Plutonic Rainbows

Lisa Graham, somewhere against a late-eighties coastline — cobalt blazer, exaggerated lapels, leather gloves, a single corsage holding it all together. Originally the cover of Marie Claire Spain, September 1988 — the kind of editorial that makes you wonder why anyone stopped dressing like this.

Somewhere around the third hour of listening I stopped taking mental notes. That's usually how I know a system is working. The Exposure 2510 integrated amplifier and ATC SCM11 standmount speakers had been running together for a few weeks by that point, and the initial critical ear — the one that listens to the equipment rather than through it — had gone quiet. What remained was just music, delivered with a directness that caught me off guard given the combined price.

The 2510 is a 75-watt Class A/B integrated from Exposure's factory in West Sussex, a company that has been building amplifiers in England since 1974 without ever chasing fashion. The design philosophy is almost aggressively simple: four controls on the front panel (input selector, volume, and two buttons for standby and mute), five line inputs, a built-in MM phono stage, and a pre-amp output if you ever want to add a power amplifier later. No DAC. No streaming. No tone controls. The circuit uses discrete components with high-quality VIMA capacitors in the signal path, and the through-hole assembly is still done by hand in England. It weighs six kilograms, which feels light until you hear what it does with those watts.

The SCM11 is ATC's entry-level standmount, though calling anything ATC makes "entry-level" feels slightly misleading. This is a sealed-box two-way design with a 150mm mid/bass driver using ATC's Constrained Layer Damping technology — a technique that reduces resonance within the cone itself rather than trying to tame it externally — and a 25mm soft-dome tweeter with a neodymium magnet and proprietary alloy waveguide. Sensitivity is 85dB, which means they need decent amplification, and nominal impedance sits at 8 ohms. ATC builds every driver in-house at their Stroud factory in Gloucestershire, winding the voice coils by hand. Both companies are separated by roughly 150 miles of English countryside and share an almost identical reluctance to overcomplicate things.

What strikes me about this pairing is how the warmth of the Exposure meets the neutrality of the ATC without either quality cancelling the other out. The 2510 has a gently rich midrange that could, with the wrong speakers, tip into softness or veil detail. The SCM11 has a forensic transparency that could, with the wrong amplifier, sound lean or clinical. Together they arrive at something I can only describe as honest warmth — the kind of presentation where a piano sounds like wood and hammers and resonating strings rather than a synthetic approximation of those things. Voices sit in the room with real weight. Not projected, not recessed. Just present.

The sealed-box loading of the SCM11 is doing important work here. Ported speakers can produce more low-end extension, but the bass they deliver often trades speed for depth. The SCM11 rolls off below 56Hz, so you're not getting subterranean weight, but what's there — everything from kick drums to upright bass to the low growl of a cello — arrives with a tightness and control that makes ported designs at this price sound flabby by comparison. I spent an evening working through some of my ripped collection via the network player and the sealed enclosure kept up effortlessly across genres. Dense electronic music, sparse acoustic recordings, everything in between. No port chuffing, no overhang. Just grip.

Seventy-five watts into 85dB-sensitive speakers might look marginal on paper, but I never came close to running out of headroom. The 2510's power supply is doing more useful work than the raw number suggests — it has current delivery that belies the specification sheet. I rarely pushed the volume past the halfway mark even in a medium-sized room with the speakers on stands well clear of the rear wall. Forum users who've paired earlier Exposure models with the SCM11 report the same experience: instruments and voices becoming discernible in a way they weren't before, an openness that invites you to keep listening without fatigue.

The midrange deserves its own paragraph. Both components prioritise this region, and it shows. ATC's CLD driver was designed specifically to clean up the critical band between 200Hz and 3kHz where most musical information lives, and the Exposure feeds it with a signal that's detailed without being etched. The result on vocal recordings is striking — not in an audiophile-cliché way where you suddenly hear a singer's saliva, but in the sense that phrasing and dynamics come through intact. You hear the intention behind a vocal performance rather than just the notes. I went back to records I'd dismissed as flat or poorly mastered and found layers of nuance sitting right there in the mix, previously masked by less resolving equipment.

I keep coming back to the value proposition. The 2510 retails for approximately £2,100 and the SCM11 sits around £1,650. Call it £3,750 for the pair before cables and stands. In a market where a single component can easily exceed that figure, what you're getting here is two British manufacturers' distilled engineering philosophy — decades of refinement expressed as restraint rather than feature creep. Neither product tries to be everything. The amplifier amplifies. The speakers convert electrical signal to sound pressure. And the narrow focus pays off in the listening chair.

There are limitations I should acknowledge. The SCM11's bass rolls off where a floorstanding speaker is just getting started, so if your musical diet demands foundation-shaking lows, you'll either need a subwoofer or different speakers. The 2510 offers no digital inputs whatsoever, which means your source needs its own DAC — a consideration I've already wrestled with when assembling this chain. And the ATC's 85dB sensitivity means they're not the right choice for a feeble amplifier tucked inside a sideboard. They want proper amplification and proper placement.

None of that diminishes what this combination achieves within its design envelope. The Exposure 2510 and ATC SCM11 share something I rarely encounter in mid-priced hi-fi: the absolute absence of a weak link. Most systems at this level have one component that's clearly punching above its weight while another holds things back. Here, the engineering intent is so closely aligned that everything arrives at the same standard simultaneously — detail, dynamics, tonal accuracy, spatial coherence. I stopped thinking about the equipment and started thinking about the music, which is the only review metric that actually matters.

Anthropic ran a Super Bowl ad mocking OpenAI for putting ads in ChatGPT. Altman fired back on X calling Anthropic "authoritarian." OpenAI poached an Anthropic safety researcher the same week it shipped a product that looks like a direct response to Claude Code. The whole thing has the energy of two restaurants opening across the street from each other and spending more time reading each other's menus than cooking.

These companies have genuinely different strengths. Anthropic owns 40% of enterprise LLM spend and builds tools developers actually trust with production code. OpenAI has consumer reach nobody else touches and multimodal ambitions that stretch well beyond text. Both positions are defensible. Neither requires obsessing over what the other shipped on Tuesday.

I wrote about speed becoming the only moat last month. The same logic applies to rivalry as a strategy — it narrows your field of vision to exactly one competitor while the market fragments around you. Meta, Google, and a dozen open-source efforts don't care about your Super Bowl ad.

GPT-5.3-Codex helped debug its own training. OpenAI said it plainly: "the first model that was instrumental in creating itself." That was ten days ago. This week, ICLR announced their first workshop dedicated entirely to recursive self-improvement, scheduled for Rio in April. Google's AlphaEvolve already discovered algorithmic improvements that beat Strassen's fifty-six-year-old matrix multiplication record. The pieces are landing on the board faster than anyone expected.

Recursive self-improvement — systems that modify their own code, weights, prompts, or architecture to become more capable, then use that increased capability to improve themselves further — has been a thought experiment for decades. Eliezer Yudkowsky warned about it. Nick Bostrom built philosophical scaffolding around it. And for most of that time it remained comfortably theoretical because the systems weren't good enough at the one thing the loop requires: writing better software than the software that already exists.

That constraint is dissolving. Not because we've achieved some sudden breakthrough in machine consciousness or general reasoning, but because the narrow version of self-improvement turns out to be enough to matter. A model doesn't need to understand itself philosophically to optimise its own training pipeline. It just needs to be good at code. And the current generation is good at code.

The METR data makes the trajectory explicit. AI task-completion horizons have been doubling every four to seven months — depending on which estimate you trust — for the past six years. If that holds for another two years, we're looking at agents that can autonomously execute week-long research projects. Another four years and it's month-long campaigns. The trend line itself isn't the alarming part. The alarming part is that the trend doesn't need to hold perfectly. Even if progress halves, the capability gap closes on a timeline measured in quarters, not decades.

Dean Ball put it starkly in his recent analysis: America's frontier labs have begun automating large fractions of their research operations, and the pace will accelerate through 2026. OpenAI envisions hundreds of thousands of automated research interns within nine months. Dario Amodei cites 400% annual efficiency gains from algorithmic advances alone. These aren't wild extrapolations from startup pitch decks. These are the people running the labs describing what they see happening inside their own buildings.

However. There's a constraint that rarely gets enough attention in the acceleration discourse. Self-improvement only generates reliable gains where outcomes are verifiable. Code that passes tests. Algorithms with measurable performance. Training runs with clear loss curves. The loop works brilliantly in these domains because you can tell whether the modification actually helped. The system generates a change, measures the result, keeps or discards. Simple evolutionary pressure.

The loop breaks — or at least stumbles badly — when it encounters domains where verification is ambiguous. Alignment research. Safety evaluation. Novel hypothesis generation. The things that arguably matter most for whether recursive self-improvement goes well or catastrophically. A system can optimise its own matrix operations all day. Whether it can meaningfully improve its own ability to recognise its blind spots is a much harder question, and I suspect the honest answer is no.

So when will genuine recursive self-improvement arrive? It depends on what you mean. The narrow version — models improving their own infrastructure, training pipelines, and deployment tooling — is already here. GPT-5.3-Codex is doing it in production. The medium version — agents that systematically discover architectural improvements and better training recipes — is probably twelve to eighteen months out, conditional on the METR trendline holding. The strong version — a system that improves its own reasoning capabilities in open-ended domains, including the ability to improve its ability to improve — remains genuinely unclear. I'm not confident it's five years away. I'm not confident it's twenty.

What I am confident about is that we'll get the narrow and medium versions before we have any serious framework for governing them. The ICLR workshop is a start — researchers trying to make self-improvement "measurable, reliable, and deployable." But the gap between academic workshops and deployed production systems has never been wider. OpenAI shipped a self-improving model before anyone published a standard for evaluating self-improving models. That ordering tells you everything about the incentive structure.

The Gödel Agent — a system that modifies its own task-solving policy and learning algorithm — climbed from 17% to 53% on SWE-Bench Verified. SICA did something similar. These are research prototypes, not products, but the delta between prototype and product in this field is about eighteen months and shrinking. Probably less now that the prototypes can help close the gap themselves.

I keep coming back to something Ball wrote: the public might not notice dramatic improvements, dismissing them as "more of the same empty promises." That feels backwards to me. The risk isn't that progress will be invisible. The risk is that it'll be visible to the people building it, acting on it, profiting from it — and invisible to everyone else until the loop is already running too fast to audit.

Nearly half of British sixteen-to-twenty-one-year-olds told the BSI they'd prefer to have grown up in a world without the internet. Forty-seven percent. Not a fringe opinion from technophobes or Luddites — a near-majority of the generation that never knew anything else.

The rest of the numbers are worse. Sixty-eight percent said they felt worse about themselves after spending time on social media. Forty-two percent admitted to lying to their parents about what they do online. Forty percent maintain a decoy or burner account. Eighty-five percent of young women compare their appearance and lifestyle to what they see on their feeds, with roughly half doing so often or very often. These aren't edge cases. This is the baseline experience.

What strikes me isn't the individual statistics — we've had versions of these figures for years. Back in 2018, Apple's own investors were pressuring the company over youth phone addiction, citing surveys where half of American teenagers said they felt addicted to their devices. Seven years later, nothing structural changed. The platforms got stickier. The algorithms got sharper. The age of first exposure dropped. And now the generation that grew up inside the experiment is telling us, plainly, that they wish the experiment hadn't happened.

Fifty percent of respondents said a social media curfew would improve their lives. Twenty-seven percent wanted phones banned from schools. Seventy-nine percent believed tech companies should be legally required to build privacy safeguards. That last number is the one I keep returning to — four out of five young people asking for regulation that adults have spent a decade failing to deliver.

The BSI's chief executive, Susan Taylor Martin, put it in corporate language: "The younger generation was promised technology that would create opportunities, improve access to information and bring people closer to their friends." The research, she said, shows it is "exposing young people to risk and, in many cases, negatively affecting their quality of life." This is what institutional understatement sounds like when the data is screaming.

There's an uncomfortable parallel with how the AI industry is repeating social media's mistakes — the same pattern of externalised harm and internalised profit, the same rehearsed contrition at hearings, the same gap between stated commitments and actual behaviour. The platforms knew what they were doing to adolescents. Internal documents confirmed it. Nothing changed because engagement metrics drove revenue, and revenue was the only number that mattered in the boardroom.

Forty-three percent of the respondents started using social media before the age of thirteen — the legal minimum. Not because their parents approved, but because the platforms made it trivially easy to lie about your age. Then those same platforms sold advertising against the attention of children who shouldn't have been there in the first place.

The generation that was supposed to be "digital natives" — fluent, empowered, connected — is telling us they'd trade it all for something quieter. We should probably listen.

Within five days of each other, Anthropic launched Opus fast mode and OpenAI shipped Codex-Spark. Same thesis, different silicon. Anthropic squeezes 2.5x more tokens per second out of Opus 4.6 through inference optimisation. OpenAI distills GPT-5.3-Codex into a smaller model and runs it on Cerebras wafer-scale hardware at over a thousand tokens per second. Both are research previews. Both are gated to developers. Both cost more than their standard counterparts.

The timing isn't coincidence. Coding agents are the first workload where latency translates directly into revenue. A developer staring at a terminal while an agent loops through forty tool calls doesn't care about cost per token — they care about wall-clock minutes. Anthropic charges six times the standard rate for fast mode. OpenAI hasn't published Spark pricing yet, but the Cerebras partnership wasn't cheap. These aren't loss leaders. They're premium tiers aimed at the one audience willing to pay for speed right now.

What interests me is the constraint both companies are accepting. Fast mode is Opus with the same weights, just served differently. Codex-Spark is a distilled, smaller model — OpenAI admits the full Codex produces better creative output. Neither approach is free. You either pay for dedicated inference capacity or you trade quality for velocity. There's no trick that makes frontier intelligence and sub-second latency coexist cheaply.

The question everyone keeps asking — will these become generally available? — misframes the situation. The technology already works. The bottleneck is economics. Anthropic can't offer fast mode to every Claude consumer at six times the compute cost without either raising subscription prices or eating the margin. OpenAI can't run every ChatGPT conversation through Cerebras wafer-scale engines. The hardware doesn't exist in sufficient quantity. Their own announcement says they're ramping datacenter capacity before broader rollout.

So the honest answer is: speed tiers will generalise, but slowly, and probably not in the form people expect. I'd bet on tiered pricing spreading across the consumer products — a fast toggle in Claude.ai, a "turbo" option in ChatGPT — before the end of the year. But it'll cost extra. The idea that baseline inference gets dramatically faster for free requires either a hardware miracle or margins that neither company can sustain.

The deeper pattern is what I wrote about last month. Speed is becoming the axis of competition because capability gains have slowed enough that users notice latency before they notice intelligence improvements. When both labs ship speed products in the same week, that tells you where the demand signal is loudest. Not smarter. Faster.

Propaganda lined up for Virgin in March 1990 with four faces calibrated for the dark — the ruffled blouse at center doing more work than any press stylist should have to admit.

Google upgraded Gemini 3 Deep Think yesterday, and the number that matters is 84.6%. That's its score on ARC-AGI-2, the abstract reasoning benchmark designed to resist brute-force pattern matching. Humans average around 60%. Claude Opus 4.6 — which landed last week to genuine excitement — scores 68.8%. GPT-5.2 manages 52.9%. Deep Think clears the human baseline by nearly 25 points and leads the next-best model by almost 16.

I'm trying to figure out what to do with that.

The Codeforces result is harder to dismiss as benchmark theatre. Deep Think hit 3,455 Elo — Legendary Grandmaster territory, better than all but seven active human programmers on the platform. No external tools. No retrieval. Just inference-time compute and whatever Google means by "parallel hypothesis exploration." The top human competitor, Benq, sits at 3,792. That gap is closing fast enough to make competitive programming feel like it has an expiration date.

What changed from the previous version: scope. Earlier iterations of Deep Think were narrowly focused on mathematics. This upgrade pushes into chemistry, physics, and engineering. Gold medals on the written portions of the International Math, Physics, and Chemistry Olympiads. A mathematician at Rutgers used it to peer-review a paper on high-energy physics structures bridging gravity and quantum mechanics. It caught a subtle logical flaw that human reviewers had missed. That's not a benchmark. That's a real research contribution, however narrow.

The architecture Google describes — they call it "Aletheia" — uses a generator, a natural language verifier, and a reviser working in concert. Parallel hypothesis exploration rather than a single reasoning chain. The interesting detail is that the system can acknowledge failure and stop rather than burning compute on dead-end paths. Most reasoning models I've used have no concept of giving up gracefully. They hallucinate forward until they hit a token limit. If Aletheia genuinely knows when it's stuck, that's a meaningful advance in how these systems manage uncertainty.

Google's approach here is fundamentally different from what Anthropic and OpenAI are doing. They're scaling inference-time compute — giving the model more time to think rather than making a bigger model. The base is still Gemini 3 Pro, not some trillion-parameter behemoth. Deep Think is a reasoning mode, not a separate model. The distinction matters because it suggests the ceiling on what you can extract from existing architectures is higher than most people assumed. You don't need a fundamentally new model. You need to let the current one actually think.

That feels right to me, intuitively. When I use extended thinking in Claude, the quality jump over instant responses is enormous — not because the model suddenly knows more, but because it has room to work through contradictions and dead ends before committing to an answer. Google is doing the same thing with significantly more compute thrown at the problem. Anthropic shows you the reasoning. Google hides it. Both approaches produce results that make the non-thinking versions look careless by comparison.

The pricing is interesting. Deep Think is included in the Google AI Ultra subscription at $249.99 per month. API access requires applying for an early programme. I keep thinking about how o3 was positioned as the reasoning breakthrough that would change everything, and then Deep Think shows up a year later scoring nearly 30 times higher on the same class of benchmark. The pace of obsolescence in this space is genuinely disorienting.

Demis Hassabis called it "new records on the most rigorous benchmarks in maths, science & reasoning." MarkTechPost ran with "Is This AGI?" — which, no. But I understand the impulse. A system that reasons better than the average human on abstract pattern recognition, codes better than 99.99% of programmers, and catches errors in peer-reviewed physics papers occupies territory that didn't exist twelve months ago.

Google DeepMind published a research impact taxonomy alongside the release, rating contributions from Level 0 to Level 4. They classify Deep Think's current output at Levels 0-2 — autonomous solutions and publishable collaborations, not landmark breakthroughs. The fact that they felt the need to temper expectations tells you something about the temperature of the conversation. When the company releasing the model is the one saying "calm down," the benchmarks have moved past what anyone's frameworks were built to accommodate.

Sodium streetlights. That's where it starts for me. Not the event or the era but the colour — that flat, amber wash that turned every pavement into something theatrical and slightly wrong. Modern LEDs render the world in full spectrum. Sodium vapour didn't. It collapsed everything into two tones and left your brain to fill in the rest. The brain, filling in, often chose unease.

Something about the 1970s and 1980s registers as faintly sinister when viewed from here, and the reaction is common enough to suggest it isn't just personal. The textures of that period — film grain, tape hiss, CRT scanlines, the particular softness of analogue video — carry an ambiguity that digital media has largely eliminated. Modern footage is sharp, bright, and hyper-legible. Older material contains noise and shadow. The brain interprets visual ambiguity as uncertainty, and uncertainty triggers a low-grade vigilance that can settle in the body as discomfort. You're not scared, exactly. You're watchful.

Pacing compounds this. Television idents, public information films, educational broadcasts — they moved slowly and left gaps. Long pauses. Static framing. Sparse dialogue. Minimal scoring or none at all. Contemporary media fills nearly every second with motion or sound because dead air is considered a failure. Confronted with silence, the viewer projects meaning into the space, and what gets projected is rarely cheerful. The Protect and Survive films weren't trying to be frightening. Their flat, institutional delivery made them more disturbing than any horror film could manage.

That institutional tone is part of it too. Public messaging in Britain during this period was formal, impersonal, authoritative. It lacked the conversational warmth that modern branding considers mandatory. Government broadcasts addressed you like a patient being told to stay calm — which, if you weren't already anxious, was a reliable way to make you start. The emotional distance reads as cold now. Cold enough to feel ominous.

The broader atmosphere didn't help. Cold War nuclear anxiety sat underneath everything like a bass frequency you couldn't quite identify but could feel in your teeth. Industrial decline, unemployment, urban decay, terrorism coverage, moral panics — even if none of this was consciously processed at the time, it shaped the cultural mood. And cinema absorbed it. Halloween, The Exorcist, Videodrome, Threads — these films used suburban quiet, analogue distortion, and institutional spaces to generate dread. Their visual language has become fused with how we perceive the entire era. I can't look at a 1970s kitchen without half-expecting something terrible to happen in it.

There's also the matter of memory without metadata. Pre-internet life left fewer searchable records. Fewer photographs, no social media archives, no instant documentation. Memories from that period feel less indexed and more dreamlike as a result. Dreamlike states carry an uncanny quality almost by default — the sense that something is present but not quite accountable. Mark Fisher called this hauntology: the persistence of lost futures, visions imagined in the past that never materialised. Media from the period can feel like a signal from an abandoned timeline. That dislocation sits somewhere between nostalgia and dread, and I'm not convinced the two are as different as we'd like them to be.

I keep a folder of screenshots from 1970s Open University broadcasts. I'm not sure why. The lecturers stand in front of beige walls explaining thermodynamics in voices that sound like they're narrating the end of the world. Nothing about the content is threatening. Everything about the atmosphere is.