Mike Burrows

Who knew @JoshuaBarczak was like the Queen in that he had TWO birthdays. Happy not birthday Joshua!

@bensig @runonthespot @MillaJovovich Love the concepts and design philosophies. Presuming this is done without imagery, usually used by humans, and leveraging just text - similar to the ways a human with aphantasia might? Really respect you releasing this as true open, permissive, source and licensing - Bravo both!

@runonthespot look, nothing is perfect... honestly there is a hybrid mode that passes the bench at 100%... but without any LLM helpers it is still at 96% which is insane. @MillaJovovich did a great job coming up with these concepts... we were stunned at the benchmark results though

My friend Milla Jovovich and I spent months creating an AI memory system with Claude. It just posted a perfect score on the standard benchmark - beating every product in the space, free or paid. It's called MemPalace, and it works nothing like anything else out there. Instead of sending your data to a background agent in the cloud, it mines your conversations locally and organizes them into a palace - a structured architecture with wings, halls, and rooms that mirrors how human memory actually works. Here is what that gets you: → Your AI knows who you are before you type a single word - family, projects, preferences, loaded in ~120 tokens → Palace architecture organizes memories by domain and type - not a flat list of facts, a navigable structure → Semantic search across months of conversations finds the answer in position 1 or 2 → AAAK compression fits your entire life context into 120 tokens - 30x lossless compression any LLM reads natively → Contradiction detection catches wrong names, wrong pronouns, wrong ages before you ever see them The benchmarks: 100% recall on LongMemEval — first perfect score ever recorded. 500/500 questions. Every question type at 100%. 92.9% on ConvoMem — more than 2x Mem0's score. 100% on LoCoMo — every multi-hop reasoning category, including temporal inference which stumps most systems. No API key. No cloud. No subscription. One dependency. Runs on your machine. Your memories never leave. MIT License. 100% Open Source. github.com/milla-jovovich…

@EpicVogel Nice breakdown. Good catch on ensuring agents can’t modify the test harness and the success criteria - most don’t expect that, though these types of constraints are critical.

I spent the last few days pushing Anthropic's VLIW performance take-home kernel optimization challenge to its limits using Claude Code as an orchestrator. The task: schedule operations for a custom VLIW SIMD architecture running a tree traversal with hashing. 256 items, 16 rounds, 5 execution engines with different slot limits. Starting point: 147,734 cycles (naive) Where Claude Code landed: 1,105 cycles — a 134x speedup The journey involved hundreds of AI agents across several dozen iterations, exploring every angle: hash algebraic merges, L4 tree caching, DAG-based list schedulers, 250K+ parameter configurations, emission order sweeps, store engine exploitation, and loop-based kernels. The last gain was 2 cycles, found by sweeping 103,000 configurations. I then had Claude write a formal lower bound proof in Lean showing the kernel cannot run in fewer than 1,081 cycles — proven from load engine capacity (2,089 ops at 2/cycle = 1,045 minimum) plus unavoidable dependency overhead. For verification I relied on Anthropic's test harness and extended it with randomized parameter testing and also extended it with Kernel Optimization Fun's output index verification. What struck me is how good Claude Code has gotten at orchestrating optimization work. It ran teams of 10 parallel agents in isolated worktrees, each exploring different hypotheses. Agents communicated findings, dead ends propagated instantly, and the system converged on proven optima. The DAG scheduler that broke through a 6-iteration plateau came from Codex (gpt-5.3) running through Claude Code's MCP integration — multi-model orchestration improving the solution. Proebsting's Law says compiler optimizations double program speed every 18 years. AI agents with the right tools are compressing that timeline dramatically — not by improving compilers, but by doing the work compilers can't: reasoning about problem structure, exploring architectural trade-offs, and proving bounds. The 134x speedup here came from algorithmic insight (merged hash stages, path-bits scheduling, bias-free C5), not instruction selection. Wild times for performance engineering.

@Rhuantavan @DID_Ltd_ @micro_prose Is that you Shaun? 🤣

F-22: Air Dominance Fighter comes to Steam! We release the 1997 classic – overhauled for modern systems with almost 100 fixes and improvements. Wishlist now! store.steampowered.com/app/3146140/F2… Shout out to @micro_prose for helping us to make this possible! FAQ below🧵

A classic. In my humble, and completely biased, opinion :)

@DID_Ltd_ @micro_prose It Lives… AGAIN! Apologies for anyone having to update threedee.asm. Most, if not all, should have been through d3d for adf. Did you update to later version of d3d? Finally, you should be able to up res the ground textures easily, iirc we used about 10m texels.

@adyaman Impressive!

With this, I was able to generate a 480p video using TurboDiffusion (Wan2.1 1.3b quantized) in about 2-3 minutes on a strix halo (framework desktop).. all on Windows!

If any of you can give please do so - great charity and very close to home! f*ck cancer!

Kostas’ presentation is well worth a look.

@DID_Ltd_ ok, so as a didy man- I appreciate the fresnel shader on the moon and atmospheric scattering though curious minds are wondering if this is Bins, Lecky and Paddy?

Definitely worth a look ;)

@tom_forsyth 13.2 is a significant step function. Reverse, in addition to forward…. Oooooh ;)

@zebedee666 Meh. My car has one computer, and it controls the spark plugs.

Elf needs FSD v13.2… badly… ;)

@VancityReynolds thanks for the Christmas card bud; came at a sad time after pup of 15 years passed over rainbow bridge, though wtf: a mocktail? ;)

Possible on both, though with WMMA will be faster on 3+

Congrats to the team getting this out there. Though perf numbers not published yet - think real real-time, not siggraph real-time ;) As always, intent is full permissive source to be released. gpuopen.com/learn/neural_s…

🎇 We're excited to share more of our latest #raytracing #research in our latest blog! 🎇 ➡️ Read about our new neural denoising and supersampling technique, where we aim to achieve real-time #pathtracing 🔥 Find out how on GPUOpen ⤵️ gpuopen.com/learn/neural_s…

Today sucks, Mike Magee passed away today. My thoughts on him and his legacy is now up. semiaccurate.com/2024/08/12/ind…

Comprehensive and useful context for understanding root signatures and hints how to glean approaches from commercial titles by leveraging pix

@jimkxa @DanraeP @TimSweeneyEpic Years. At least for separate scene graph and simulation(s), to generate such, first. Full representation with emergent behaviors may be a decade, though I doubt it will be that long. Few 1000x problems to solve…

@DanraeP @TimSweeneyEpic decades or years ?

I think this will happen sooner than you think but will be too expensive for wide use. Sort of like high end pro graphics were before PC gaming GPUs twitter.com/TimSweeneyEpic…

@Peter_shirley @2Sexy4MyGPU I remember when Peter-Pike first started talking SH, then the benefits of dynamic lighting when using prt, with its use of principal component analysis to “simplify”… 🤯

@2Sexy4MyGPU You and me both brother.

I am having fun preparing our #I3D2024 talk and there are lots of choices with how to display spherical harmonics and I am browsing all these from #wikipedia