Amichai Mantinband retweeted
not sure why this blew up, but here’s my app jayer.app
not sure why this blew up, but here’s my app brilliant.design
0
1
2
266
not sure why this blew up, but here’s my app brilliant.design
not sure why this blew up, but here is my app supabird.io
1
2
13
1,499
we have ~100 agents running at any single moment, sometimes 200+ across 4 different machines attempting to push our company's needs full force
we keep hitting different bottlenecks but the most surprising bottleneck is NOT inference speed (token/sec) or even decision making anymore
it's actually... COMPUTE
CPU/GPU load at full capacity across the various machines is, for the past week, consistently the bottleneck we're trying to work around
the load from building, running & testing the app is what's slowing down the pipeline and queue of changes from landing
for the first time, I'm considering moving our entire agent infrastructure to the cloud so it scales with the load
anyone facing this issue too? if yes, what's your solution?
we keep hitting different bottlenecks but the most surprising bottleneck is NOT inference speed (token/sec) or even decision making anymore
it's actually... COMPUTE
CPU/GPU load at full capacity across the various machines is, for the past week, consistently the bottleneck we're trying to work around
the load from building, running & testing the app is what's slowing down the pipeline and queue of changes from landing
for the first time, I'm considering moving our entire agent infrastructure to the cloud so it scales with the load
anyone facing this issue too? if yes, what's your solution?
3
0
9
640
meanwhile opus 5.5 high in @usebrilliant (4k closeups in comments)
What is the explanation for why models still cannot one-shot truly top notch UI design?
3
3
25
1,313
is everyone sleeping on Gemini 4 Argon for design? 🧵
I know, I know, making fun of gemini is extremely fun.
HOWEVER
Example 1: check out the detail and structure of the groovebox
I know, I know, making fun of gemini is extremely fun.
HOWEVER
Example 1: check out the detail and structure of the groovebox
5
5
31
2,488
Google, you cooked. Congratulations on a genuinely good model for design
Gemini 4 Argon, now available in Brilliant!
Try it here, free, no sign in required: brilliant.design/playground
1
1
18
1,580
Amichai Mantinband retweeted
Gemini 4 Argon, now available in Brilliant!
Try it here, free, no sign in required: brilliant.design/playground
Try it here, free, no sign in required: brilliant.design/playground
3
4
25
2,405
Why can't HTML/CSS-based canvases pose a threat to Figma?
The three rendering-engine choices and where each one breaks:
The three rendering-engine choices and where each one breaks:
Fine, I'll bite.
Paper will never, ever, eat Figma's lunch.
Not Paper, not Claude Design, not Google Stitch, not V0, bolt, lovable, base44 or any of the HTML/CSS-based canvas.
Why? The number one reason:
Performance.
Paper struggles at 2-5K elements of plain rectangles
For ched video: a small, early-stage exploration canvas with 13K elements.
Real-world projects are tens of thousands sometimes hundreds of thousands of elements.
Why do you think Figma has a hyper-optimized webGL rendering engine? It would've been way easier to have an HTML/CSS-based canvas if that was feasible.
Figma's money comes from tens of thousands of organizations where *at the very minimum* need to support these canvas sizes.
Does it hurt Figma that many individuals and small teams are switching to other tools? arguable.
Does Paper have better *MCP support* than Figma*? 1000 percent. Not even close.
but MCP support != better overall or even an option for most.
I'm not saying Figma's lunch won't be eaten.
I have a huge horse in that game, betting it will be.
I'm saying *HTML/CSS-based canvases, specifically* won't.
8
4
41
3,410
Disclaimers and small fixes after rewatching:
1. I did a poor job of communicating this: you can absolutely render a static page with 100K DOM nodes at 120 Hz as long as nothing invalidates.
The issue is that in a general purpose canvas: the invalidated set per-frame can be huge and THAT is what will eat up the budget
2. I used "rendering engine" as a umbrella word to describe 3 different things: layout engines (Blink, WebKit), rasterizers (Skia, Impeller, Core Graphics) and GPU scene renderers (what Brilliant and Figma built).
3. The video is far from covering all the architectures and workarounds for squeezing more performance out of the canvas.
4. probably other stuff. Please point out anything I got wrong or did a poor job explaining :)
1. I did a poor job of communicating this: you can absolutely render a static page with 100K DOM nodes at 120 Hz as long as nothing invalidates.
The issue is that in a general purpose canvas: the invalidated set per-frame can be huge and THAT is what will eat up the budget
2. I used "rendering engine" as a umbrella word to describe 3 different things: layout engines (Blink, WebKit), rasterizers (Skia, Impeller, Core Graphics) and GPU scene renderers (what Brilliant and Figma built).
3. The video is far from covering all the architectures and workarounds for squeezing more performance out of the canvas.
4. probably other stuff. Please point out anything I got wrong or did a poor job explaining :)
2
0
5
356