Naty Hoffman

🎉 Jellyfish AI Impact now supports Windsurf, a leader in AI-powered coding solutions, offering tools like the Windsurf Editor that blend collaborative, continuous AI assistance with developers' workflows. Together, Windsurf and Jellyfish are changing how AI and humans work together.  Through our robust integration with Windsurf, Jellyfish AI Impact allows joint customers to:  🤖 Measure Windsurf’s impact on overall developer productivity: Analyze how Windsurf solutions like Windsurf Editor impact different engineering workflows and use cases – whether it's building new features, reducing technical debt, or reviewing for code quality and security improvements. 📈 Scale Windsurf adoption across the organization: Identify teams with high engagement to further accelerate development, automate testing, and streamline deployment, then replicate these best practices and implementation strategies company-wide. Learn more about what this integration means for your engineering organization here: https://lnkd.in/ghNguE47

54

GDC 2026 Wrapup (2 of 2): A Shift in the Craft In Part 1 I covered what I learned at GDC this year. This one's more personal. I've spent 30+ years working professionally on traditional game AI. Behavior trees, GOAP, utility AI, finite state machines, genetic programming, neural networks, battle pattern matching and tactics, etc. That was my craft. GDC used to include a two-day AI Summit with multiple overlapping sessions to choose from throughout. This year you could count the traditional game AI sessions on one hand and still have fingers to spare. In the Game AI Roundtable, one of the few surviving AI sessions, people kept steering toward ML and LLMs while the host kept pulling it back to traditional game AI. It felt like watching our community process the loss in real time. I know I was. The Fireside Chat with Damian Isla and Jeff Orkin hit the same note. Veterans reminiscing about what felt like the golden age of traditional game AI. I knew some of those people from previous GDCs. Worked at some of the same companies. Bittersweet is the right word. But the PUBG Ally session I mentioned in Part 1? An AI buddy that learns your play style, remembers your favorite weapons, talks to you in combat through two-way voice? That's where it's all heading. And honestly, it's exciting even for someone who grew up in the old school. What I'm Taking Home Between the two posts, here's what I'm bringing back to the studio for The Realm: Return Home: - AI companions and villagers you can actually talk to and who learn from you - Adaptive music that matches gameplay, layering tracks and crossfading between themes so the soundtrack reacts to what's happening in real time - Polish and marketing fundamentals so our players can actually find us on Steam - VR gesture-based interactions that skip the controller entirely - Steam localization to reach a wider audience - And a personal favorite throwaway: texture-trained neural network shaders that dramatically improve both compression and fidelity at the same time 25 sessions. 20-30 pages of notes. A head full of ideas and a craft that's changing under our feet. It's a strange place to be, but exciting at the same time. What sessions stood out to you? #GDC2026 #GameDev #GameAI #IndieGame #VR

3

We've been implementing Neural Texture compression in #Assassinscreedmirage. The overall benefit results in better quality textures for less GPU memory usage (links in the first comment). If you followed the topic since our GDC2024 presentation ,this a very practical application of La Forge's mission to bridge the academic world with Ubisoft. #ubisoft #ubisoftlaforge #innovation 👇 "If you've been thinking, like I have, that we're still some way off seeing AI being used in games for more than just upscaling or frame generation, then it's time to think again" (...) "Enter stage left, neural texture compression to cut down on that memory demand. Given the state of the DRAM and SSD market right now, you can bet your last dollar that a whole heap of game developers will be looking closely at what Ubisoft has done to see if it can help out in their games, too"

203

9 Comments

Chapter 7 of my book, Data-Oriented Design for Games, published by Manning Publications Co., is now available in the Manning Early Access Program (MEAP). This chapter discusses how to reduce code complexity and enhance the game development workflow by separating dynamic and static asset data. The example in this chapter is how to add a menu system to a survivor/bullet-heaven style game. You can check out the book and read the first chapter for free in the link below: https://lnkd.in/dYvM32zc

25

6 Comments

A couple days ago I gave Claude Code the ability to debug GPUs with RenderDoc. Now I taught it Apple's Metal ecosystem.                                                I built a second skill — metal-ai-skill — that connects Claude Code to Apple's GPU toolchain: xctrace, Metal validation layers, shader compilation, .gputrace frame capture, and screenshot-based visual  inspection. When I describe a rendering bug on macOS, Claude can:  - Screenshot the app and actually see what's wrong (multimodal vision)  - Capture a .gputrace frame and inventory every buffer, texture, and shader  - Compile shaders with -Weverything to catch static issues  - Read source code to trace data flow from CPU to GPU  - Fix bugs and re-screenshot to verify — in a loop until it's correct  The interesting challenge: Metal's CLI tooling doesn't give you buffer data from programmatic captures (unlike RenderDoc's rdc-cli which gives you everything). So I built an autonomous debugging workflow  — Claude gathers signal from multiple sources in parallel, checks what data is actually available, and falls back to source code analysis when GPU data isn't there. Every path leads to a diagnosis.  In the demo, I give Claude a Metal cube renderer with 4 intentional bugs: collapsed geometry, flipped Y axis, swapped color channels, and zeroed alpha. Claude screenshots the broken output, reads the  shaders and Swift code, diagnoses all 4 bugs, fixes them, rebuilds, and verifies the output — all autonomously.  The RenderDoc skill taught me the pattern. The Metal skill taught me the fallbacks. The real insight: when you can't inspect GPU state directly, combine enough indirect sources and you still get there.  Open source (MIT).  GitHub: https://lnkd.in/g8n8qb6J  Video: https://lnkd.in/gryAY4YR  #ClaudeCode #GPU #Metal #Apple #GraphicsProgramming #AI #GameDev #OpenSource

114

11 Comments

The layoffs at Epic are so brutal to the industry. We are still full-on with the development of Puget Bench for Unreal and plans for UnrealFest, but it gives me an uneasy feeling. My thoughts are with all those impacted. I believe in the potential of the game development community and am hopeful that this industry has a strong future.

17

4 Comments

Game art asset factories aren’t enough anymore. Pretty models mean nothing if the shaders won’t compile. True co-development requires technical art literacy. If an external team can’t debug a pipeline, they aren’t helping. They’re generating high-fidelity technical debt. At Artchop!, we build with integration in mind from day one. The real question is whether studios are looking for true integration partners or just capacity. #GDC #CoDev #TechnicalArt

4

Sound design with robots: My new SFX workflow (listen to hear the chimes)... In the cyan window, Claude Code is generating and running sox commands on my Mac. In the right window, I'm previewing them. In seconds, Claude generates code to make melodies from pitch and timing parameters - entirely replacing a DAW or MIDI workflow. Describe what you want, listen, iterate. And meanwhile, the underlying sounds were generated by ElevenLabs and Adobe Firefly.

4

New paper alert about Renderman XPU released by Julian Fong & Per Christensen from Pixar. For those keen on technical breakdowns, this is a must-read. Dive into the details here: https://lnkd.in/eK6vDuwp #HPG2025 #Renderman #Pixar #Technical_Art

10

1 Comment

The aim for us, McCarthy Brothers, is to remove uncertainty starting at pre production and a focused forward momentum, design justifications and reasonings so teams designing from our architecture know intent, direction, complexity, all before adding a single line of code. Our aim is to help teams avoid production hell and shorten time and money spent in their productions. Years we have invested into solving these problems in our own work and its now mature enough that we can scale it into others. We have typically been known for our 3D and Tech art work in the games industry and now we have more to share in new roles in the industry.

1

There was a really neat demo of a student project at the vulkanized conference called "vkDuck". https://lnkd.in/g24yHhyF Much like Gigi, it's a visual render graph editor / viewer. I really think rendergraphs as data (instead of code) is the next evolution for rendering engines. There are so many benefits to moving render graphs from code to data, including: * Faster iteration time * Ability to have render graphs for individual games or platforms, instead of having an Uber graph that has to solve all concerns simultaneously. * Better built in debugging and profiling * Letting tech artists and other non engineers make changes to the render pipeline, or do experiments around that, without needing engineering support. Also, by moving render graphs to data, we could get away from needing graphics API abstractions (RHIs) because you can translate the render graph nodes directly to target API actions. This allows the render graph to be re-interpreted, such as fusing VS/PS/CS on tiled platforms, while leaving them separate on other platforms.

78

5 Comments