One of the enduring myths of audiophilia is the concept of the “end-game” system. No matter the quality of the system you have, there always seems to be a missing piece.

“If I can only add a (insert next-level piece of equipment here) to my system I will be forever content and can die a happy (wo)man.”

I have achieved what I thought was end-game several times – for a few days or several weeks – but each time a new siren song emerges. Klipsch Forté, Lenco L70 and Sansui AU-777 are calls of the past now silenced. Current objects calling with varying urgency include blue-baffle JBLs, concentric-driver Tannoys, Thorens TD-124 or 125, and the Sansui AU-111.

With a planned move back to Japan in two or three years, the question of whether any of these will be achieved is on hold. I am in an enforced end-game state, knowing I will sell everything I have before I move, and will start again when I land on the other side of the Pacific.

Upgrading my Cleaning Game

This continual desire to upgrade and improve the system is about more than just equipment. It applies also to furniture, storage, cables, accessories, and record cleaning.

Four years ago, as a fairly new vinyl collector with a few hundred records in my collection, I wrote about budget cleaning solutions that did the job and kept the wallet (and Mrs. Audiolove) happy.

Today, with nearly 1,800 albums, I’ve become pickier about cleaning; I won’t cut corners and am more willing to drop some coin on quality. I’ve replaced records that didn’t cut the mustard (including grey-market EU “Public Domain” reissues) with modern audiophile or early pressings, and I want to show these the respect they deserve so they play clean and clear for my remaining decades.

And so ladies and gentlemen I present my 2026 cleaning arsenal, with medium-of-choice dependent on dirt levels and apparent vinyl condition.

Dust and Static – Ramar Berlin Record Brush (Red)

Ramar record brushes are made with a combination of carbon fibre (six double rows) and two rows of goat hair to penetrate every groove and remove fine dust and larger dirt particles while dissipating electrostatic charges.

The body and protective case of my brush are made with cherry wood, fashioned from a single block. The case protects the brush fibres from damage and dirt. A range of handle and case styles are available, including other wood variants and metal finishes.

Brushes come with a natural felt cleaning pad for removing any dust or dirt caught between the fibres during use, and Ramar offers after-market renewal and repair services.

The Ramar brush replaced my $20 Audio-Technica anti-static brush, and the difference was obvious. It feels far more substantial and better made, and it delivers noticeably better dust and static removal. At €360, it had better be excellent, and it is. For most records, this is the only cleaning solution I need, which makes the expense easier to justify.

Minor Dirt – GrooveWasher Hardwood Record Cleaning Kit

GrooveWasher makes a variety of cleaning accessories and kits, including record and stylus fluids and brushes. They also make anti-static record sleeves.

This was my first cleaner “upgrade,” replacing the cheapie DiscWasher. The look and feel of the two cleaners are similar, but the heft of the GrooveWasher’s wooden handle and the cleaning performance of the black Terry microfibre pad are a step up in quality.

The Hardwood Kit costs around $50 and comes with a 4 oz spray bottle of G2 high tech record cleaning fluid. This combination effectively removes minor grime like errant fingerprints or other sticky dirt that the Ramar can’t tackle. I use this brush for a first clean of used records that look to be in very good condition, and every few plays for records I’ve had in the collection for some time.

Over time the plush terry cloth pad does wear down and flatten out, and I’m currently eyeing up a replacement pad. The cleaning pads are easily removable, and replacements adhere solidly to the wood handle by way of Velcro fasteners.

Embedded Dirt and Persistent Crackle – HumminGuru HG01 Ultrasonic

The HumminGuru HG01 ultrasonic replaced a Spin Clean Mk. II a couple of years ago. The Spin Clean manual water-and-brush system worked well initially, but for some reason had begun jamming, even after replacing the brushes. While it still did a good job of cleaning, using it became frustrating and an upgrade was called for.

I investigated various vacuum and ultrasonic systems and the HumminGuru seemed to offer a good balance between results and financial outlay (Yes, I’m willing to drop some coin, but my pockets are not bottomless).

Ultrasonic cleaning uses high-frequency sound waves to create cavitation bubbles in water, requiring zero physical contact with the record. The HumminGuru automates the process of both cleaning and drying. All that’s required of the user is to add distilled water to the bath area on top, insert the record vertically into the cleaning slot, and hit a few buttons to set cleaning and drying time and start the cleaning process.

The record spins in the bath for several minutes while the ultrasonics remove dirt, the bath auto-drains into a removable water receptacle at the bottom of the machine, and then dual warm air fans dry the record. After 7-10 minutes, remove the record and it’s ready to store or play.

Water can be re-used to clean multiple records, and HumminGuru recommends using a few drops of alcohol-free cleaning formula to reduce surface tension and facilitate better penetration into the record grooves and to enhance drying in humid environments. Adaptors are available for 7” and 10” records.

I’ve been very impressed with results from the HumminGuru, with big improvements in grading quality post cleaning. I also noted improvements for records already cleaned with the old Spin Clean (which was no slouch, even with the jamming issues I experienced).

At time of writing the HumminGuru HG01 costs around $400 direct from the manufacturer, which is down significantly from my original purchase price.

Since I purchased my HG01, Humminguru has introduced an advanced model (the Nova) which features quieter cleaning, faster drying and automatic adjustment for different record sizes. The Nova runs about $700.

HumminGuru also introduced an Automatic Water Dispenser ($159.99 at Amazon) unit which eliminates my one gripe with the HG01 (and Nova), that being the somewhat inelegant process of removing the water receptacle to refill the water bath. The water dispenser costs around $160, and is very definitely on my to-buy list (and an exception to enforced end-game status).

The Bottom Line

And there we have the three arrows in my cleaning quiver, with all needs and bases covered. A final mention goes to the Ramar brush, which elicits frequent comments on Instagram regarding what many see as an exorbitant price (about the same as the HumminGuru).

No, there are no moving parts. Yes, it’s just a brush. But what a brush! As mentioned, this is my main cleaner and so on a per-use basis the cost is not so high. Factor in craftsmanship and precision – hand crafted, grain-matched wooden handle and holder, exquisitely layered brush fibres – and an obviously time-intensive build process, and it all makes sense. In my mind the juice is worth the squeeze.

What do you think? Leave your thoughts in the comments, or message me on the ‘Gram at @audioloveyyc.

Where to buy:

Related Reading:

Lachy Groom, one of Silicon Valley’s most closely watched solo investors, decided to back Indian startup Pronto just 20 minutes into his first meeting with its 24-year-old founder.

The meeting, which took place in February through a mutual connection, led to Groom investing $20 million in Pronto as an extension of its Series B round, valuing the startup at $200 million after the investment — double its valuation just over two months earlier, as TechCrunch had previously reported. The deal came together within weeks, bringing the solo investor on board as the Bengaluru-based startup expands to meet growing demand for on-demand home services in India.

Groom said he was drawn to Pronto’s ambition to build what he called the world’s largest platform for organizing domestic labor, starting with India’s vast and largely unstructured workforce. “The work underneath that is genuinely hard, and most attempts in adjacent categories have struggled with the operational discipline,” he said, adding that Pronto founder Anjali Sardana (pictured above) and her team were operating “at a level I haven’t seen elsewhere in this space.”

Before founding Pronto in 2025, Sardana worked at Bain Capital and venture firm 8VC, where she gained early exposure to investing and high-growth startups. The startup connects households with workers for everyday tasks such as cleaning and basic home services.

The introduction was arranged through Paul Hudson, founder of Glade Brook Capital, who connected Groom and Sardana during her trip to San Francisco earlier this year. Glade Brook has backed startups founded by both: Pronto, which Sardana leads, and Physical Intelligence, where Groom is a co-founder. Hudson and Groom have also backed Indian quick-commerce startup Zepto.

Sardana said Groom’s investment approach is heavily founder-driven. “He indexes two things. One is the founder, and that’s 95% of it. If he loves the founder, then he will invest,” she told TechCrunch, adding that the rest comes down to the scale and potential of the business.

Groom’s bet comes as a clutch of startups in India race to build instant home services platforms, a category that is seeing rapid adoption among urban households as more consumers turn to on-demand help for everyday tasks.

The opportunity is significant. A recent Bank of America note, reviewed by TechCrunch, estimates the instant home services market in India could grow into a $15 billion to $18 billion industry by the end of the decade, as companies including Pronto, Snabbit, and Urban Company’s InstaHelp compete for share in the fast-growing category.

Competition is intensifying, with heavy capital inflows and aggressive pricing, particularly to attract first-time users. Bank of America estimates that Snabbit and Urban Company’s InstaHelp each account for about 40% of the market, while Pronto has around a 20% share, even as it scales rapidly. The category is expected to remain “burn-heavy” over the next two to three years.

Despite trailing larger rivals, Pronto has been scaling rapidly, growing from around 18,000 bookings a day to 26,000 in just over a month. The startup is focused on driving repeat usage, betting that turning occasional demand into frequent, habit-driven usage will be key to winning the category, with its top 10% of users accounting for about 40% of bookings.

This growth has also brought challenges, particularly in building out supply. Pronto has expanded its network of service workers to 6,500, up from 1,440 in January. But Sardana said demand continues to outpace supply, making forecasting and capacity management key challenges as the startup grows.

from the MAGA-just-means-hating-American dept

The content of their character was never up for consideration. Under Donald Trump, the only thing that matters is the color of their skin. That’s why almost every single person granted asylum since Trump took office has been white. That’s why Trump has been asking (out loud!) why we keep getting migrants from “shithole” countries (like those located in South America, Africa, and Latin America) rather than blond haired, blue eyed expats from Scandinavian countries whose residents’ lives would become noticeably worse if they chose to move to the US.

The president wraps himself in the flag, delivers a lot of garbled Team USA jingoism, and routinely proclaims we have the best military in the world. But even the people most directly responsible for keeping the US on top of the military game aren’t allowed to remain here if they’re not white.

Jose Serrano, an active duty soldier who served three tours in Afghanistan, said immigration agents arrested his wife April 14 as they attended an appointment with immigration services to take steps toward her permanent residency.

“A person opened the door, escorted us through the hallway, and at the end of the hallway, my wife got arrested,” Serrano said. “Arrested without any order, any warrant … They took away my wife. They don’t tell me anything.”

On top of all this awfulness, this incident shows ICE isn’t actually shifting away from immigration court arrests despite (1) officials saying otherwise, and (2) more importantly, ICE itself supposedly letting officers know that court arrests like these are not allowed under current ICE policy.

The regular awfulness is this: the Trump administration is willing to attack its own military if it means racking up a few more arrests and deportations:

[L]ast April, DHS eliminated a 2022 policy that considered military service of an immediate family member to be a “significant mitigating factor” in deciding whether or not to pursue immigration enforcement. The administration’s new policy states that “military service alone does not exempt aliens from the consequences of violating U.S. immigration laws.”

It’s not just this nation’s relationship with its own military that’s being permanently damaged by Trump’s bigoted war on non-white people. It’s also any future relationships we might have in countries where we’re engaged in combat. When the US began its full withdrawal from Afghanistan, it promised protections to Afghans who worked with the military to provide intelligence or otherwise aided in the US in the decades-long war.

That’s all being tossed aside by Trump because he and his administration simply just don’t like non-white people.

After halting a U.S. resettlement program for Afghans who helped the American war effort, President Trump is in talks to send as many as 1,100 of them to the Democratic Republic of Congo, an aid worker briefed on the plan said Tuesday.

The group includes interpreters for the U.S. military, former members of the Afghan Special Operations forces and family members of American service members. More than 400 children are among them.

Advertisement

The Afghans have been living in limbo in Qatar for over a year. They were taken there after being evacuated by the United States for their own safety because they supported American forces during the war against the Taliban that began in 2001.

Thanks for your help. Now, go fuck yourselves. That’s the message the US is sending to people who aided the US during this war. It’s the kind of message that isn’t likely to score it any allies as it resumes hostilities in the Middle East.

This report says Trump is “in talks” with DRC to pursue this “resettlement” of Afghan allies — one the administration pursues despite the protests of the people who risked their own lives to assist the US during the Afghanistan war.

It’s hard to believe Trump is actually engaged in anything. DRC already has a refugee problem of its own.

More than 600,000 refugees, mostly from the Central African Republic and Rwanda, are currently in Congo, according to the United Nations. Human rights activists say that the country is not equipped to take in more in the midst of fighting with neighboring Rwanda that has displaced even more people because of attacks on refugee camps.

On top of this, many Afghan allies already have family members living in the United States due to previous efforts made by the Biden administration to protect those who aided the US. This forced resettlement in, well, pretty much any African country that agrees to take them divides even more families. It also demonstrates the United States is not to be trusted when it offers favors in return for assistance. All it takes is an election cycle to roll back guarantees and turn trusted allies into just another set of people being moved from “shithole country” to “shithole country” by a bunch of bigots who would rather destroy America than allow any more non-white people to become residents of what used the be the world’s “melting pot.”

At least for now, Trump has seemingly found a willing dumping ground for people he doesn’t want in this country:

On April 17, the U.S. government deported 15 people to the capital of the Democratic Republic of Congo, a deeply impoverished African country that’s been scarred by years of conflict.

The group—comprising men and women from Colombia, Ecuador and Peru—is the first to arrive as part of a secretive migration deal brokered with the Trump administration.

“They took us, they put us on a plane, and they chained us by our hands and feet,” said one Colombian man, sitting on a plastic chair in a shabby hotel near Kinshasa’s airport. The deportees didn’t know their final destination until they were on the plane, he added.

Advertisement

Like El Salvador, I’m sure the DRC is more than happy to take our money to take some people off our hands. And like El Salvador, I’m sure the DRC government doesn’t actually care what happens to any of these people being shoved out of DHS charter flights like so much human refuse. If the US can’t be bothered to care, why should some third party in a developing nation do anything more than allow planes to land so long as the checks keep clearing?

This is what America is now: a place where human rights, civil liberties, and basic human morality are no longer weaved into the fabric of the nation. America is no longer the world’s policeman. It is now the world’s corrupt, racist sheriff.

Filed Under: afghanistan, bigotry, cruelty, dhs, ice, mass deportation, pete hegseth, trump administration, us military

Looking for the most recent regular Connections answers? Click here for today’s Connections hints, as well as our daily answers and hints for The New York Times Mini Crossword, Wordle and Strands puzzles.

Today’s Connections: Sports Edition is a tough one, but fun for movie fans. If you’re struggling with the puzzle but still want to solve it, read on for hints and the answers.

Connections: Sports Edition is published by The Athletic, the subscription-based sports journalism site owned by The Times. It doesn’t appear in the NYT Games app, but it does in The Athletic’s own app. Or you can play it for free online.

Read more: NYT Connections: Sports Edition Puzzle Comes Out of Beta

Hints for today’s Connections: Sports Edition groups

Here are four hints for the groupings in today’s Connections: Sports Edition puzzle, ranked from the easiest yellow group to the tough (and sometimes bizarre) purple group.

Yellow group hint: If the shoe fits.

Green group hint: Fore!

Blue group hint: Take me out to the ball game.

Purple group hint: Cinema titles.

Answers for today’s Connections: Sports Edition groups

Yellow group: Sneaker brands.

Green group: Golf courses to host the U.S. Open.

Blue group: Famous nicknames for MLB teams.

Purple group: Movies that contain an NFL team name.

Read more: Wordle Cheat Sheet: Here Are the Most Popular Letters Used in English Words

What are today’s Connections: Sports Edition answers?

The yellow words in today’s Connections

The theme is sneaker brands. The four answers are Converse, New Balance, Saucony and Under Armour.

The green words in today’s Connections

The theme is golf courses to host the U.S. Open. The four answers are Pebble Beach, Shinnecock Hills, Torrey Pines and Winged Foot.

The blue words in today’s Connections

The theme is famous nicknames for MLB teams. The four answers are Amazin’ Mets, Big Red Machine, Gas House Gang and Murderers’ Row. 

The purple words in today’s Connections

The theme is movies that contain an NFL team name. The four answers are Little Giants, Raiders of the Lost Ark, Remember the Titans and The Bad News Bears.

A little-known Miami-based startup called Subquadratic emerged from stealth on Tuesday with a sweeping claim: that it has built the first large language model to fully escape the mathematical constraint that has defined — and limited — every major AI system since 2017.

The company claims its first model, SubQ 1M-Preview, is the first LLM built on a fully subquadratic architecture — one where compute grows linearly with context length. If that claim holds, it would be a genuine inflection point in how AI systems scale. At 12 million tokens, the company says, its architecture reduces attention compute by almost 1,000 times compared to other frontier models — a figure that, if validated independently, would dwarf the efficiency gains of any existing approach.

The company is also launching three products into private beta: an API exposing the full context window, a command-line coding agent called SubQ Code, and a search tool called SubQ Search. It has raised $29 million in seed funding from investors including Tinder co-founder Justin Mateen, former SoftBank Vision Fund partner Javier Villamizar, and early investors in Anthropic, OpenAI, Stripe, and Brex. The New Stack reported that the raise values the company at $500 million.

The numbers Subquadratic is publishing are extraordinary. The reaction from the AI research community has been, to put it mildly, mixed — ranging from genuine curiosity to open accusations of vaporware. Understanding why requires understanding what the company claims to have solved, and why so many prior attempts to solve the same problem have fallen short.

The quadratic scaling problem has shaped the economics of the entire AI industry

Every transformer-based AI model — which includes virtually every frontier system from OpenAI, Anthropic, Google, and others — relies on an operation called “attention.” Every token is compared against every other token, so as inputs grow, the number of interactions — and the compute required to process them — scales quadratically. In plain terms: double the input size, and the cost doesn’t double. It quadruples.

This relationship has shaped what gets built and what doesn’t. The industry standard is 128,000 tokens for many AI models and up to 1 million tokens for frontier cloud models such as Claude Sonnet 4.7 and Gemini 3.1 Pro. 

Even at those sizes, the cost of processing long inputs becomes punishing. The industry built an elaborate stack of workarounds to cope. RAG systems use a search engine to pull a small number of relevant results before sending them to the model, because sending the full corpus isn’t feasible. Developers layer retrieval pipelines, chunking strategies, prompt engineering techniques, and multi-agent orchestration systems on top of models — all to route around the fundamental constraint that the model itself can’t efficiently process everything at once.

Subquadratic’s argument is that these workarounds are expensive, brittle, and ultimately limiting. As CTO Alexander Whedon told SiliconANGLE in an interview, “I used to manually curate prompts and retrieval systems and evals and conditional logic to chain together the workflows. And I think that that is kind of a waste of human intelligence and also limiting to the product quality.”

Subquadratic’s fix is deceptively simple: stop doing the math that doesn’t matter

The company’s approach, called Subquadratic Sparse Attention or SSA, is built on a straightforward premise: most of the token-to-token comparisons in standard attention are wasted compute. Instead of comparing every token to every other token, SSA learns to identify which comparisons actually matter and computes attention only over those positions. Crucially, the selection is content-dependent — the model decides where to look based on meaning, not on fixed positional patterns. This allows it to retrieve specific information from arbitrary positions across a very long context without paying the quadratic tax.

The practical payoff scales with context length — exactly the inverse of the problem it’s trying to solve. According to the company’s technical blog, SSA achieves a 7.2x prefill speedup over dense attention at 128,000 tokens, rising to 52.2x at 1 million tokens. As Whedon put it: “If you double the input size with quadratic scaling laws, you need four times the compute; with linear scaling laws, you need just twice.” The company says it trained the model in three stages — pretraining, supervised fine-tuning, and a reinforcement learning stage specifically targeting long-context retrieval failures — teaching the model to aggressively use distant context rather than defaulting to nearby information, a subtle failure mode that quietly degrades performance in existing systems.

Three benchmarks paint a strong picture, but what they leave out may matter more

On the surface, SubQ’s benchmark numbers are competitive with or superior to models built by organizations spending billions of dollars. On SWE-Bench Verified, it scored 81.8% compared to Opus 4.6’s 80.8% and DeepSeek 4.0 Pro’s 80.0%. On RULER at 128,000 tokens, a standard benchmark for reasoning over extended inputs, SubQ scored 95% — edging out Claude Opus 4.6 at 94.8%. On MRCR v2, a demanding test of multi-hop retrieval across long contexts, SubQ posted a third-party verified score of 65.9%, compared with Claude Opus 4.7 at 32.2%, GPT-5.5 at 74%, and Gemini 3.1 Pro at 26.3%.

But several details warrant scrutiny. The benchmark selection is narrow — exactly three tests, all emphasizing long-context retrieval and coding, the precise tasks SubQ is designed for. Broader evaluations across general reasoning, math, multilingual performance, and safety have not been published. The company says a comprehensive model card is “coming soon.”

According to The New Stack, each benchmark model was run only once due to high inference cost, and the SWE-Bench margin is, as the company’s own paper acknowledges, “harness as much as model.” In benchmark methodology, single runs without confidence intervals leave room for variance. There is also a significant gap between SubQ’s research results and its production model. On MRCR v2, the company reported a research score of 83 — but the third-party verified production model scored 65.9. That 17-point gap between the lab result and the shipping product is notable and largely unexplained.

Subquadratic also told SiliconANGLE that on the RULER 128K benchmark, SubQ scored 95% accuracy at a cost of $8, compared with 94% accuracy and about $2,600 for Claude Opus — a remarkable cost claim. But the company has not publicly disclosed specific API pricing, making it impossible to independently verify the cost-per-task comparisons.

The AI research community’s verdict ranges from ‘genuine breakthrough’ to ‘AI Theranos’

Within hours of the announcement, the AI research community erupted into a debate that crystallized around a single question: Is this real?

AI commentator Dan McAteer captured the binary mood in a widely shared post: “SubQ is either the biggest breakthrough since the Transformer… or it’s AI Theranos.” The comparison to the infamous blood-testing fraud company may be unfair, but it reflects the scale of the claims being made. Skeptics zeroed in on several pressure points. Prominent AI engineer Will Depue initially noted that SubQ is “almost surely a sparse attention finetune of Kimi or DeepSeek,” referring to existing open-source models.

Whedon confirmed this on X, writing that the company is “using weights from open-source models as a starting point, as a function of our funding and maturity as a company.” Depue later escalated his criticism, writing that the company’s O(n) scaling claims and the speedup numbers “don’t seem to line up” and called the communication “either incredibly poorly communicated or just not real.”

Others raised structural questions. One developer noted that if SubQ truly reduces compute by 1,000x and costs less than 5% of Opus, the company should have no trouble serving it at scale — so why gate access through an early-access program? Developer Stepan Goncharov called the benchmarks “very interesting cherry-picked benchmarks,” while another commenter described them as “suspiciously perfect.”

But not everyone was dismissive. AI researcher John Rysana pushed back on the Theranos framing, writing that the work is “just subquadratic attention done well which is very meaningful for long context workloads,” and that “odds of it being BS are extremely low.” Linus Ekenstam, a tech commentator, said he was “extremely intrigued to see the real-world implications” particularly for complex AI-powered software.

Magic.dev made strikingly similar claims two years ago — and then went quiet

Perhaps the most pointed critique of SubQ’s launch comes not from its specific claims but from recent history. Magic.dev announced a 100-million-token context-window model in August 2024, with a claimed 1,000x efficiency advantage, and raised roughly $500 million on the strength of those claims. As of early 2026, there is no public evidence of LTM-2-mini being used outside Magic.

The parallels are uncomfortable. Both companies claimed massive context windows. Both touted roughly 1,000x efficiency gains. Both targeted software engineering as their primary use case. And both launched with limited external access.

The broader research landscape reinforces the caution. Kimi Linear, DeepSeek Sparse Attention, Mamba, and RWKV all promised subquadratic scaling, and all faced the same problem: architectures that achieve linear complexity in theory often underperform quadratic attention on downstream benchmarks at frontier scale, or they end up hybrid — mixing subquadratic layers with standard attention and losing the pure scaling benefits.

A widely cited LessWrong analysis argued that these approaches “are all better thought of as ‘incremental improvement number 93595 to the transformer architecture’” because practical implementations remain quadratic and “only improve attention by a constant factor.”

Subquadratic is directly aware of this history. Its own technical blog specifically addresses each prior approach — fixed-pattern sparse attention, state space models, hybrid architectures, and DeepSeek Sparse Attention — and argues that SSA avoids their tradeoffs. Whether it actually does remains an empirical question that only independent evaluation can settle.

A five-time founder, a former Meta engineer, and $29 million to prove the doubters wrong

The team behind the claims matters in evaluating them. CEO Justin Dangel is a five-time founder and CEO with a track record across health tech, insurancetech, and consumer goods, and his companies have scaled to hundreds of employees, attracted institutional backing, and reached liquidity. CTO Alexander Whedon previously worked as a software engineer at Meta and served as Head of Generative AI at TribeAI, where he led over 40 enterprise AI implementations.

The team includes 11 PhD researchers with backgrounds from Meta, Google, Oxford, Cambridge, ByteDance, and Adobe. That is a credible collection of talent for an architecture-level research effort. But neither co-founder has published foundational AI research, and the company has not yet released a peer-reviewed paper. The technical report is listed as “coming soon.”

The funding profile is unusual for a company making frontier AI claims. Subquadratic raised $29 million at a reported $500 million valuation — a steep price for a seed-stage company with no publicly available model, no peer-reviewed research, and no disclosed revenue. The investor base, led by Tinder co-founder Mateen and former SoftBank partner Villamizar, skews toward consumer tech and growth investing rather than deep technical AI research. The company is not open-sourcing its weights but plans to offer training tools for enterprises to do their own post-training, and has set a 50-million-token context window target for Q4.

The real test for SubQ isn’t benchmarks — it’s whether the math survives independent scrutiny

Strip away the marketing language and the social media drama, and the underlying question Subquadratic is asking is genuinely important: Can AI systems break free of quadratic scaling without sacrificing the quality that makes them useful?

The stakes are enormous. If attention can be made truly linear without degrading retrieval and reasoning, the economics of AI shift fundamentally. Enterprise applications that today require elaborate retrieval pipelines — processing entire codebases, contracts, regulatory filings, medical records — become single-pass operations. The billions of dollars currently spent on RAG infrastructure, context management, and agentic orchestration become partially redundant. 

Whedon’s willingness to engage publicly with technical criticism — posting a technical blog within hours of pushback — suggests a team that understands it needs to show its work, not just describe it. And to its credit, the company acknowledged openly that it builds on open-source foundations and that its model is smaller than those at the major labs.

Every frontier model in 2026 advertises a context window of at least a million tokens, but almost none of them are actually great at making use of all that information. The gap between a nominal context window and a functional one — between what a model accepts and what it reliably reasons over — remains one of the most important unsolved problems in AI. Subquadratic says it has closed that gap. If independent evaluation confirms that claim, the implications would ripple far beyond a single startup’s valuation. If it doesn’t, the company joins a growing list of long-context promises that sounded revolutionary on launch day and unremarkable six months later.

In computing, every fundamental constraint eventually falls. When it does, the breakthrough never comes from the direction the industry expected. The question hanging over Subquadratic is whether a team of 11 PhDs and a $29 million seed round actually found the answer that has eluded organizations spending thousands of times more — or whether they just found a better way to describe the problem.

There’s an app for nearly every imaginable user and use case these days, but one thing they all have in common is that they’re centered around one device: the smartphone.

That changes today as Hugging Face, the 10-year-old New York City startup best known for being the go-to place online to host and use cutting-edge, open-source AI models, agents and applications, launches a new App Store for Reachy Mini, its low-cost ($299) open-source physical robot that debuted back in July 2025 (itself the fruit of Hugging Face’s acquisition of another startup, Pollen Robotics).

The new Hugging Face Reachy Mini App Store already hosts a library of over 200 community-built applications, and Reachy Mini owners will be able to download any of these free of charge to start (unlike smartphone apps, there’s no monetization option for app creators on this store — yet).

The Reachy Mini App Store will also offer Reachy Mini owners — around 10,000 units have been sold so far since last year — an easy means of building their own custom apps for the tiny, stationary desktop robot with built-in camera eyes, speaker, and microphone, via Hugging Face’s existing, AI-powered agent called “ML Intern.”

The significance lies not just in the hardware, but in the removal of the “roboticist” barrier; for the first time, individuals without a background in engineering or coding are shipping functional robotics software in under an hour.

“Anyone can build the apps,” said Clément Delangue, CEO and co-founder of Hugging Face, in a video interview with VentureBeat. “My intuition is that more and more [AI] model builders will release on Reachy Mini as a way to test the robotics ability of new models.”

Make robots as accessible to laypeople as PCs and smartphones

The technical bottleneck in robotics has historically been the scarcity of high-quality training data.

While Large Language Models (LLMs) have mastered general-purpose coding by training on massive repositories like Microsoft’s GitHub, the volume of code specific to robotics remains “tiny” by comparison (though Github does contain likely the largest existent, publicly accessible library of robotics code to date, with more than 17,000 different repositories or “repos” dedicated to the field).

This lack of data has meant that, until now, AI agents were relatively poor at understanding the physical abstractions and firmware requirements of hardware.

Hugging Face’s solution is an agentic toolkit that acts as an intermediary. Rather than forcing a user to learn a specific robotics SDK or master the nuances of a robot’s firmware, the toolkit allows a user to describe a desired behavior in plain English—for instance, “wave when someone says good morning”.

An AI agent then handles the heavy lifting: it writes the code, tests it against the robot’s specific constraints, and ships the final package

“Historically, it’s been extremely hard,” Delangue told VentureBeat of building robotics applications. “But we’ve worked really hard on the topic with a mix of open sourcing everything we do, working on the right abstractions for robotics, and making it easier for agents to understand and use it.”

The platform is model-agnostic, supporting a wide range of leading intelligence engines. Users can build apps using Hugging Face’s own ML Intern agent or leverage external models including GPT-5.5, Claude Opus 4.6, Kimmy 2.6, Mini Max GM5, and Deep Sig V4 Pro.

For real-time interaction, the official conversation apps utilize OpenAI Realtime and Gemini Live. By providing these high-level abstractions, Hugging Face has collapsed the traditional “integration weeks” of robotics work into a process that takes minutes.

Low-cost Reachy Mini is a hit

In order to take advantage of the new Hugging Face Reachy Mini App Store, users are encouraged to purchase Reachy Mini, a cute desktop robot Hugging Face launched back in July 2025 as an affordable, open-source alternative to the existing, commercially available robots from the likes of Boston Dynamics, whose infamous Spot robot dog retails for around $70,000. Even Chinese competitors start at $1,900+.

In contrast, the Reachy Mini is accessibly priced for hobbyists and developers. It comes in two variants:

• Reachy Mini Lite ($299 plus shipping): A tethered version that connects via USB and uses an external computer for processing.

• Reachy Mini Wireless ($449 plus shipping): A standalone version featuring an on-board Raspberry Pi CM 4 and Wi-Fi connectivity.

Delangue said that of the 10,000 Reachy Mini units sold so far, 3,000 were sold in just the past two weeks. Hugging Face expects to ship another 1,000 units within the next 30 days.

Even those who don’t own a Reachy Mini can still develop apps for it, however, using the Reachy Mini App Store and the Reachy App, which contains a 3D simulation of the robot and its responses.

The App Store itself is hosted on the Hugging Face Hub. It functions much like a standard software repository but for hardware behaviors:

• Search and Install: Users can find apps, click a button, and install them directly to their robot.

• Forkability: Every app is “forkable,” meaning a user can duplicate an existing app and ask an AI agent to modify it (e.g., “make it answer in French”).

• Simulation Mode: Crucially, the store includes a browser-based simulator. This allows users who do not own a physical Reachy Mini to build, test, and play with the catalog in a virtual environment.

Both are part of Hugging Face’s ongoing “Le Robot” effort — a project that began in 2024 with Hugging Face researchers specializing in robotics and AI developing and publishing on the web their own open-source code, tutorials, and hardware to make robotics development more accessible to a wider audience.

And unlike Github, which is designed for a developer audience, the Hugging Face Reachy Mini App Store is designed for robot owners and users who may have no technical experience or training whatsoever.

Continuing with the open-source ethos and practice

Hugging Face’s strategy is rooted in the belief that closed-source hardware and software are “almost impossible” to build for at scale.

Delangue notes that closed systems prevent the training of agents and limit the ability of the community to innovate. Consequently, the entire Reachy Mini platform is open-source.

This open licensing model has two primary implications for the ecosystem:

1. Accelerated Development: Because the code is public and integrated with the Hugging Face ecosystem via “Spaces,” Hugging Face’s feature for hosting AI-powered web apps launched in 2021, agents can more easily learn how to interact with the hardware.

2. Community Sovereignty: Apps are not locked behind a proprietary wall. Currently, all 200+ apps on the store are free, though the platform’s foundation on “Spaces” provides the flexibility for creators to potentially monetize their work in the future.

“For the moment, all the apps are free,” Delangue noted. “It’s flexible, it’s built on [Hugging Face] Spaces, so at some point maybe people are going to make them paid.”

Robotics enters its accessible hobbyist era

Hugging Face’s Reachy Mini App Store is launching with 200 apps already available.

So who built them, and how did they do it without this platform existing prior?

Delangue told VentureBeat that more than 150 different creators have contributed to the store, most of whom had never written a line of robotics code before.

Yet, they have been able to do so thanks to Hugging Face’s ML Intern and Github. The new Hugging Face Reachy Mini App Store now puts the tools and existing apps into one place for easier accessibility.

Delangue was keen to highlight one of the early Reachy robotics app developers in particular to VentureBeat: Joel Cohen, a 78-year-old retired marketing executive.

Cohen, who is colorblind and has no technical background, spent two weeks assembling his Reachy Mini Lite (a task that usually takes three hours). Despite these physical challenges, he used an AI agent to build a “VP of Future Thinking” facilitator for his Zoom-based CEO peer groups. The app enables the robot to:

• Greet 29 members by name.

• Fact-check discussions in real-time.

• Summarize key themes and push back on surface-level answers.

“I built this by describing what I needed in plain English,” Cohen stated in a press release provided to VentureBeat ahead of the launch. “No SDK. No robotics background. No developer experience”.

Other community-driven applications include:

• Emotional Damage Chess: A robot that plays chess and mocks the user’s blunders.

• Reachy Phone Home: An anti-procrastination tool that detects when a user picks up their phone and tells them to get back to work.

• Language Tutor: A physical companion that listens to speech and corrects accents.

• F1 Race Commentator: A desk companion that calls Formula 1 races live as they happen.

Delangue himself related to VentureBeat that in only a few hours, he built an app for his own Reachy Mini robot at the Hugging Face Miami office to have the robot act as a receptionist.

“It basically does face recognition to detect when you arrive in the office, and then it looks at you and onboards you,” Delangue related. “It says, ‘Hey, welcome to the office. Who are you here to see?’ Then it sends me a message: ‘Carl just arrived at the office. He’s here to meet you, and for these reasons.’ It works a little bit as my welcoming booth at the office, and it took me less than two hours to build that.”

Even for an experienced founder and developer as Delangue, building apps for a robot was out of the question until the combination of Reachy Mini and ML Intern.

“For me, it would have been impossible,” the Hugging Face CEO said. “If you weren’t a robotics developer, it probably would have been impossible, or it would have taken a few months.”

Democratizing robotics

The launch of the agentic App Store signals a fundamental shift in how we interact with machines. For sixty years, the field was gated by the requirement for deep technical expertise.

By combining low-cost open hardware with the reasoning capabilities of modern AI agents, Hugging Face is moving toward a future where the hardware is a commodity and the behavior is limited only by what a user can describe.

As Delangue noted during the launch, the goal was to provide a platform for people who “want to get into robotics but don’t have the hardware or the skills”.

With nearly 10,000 robots now “in the wild” and a burgeoning store of agent-written apps, the Reachy Mini has become the most widely deployed open-source desktop robot in history.

The question is no longer how to build a robot, but what—now that the gate is open—we will ask them to do.