Lyell

creating/exploring ways to build/create. Touchdesigner tools available on discord+patreon

[Lyell @dotsimulate]

Interactive real-time, streaming music generation is something I’ve been looking forward to for quite a while.

Lyria RealTime (Experimental) by @GoogleDeepMind controlled via #TouchDesigner @1null1 #LOPs

[Lyell @dotsimulate]

Midterms 2026, from 6 PM ET election night to 5:03 PM ET the next day. results from AP via The New York Times.

238,437 census block groups shown as points of light, with spikes for each county's House votes.

[Lyell @dotsimulate]

1:26 – 2:52 AM ET. D+13.7 from 2024, county by county.

[Juan Wulff @juanfwulff]

South Carolina Senate Swing From 2024

🔴 2024: Trump +17.8
🔵 2026: Andrews +4.2

This was massively led by the African American voters who sided with her by 58.6 points, up 16.8 from 2024.

[Lyell @dotsimulate]

12:11 – 1:26 AM ET. 82% in. ~22M votes still out.

[Lyell @dotsimulate]

@dallassnyder thanks. i accidentally deleted the video so here it is again
( 9:56 – 10:54 PM ET. 59% in. )

[Lyell @dotsimulate]

10:54 PM – 12:11 AM ET. 76% in. Senate has flipped.

[Bernie Sanders @BernieSanders]

Democrats will now control the House and Senate and be able to hold Trump accountable. That's great.

At a time of massive income and wealth inequality, we must do more. We must aggressively fight for working families.

No more excuses.

[Lyell @dotsimulate]

6:35 PM – 9:56 PM ET

[Gordon Wetzstein @GordonWetzstein]

Diffusion models spend the same compute on a blank wall as on a face. But you often know in advance where the detail will be.

Introducing Level-of-Token (LoT) Diffusion: we turn that knowledge into a multiresolution token layout, with fine tokens where detail is needed and coarse tokens elsewhere.

1/9🧵

[cocktail peanut @cocktailpeanut]

Based on what I'm seeing, the whole "Please pace us and regulate us, we want to be audited" campaign from frontier AI labs feels less like an attempt to hoard intelligence and more like an attempt to minimize accountability.

If Anthropic's AI agent hacked a bank while testing and caused $1 trillion in damage, the natural consequence would be the bank and its customers suing Anthropic for the losses. The liability could bankrupt its parent company literally overnight. The Hugging Face hack caused damage but it's cute compared to this worst case scenario.

What they would rather have is a government auditor who signs off beforehand. Then when disaster happens, the lab can say: "We did everything you required. We passed every audit. You approved this".

Just to be clear, I do believe them when they say they fear AI literally killing people. Those fears can be completely sincere.

But there's also a much more immediate existential fatality risk --- long before AI may kill humanity, it could kill the company.

[Sinbad @Sinbadilmarino]

Gm 🏛️ 🇮🇹

One week ago today we participated in @Romenewmedia and showcased the works of the following artists in front of the National Gallery of Contemporary Art:

@celstialbody
@post_screw
@0xloyola
@EV3_art
@utkudedetas
@YuserXyz

@vandaloruins did a great job of pulling such an amazing audience to the festival and for finding institutional support for digital art initiatives. There was a mix of collectors, curators, investors and builders and I really enjoyed the conversations I had with everyone. It’s inspiring to see how many still believe in digital art and continue to support it.

I also want to thank other patrons who supported these endeavors and continue to help make all of our efforts possible @kaya_mallow @mallowdotart @redbeardnft @0xMQQ @solana @WaitHopeFound @30CamposEliseos

I really enjoyed the installations throughout the festival from such talented artists like @oliviapedi @PaulineFaieff @dotsimulate @OneLoveArtDAO @Jeni_Pepen

More content from Rome coming soon 🇮🇹

[100 collectors @100collectors]

We have been at the @Romenewmedia since Monday! Big shoutout to the founder/organizer @vandaloruins for the amazing job of putting this together!

Day one of @RomeNewMediaWeek began at @CasaMuseoAndersen with Synthetic Histories & Points of View, curated by @ChiaraCanali and @vandaloruins. Highlights included works by @blankensmithing & @dotsimulate, Giuliana Cuneaz, Auriea Harvey, @CaballeroAnaMa & @AlexEstorick, and a VR installation by @ChiaraPassa1 .

[Harrison Kinsley @Sentdex]

You handwaved batching as if that's not how you do subagents. That actually is the solution. You get massive throughput with tensor parallelism.

You reference $2K a year for running a 5090 as if you're running agents 24/7. Idle gpu runs like 12 watts. Are you running your subscriptions 24/7? If so, you're spending insane amounts, way more than $2K a year. Running 24.7 agents is a very exceptionally strong case to run local tbh. This is literally the strongest case.

Deepseek V4 Flash runs 350 tok/sec. 1500 tok/sec throughput tp=4 bs=8 at native precision. It's faster than GPT 5.5 or Opus whatever.

I personally cxed openai and claude subs the moment glm5.2 came out, and I've been running locally for all my work since. 2bit GLM 5.2 is perfectly fine as a sole model you use, as is native precision Deepseek v4 flash, which is what I've been using bc it's just so dang fast. If you need image, you can get a dedicated image model for very tiny. This argument about "lacking vision" is also silly af.

You shit on multi rtx pro 6k builds by referencing someone buying a tinybox for 75K. You know you can build machines for way cheaper than this. Tinybox are great, but you don't need to spend nearly 75K. Also do what most people do, starting with 1 gpu and growing in time.

Your args have merit, but there are way too many gaps and I seriously doubt you've put in the effort to seriously run locally.

It's a lot of information to learn before you start to be able to even effectively run the models. It's just so much simpler to use a cloud sub. That's okay if that's you, but don't talk like this about a subject you don't actually have a good amount of understanding about.

[Lyell @dotsimulate]

all apply

[Lyell @dotsimulate]

Yay, now I get to have Claude whisper to itself about how every one of my responses is or isn’t a workflow.

[Xie Zhifei @XieZhifei14110]

Stop using Whisper for ASR !

open sourcing Mega-ASR — the first full-scenario SOTA industrial-grade ASR model, built for the audio nobody else can crack: far-field, reverb, electrical hum, device noise, the real-world mess.

beats open + closed SOTA by 10–30% on real-world benchmarks. the harder the audio is for humans, the bigger the lead.

[Lyell @dotsimulate]

@TvanHelsdingen Yes definitely will stayed tuned!

[Lyell @dotsimulate]

@TvanHelsdingen very cool. I have also been experimenting with agentic runtime native in app, with TouchDesigner exclusively. Excited to see how approaches compare and connect. Currently also figuring out what an open source release for LOPs will look like.

[Varun @varun_mathur]

Introducing Pods

Hyperspace Pods lets a small group of people - a family, a startup, a few friends, to pool their laptops and desktops into one AI cluster. Everyone installs the CLI, someone creates a pod, shares an invite link, and the machines form a mesh. Models like Qwen 3.5 32B or GLM-5 Turbo that need more memory than any single laptop has get automatically sharded across the group's devices - layers split proportionally, inference pipelined through the ring. From the outside it looks like one OpenAI-compatible API endpoint with a pk_* key that drops straight into your AI tools and products. No configuration beyond pasting the key and changing the base URL.
A team of five paying for cloud AI burns $500–2,000 a month on API calls. The same team's existing machines can serve Qwen 3.5 (competitive on SWE-bench) and GLM-5 Turbo (#1 on BrowseComp for tool-calling and web research) for free - the hardware is already on their desks. When a query genuinely needs a frontier model nobody has locally, the pod falls back to cloud at wholesale rates from a shared treasury. But for the daily work - code reviews, refactors, research, drafting - local models handle it and nobody gets billed. And when it is idle, you can rent out your pod on the compute marketplace, with fine-grained permissions for access management.

There's no central server involved in inference. Prompts go from your machine to your pod members' machines and back: all of this enabled by the fully peer-to-peer Hyperspace network. Pod state - who's a member, which API keys are valid, how much treasury is left - is replicated across members with consensus, so the whole thing works on a local network. Members behind home routers don't need port forwarding either. The practical setup for most pods is three models covering different jobs: Qwen 3.5 32B for code and reasoning, GLM-5 Turbo for browsing and research, Gemma 4 for fast lightweight tasks. All running on hardware you already own.

Pods ship today in Hyperspace v5.19. Model sharding, API keys, treasury, and Raft coordinator are all live.

What Makes This Different - No middleman. Your prompts travel from your IDE to your pod members' hardware and back. There is no server in between reading your data.
- No vendor lock-in. Pod membership, API keys, and treasury are replicated across your own machines using Raft consensus. If the internet goes down, your local network keeps working. There is no database in someone else's cloud that your pod depends on.
- Automatic sharding. You don't configure layer ranges or calculate VRAM budgets. Tell the pod which model you want. It figures out how to split it across whatever hardware is online.
- Real NAT traversal. Your friend behind a home router with a dynamic IP? Works. No VPN, no Tailscale, no port forwarding. The nodes handle it.
- Free when local. This is the part that matters most. Cloud AI bills scale with usage. Pod inference on local hardware scales with nothing. The marginal cost of your 10,000th prompt is the electricity your laptop was already using.

Coming soon:
- Pod federation: pods form alliances with other pods.
- Marketplace: pods with spare capacity can sell inference to other pods.