Plutonic Rainbows

Google released Gemma 4 today. Four model sizes, multimodal across the board, and a license change that matters more than any benchmark number on the page.

The headline spec is a family of open-weight models built from the same research as Gemini 3. There is a 31B dense model, a 26B mixture-of-experts variant that activates only 4 billion parameters at inference time, and two edge-optimised models (E4B and E2B) small enough to run on a Raspberry Pi 5. The context windows stretch to 256K tokens on the larger models and 128K on the smaller ones. All four handle images and text natively. The edge models add audio input. The larger two process video.

None of that is the story.

The story is Apache 2.0.

Gemma 3 shipped under a custom license, Google's own "Gemma Terms of Use," which imposed restrictions that made legal teams nervous and hobbyists indifferent. It was open in the way that a restaurant with a dress code is open. You could walk in, but the terms reminded you this was someone else's house. Gemma 4 drops all of that. Apache 2.0 means no usage caps, no commercial restrictions, no "contact us if you exceed 700 million monthly active users" clause like Meta's Llama license carries. Fork it, ship it, sell it, modify it without asking. The freedom is unconditional.

This is Google choosing to compete on capability rather than control. And the capability argument is strong. The 31B dense model ranks third on the Arena AI text leaderboard. The 26B MoE variant, running on just 4 billion active parameters, sits sixth, outperforming models with twenty times its effective compute budget. Google's own framing is "intelligence per parameter," and the numbers back it up. A model that small matching frontier-class open weights running at 100B+ parameters is not incremental progress.

The architecture has some genuinely interesting choices. Alternating attention layers split work between local sliding-window attention (512 or 1024 tokens depending on model size) and global full-context layers. Each attention type gets its own RoPE configuration: standard for local, proportional for global. A feature called Per-Layer Embeddings feeds a secondary signal into every decoder layer, combining token identity with contextual information, which seems to be how they squeeze so much quality out of fewer parameters. The shared KV cache reuses key-value tensors from earlier layers in later ones, cutting memory without obvious quality loss. It is a dense collection of efficiency tricks that compound.

The on-device numbers are where this gets practical. On a Raspberry Pi 5, the E2B model hits 133 tokens per second on prefill and 7.6 tokens per second on decode, using less than 1.5GB of memory with 2-bit quantization. Four thousand input tokens across two distinct skills process in under three seconds on mobile GPU. These are not synthetic benchmarks designed to flatter a press release. Raspberry Pi inference is the kind of thing people will actually try within hours of a release, and if those numbers hold, this becomes the default local model for a lot of embedded and mobile work.

I keep circling back to the agentic framing. Google is not positioning Gemma 4 as a chatbot engine. The marketing language says "purpose-built for advanced reasoning and agentic workflows," and the tooling reflects it: constrained decoding for structured outputs, multimodal function calling, GUI element detection, object detection and pointing. These are the primitives you need for an AI agent that can look at a screen, understand what it sees, decide what to do, and call the right function. The fact that it works offline, on a phone, without phoning home to a cloud endpoint, makes the agentic pitch credible in a way that server-dependent agents never quite were.

The ecosystem support at launch is unusually comprehensive. Day-one availability across Hugging Face Transformers, llama.cpp, MLX for Apple Silicon, Ollama, mistral.rs, ONNX, and browser-based inference through WebGPU via transformers.js. Google clearly pre-coordinated with the major frameworks. When I wrote about model discovery and pricing a couple of weeks ago, the friction was still in finding and deploying the right model. Gemma 4 arrives already integrated into every tool people actually use.

What Google is doing here has a clear strategic logic. The Gemini 3.1 Pro updates showed them closing the gap with Claude and GPT on their proprietary side. Now the open side gets a model built from the same research foundations, under the most permissive license in the major open-weight landscape. Meta's Llama has its commercial threshold. Mistral has been ambiguous about which models are truly open. Google just removed every legal obstacle at once.

The 140+ language support is quietly significant. Most open models optimise for English with a handful of other languages bolted on. Google's multilingual training infrastructure, built for Search over two decades, gives Gemma 4 a natural advantage here. For developers building products outside the English-speaking world, this might be the deciding factor regardless of benchmark position.

I'm less certain about the video capabilities in the larger models. Processing video natively is useful, but the context window arithmetic gets expensive fast. A few minutes of video at reasonable frame rates will consume a large fraction of that 256K window, leaving limited room for reasoning about what was seen. The image and audio capabilities feel more immediately practical, especially on the edge models where audio input enables real-time speech understanding directly on device.

The competitive pressure this creates is substantial. Llama 4 from Meta is the obvious comparison, and Meta's response will need to address both the licensing gap and the efficiency gap. A 4B active parameter model matching 100B+ models on quality is the kind of result that forces everyone else to rethink their architecture, not just their marketing. Qwen, Phi, and the rest of the open-weight field now have a new bar to clear, set by a company with functionally unlimited compute and training data.

Whether Gemma 4 becomes the default open model depends on what happens in the next few weeks as developers actually stress-test these claims. Arena scores and launch-day benchmarks are one thing. Sustained performance across real workloads, fine-tuning stability, and the texture of outputs on tasks that benchmarks do not measure will determine if this is the model people reach for by default or just another strong option in an increasingly crowded field.

The Apache 2.0 move, though, is irreversible. Google cannot walk that back without destroying trust. And for every developer who avoided Gemma 3 because of licensing uncertainty, the door is now wide open.

Pankaj Gupta built a product that let 1.3 million people vote on which AI model gave the best answer. Jeff Dean invested. Biz Stone invested. The CEO of Perplexity invested. a16z crypto's Chris Dixon led a $33 million seed round. On Tuesday, Gupta announced Yupp.ai is winding down, less than ten months after launch. Platform access ends April 15.

The stated reason is the one every failed startup reaches for: product-market fit. "The AI model capability landscape has changed dramatically in the last year alone," Gupta wrote. Which is a polite way of saying the product was a leaderboard for a race where the runners kept swapping positions between refreshes.

Yupp's premise made a kind of sense when it launched in June 2025. Back then, picking between Claude and GPT and Gemini and whatever Mistral was calling itself that week felt consequential. You'd paste a prompt into three chat windows, squint at the results, and develop superstitions about which one "got you." Yupp crowdsourced that process across 800 models. Millions of preference signals a month, all feeding into a ranking system that was supposed to help ordinary people navigate the model landscape.

The problem is that ordinary people stopped caring. Not because the models got worse, but because they got interchangeably good enough. When the gap between first place and eighth place on a benchmark is statistical noise, a consumer taste-test platform becomes a thermometer for a room that's already at temperature.

There's a crueller reading. AI labs figured out that crowdsourced preferences from casual users are a blunt instrument. The shift toward agentic workflows meant models needed to impress other models, not people scrolling on their phones. For the kind of reinforcement learning that matters now, labs hire domain experts and run evaluations against PhD-level feedback. The crowd was never going to be precise enough.

Forty-five angel investors. DeepMind's chief scientist. A $33 million cheque from one of the most connected funds in Silicon Valley. And the thing it bought was ten months of server time and a blog post titled "winddown." The economics of wrapping someone else's API haven't changed since Anthropic started enforcing its terms of service. If anything, the lesson has sharpened. The thinner your layer, the faster the substrate makes you irrelevant.

Some of Yupp's employees are reportedly joining a "well-known" AI company. Which sounds like a soft landing until you consider that it's the same trajectory the product followed: absorbed back into the infrastructure it was built to evaluate.

Most of what humanity has written, recorded, and published does not exist on the internet. Not even close. Large language models, search engines, recommendation algorithms: they all treat the web as though it were a reasonable proxy for human knowledge. It is not. It is a shallow, recent, and spectacularly incomplete sample.

Google has scanned tens of millions of books, but most sit behind copyright walls, neither fully searchable nor publicly readable. The rest exist on shelves, in basements, in charity shops where nobody is looking. The vast majority of the world's cultural heritage has never been digitized in any form. Not suppressed, not restricted. Just absent.

The pre-internet age was not merely analogue. It was geographically bounded. John Holbo, writing on Crooked Timber, described it as a kind of epistemic accident: you knew what the six people around you knew, what your local library stocked, what your local record shop carried. A left-handed guitarist might never discover that left-handed guitars existed. That accidental ignorance, that texture of ordinary life, was never documented in a form that any crawler could find. It was the water, not the fish.

The physical record is vanishing too. When the Chicago Sun-Times consolidated its suburban papers, photographs from the Aurora Beacon-News and Elgin Courier-News were thrown in the bin. The Louisville Courier Journal's archive of roughly ten million photographs nearly followed before the University of Louisville negotiated a last-minute donation. These aren't edge cases. They are the norm for local journalism across America and, by extension, for any community record that depended on newsprint.

Meanwhile, born-digital content fares no better. Pew Research found in 2024 that a quarter of all web pages that existed between 2013 and 2023 have already disappeared. MySpace's 2019 migration destroyed millions of songs, videos, and photographs in what the Long Now Foundation described as irreversible data loss. Andy Warhol's digital artwork from the 1980s sat stranded on obsolete Commodore hardware for decades.

The gap is self-reinforcing. If knowledge isn't online, AI can't learn it. If AI can't surface it, fewer people encounter it. If fewer people encounter it, there's less incentive to digitize it. The loop tightens and the memory without metadata that defined most of human experience drifts further from retrieval.

I think about this when people describe AI as a knowledge tool. It is a tool for a particular kind of knowledge, overwhelmingly English-language, overwhelmingly post-1990s, overwhelmingly sourced from the kind of person who publishes on the internet. Everything else, the vast majority of what humans have thought and made and recorded, sits in formats that no model will ever ingest. Not because the technology couldn't handle it, but because nobody is going to scan it.

Three hundred looks. One hour. A 150-year-old circus in the 11th arrondissement. Thierry Mugler's Autumn/Winter 1995 haute couture show at the Cirque d'Hiver wasn't a runway presentation. It was a siege.

March 1995 was Mugler's twentieth anniversary, and he chose to mark it by staging the single most excessive fashion event of the century. French television broadcast it live during primetime. The set was a multi-levelled white stage with two runways and a spiral staircase, the Angel perfume star logo hovering above like a corporate halo. Male go-go dancers flanked the walkways. The soundtrack veered between house music and classical before James Brown walked out and started screaming.

The cast list reads like someone ransacked three decades of fashion and film. Naomi Campbell. Kate Moss. Claudia Schiffer. Eva Herzigova in a volcanic eruption of red ostrich feathers. Jerry Hall in a crystal-encrusted catsuit so sheer it functioned more as a dare than a garment. Linda Evangelista wore a powder-blue gown with a jewelled salamander headpiece. Elle Macpherson carried a dog. Tippi Hedren, thirty-two years after The Birds, walked in a black satin avian dress that referenced the role Hitchcock made her famous for. Julie Newmar, television's original Catwoman, appeared in black rubber lace. Patty Hearst performed a striptease.

Then came the robot.

Nadja Auermann emerged encased in an articulated silver bodysuit made from metal and Plexiglas, six months in construction, built by corsetier Mr. Pearl, artist Jean-Jacques Urcun, and aircraft specialist Jean-Pierre Delcros. It was Fritz Lang's Maschinenmensch remade as couture, the kind of garment that doesn't belong on a runway because it belongs in a museum or possibly an armoury. Helmut Newton photographed it. Zendaya wore it in 2024.

The Birth of Venus dress was quieter but stranger, a translucent bodysuit embroidered with paillettes and pearl beads, Botticelli translated through the logic of a Parisian atelier. Cardi B wore it to the 2019 Grammys, which tells you something about its shelf life.

Tim Blanks called it one of the greatest fashion shows ever staged. That feels about right. The show existed in opposition to everything the mid-nineties was supposed to be about. Calvin Klein was selling silence. Ralph Lauren was selling tasteful restraint. Mugler was selling chrome exoskeletons and live James Brown and Patty Hearst taking her clothes off in a circus. The minimalists won the decade, eventually. But nobody remembers their shows.

I keep thinking about what it meant to watch this on television. Not a highlight reel, not a documentary years later. Live, in your living room, between whatever else was on that night. Fashion as an event that happened to you whether you cared about fashion or not.

Eva Herzigova in red feathers at the Cirque d'Hiver

Alibaba released Qwen3.5-Omni on Monday and the most interesting thing about it is not what the model can do. It is what Alibaba chose to keep.

The Qwen family has been downloaded over 700 million times on Hugging Face, with more than 100,000 derivative models. That makes Alibaba the most-downloaded open-weight AI provider on the platform, and it was deliberate, a land grab disguised as generosity. Now, with Qwen3.5-Omni, the generosity has limits.

The model splits into two components the team calls the Thinker and the Talker. The Thinker handles reasoning across text, images, audio, and video. The Talker converts that reasoning into streaming speech, frame by frame, through a lightweight convolutional renderer called Code2Wav. The separation is not just clean design. It means external systems (safety filters, retrieval pipelines, function calls) can intervene between cognition and output. Enterprise deployment teams will notice.

The numbers are aggressive. A 256,000-token context window that can absorb ten hours of continuous audio or four million frames of 720p video. Speech recognition in 113 languages. Voice cloning via the API. An emergent capability the team calls audio-visual vibe coding: the model writes functional code by watching screen recordings with spoken instructions, without having been trained on that task. That last detail sounds like marketing until you remember that emergent capabilities in large models have a track record of being real and unsettling in equal measure.

On benchmarks, it outperforms Gemini 3.1 Pro on music understanding (72.4 to 59.6) and edges it on audio comprehension. Voice stability scores undercut ElevenLabs by an order of magnitude. These are not incremental wins.

But only the Light variant ships as open weights. Plus and Flash, the versions you would actually deploy, are API-only through Alibaba's DashScope. No technical paper has been published. No weights to inspect. The 700 million download count was built on open licensing, and the moment the Qwen team produced something genuinely frontier in multimodal, they pulled it behind a paywall.

This is not hypocrisy. It is strategy. Open-weight text models seed the ecosystem, create dependency, train a generation of developers on your API surface. Then, when voice and video become the competitive edge, you charge for access. Alibaba built the largest open-source AI distribution network in history specifically so they could close it at the right moment.

The Thinker reasons for free. The Talker costs money. That might be the most honest thing about the whole architecture.

Room 13, BBC Maida Vale Studios. Before it held oscillators and tape machines, the building was a roller skating palace. Opened in 1909 on Delaware Road, converted by 1934, given to a handful of BBC engineers in 1958 with two thousand pounds and whatever surplus military electronics they could find at Portobello Market.

Delia Derbyshire joined the Workshop in 1962 with a mathematics and music degree from Cambridge and a rejection letter from Decca Records, who did not employ women in their studios. In eleven years she created sound for roughly 200 programmes. The Doctor Who theme remains the most famous: Ron Grainer handed her a single sheet of A4 manuscript paper with annotations like "wind bubble" and "cloud," and she realised it from tape-spliced fragments of a plucked string, white noise, and test-tone oscillators meant for calibrating equipment. When Grainer heard it he asked, "Did I really write this?" She said, "Most of it." The BBC would not credit her for another fifty years.

None of this is news. The Workshop's history has been thoroughly documented. What interests me is what those sounds have become now that the context they were made for no longer exists.

The Radiophonic Workshop did not just make television themes. It soundtracked a specific institutional vision of Britain: Open University lectures, schools broadcasts, public information films. The BBC under its post-war mandate believed that educating the nation was a public good, and these electronic textures were the sonic furniture of that belief. Mark Fisher identified this precisely. Hauntological music, he wrote, constitutes "an oneiric conflation of weird fiction, the music of the BBC Radiophonic Workshop, and the lost public spaces of the so-called postwar consensus." That consensus ended in 1979.

The Workshop itself held on until 1998, killed by John Birt's internal market policies. Elizabeth Parker, the last remaining composer, switched off the lights. The archive was nearly discarded.

When Derbyshire died in 2001, 267 reel-to-reel tapes were found in her attic. They sat there like letters from someone who had stopped writing decades earlier. She left the BBC in 1973 and abandoned music entirely by 1975.

Julian House of Ghost Box Records described the Workshop's older material as "the reverb of a reverb of a reverb." That phrase captures how these sounds circulate now. They are not nostalgic. Nostalgia implies you want to go back. This is different. The sounds point forward, toward a public future that was defunded and dismantled, and the fact that they still sound futuristic is the cruel part. They describe a destination cancelled while the signal was still transmitting.

Simon Reynolds called the tension in Ghost Box's work a pull between "heathen heritage" and "modernizing socialism." The Workshop operated at the intersection of state-funded infrastructure and radical experimentation, and both feel equally impossible now.

I keep returning to those 267 tapes in the attic. An entire career's parallel output, boxed and unlabelled, surviving because nobody thought to throw them away.

Fiera Milano, March 1991. An exhibition hall on the city's outskirts, a fifteen-metre marble runway, and a U-shaped seating plan that separated press from celebrities from international buyers. Gianni Versace had staged shows before, obviously. But nothing like what happened at the end of this one.

The collection itself was pure Versace at full volume. Boxy cropped jackets over Lycra catsuits printed with baroque scrollwork. Studded leather cut alongside pleated skirts. Thigh-high boots that had no business being paired with silk but somehow were. The colour ran from black through to saturated reds, greens, oranges, and yellows, all of it rendered in that specific register Versace owned: sexy, loud, and entirely uninterested in apology.

Then the finale. George Michael's Freedom! '90 hit the speakers and out came Linda Evangelista, Cindy Crawford, Naomi Campbell, and Christy Turlington. Not walking individually. Not one after another. Arm in arm, four across, lip-syncing the lyrics, laughing, mugging for the front row. They wore dresses in red, yellow, and black. George Michael watched from his seat.

The four supermodels at the Versace AW91 finale

The previous October, David Fincher had released the music video for the same song, starring all four (plus Tatjana Patitz). No George Michael in frame, just supermodels lip-syncing in a stripped-down loft while a jukebox exploded. The video made them icons outside fashion. The Versace finale made that iconography physical, live, happening in a room full of people who understood they were watching something that couldn't be repeated.

The backstory matters. Liz Tilberis, then editor of British Vogue, had told Versace to stop splitting the top models across different slots. Book them together. Let their combined weight collapse the room. He listened. And the result was not just a fashion show but a proof of concept: the runway could function as spectacle, as cultural event, as something people who had never touched a copy of Vogue would eventually see and remember.

Before this night, runway shows were trade events. After it, they were content. Every designer who stages a celebrity-packed front row, every brand that livestreams its collection, every fashion week headline that leads with a name rather than a garment owes a debt to what happened at Fiera Milano. Versace understood something his contemporaries didn't, or wouldn't admit: the models were the collection. The clothes were spectacular. But four women walking in sync to a pop song, grinning like they owned the building (they did), turned a presentation into a cultural marker that outlived the season, the decade, and eventually the designer himself.

Cindy Crawford later said it felt like all the stars had aligned. She wasn't wrong. But stars don't align by accident. Someone has to set the stage.

Newsom signed Executive Order N-5-26 on March 30. AI companies seeking California state contracts must now certify safeguards against bias, CSAM distribution, and civil rights violations. The Department of General Services and the California Department of Technology have 120 days to build the certification framework. State-level AI watermarking guidance comes bundled in.

This matters for what it isn't. It isn't legislation.

Ten days earlier, the Trump administration published its National Policy Framework for Artificial Intelligence. Seven pillars, a call for Congress to create a single federal standard, and explicit language about preempting what it called a fragmented patchwork of state AI regulation. The message: states should stop. But the framework's own text carves out state government procurement from preemption. The administration conceded, in writing, that it cannot dictate how states buy AI.

Newsom's team read the fine print more carefully than the people who wrote it.

California hosts 33 of the world's top 50 privately held AI firms. It captured over half of U.S. AI startup funding between Q3 2024 and Q2 2025. The state budget runs to roughly $300 billion annually. When California tells vendors they need certification to bid, that isn't a suggestion. Neil Shah at Counterpoint Research called it "a benchmark for de facto AI standards when it comes to procurement, safety, and ethics." Smaller vendors face a heavier compliance burden. The ones who pass have something to show for it elsewhere.

The sharpest provision targets the Anthropic situation. After the company refused to strip safeguards preventing autonomous weapons deployment and mass domestic surveillance, the Pentagon designated it a supply-chain risk to national security. A federal judge called the move "Orwellian." California's order now gives the state CISO authority to review such designations independently and, where warranted, override them for state procurement.

So California is building a parallel regulatory structure through spending power alone. No legislation required. No direct challenge to federal preemption. Just a procurement policy that sets safety standards the federal government specifically declined to set.

The ACLU's Cody Venzke described the administration's preemption strategy as "a hodgepodge of faulty legal theories." Even Republican legislators pushed back, with Utah and Texas among the states objecting to federal overreach. The administration retains leverage. The DOJ can sue. They can threaten broadband funding. But procurement sits in the carve-out. The federal government's own policy document says so.

California didn't break through the wall. It walked through the door that was already open.

Rose Glass made Saint Maud for roughly $2.5 million, which is less than the catering budget on most studio horror. You'd never know it. The film looks like it cost ten times that, partly because Ben Fordesman's cinematography treats a bleak Scarborough beachfront like it's the edge of the world, and partly because the production design understands that a lonely bedsit can be more frightening than any haunted house if you shoot it correctly.

Morfydd Clark plays Maud, a palliative care nurse who has recently converted to Catholicism after something went wrong with a previous patient. She's assigned to care for Amanda, a terminally ill choreographer played by Jennifer Ehle with the precise detachment of someone who has already made peace with dying and finds Maud's earnestness first curious, then entertaining, then repulsive. The power dynamic between them is the engine of the film. Amanda has money, sophistication, a history of artistic achievement. Maud has God. For a while, God seems like enough.

The possession question is handled with more ambiguity than most horror films would tolerate. Maud experiences physical sensations she interprets as divine. Her body arches. Her eyes roll back. Whether this is ecstasy or seizure depends entirely on which character you believe, and Glass refuses to resolve the tension. She cited Taxi Driver as an influence, which tracks: Maud shares Travis Bickle's conviction that she has been chosen for a sacred mission, and the same inability to recognise that the mission is the disease.

I keep returning to Adam Janota Bzowski's score. Also a debut. He built what he called a Colourbox, a folder of processed sounds made by hitting objects with a drumstick and running the recordings through effects chains until they became something between music and industrial noise. The result sits underneath the film like a migraine, present even when you can't quite identify it. There's a click-clack sound that recurs, something straining and ready to snap. It won an Ivor Novello nomination, which felt overdue by the time it happened.

Glass joins a line of directors who understand that faith and horror share a border. The same territory The Blackcoat's Daughter occupies, where the supernatural isn't the threat but the comfort, and the real horror is what happens when it withdraws. Saint Maud takes that idea further. Maud's self-mortification scenes, nails pressed into the soles of her shoes, kneeling on broken glass, are shot with a tenderness that makes them harder to watch than if they were played for shock. She isn't being punished. She's trying to feel something she felt once and can't find again.

The final image is the cruelest thing A24 has put on screen. We see Maud's apotheosis through her own eyes first: wings, a crowd of worshippers, transfiguration. Then a smash cut to reality. An 84-minute film and Glass saves her most devastating technique for the last three seconds. The entire audience at Toronto reportedly gasped. I believe it. Some images you can't unsee, not because they're graphic but because they contain two contradictory truths at once and force you to hold both.

Sources

• Saint Maud Review - Variety

• Saint Maud - BFI Sight and Sound

• Composer Interview - SPIN