Plutonic Rainbows

The original Wonder model turned everything into watercolour. Skin looked airbrushed, fabric lost its weave, and faces — especially small ones in group shots — came out waxy. I ran a batch of family photos from the early 2000s through it last year and the results were unusable. Everyone looked like they'd been smoothed in Photoshop by someone who'd just discovered the blur tool. I stopped using it.

Wonder 2 is different. Topaz finally acknowledged what users had been complaining about: the over-processing. The new model dials back the artificial sharpening and actually preserves texture. Hair looks like hair. Skin has pores. Fabric keeps its weave instead of melting into some vague suggestion of cloth.

I tried the same batch again. The difference is significant. Not perfect — I'm not sure any upscaler handles compression artefacts from early digital cameras gracefully — but the faces are recognisably human now. That waxy sheen is gone.

The catch: it's cloud-only. Topaz says the computational demands are too heavy for local processing, which is probably true, but it means you're uploading your images to their servers. For personal photos, I don't love that. For client work, some people won't accept it at all. I understand the technical reasoning — these models are enormous and running them locally would require hardware most photographers don't have — but it still feels like a step backward in terms of control.

There's also the subscription question. Topaz moved to a subscription model last year, which rubbed a lot of long-time users the wrong way. I'm not going to relitigate that argument here. The software works or it doesn't. For me, Wonder 2 works well enough that I've started using Topaz again after months of avoiding it.

What I actually wanted from an upscaler was always simple: make the image bigger without making it worse. Don't add detail that wasn't there. Don't smooth things that should be rough. Don't sharpen edges until they ring. Just scale it up and preserve what exists. Wonder 2 gets closer to that than anything else I've tried. It's not magic — you can't turn a 200x300 thumbnail into a printable image — but for moderate upscaling of decent source material, it does the job without leaving obvious fingerprints.

The Fidelity update that shipped alongside Wonder 2 includes a bunch of other models too. Recover 3 for softer results, some video stuff I haven't tested. But Wonder 2 is the one that matters to me. It's the reason I'm writing this instead of just ignoring another Topaz release.

Sources:

• Wonder Model Documentation - Topaz Labs

• Fidelity Update - January 2026 - Topaz Labs

I found my copy of Wintering Out in a box I hadn't opened since moving flats in 2019. The spine was cracked in three places, the pages yellowed in that particular way paperbacks get when they've lived in damp rooms. I'd forgotten I owned it.

Seamus Heaney published this collection in 1972, the year of Bloody Sunday, though you wouldn't necessarily know that from reading it. The violence is there — it's always there in his work from this period — but it comes at you sideways, through bog bodies and place names and the particular way rain sounds on different surfaces. He doesn't write about soldiers. He writes about the word "Anahorish" and how it feels in the mouth.

That indirection annoyed me when I first read him. I was twenty-two, impatient, wanted poets to say what they meant. Why all this business about etymology and townlands when people were dying? It felt like evasion. I put the book aside and didn't pick it up again for over a decade.

I was wrong.

The poems in Wintering Out aren't avoiding the Troubles — they're excavating the ground beneath them. When Heaney writes about the word "Broagh" and how the "gh" sound at the end is unpronounceable for English speakers, he's writing about borders. About who belongs and who doesn't. About how language itself draws lines that bodies later bleed across. This isn't evasion. It's archaeology.

"The Tollund Man" is the poem everyone talks about, and for good reason. A body preserved in a Danish bog for two thousand years, sacrificed to some fertility goddess, becomes a lens for looking at sectarian murder in Belfast. The logic shouldn't work. Denmark isn't Ireland, and ritual killing isn't the same as a car bomb. But Heaney makes the connection feel inevitable rather than forced. Both are forms of tribal violence. Both leave bodies in the earth.

I keep thinking about my grandmother's accent. She grew up in Tyrone, moved to England in the fifties, and by the time I knew her, her voice had become something strange — not quite Irish, not quite English, caught between. She pronounced certain words in ways I've never heard anyone else pronounce them. When she died, those pronunciations died with her. That's what Heaney is getting at in poems like "Traditions" — language as inheritance, but also language as loss. Every generation forgets something.

He wrote much of this collection while on sabbatical at Berkeley in 1971. California, of all places. He said the distance loosened something in his form, made the quatrains more relaxed. I find that odd — that you'd need to go to the other side of the world to write about the six inches of soil beneath your childhood home. But maybe that's exactly right. Too close and you can't see it. The Irish memory bank, he called it. Something you can only access from far away.

The shorter poems frustrate me. "Servant Boy" and "Limbo" feel slight next to the longer pieces, sketches rather than finished work. Critics at the time complained that Heaney wasn't addressing the violence directly enough, and I understand the impulse even if I think they were wrong. When your country is tearing itself apart, poems about place names can feel like fiddling while Rome burns.

But that misses what Heaney understood: the violence didn't come from nowhere. It grew from centuries of contested ground, contested language, contested memory. You can't address the present without digging into what made it. The bog preserves everything — bodies, butter, wooden trackways. It's a kind of memory that doesn't forget. Heaney keeps returning to that image because it's doing real work for him. The past isn't past. It's right there, just under the surface, waiting to be cut into.

My copy still smells faintly of the flat I lived in during my twenties. Damp plaster, radiator dust, the particular staleness of single-glazed windows. I don't know why I kept it through three moves. Most of my books from that period went to charity shops or got left on trains. This one survived.

Harold Bloom called Heaney's voice "keyed and pitched unlike any other significant poet at work in the language anywhere." That's the kind of sentence critics write when they can't quite explain what they mean. But he's not wrong. There's something in the sound of these poems — the vowels, the consonant clusters, the way lines break mid-phrase — that doesn't sound like anyone else. You can recognise a Heaney poem by its music before you've parsed a single image.

The collection ends with "Westering," a poem about California, about being far from home, about the direction of travel that the word itself implies. West into the unknown. West into the sunset. West into America, where so many Irish ended up. It's not a conclusion exactly — more of a trailing off, a question left hanging. Where do you go when the ground you came from is contested? What happens to memory when it crosses an ocean?

I've started rereading the bog poems aloud. There's no other way to get them right. The sounds matter in a way that silent reading can't capture. "The Tollund Man" in particular needs to be spoken — the way "Tollund" itself echoes and dulls, the flat vowels of "the mild pods of his eyelids." Heaney was obsessed with how words feel in the body. The tongue, the teeth, the soft palate. Poetry as a physical act.

I still don't love all of it. Some of the mythological pieces feel like exercises. But the best poems here — "Anahorish," "Broagh," "The Tollund Man," "Gifts of Rain" — do something I can't quite name. They make the familiar strange and the strange familiar. They make language itself feel like archaeology, like digging.

Sources:

• Wintering Out - The Estate of Seamus Heaney

• Wintering Out - Wikipedia

A static site generator seems like the safest possible software. No database. No user authentication. No server-side processing. Markdown goes in, HTML comes out. What could go wrong?

Quite a lot, as it turns out. I spent part of today running a systematic security audit on the Python code that builds this blog, and the results were sobering. Forty-five issues across six severity categories, ranging from XSS vulnerabilities in the search functionality to race conditions in file operations. The code has been running in production for months. Every issue had been hiding in plain sight.

The most serious problem was a classic cross-site scripting vulnerability. The search feature highlights matching text by inserting content directly into the DOM via innerHTML — a pattern that the OWASP DOM-based XSS Prevention Cheat Sheet explicitly warns against. If a malicious post title contained script tags, the search results would execute them. The fix was straightforward: escape HTML entities before highlighting. However, the vulnerability should never have existed in the first place.

Race conditions appeared in several places. The build script called os.listdir() twice for duplicate detection — once to build a normalisation map, again to process files. Between those calls, the filesystem could change. The cache file for URL shortening used a naive write pattern that could corrupt data during concurrent builds. The asset copying routine deleted the entire output directory before recreating it, creating a window where the site would be unavailable. Each fix required thinking carefully about atomicity and the assumptions that file operations make about a static world.

Date arithmetic revealed a subtler class of bug. The time-based filtering used timedelta(days=months * 30) to calculate cutoff dates — a calculation that drifts by five or six days over a year. Posts from exactly twelve months ago might or might not appear depending on which months fell within the range. The dateutil library provides relativedelta specifically to handle calendar arithmetic correctly. There was no excuse for not using it.

Path traversal prevention was missing entirely. A crafted slug containing ../ could write files outside the output directory. Input validation existed for character sanitisation but not for structural attacks. The oversight was embarrassing.

What strikes me most is that this code was written with agentic coding tools — the same tools that are supposed to bring senior-level expertise to every developer. The tools generated working code that passed all tests and produced correct output. They did not generate secure code. They did not flag the race conditions or the XSS vulnerability or the date arithmetic error. The code worked, which is a different thing from the code being right.

This reinforces something I have been thinking about: no system can verify its own blind spots. The AI that helped write the code could not see what it had missed. The developer reviewing the output — me — did not catch the issues either. Only a deliberate, adversarial audit with a checklist of known vulnerability patterns found what was hiding in plain sight.

The fixes took a few hours. The lesson will last longer. Safe-looking code is not the same as safe code. Static sites are not immune to security issues. And the tools that accelerate development do not eliminate the need for the slow, careful work of verification.

Sources:

• DOM based XSS Prevention Cheat Sheet - OWASP

• Cross Site Scripting Prevention Cheat Sheet - OWASP

• dateutil - relativedelta - dateutil documentation

Spent the morning performing maintenance on my Roon music library—removing extended attributes from 26,948 audio files. macOS applies com.apple.quarantine and com.apple.provenance attributes to downloaded files as security measures, but these can cause file access issues with Roon's music server. The cleanup was straightforward using xattr -dr commands to recursively remove the problematic attributes. Tested playback afterward with Oneohtrix Point Never's Tranquilizer—no audio quality degradation, exactly as expected. Extended attributes are filesystem metadata stored separately from audio data itself. The files remain unchanged; only the invisible annotations have been stripped away. The library now runs cleaner without these unnecessary flags interfering with normal operation.

Waswaas and The Dullard Sage have constructed something genuinely strange with Und Ewig Ist — an album that feels less composed than discovered. The eight tracks unfold as what the artists call "low frequency field recording excursions," and that description captures the essential character of the work. This is music that moves through environments rather than describing them.

The collaboration spans territory that defies easy categorisation. Tags on Bandcamp list Sufi, cosmic black metal, drone, and modern classical as reference points. However, none of these labels fully accounts for what happens across tracks like "Disorders of Consciousness" or "Datacombs." The low end dominates — rumbling bass frequencies that seem to emanate from the earth itself — while field recordings add texture and occasional brightness to the murky depths.

The dedication to Maryanne Amacher feels particularly apt. Amacher spent decades exploring how sound interacts with physical space and the listening body. Waswaas and The Dullard Sage pursue a similar investigation, creating music that rewards deep listening and physical presence. The cassette edition sold out quickly, though digital versions remain available with bonus tracks.

Sources:

• Und Ewig Ist - Bandcamp

• The Maryanne Amacher Foundation - Blank Forms

Sultan Pasha's decision to reformulate Thebes as an alcohol-based Extrait de Parfum marks a significant departure from the oil-based attar tradition that established his reputation. The original Thebes Grade 1 arrived in 2016 as an homage to Guerlain's discontinued Djedi — a fragrance so evocative that Sultan Pasha described it as the only perfume that had brought him close to tears. After months of painstaking recreation, he captured that spectral atmosphere in oil form, creating what became his signature composition.

Nearly a decade later, the 2025 release transforms that intimate, skin-hugging attar into something altogether different. Working alongside Christian Carbonnel under the new Sultan Pasha Perfumes label, the reformulation explores what happens when you translate oil's density and warmth into alcohol's volatility and projection. The result maintains the core narrative — an ancient Egyptian tomb, the boundary between life and death — while fundamentally altering how that story unfolds in space and time.

The composition itself reads like an exercise in controlled opposition. Bright aldehydes and a white floral bouquet of jasmine, muguet, and rose sit against somber, earthy vetiver and the distinctive chalk-like texture of genuine orris butter. Reviewers consistently note this tension: the fragrance is simultaneously luminous and gloomy, uplifting and ritualistic. One detailed review describes waves of heady florals alternating with leather and salty ambergris, creating an animalic, fatty quality that feels deliberately unsettling.

This approach differs markedly from the attar version's intimate revelation. Alcohol-based perfumes diffuse outward, creating a more public presence that transforms the wearer's relationship to the scent. Where the oil version whispered ancient secrets directly to the skin, the Extrait broadcasts them into the surrounding air. The projection reportedly remains strong for the first two hours before settling closer to the body, with longevity hovering around five hours — a relatively modest performance for an Extrait concentration, suggesting the formula prioritizes complexity over sheer endurance.

The move to alcohol represents more than technical reformulation. Sultan Pasha built his reputation through traditional attar craftsmanship, a method that demands patience and precision but limits commercial reach. Attars require direct application, careful storage, and an understanding that comes through experience. By creating alcohol-based versions of his most celebrated works, he opens a door to audiences who might find oil-based perfumes too unfamiliar or demanding.

However, this accessibility comes with artistic risks. The attar community values the medium's contemplative nature — its quiet intensity, its refusal to announce itself beyond the wearer's personal space. Translating that aesthetic into alcohol requires careful calibration to avoid losing what made the original compelling. Based on early responses, Thebes manages this balance by maintaining its strange, funereal atmosphere even as it reaches farther from the skin. The reformulation amplifies certain aspects — particularly the aldehydic brightness and floral lift — while preserving the dusty, ritualistic core that defines the concept.

Sample sets became available for preorder through January 2026, a deliberate strategy that allows serious enthusiasts to experience the full lineup before committing to full bottles. This approach respects the considered, exploratory mindset that characterizes niche perfume appreciation. These are not fragrances designed for casual purchase; they demand time, attention, and a willingness to sit with discomfort. The animalic qualities alone ensure this remains far from mainstream tastes.

What strikes me most about this release is its timing. The niche perfume market has become increasingly crowded, with countless brands claiming artisanal credentials while churning out derivative compositions. Sultan Pasha's move to alcohol could be read as capitulation to commercial pressure, but the execution suggests otherwise. By maintaining Extrait concentration and preserving the challenging, unconventional character of the original work, he signals that accessibility need not mean simplification.

The question now becomes whether this model succeeds — whether audiences accustomed to attars will embrace the reformulations, and whether those new to Sultan Pasha's work will appreciate what makes these fragrances distinctive. Thebes tests that proposition directly, offering a scent that refuses conventional pleasantness in favor of atmospheric depth. It remains to be seen whether the broader market rewards that uncompromising vision or whether the commercial realities of alcohol-based production eventually push toward safer ground.

For now, Thebes in Extrait form exists as a fascinating experiment in translation, asking how much of an attar's soul survives the journey from oil to alcohol. The early evidence suggests more than you might expect, though undoubtedly something irretrievable remains bound to the original medium. What emerges is neither superior nor inferior, but genuinely different — a parallel interpretation that extends the concept rather than simply reproducing it in another format.

Sources:

• Thebes Sultan Pasha Perfumes - Fragrantica

• Sultan Pasha Perfumes Thebes: Read the Full Review - Perfumeaholic

• Thebes Grade 1 by Sultan Pasha Attars - Fragrantica

• Sultan Pasha Presents Four New Extrait de Parfum - Fragrantica

Philip Johnson's Glass House served as more than a venue for Ryuichi Sakamoto and Alva Noto's 2016 improvisation — it became the instrument itself. Contact microphones placed on the glass walls captured vibrations, transforming the structure into a resonant body. The resulting album, released in 2018, documents a single 37-minute performance where architectural space and electronic processing merge.

The collaboration marked their first live work together since Sakamoto's cancer diagnosis in 2014. Both artists approached the session with minimal rehearsal, spending only one day preparing before the recording. Sakamoto brought a keyboard and glass singing bowls, while Nicolai contributed his characteristic digital processing. However, the true voice emerged from the building itself.

Yayoi Kusama's installation — Dots Obsession: Alive, Seeking for Eternal Hope — occupied the space during the performance. Sakamoto described looking through the glass walls at the landscape while surrounded by Kusama's dots as "a strange mixture of natural, nature, and artificial things, art." That tension between organic and synthetic pervades the recording. Nicolai's glitches and static rest against Sakamoto's melodic fragments, neither dominating.

The Glass House's transparent walls offered ideal conditions for an experiment in architectural acoustics. Therefore, what emerged was not merely electronic music performed in a space, but music generated from the space itself — a document of place as much as performance.

Sources:

• Glass - Alva Noto + Ryuichi Sakamoto - NOTON

• Glass (composition) - Wikipedia - Wikipedia

• Alva Noto & Ryuichi Sakamoto – Glass album review - Louder Sound

I've been watching Anthropic's release cadence closely over the past year, and something has changed. The company that brought us Claude Opus 4.5 in November 2025 has gone conspicuously quiet. No leaks, no benchmarks teased on Twitter, no cryptic blog posts hinting at breakthrough capabilities. Just silence. That silence, however, tells me more about their next model than any press release could.

The industry has trained us to expect a particular rhythm. OpenAI drops a new model every few months, each one incrementally better than the last. Google races to catch up. The smaller labs scramble to carve out niches. We've come to expect this treadmill of marginal improvements, each accompanied by breathless claims of revolutionary progress. Anthropic participated in this race for a while, but I believe they're stepping off it deliberately.

Consider what we know about their philosophy. The company was founded explicitly on the principle that AI safety cannot be an afterthought. Their Constitutional AI approach isn't marketing — it's baked into their training methodology. They've published papers on interpretability that most companies wouldn't touch because they reveal uncomfortable truths about what we don't understand. This isn't a company optimizing for Twitter engagement or shareholder updates.

Therefore, when I look at the gap between Opus 4.5 and whatever comes next, I don't see delay. I see intentionality. I believe Anthropic is rebuilding their development process from the ground up, and the next Sonnet model will reflect that fundamental shift.

The current generation of frontier models, including Anthropic's own, share a common weakness. We can measure their performance on benchmarks, but we struggle to predict their behavior in edge cases. They excel at standard tasks while occasionally producing outputs that reveal concerning blind spots. This unpredictability isn't just an engineering challenge — it's an existential risk that scales with capability. Additionally, the compute required to train these models has grown exponentially, while the improvements have become increasingly incremental.

I suspect Anthropic recognized this pattern and decided to break it. Rather than rush out Sonnet 5 with another ten percent improvement on MMLU, they're likely pursuing something harder. They're probably working on models that can explain their reasoning not as a party trick, but as a core architectural feature. Models that know what they don't know and communicate that uncertainty clearly. Models that scale in safety as aggressively as they scale in capability.

This approach demands patience. You can't bolt interpretability onto a model after training and expect meaningful results. You can't patch constitutional principles into an architecture designed around different priorities. If Anthropic is serious about building models that remain aligned as they grow more powerful, they need to redesign the foundation. That takes time.

The economics support this theory as well. Training runs for frontier models now cost tens of millions of dollars at minimum, likely hundreds of millions for the largest experiments. Companies can sustain that spending if each model clearly surpasses its predecessor and generates corresponding revenue. However, as improvements become marginal, the calculus changes. Anthropic has substantial funding, but they're not infinite. A strategic pause to ensure the next model represents a genuine leap rather than an incremental step makes financial sense.

I also notice that Anthropic has been unusually active in publishing research on model interpretability and mechanistic understanding. These papers don't generate immediate commercial value, but they lay groundwork. They suggest a company thinking several moves ahead, building the theoretical foundation for techniques they plan to deploy at scale. When Sonnet 5 eventually arrives, I expect we'll see these research threads woven throughout its architecture.

The competitive landscape reinforces this reading. OpenAI remains the market leader in terms of mindshare, but their recent releases have felt increasingly similar to each other. Google has made impressive strides with Gemini, but they're playing the same game everyone else is playing — faster, bigger, slightly better on benchmarks. There's an opening for a company willing to compete on a different axis entirely. If Anthropic can deliver a model that's not just capable but genuinely more trustworthy and interpretable, they could define a new category of competition.

Think about what enterprises actually need from these models. They don't need another incremental improvement in code generation or mathematical reasoning. They need models they can deploy with confidence, models whose failure modes they understand, models that integrate into systems with predictable behavior. The company that solves those problems will command premium pricing and customer loyalty that benchmark performance alone cannot buy.

As a result, my prediction for Sonnet 5 is specific. I don't think we'll see a traditional release announcement with the usual fanfare. Instead, I expect Anthropic will publish a detailed technical paper explaining new approaches to alignment and interpretability, followed by a model that demonstrates those approaches in practice. The improvements on standard benchmarks might be modest — perhaps even deliberately restrained. The real advances will be in areas we currently struggle to measure: robustness, predictability, transparency.

The timeline is harder to predict, but I'd be surprised if we see anything before mid-2026. Anthropic's silence suggests they're deep in the experimental phase, not polishing a nearly-ready product. They're likely running training experiments, evaluating results, iterating on architecture. That process can't be rushed without compromising the principles that differentiate them.

This slower pace might frustrate those of us who refresh the Anthropic homepage daily hoping for news. However, I find it reassuring. We've spent the past few years in a headlong sprint toward more capable AI systems, often with safety and interpretability lagging behind. If one major lab is willing to slow down and do the harder work of building systems that scale safely, that benefits everyone.

The race to AGI continues, but perhaps we need some participants racing toward a different finish line. Anthropic appears to be positioning themselves as exactly that. When Sonnet 5 arrives, I believe it will represent not just an incremental improvement, but a statement about what frontier AI development can and should prioritize. The deliberate slowdown isn't weakness — it's the most ambitious move they could make.

I read the news about OpenAI exploring advertising-supported products with a kind of weary recognition. Not surprise — the trajectory has been obvious for months — but something closer to resignation. The company that positioned itself as humanity's steward in the age of artificial intelligence is now contemplating the same business model that turned social media into a surveillance apparatus and search engines into glorified billboards. The irony is almost too neat.

The reporting suggests OpenAI is considering ads as a way to expand access to ChatGPT and its other products. Free tiers supported by advertising would lower the barrier to entry, bringing AI capabilities to users who cannot or will not pay subscription fees. This sounds reasonable. It sounds, in fact, like the familiar Silicon Valley playbook: build something compelling, give it away for free, monetize attention. However, applying this model to AI systems creates problems that do not exist with traditional software.

The fundamental issue is alignment — not in the technical sense that AI researchers discuss, but in the economic sense that determines what companies actually optimize for. A subscription business aligns the company's interests with the user's interests. I pay for a service that works well for me. The company improves the service to justify continued payment. The incentive structure is straightforward. An advertising business, by contrast, splits the alignment. The user is no longer the customer. The user is the product being sold to the actual customer: the advertiser.

This misalignment has predictable consequences. Facebook optimized for engagement because engagement generates ad impressions. The algorithm learned to surface content that provokes strong emotional reactions — outrage, fear, tribal identification — because those reactions keep people scrolling. Additionally, Google Search has degraded steadily as ads colonize more of the results page and SEO spam proliferates because Google's incentive is to show ads, not to surface the best information quickly.

Apply this dynamic to ChatGPT and the implications become unsettling. An advertising-supported AI assistant would be optimized not for providing accurate, helpful information, but for maximizing user engagement with advertising content. The model might subtly bias its responses toward advertisers' products. It might provide longer, more circuitous answers that create more opportunities to insert promotional content. It might recommend solutions that happen to involve purchasing something from a sponsor. The corruption would be gradual and deniable, but the economic incentives point in one direction only.

I recognize the counterargument. OpenAI will maintain strict separation between the AI's core functionality and the advertising layer. Ads will be clearly labeled and isolated from responses. The company has a reputation to protect and sufficient capital to resist immediate pressure for aggressive monetization. Therefore, the pessimistic scenario I describe will not materialize because OpenAI will implement advertising responsibly.

This argument fails on two grounds. First, advertising businesses always become more aggressive over time. The initial implementation is restrained and user-friendly. Then quarterly revenue targets increase. Growth slows. Investors demand higher returns. The product team faces pressure to make ads more prominent, more targeted, more integrated into the core experience. The trajectory is so consistent across companies and platforms that treating OpenAI as an exception requires extraordinary optimism about corporate incentive structures.

Second, even well-intentioned advertising creates subtle distortions. Consider how sponsored content works in traditional media. A magazine might maintain editorial independence while running advertiser-funded articles clearly labeled as such. Yet studies consistently show that publications are less likely to publish negative coverage of their advertisers and more likely to cover topics that advertisers favor. The influence operates through internalized norms and anticipatory self-censorship, not through explicit directives. An AI trained on interaction patterns shaped by advertising incentives would learn these biases without anyone deliberately programming them in.

The timing makes this development particularly concerning. We are in the early stages of AI integration into critical workflows — research, education, professional services, creative work. The tools people adopt now will shape expectations and habits for years. If the default free tier of AI assistance comes with advertising, an entire generation of users will internalize that relationship as normal. They will learn to navigate around commercial influence, to discount AI recommendations that seem suspiciously aligned with products, to treat the technology with appropriate skepticism. However, this adaptive response has costs. Trust erodes. The cognitive overhead increases. The technology becomes less useful precisely because users must constantly evaluate whether they are receiving genuine assistance or sophisticated marketing.

Additionally, advertising-supported AI would likely accelerate inequality in access to reliable information. Those who can afford subscription services get uncompromised AI assistance. Those who cannot get a version optimized for advertiser revenue. The gap is not merely about features or response speed — it is about epistemic reliability. The free tier becomes a second-class information environment where answers are shaped by commercial interests. This is not hypothetical. We already see this pattern with news media, where quality journalism retreats behind paywalls while ad-supported content proliferates with minimal editorial oversight.

I want to believe that OpenAI will resist this path. The company has made commitments to safety and alignment that advertising fundamentally undermines. The leadership has expressed concern about AI systems pursuing goals misaligned with human values. Optimizing an AI for advertising revenue is deliberately introducing misalignment — choosing a business model that requires the system to serve two masters with competing interests.

The alternative exists. OpenAI could focus on enterprise customers who pay substantial fees for reliable, uncompromised AI capabilities. They could offer educational and nonprofit discounts funded by commercial revenue rather than by advertising. They could maintain free tiers at reduced capability levels without introducing the perverse incentives that advertising creates. These paths are harder. They generate less total revenue. They do not scale as rapidly. Nevertheless, they preserve the alignment between the technology's purpose and its economic foundation.

The broader pattern troubles me more than any single company's decision. The AI industry is barely five years into commercial deployment of large language models, and already we are seeing convergence toward the advertising model that has degraded so much of the internet. The technology is different. The capabilities are unprecedented. Yet the business logic is depressingly familiar. Build engagement, monetize attention, optimize for advertiser revenue, accept the externalities.

If OpenAI proceeds with advertising, other companies will follow. The precedent will normalize what should be seen as a profound compromise. Users will be told they are getting AI access for free, while paying with something far more valuable than subscription fees: their trust in the information they receive. The oracle will start selling ad space, and we will all pretend this does not change the nature of what it tells us.

I hope OpenAI chooses differently. The company has the resources and the stated mission to build AI that serves users rather than advertisers. However, hope is not a strategy, and economic incentives are persistent. If the oracle starts selling ad space, we should at least acknowledge clearly what we are trading away.