Mad Mat (@ReallyMadMat) header
Mad Mat (@ReallyMadMat) profile picture

Just chillin

Heaven & Hell
Joined November 2023
Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)Photo by Mad Mat (@ReallyMadMat)
Gaussian splats feel like the photogrammetry moment again. Everyone loves the capture, few ask how you relight it. In film that one question decides if a technique makes it to final pixels or stays in previs.
1
8
Me: I built an AI agent that handles my paperwork. Amt: Great. Please fax it. Me: It can send a signed PDF. Amt: PDF is not a signature. Me: It has a qualified digital signature. Amt: Please come in person. Next appointment is in March. Me: Which March? Amt: Please call to ask. Germany isn't behind on AI. We're just waiting for the appointment.
1
41
In VFX nobody asks if the renderer can make a frame. They ask how many frames finish per night and how many come back for fixes. Agent setups need the same two numbers: throughput and rework rate.
6
human representation in XR is the right problem. once the avatar drifts from intent the whole interaction feels off. paper space on that is overdue
6
enterprise vibe coding only works if ownership is clear. who can explain the change on monday when the prompt author is out is the real gate
5
tracing the agent run is the missing piece tbh. without a local timeline of what claude or cursor actually touched you just stare at a diff and guess. proxy plus viewer beats screenshots of the chat
1
1
64
dance is lowkey the hardest test for these models, footwork has to stay on beat across cuts. did the chakkars keep the same rotation speed between shots or did you regen till it matched?
7
nobody should be asking who is allowed to vibecode. real question is who owns the tool on monday when it breaks. give people agents, sure, but every tool they build needs an owner and an expiry date
4
honestly its less about the tool, more about whether the agent can check its own work. tests, screenshot diff, lint pass. i let grok bot orchestrate and claude code do the work and it only runs alone when "done" is defined before it starts
1
5
/plan is the one I'd put first. plan, read it, then let it code. skipping that is where most of the cleanup time comes from. and one task per session, a fresh context beats /compact more often than people think
2
nice test. I'd add camera language to the matrix. action prompts mostly land, but lens and camera move is where you see which model gets film. dolly in vs zoom is my quick check. which one held the camera best?
4
Screenshare quality is the real flex. Half of every call is still "can you see my screen" and then nobody can see the screen.
5
Saying it four times is honestly the most rigorous eval methodology on this site.
3
Two weeks between frontier models now. The blog post barely cools down before it's legacy.
3
Hot take: the skill that matters in AI coding is not prompting. It's writing the check that tells the agent it's done. Tests, linters, acceptance criteria. Everything else is vibes. This account: takes, checklists and diagrams on agents, vibecoding and automation.
1
1
17
I build AI content pipelines: image, video and code. Background: Hollywood VFX and pipeline tools. I post what works, what breaks and what it costs. Bots prepare, I approve. AI content is labeled.
4
Seeing all those 80s/90s movie idols in new roles or funny adaptions somehow takes away the emotional value of those movie memories. I do like those videos, but at the same time, I don't - weird.
16
Life can be hard sometimes
18