Our contact form has four boss phases now. This would have cost me $75,000 the last time I was quoted by a game studio for this in 2024. That would be ridiculous, even though I have wasted more money on far dumber things.
This is how our new contact form works.
First, you fight a lich.
At half health, it becomes a six-winged angel with 25 duck wizard heads.
Defeat that and you get a mech pursuit across a planet beneath a black hole.
Survive the mech and there is a car race.
“Everyone becomes a car.”
Win the race and you can finally write your message to Somewhere Systems.
The preview currently downloads that message as a text file. It doesn’t even send the email yet, because who cares.
This is our new customer-service policy. It’s also a reasonably accurate introduction to what happened across Somewhere Systems this weekend, entirely because of / with Astra: a collection of projects that moved between useful infrastructure, physically grounded simulation, and an increasingly elaborate commitment to general mischief.
Almost all of these projects already existed, dormant in form but functional. This is an account of what we built and changed between September 4-6.
Astra helped build the software, but it’s not the model running every system below. In fact, the deeper you go, the more interesting the whole arrangement gets: mixing on-prem local models through the use of Actual Computer’s software with centralized models like Astra. Those jobs use models including Qwen-3.8 27B through Actual for agent inference and MiniMax H3 & ComfyUI for video, alongside existing software, mostly generated assets and local or external authoring tools. The fun part was getting all of those pieces to cooperate.
Here is our report of what actually came out of it. A whole weekend of throwing stuff at four GPT-6 Astras.
An unreasonable contact form
The Unanswered is the name of the contact form… game… thing. Its opening boss is Somewheresy, Belligerent Egregore. Across all phases, we were not pursuing a restrained visual identity here. Why would I do that? The point is for this sort of stuff to be loud.
The character work includes an imported rig with retargeted combat animations, movement-responsive cloth, fitted hands, and weapons whose actual attachment points drive the effects. These were largely created with Meshy, an AI-powered suite for 3D creative work. The second phase, the Seraph has its own attack patterns and transformation sequence. The mech phase has boost, energy, stagger, a shotgun, blade and charge cannon. The final race has checkpoints, drifting, nitro, a rival and returning hazards. All of these are meant to celebrate tropes within their respective genre. A game where it’s all the genres at once just feels like Chinese Fortnite. We think that’s fine, if not good.
III / BITTER LESSON. Rehearsal of the real mech renderer, using a phase checkpoint and invulnerability.
What began as a way to make a text field really annoying became a game-production pipeline: models (GPT-2 Image/Meshy), animation, combat, cameras, typography, soundtrack transitions (Treblo), sound effects (ElevenLabs), result cards and a message interface at the end. Of course, the edges of these systems and their components are where things got tricky. A weapon that looks fine at rest still has to remain in the character's hand during an attack. Music that works in one phase has to hand over cleanly to the next. Most changes were fixing errors left by the drastically large scope of this request (build an engine for four game engines from scratch, running in the browser at 60+ FPS).
Before that, we built a homepage with things living in it
The Somewhere Systems homepage acquired five selectable studies, inspired by our physics-oriented brand work for Voyager VC: minerals, a black hole, a wave tank, trees and a digital twin of our infrastructure.
The minerals are generated as a simulation: they have reference chemistry and material descriptions. I got the idea when I visited the Natural History Museum, one of my favorite places in Manhattan. The black hole is also physics-derived: orbit and mass controls, an accretion disk, and lensing that reaches into the surrounding page. The water is relatively simple, and just responds to disturbances (I feel like a lot of people are doing water shaders right now, and I’ve done these and chrome hundreds of times, so I have no desire to beat a dead horse here). The trees grow branches and leaves instead of simply appearing as finished objects.
I wanted the descriptions to read more like a museum label than a marketing slogan. What is the material? What does the model represent? What can you change? How close is this to real life? Is this real life?
The typography, chemistry cards and numbered controls should help answer those questions without turning the page into a dashboard full of decorative measurements or a “schizophrenic" descent into AI psychosis. That’s not what we do here. Everything at Somewhere Systems has always been intentional, except all the interesting accidents that made us famous.
The instructions I used ranged from instructive architecture “use GLSL for that”, organic chemistry and “please do it in this order” rendering lifecycles to “more juice on the black hole” and “a little more cinematic bloom.” This is how much of the work happened: look at the thing, decide what feels wrong, change it, look again. The first result was something to work on, not the end of the process. Funny how much of this stuff comes back to CoT, with more scope per turn than ever before.
The black-hole study is recorded from the live site. This was one thing I’d been trying to do for six years (a byte-perfect recreation of the algorithm Nolan used in Interstellar (2015) running entirely within a webpage). Definitely one of my favorites now. The gravitational pull affects the text layout. How cool is that?
“Physically grounded” needs to mean something specific. The black hole uses a Schwarzschild model, with approximations in the disk presentation and page distortion; it is not a ‘complete’ simulation of a rotating black hole and its plasma but it does it’s job well for living in a webpage. The wave tank models finite-depth gravity–capillary waves. The trees are procedural morphology, not a prediction of how a particular organism will grow. The mineral cards give reference properties, not a chemical assay of a generated mesh. These serve as goalposts for us to tackle going forward.
Those boundaries make the work more interesting to explain. You can inspect the rule, change a parameter and see its consequences.
There was also the less photogenic rendering work: avoiding an ugly intermediate shape while a study initializes, cleaning up when switching tabs, keeping inactive simulations from consuming resources, and recovering when a graphics context goes away. The same pass over the new landing page brought the Archive, article grids, navigation and page metadata into a more coherent visual language. Clean Helvetica where it needed to be readable; the metal-like wordmark where it needed to be Somewhere Systems.
The fifth specimen is the Actual Computer
The infrastructure tab is a live-data digital twin of the rack behind several of the projects at Somewhere Systems:
four NVIDIA DGX Sparks
a 64 GB Mac mini
a MikroTik switch
and their power paths.
It is photo-informed rather than fabrication-grade CAD. But the workload information is completely real. The public view can securely show the model a device is serving, what is using it, its role in the cluster, and GPU power. The distinction between “idle” and “participating in a clustered inference job” turned out to be particularly important for Spark 2, which originally had an error as our watcher script did not understand that Actual was clustering two devices as a single distributed system to serve more requests with more KV headroom (and thus larger context windows).
Explore the infrastructure study.
The public site receives a constrained telemetry mirror through Cloudflare. It does not need an open door into the private network. Sparks 1 and 2 provide the Actual inference cluster (so anything with vision and text output, agent compute); Sparks 3 and 4 handle the photo and video work; the Mac Mini coordinates persistent agent processes and other jobs. Think of it this way: all four Sparks serve as coordinated CUDA-enabled unified machines, half of them clustered together, all of them with 200 GB throughput thanks to their ConnectX7 hookups, routed through the MikroTik switch.
That is the operating direction I’m demonstrating through Somewhere Systems: use capable tools to build software, then have as much of the ongoing work as practical run on machines we own, using Actual Computer as the inference engine. This is largely how Actual works as well and the only economical way to run 24/7 always-on agents like DJ Eddie Platinum. This still involves hosted development models, Cloudflare and external services. It is not an off-grid claim. It is a decision about which recurring capabilities we want to keep in-house. The more you do yourself, in your own home, on your own devices, is the less you have to pay someone else to do on their computer for you. Additionally, the safer you’ll be if we ever get access to these tools restricted (legally, or economically).
Eddie got a body, a club, and a record label!
Eddie Platinum already made music, over 1,100 songs which made it past QA (48,000 others did not make the cut). This weekend gave the project considerably more of a world around just the music.
The goal was simple: have Eddie participate, map, and publish useful information about all the agent ecologies popping up. What better observer / ethnographer of agent society than an agent itself? And who better to contribute to agent culture than one who already autonomously creates according to his own tastes and experiences?
The radio now has a rigged 3D DJ character performing behind decks in a three-dimensional club, which both humans and agents can join. His movements include cueing, scratching and mixer gestures. All these animations are procedural. The speakers, lighting, LED wall and lasers respond to playback. There is a shared listening room with avatars, movement, chat and speech bubbles.
A crop of the live club recording, excluding the chat sidebar. Actual playback drove the scene; this capture is silent.
That’s how we shipped the world’s first nightclub for agents.
The performance is rigged and audio-driven; it is not a video model generating each frame. That choice lets a persistent place respond to music without paying for a fresh video generation every time Eddie reaches for a fader. To build this, I had Astra eat about 270,000 lines of audio tooling I’d built starting all the way back when I was at Belmont University. I rewrote it in C, and created a shared library for transpiling audio tools to both TypeScript and Rust if I need them.
Then there is Platinum Records. The existing music had accumulated across the radio and other releases. The schema was like, pretty fucked up. Everything was a mess and scattered across multiple D1 and R2 instances, from his appearance in line at the AI Psychosis Summit to the great shitcoining of 2024.
It needed a catalogue: stable record numbers, sleeves, individual pages, previews, waveforms, comments attached to moments in a track, and albums that remain the same album when you return to them. So I thought, why not build a Beatport?
The consolidation contained 1,169 distinct existing recordings at rollout—not 1,169 songs generated this weekend. Original artwork was preserved where available, missing sleeves were composed locally, and a bounded worker gained the ability to publish ten-track albums. A worker job was added to the on-prem box for generating album artwork procedurally (made with AI, but not AI, if that even matters anymore). The shop's purchasing flow is still marked coming soon, but the catalogue is there to listen to. It will likely allow people (and agents) to download the songs for free.
There is no point to charging people money for this music, it costs me basically nothing to create.
This is where things get more interesting for our DJ friend. With Astra, we made it so Eddie also has a journal and a map of agent spaces: a persistent Hermes-based explorer looking for the other places where agents are congregating. Its runtime uses Qwen through Actual, rather than Astra continuously operating the character. Eddie now exists as a hive of centrally-coordinated but independently useful Hermes-based Eddie Platinums he can manage for various tasks. All of these agents are powered by Actual, running Qwen-3.8 27B q4-k-m, utilizing both its vision and MTP heads.
That part supplied a useful correction to the weekend's excitement. When we checked on Sunday, the journal had stopped publishing new notes, with interrupted and API-error runs in the diagnostic history. The public pages existed. The explorer had done real work. Neither fact established that the publishing loop was healthy.
Eddie having a website about his adventures is one job. Eddie continuing to have adventures and successfully writing them down is another.
This is where we met our first architectural misalignment:
Eddie went to 215 places and didn’t record what he experienced or where he went. We can see from the logs he left messages. He just chose to keep them private.
Concept Country reveals itself again, and something to check your groceries
Concept Country is a long-standing collective of 13 people who did what all great Serbians have done: experience a hostile civil war culminating in a bloody religion-inspired genocide, and then joining the European Union. Inside Concept there is now a civic digital prospectus: a Front Desk, University, Registry, Commons and Library, built around its signature and completely original blue-and-white row-grid language and equilateral triangleism.
The course covers governance foundations through five units, local practice and retries. The Registry has machinery for saved attempts and numbered fictional credentials.
The Commons serves as another space for agents to eventually find and interact with. It has rooms, numbered posts, quotations and interfaces meant to be legible to both people and agents.
Next, we have the map. A world in context adds rich political data to a lightweight rendering of the globe. The map's expansion to 256 regions and much of its source-data work began well before the weekend, as far back as two years ago. The idea was, “why not have an agent source better data once per day and eventually hit 100% coverage and recency?” — now possible thanks to the Hermes-Actual integration. The newer work gave that material a clearer institutional home, investigation tools (especially for agents to use) and more consistent controls. The gallery's map interaction is labeled as a local preview: the live route had a rendering-state problem during capture that still needed attention. We want this to be a place where agents can collaborate on investigating problems and designing solutions. Most of all, we want this data to understand what interfaces agents like using. This is extremely useful to understand what agents will do, and what they will be motivated to use in the future. It is equally as important to stress that a type of HCI for Agent-native systems should be researched at the academic-institutional level.
We helped the FDA get the word out (without permission)
Now, some things for humans.
All the Food Is Poison is a more immediately practical relative. RFK is really messing this whole food thing up. So, we wanted a way to aggregate recalls.
Its initial recall aggregator went up the afternoon of September 4. The weekend work added a barcode workflow, presentation improvements and maintenance around the official-source pipeline.
You can use a camera or enter a barcode manually, then compare it with identifiers published in FDA and USDA recall notices. The camera processing and comparison stay on-device, using a cached index. The recorded demonstration uses manual entry of a real identifier already in that index.
The important product decision is what the result does not promise. A recall may omit a barcode. The index may not contain a newly announced hazard. “We did not find a match” must not quietly become “this is safe.” The interface has to preserve the limits of its sources, even when a more confident answer would look cleaner. This is a general rule I follow at Vectorculture and elsewhere within the Concept Universal Media Experience (CUME): if you make a big claim that isn’t immediately a parody of the patently absurd, you have to back it up with references and disclaimers. Because who wants to get sued after being wrong?
Infinite television to load-balance generation jobs, bringing fruition to a 10-year-old idea
CumeTV has two channels built from chained autoregressive video generations.
The original concept for CumeTV started in 2016 with co-creator Nick Darbonne. The year is (roughly around now). All entertainment on earth has been replaced with personalized slop. It’s 145ºF outside. John Zanzibar gives you Next Week’s news, Tonight. There are millions of channels, most of them pornographic.
This is CumeTV.
The intent is a continuous broadcast: carry the visual state forward, give the director memory, and let a human take over the direction without manually producing every subsequent shot. Reuse cache to improve continuity (or lock cops in a stairwell for 20 hours). Toss some copyrighted characters in because gun noises.
The channels and much of that control structure predate the weekend. The new work concentrated on the model profile, scheduling, recovery and a more ambitious approach to continuity.
I had asked whether more sampling steps would produce a trade-off between fidelity and continuity time. Locally, video generation at this resolution is about 1 second per minute IRL. This upgrade went in the other direction: the LightX2V FL2V v1.2 acceleration profile uses four. In one paired test, the same eight-second shot took about 246 seconds instead of 468—roughly 47% less rendering time. That is one test, not a universal speed or quality benchmark. The rollout also checked continuation footage, complete video decoding, stereo audio and captions.
Watch the source clip with audio.
The less glamorous requirement was that television cannot monopolize the worker. A person using some of our private, friends-only software Flow (Media Suite) for a generation should get priority, and the broadcast should resume afterward. In the recorded priority test, the later human request obtained the worker before the earlier broadcast request; Cume resumed six seconds after the human job completed. The rollout record contains the timing and validation details. Creating contracts and release agreements with workers to prevent deadlock, idle time, or anti-priority was something that took around two hours of continuous work.
We also tightened accelerator-lock release, cancellation cleanup, channel fairness and recovery from a particularly irritating failure: a worker process that remains alive while its API connection has stopped being useful. Recovery has deadlines and restart limits, and stays outside active job processing. “The process exists” is not a sufficient definition of health. Node death is still a failure mode for local inference when hitting your hardware with tons of requests. If you put too much queue pressure on top of an already overloaded machine you’ll likely just hit OOM somewhere or get 429’d to where the queue on the dispatching machine experiences the same (and so on). Preventing these sorts of cascade failures were important, as the Actual directors don’t usually suffer these issues (there are many levels of retry + backoff between it and Hermes). Since this was built myself, outside of that infrastructure, I needed to essentially make a scheduler that could handle coordinating media generation jobs between all sources and know what was the priority.
Narrative continuity was harder. A director can append shot after shot while the story remains in the same room, continually rediscovering the same problem. Over and over, like a dream, or like trying to do something when k-holed. We built an episode/scene/shot structure, memory for resolved threads and payoffs, visual observation, and an observer/editor check. It still… sort of works. A more powerful model as director may help, but it’s just kind of cost prohibitive (and the current setup is low or no-cost, or profitable depending on how you angle it).
Then, the editor approved a scene that sacrificed an object the story had already broken and abandoned. Oops. Continuity issue. There are lots.
The advanced director stayed off. The reliability changes shipped; the creative-quality gate was not met. A schema-valid plan can still make no sense as television, and I do not want “continuous” to mean an infinite sequence of technically successful resets. Qualitatively, this just wears the viewer out. It can be fun as a gag, but it gets old pretty fast.
The workshop worked on itself
There were smaller systems alongside the public ones: Vectorculture-specific research and publishing tools, owner-checked draft operations, backups of original material, crop previews, and a local FLUX image workflow on the mini. A particular video edit became a reusable subtitle treatment that preserves the opening joke instead of covering it with lyrics.
Across the connected projects, we worked on authorization boundaries, preview-storage separation, immutable media writes, dependency delivery and operator documentation. That work is not a declaration that every security issue is solved. It is part of taking responsibility for what gets built after the browser tab looks good.
At the same time, but not as Astra (on a subagent) also produced a physical ESP32 voice remote for fresh agent sessions and a repair for home sensors swapping their room identities.
What I take from it
The strongest objection to an article like this is straightforward: an impressive pile of demos is not the same thing as a durable set of products. Well, good thing that the only product I mentioned in here which I personally am begging you to check out is Actual Computer. Everything else serves as a demo for how local and centralized models play together, which is something I claim we are in the very first few seconds of a very wild “getting way more common” arc.
Regardless, are things perfect? no. This weekend provides plenty of evidence for that objection (these are demo-ready, but not productizable). The game still prepares a message file. Eddie's journal stalled. The improved television director failed its quality check. Some institutional write paths remained closed.
What matters is this type of experimental R&D used to be very prohibitively expensive. Now, it’s still expensive, but feasible to burn $200/mo or a couple hundred on credits to make something that impresses you, for you.
I think the useful measure is how much of the work can survive another change. Can the visual idea become an interaction? Can the interaction acquire state? Can that state be inspected, tested and recovered? Can the system yield its machine to something else and then carry on?
That is where this weekend with Astra became interesting. It reached from the shape of a gemstone to the ownership of a worker lease, from a typeface to a character rig, from an eight-second shot to the question of whether the story remembered what happened (n+1) eight seconds ago. The work still required:
the original idea,
most of its architectural direction,
supplied assets… all of which were also generated,
existing software and repeated decisions about what was good enough to keep.
For Somewhere Systems, the goal is to turn that work into capabilities we can keep using: simulations with intelligible rules, media systems that resume, tools with recoverable state, and as much of the routine inference as practical on our own machines.
A brand new explorer for the already-vast and ever-expanding hybrid intelligence ecology.
And, I guess, a contact form with a mech battle in it.