultra has long been rumored (first in 2022, and again in 2023), but it did not make sense as to why an iPhone with a lot more power than the Pro line needed to exist. I cannot think of a single person that thinks that the current iPhones are compute-restricted. With the advent of more demanding compute for AI and AR, I think it is the right time for this model. iPhone Ultra can be heavier and pack the most powerful chip (maybe exclusive to it), and be the only phone to support AR capabilities in the initial years. As AR-related compute gets cheaper, compatibility and features could trickle down to the regular and Pro lines, while Ultra still gets the newest and best features, similar to what is currently happening with the Pro line. If the glasses are anything like the Vision Pro, in that they are extremely expensive at launch, and only target the higher end of the market, I could see Apple limiting their compatibility to the most powerful iPhone model to start with.

Here's my prediction for how the lineup will evolve over the next few years.

iPhone Air: Two screen sizes, and is the default phone most people should buy.
iPhone Pro: Same as now, with cutting edge photography and video features, while maintaining a reasonable balance of weight and battery life.
iPhone Ultra: The highest end, no compromises iPhone, that gets the most powerful A-series chip, and is the only one initially to support the AR product.

The Road Forward

Whether or not Meta's intent was to pressure Apple with Orion, it has undoubtedly set a new benchmark for future AR devices. If Meta's vision holds, AR might just break free from its current limitations, propelling the industry into a new era of personal computing--one in which the "puck" defines the play.