[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

@SebAaltonen Avatar @SebAaltonen Sebastian Aaltonen

Sebastian Aaltonen posts on X about voxels, voxel, gpu, instead of the most. They currently have XXXXXX followers and XXX posts still getting attention that total XXXXXXX engagements in the last XX hours.

Engagements: XXXXXXX #

Engagements Line Chart

Mentions: XX #

Mentions Line Chart

Followers: XXXXXX #

Followers Line Chart

CreatorRank: XXXXXX #

CreatorRank Line Chart

Social Influence #

Social category influence nfts XXXX% technology brands XXXX% stocks XXX% finance XXXX% cryptocurrencies XXXX% luxury brands XXXX% automotive brands XXXX% gaming XXXX% social networks XXXX% celebrities XXXX%

Social topic influence voxels 9.63%, voxel #47, gpu #1062, instead of 2.96%, triangle #127, plug 2.22%, want a #19, generated 1.48%, faster 1.48%, gaming XXXX%

Top accounts mentioned or mentioned by @agilejebrim @simplex_fx @sebasti66855537 @jbrooksbsi @oswald_sn @rouanijihad @atomlib @tomcr2100 @faintmaster @theonecid @boenau @3mar445 @therealdecross @treylorswift @blurbusters @opinali @never_released @tylerglaiel @joaobapt @lafadhel

Top assets mentioned Voxels (voxels) Perlin (PERL) Alphabet Inc Class A (GOOGL)

Top Social Posts #

Top posts by engagements in the last XX hours

"AI generated C is a real deal. C coders wrote fast & simple code. No high freq heap allocs no abstractions slowing the compiler down. Lots of good C example code around. Ai workflows need a language with fast iteration time. Why waste compile time and perf on modern languages"
X Link @SebAaltonen 2025-10-23T22:00Z 46.1K followers, 70.4K engagements

"And why would you even think about generating Python with AI Why would you choose a 100x slower language if AI is writing it instead of you Same applies to Javascript and other common internet languages. Just generate C and compile to WASM. Nothing runs faster"
X Link @SebAaltonen 2025-10-23T22:09Z 46.1K followers, 28.5K engagements

"Bug report: Car wheels don't rotate. Looks bad. Repro steps: X. Pick the Phantom X. Drive it. Expected behavior: Wheels should rotate"
X Link @SebAaltonen 2025-10-24T14:21Z 46.1K followers, 19.6K engagements

"Thus 165Hz gaming phones aren't really important. Most games are 60Hz locked and even those few games that support 165Hz don't benefit much from it"
X Link @SebAaltonen 2025-10-26T19:37Z 46.1K followers, 3148 engagements

"Digital audio has better quality. DAC inside the phone needs to be small. It's a compromise. And there's inference issues. Better to have a good quality DAC inside the headphones. Move the audio digitally over the cable. USB-C cable supports digital audio. It is better"
X Link @SebAaltonen 2025-10-28T08:30Z 46.1K followers, 2864 engagements

"My favourite meme picture :)"
X Link @SebAaltonen 2025-10-21T15:57Z 46.1K followers, 90.6K engagements

"Why do we need a separate headphone plug Apple finally uses the USB-C standard. USB-C can carry the audio signal (both ways) and power the headphones (noise cancelling) transfer volume/next button events from headset to the phone. USB has become the standard audio connector"
X Link @SebAaltonen 2025-10-27T13:56Z 46.1K followers, 57.5K engagements

"Why would somebody today still want a separate audio plug that only supports one way audio (no microphone) and doesn't provide enough power for noise cancellation and can't carry volume/track button events Do you have legacy headphones"
X Link @SebAaltonen 2025-10-27T13:58Z 46.1K followers, 12.3K engagements

"I told the AI that I don't want a global variable because I need to multithread the algorithm. It wrote a thread local variable instead :D"
X Link @SebAaltonen 2025-10-28T17:36Z 46.1K followers, 16.3K engagements

"I tried to convince the AI too many times to continue the big refactoring. It simply didn't want to do it. Asked me to do it instead. We had an argument :) "Codex ran out of room in the model's context window. Start a new conversation or clear earlier history before retrying.""
X Link @SebAaltonen 2025-10-28T20:44Z 46.1K followers, 5628 engagements

"Voxels are so easy. uint64 stores a 4x4x4 brick. uint64 a b; // two 4x4x4 bricks Union (add): a b Intersection (cut) a & b Both are a single full rate instruction on most CPUs and GPUs today. As @mmalex said a decade ago: I don't like polygons :)"
X Link @SebAaltonen 2025-10-01T05:58Z 46K followers, 26.5K engagements

"Perf matters outside AAA too. Fall Guys barely hits XX fps on Nintendo Switch with sacrifices: 10Hz opponent animations and player count 100-60. Down to XX in later patch to improve performance. Meanwhile HypeHype runs XXX player MP at 60Hz (all anims) on 2x slower phone"
X Link @SebAaltonen 2025-10-11T08:26Z 46.1K followers, 8707 engagements

"Spent an extra hour on my WebGPU/WASM prototype. Compute shader fills the pixels to offscreen storage texture then there's a blit pass (full screen triangle)"
X Link @SebAaltonen 2025-10-11T11:34Z 46K followers, 15.8K engagements

"Sense of scale plays a huge role in perception. 81923 map at 1m3/voxel is 8km3. 12.5cm3 voxels = 1km3. Why do they look different I've been traveling seeing natural and architectural wonders. Would be nice to build a game with this kind of scale. Thread"
X Link @SebAaltonen 2025-10-17T12:30Z 46K followers, 17.1K engagements

"You need reference to sense scale. Above tiny humans are a good reference of scale. Birds offer a good reference for tall structures. In this legendary Jurassic Park scene the trees are a good reference. The dinosaur pulling/snapping the trees shows a physical response"
X Link @SebAaltonen 2025-10-17T12:33Z 46K followers, 1831 engagements

"If we generate a voxel terrain with some fancy perlin noise how do we know how big it is If we add trees or rocks we start sensing some scale. If we add 100000 players (see my previous MMO thread) we can use them as reference of scale. Animals vehicles etc also work well"
X Link @SebAaltonen 2025-10-17T12:39Z 46K followers, 1748 engagements

"Apple M5 GPU has tensor cores similar to Nvidia and AMD's professional CDNA. Older Apple GPUs already had tensor intrinsics but these intrinsics ran using the vector pipelines similar to AMD RDNA3 WMMA. AMD UDNA is also expected to have tensor cores"
X Link @SebAaltonen 2025-10-18T08:48Z 46K followers, 25.5K engagements

"People often think voxels take a lot of storage. Let's compare a smooth terrain. Height map vs voxels. Height map: 16bit 8192x8192 = 128MB Voxels: 4x4x4 brick = uin64 = X bytes. We need 2048x2048 bricks to cover the 8k2 terrain surface = 32MB. SVO/DAG upper levels add 10%"
X Link @SebAaltonen 2025-10-18T09:17Z 46K followers, 22.6K engagements

"The above estimate is optimistic. If we have a rough terrain we end up having two bricks on top of each other in most places. Thus we have 64MB worth of leaf bricks. SVO/DAG upper levels don't increase much (as we use shared child pointers). Total is 70MB. Still a win"
X Link @SebAaltonen 2025-10-18T09:19Z 46K followers, 4144 engagements

"The interesting realization is that octtree stores bits recursively in tree levels. If each level stores 4x4x4 bricks you get X bits of precision per level. Thus X levels = XX bits of precision in height. But the high bits are shared between lots of voxels"
X Link @SebAaltonen 2025-10-18T12:06Z 46K followers, 5463 engagements

"2x2x2 = uint8. That's one byte. Link = uint32 = X bytes. Total = X bytes. 4x4x4 = uint64. That's X bytes. Total (with link) = XX bytes. 4x4x4 tree is half as deep as 2x2x2 tree. You get up/down twice as fast"
X Link @SebAaltonen 2025-10-18T19:09Z 46.1K followers, 4193 engagements

"Voxel mask (in non-leaf nodes) tells us which children are present. You can do a popcnt (full rate instruction on GPU) to count the children. Children are allocated next to each other. Children sub-address can be calculated with binary prefix sum (= AND prev bits + popcnt)"
X Link @SebAaltonen 2025-10-18T19:12Z 46.1K followers, 3946 engagements

"This is the same reason why 4x4x4 tracing ends up being faster. It's touches less different cache lines and transitions faster to near/far trace (fast empty space skip - near surface root finding)"
X Link @SebAaltonen 2025-10-18T19:19Z 46K followers, 3003 engagements

"2nd reason why 4x4x4 tree is better is the amount of metadata compared to payload. For each XX bits of voxels we have XX bits of link (offset). That's XX% payload / XX% link. 2x2x2 tree has XX% payload / XX% link. In data structure design you always want to minimize metadata"
X Link @SebAaltonen 2025-10-18T19:22Z 46K followers, 2850 engagements

"Thus even if we have all X children present the octree still doesn't utilize the cache lines fully. Partially used cache lines will be evicted from the tiny L1$ and you pay for data you didn't use. In realistic case we have less than X children so this issue is even worse"
X Link @SebAaltonen 2025-10-18T19:28Z 46K followers, 2612 engagements

"For the 4x4x4 tree the children often span multiple cache lines (up to 6) since they are continuously allocated. This is great for cache access pattern especially in cases of nearby voxels representing a continuous surface which is a common pattern for real content"
X Link @SebAaltonen 2025-10-18T19:30Z 46K followers, 2424 engagements

"Why not 8x8x8 or 16x16x16 trees If you already allocate children next to each other 4x4x4 is good enough as it puts children in the same cache line implicitly. 4x4x4 fits nicely in a single uint64 so it's efficient to load and process in a single register. 8x8x8 = XXX bits"
X Link @SebAaltonen 2025-10-18T19:33Z 46.1K followers, 2291 engagements

"Since voxel octrees often encode just the surface a more narrow region is more storage efficient. Bigger nodes increase the ratio of payload:metadata but the ratio is already pretty good in 4x4x4 tree. Most nodes are leaf nodes and leaf nodes don't need the link"
X Link @SebAaltonen 2025-10-18T19:36Z 46K followers, 2864 engagements

"Of course we have to ask: Why not a single cache line aligned nodes. That's iffy because A) are storing sparse data (surface near band is not a tile) B) GPU SIMD branch/loop coherency issues (homogenerous tiles are important). Allocating children contiguously is almost as good"
X Link @SebAaltonen 2025-10-18T20:38Z 46K followers, 2108 engagements

"@TylerGlaiel For tree structures on CPU side wider nodes are often a big win. Both because of cache improvements but also because of OoO CPUs like seeing multiple links at the same time. Allows them to load early to hide latency. Linked list is bad as it's only one link. More = better"
X Link @SebAaltonen 2025-10-18T20:41Z 46.1K followers, XXX engagements

"@TylerGlaiel But of course the devil is in the details. CPU = GPU. SIMD branch/loop coherence is important for the GPU. I will implement a quick 4x4x4 voxel tracer soon in my WebGPU prototype. I will benchmark it of course"
X Link @SebAaltonen 2025-10-18T20:43Z 46.1K followers, XXX engagements

"And most important last: Do not mix material/texture data in your ray-tracing data structure. Keep it minimal. Find the ray hits first. Then do material/texture stuff. You don't want to load material/texture data to L1$ for voxels that you miss"
X Link @SebAaltonen 2025-10-18T20:47Z 46K followers, 4845 engagements

"@JoaoBapt Link is just a 32-bit aligned offset if your data is 4-byte aligned then a 32-bit link it can address up to 16GB of voxel data. 16GB address space is fine for gaming purposes"
X Link @SebAaltonen 2025-10-19T11:57Z 46K followers, XXX engagements

"@AgileJebrim The idea is to have a fixed cost regardless of overdraw. This requires very fine grained occlusion culling and LOD. If you put a new object in the view that object occludes as many pixels behind it. Thus the total amount of pixels always stays the same. That's the main idea"
X Link @SebAaltonen 2025-10-19T12:38Z 46K followers, XXX engagements

"@AgileJebrim Aren't you splatting (scattering) all voxels to the froxel viewport If you scatter 100x overdraw worth of buildings (city street view) that's 100x more expensive than scattering only the first visible surface"
X Link @SebAaltonen 2025-10-19T12:43Z 46K followers, XXX engagements

"@AgileJebrim Voxels and triangles behave similarly if you cluster the meshes to small spatially local clusters. Voxel bricks (let's say 8x8x8 bricks) are similar to geometry clusters and can utilize similar occlusion culling strategy such as a hierarchical Z buffer"
X Link @SebAaltonen 2025-10-19T12:46Z 46K followers, XX engagements

"Another hour of progress in my WebGPU compute shader voxel ray-tracer: Hint: There's no ray-tracing or voxels yet :) It works fine on my iPhone too iOS XX finally added WebGPU support. Seems to be working great so far"
X Link @SebAaltonen 2025-10-19T15:15Z 46K followers, 7920 engagements

"@Sebasti66855537 Nvidia's geometry pipelines are super awesome. Always been and Turing also bumped those up drastically (both the classic pipes and added mesh shaders). Nvidia can achieve awesome occupancy in vertex-only workloads. Nobody can touch them when it comes to pure triangle throughput"
X Link @SebAaltonen 2025-10-19T15:22Z 46K followers, 1290 engagements

"@Sebasti66855537 Tile based systems have extra overheads for draw calls and triangles. They optimize for pixel processing and pixel memory traffic. Modern AAA geometry can be 1:1 triangles:vertices. TBDR isn't the perfect fit for that kind of content"
X Link @SebAaltonen 2025-10-19T15:24Z 46K followers, XXX engagements

"@Sebasti66855537 Maybe we should write compute-shader only rendering for Apple. Sidestep the whole TBDR stack. Nanite kind of does that already with their SW raster path. But I don't know if Apple has super fast 64-bit atomics. Need to measure it. SDF/voxel ray-tracing also should be nice on M5"
X Link @SebAaltonen 2025-10-19T15:26Z 46.1K followers, XXX engagements

"@JBrooksBSI Yes if it is axis aligned (best case). The maximum amount of touched voxels is x+y+z: X for a 2x2x2 brick and XX for a 4x4x4 brick. 2x2x2 unused bits = X. 4x4x4 unused bits = XX. However both have XX bits of metadata which is a waste"
X Link @SebAaltonen 2025-10-19T18:00Z 46K followers, XXX engagements

"@JBrooksBSI Also the worst case 4x4x4 traversal = XX voxels would touch six 2x2x2 bricks in the same region thus we load 6x 32-bit metadata instead of just a single 32-bit metadata. This is a big win. It's not just about voxels the metadata is also a waste"
X Link @SebAaltonen 2025-10-19T18:03Z 46K followers, XXX engagements

"@JBrooksBSI The scaling is indeed bad since we touch less and less voxels as the brick size increases. According to my math the sweet spot should be 4x4x4. It nicely fits into a single 64-bit register (efficient mem load efficient addressing with bit ops etc)"
X Link @SebAaltonen 2025-10-19T18:06Z 46K followers, XXX engagements

"Seems that MacOS XX broke screenshot pasting to Slack. Previously you could see the screenshot directly in Slack as an image now you get a PNG file link instead (no visible image have to download to see it). Anybody figured out how to workaround"
X Link @SebAaltonen 2025-10-20T08:16Z 46.1K followers, 7569 engagements

"@LAFadhel Didn't they just do a big deal with Qualcomm Snapdragon Elite for Galaxy"
X Link @SebAaltonen 2025-10-20T13:20Z 46K followers, 1454 engagements

"I just realized that I didn't actually want an iPhone Mini. I wanted a more lightweight phone. Tested iPhone Air in the local store and it feels awesome in my hand. 120Hz display too (which is a must have for me). Seems like a perfect phone for me. Managed to resist the urge :)"
X Link @SebAaltonen 2025-10-20T19:03Z 46K followers, 13.4K engagements

"A common problem for tech companies: When shit hits the fan (technical) artists are fired. When Unity layoffs started the content production team was the first to go. Ben is not the first casualty. You can't dogfood an engine without content. Unity editor has scaling issues"
X Link @SebAaltonen 2025-10-21T12:19Z 46K followers, 24.3K engagements

"iPhone Air has almost 2x screen area compared to iPhone X (8 years old flagship) but is only XX% heavier. They also crammed in a XX% bigger battery and a 4x brighter screen at 2x refresh rate. That's progress"
X Link @SebAaltonen 2025-10-21T19:09Z 46K followers, 13.2K engagements

"This is the thing with AI coding. It's fine for unimportant leaf code but when you ask AI to do crucial performance critical inner loops it will not do what you want. I asked it to do sparse SDF brick tracer too and it had horrible SIMD loop/branch coherency perf issues"
X Link @SebAaltonen 2025-10-25T21:02Z 46.1K followers, 4912 engagements

"8192x8192 terrain. 64M grid cells. X triangles per grid cell. 128M triangles. Non-indexed = 128M * X = 384M vertex shader lanes Indexed = 64M vertex shader lanes Savings = XXX million vertex shader invocations"
X Link @SebAaltonen 2025-10-26T08:49Z 46.1K followers, 2741 engagements

"Assuming perfect vertex reuse. Instancing plays a big role here. First reason for instancing is to avoid having a 8k223*4 = 1.5GB index buffer. 32x32 patch = 1024x2 triangles = 6K indices. 16-bit indices are fine here so that's 12KB. Fits in L1$"
X Link @SebAaltonen 2025-10-26T08:54Z 46.1K followers, 2520 engagements

"Second reason for instancing is vertex reuse. If you render a big terrain (such as 8192x8192) at scanline order the second row vertices are already out of the vertex cache when you start rendering them. Thus you have to transform that row again. Which amortizes to X vx per tri"
X Link @SebAaltonen 2025-10-26T08:55Z 46.1K followers, 4567 engagements

"If you however have short enough vertex stride or if you do morton order (or similar) to your vertices the scanline is still in cache for the next row. This way you only need to transform X vertex per X triangles = XXX per triangle. Which is the optimal for vertex grids"
X Link @SebAaltonen 2025-10-26T08:56Z 46.1K followers, 6501 engagements

"@Sebasti66855537 There's no vertex buffer at all in any version. Vertices are generated in the vertex shader by math. And heightmap texture sampling. The hmap texture sampling is indeed worse memory locality if your terrain has scanline ordering (8192 long scan lines)"
X Link @SebAaltonen 2025-10-26T14:34Z 46K followers, XXX engagements

"@JohnnyB_222 Strip is X vertex shader invocation per triangle. Indexed is XXX. For a ordered grid (2 triangles per grid cell). Non-indexed is 3"
X Link @SebAaltonen 2025-10-26T20:47Z 46K followers, XXX engagements

"@2Sexy4MyGPU @JohnnyB_222 Yeah you get a degenerate triangle for each row in each terrain tile. Not a massive deal. But today indexed is of course preferable over strips as it's around half the invocations per triangle (if tile row fits in post transform vertex cache)"
X Link @SebAaltonen 2025-10-26T20:51Z 46.1K followers, XXX engagements

"@jonathanperos @IftodeBogdan The best quality is achieved by putting the DAC inside the speakers/headphones. Most speakers have digital inputs nowadays. The reason is better audio quality. Less inference and other issues"
X Link @SebAaltonen 2025-10-28T08:34Z 46.1K followers, XXX engagements

"@therealdecross Genelec speakers have won countless of audio magazine tests and they have the DAC inside them. Digital signal to speakers. No inference. Use a top quality DAC manufactured by a real audio company instead of a phone company"
X Link @SebAaltonen 2025-10-28T10:54Z 46.1K followers, XXX engagements

"@treylorswift I never trust AI. I am using GPT5-Codex with maximum effort settings. This is basically a glorified search and replace operation. Using AI for it because it would take me couple of days to do because it's not a direct search and replace"
X Link @SebAaltonen 2025-10-28T14:40Z 46.1K followers, XXX engagements

"Let's discuss why I think 4x4x4 tree is better than 2x2x2 (oct) tree for voxel storage. It all boils down to link overhead and memory access patterns. L1$ hit rate is the most important thing for GPU performance nowadays. Thread"
X Link @SebAaltonen 2025-10-18T19:07Z 46.1K followers, 49.9K engagements

"WTF is this Some windows in MacOS XX have different curve radius on the corners. Steve would have never approved this"
X Link @SebAaltonen 2025-10-20T11:51Z 46.1K followers, 12K engagements

"Awesome new feature in Xcode GPU debugger. You can expand passes to see stage dependencies now. As you can see our G-buffer pass vertex shading has no dependencies to prior GPU work meaning that it can already start during the previous frame. This gives XX% perf increase"
X Link @SebAaltonen 2025-10-20T18:40Z 46.1K followers, 15.2K engagements

"C++ ranges are safe Avoids buffer overflows and all raw loop issues Not exactly. The nice thing with raw loops is that I can easily debug them and our vector class asserts if operator is OOB. I prefer loud crash ASAP instead of silent memory corruption"
X Link @SebAaltonen 2025-10-21T10:02Z 46.1K followers, 29.8K engagements

"Sad news. I worked at RedLynx during my early career for XX years. Now they fire XX people and stop making console / PC games and focus only on mobile games. Ubisoft hasn't been doing well recently so changes like this were expected of course"
X Link @SebAaltonen 2025-10-23T06:42Z 46.1K followers, 6383 engagements

"I chose OLED for my phone TV and PC monitor. Laptop is still LCD (but has 2010 dimming zones). Waiting for M6 Max with dual-layer OLED. Old iPad Pro is still LCD. Reason: I have never personally seen burn-in in any of my devices. Used a Panasonic Plasma for XX years too"
X Link @SebAaltonen 2025-10-23T12:55Z 46.1K followers, 12.6K engagements

"Seems that people buy a new computer every X years. Also M4 was a great jump for entry level models. 16GB RAM and nice CPU/GPU bump. M3 was bigger jump for Max models (2x MT perf vs M1 Max) but entry level models are much more popular"
X Link @SebAaltonen 2025-10-24T13:17Z 46.1K followers, 9883 engagements

"Asked AI (Codex) to write me a 8k2 perlin noise heightmap generator and a simple triangle terrain renderer. I will use this as a comparison to develop my sparse voxel ray-tracer. AI written terrain shader calculates xy from vertex index as I asked but it's not optimal at all"
X Link @SebAaltonen 2025-10-25T20:56Z 46.1K followers, 12.6K engagements

"Indexed geometry still matters. Indexed = XXXX fps Non-indexed = XXXX fps Naive non-indexed terrain renderer does 6x vertex transforms per grid cell (2 triangles). Indexed does just one. That's 6x less vertex shader work. My indexed uses 32x32 terrain patches (instanced)"
X Link @SebAaltonen 2025-10-26T08:46Z 46.1K followers, 26.2K engagements

"Second day WebGPU hobby proto: Asked AI to improve the terrain generator + add exponential fog. Now it starts to look a bit like a real terrain. I am writing a voxel renderer. Used AI to write a quick hmap generator + triangle raster comparison"
X Link @SebAaltonen 2025-10-26T21:03Z 46.1K followers, 16.5K engagements

"I don't think AI is a bubble. Latest AI tools offer a real productivity boost. Not just for programmers but for everybody. X billion consumers paying $20/month = $XXX billion/year. Companies are willing to pay much more. 100$/month for professional AI is considered cheap"
X Link @SebAaltonen 2025-10-27T13:35Z 46.1K followers, 21.7K engagements

"AI writing idiomatic C++ as always. Uses a lambda in low level code. Lambda does just a BSF/CTZ + bitmask. The lambda is used in one place so it doesn't even reduce repetition. When I generate pure C with AI there's no problems like this"
X Link @SebAaltonen 2025-10-27T13:40Z 46.1K followers, 13.3K engagements

"Testing whether AI can write me a good changed (dirty) bitmask generator for a 64-byte struct (16x 4-byte elements). Apparently NEON doesn't have _mm_movemask_epi8 equivalent. That sucks. Also AI codegen extracts lanes X times. Could do it once at the end (bitpack in NEON)"
X Link @SebAaltonen 2025-10-28T08:21Z 46.1K followers, 9646 engagements

"@therealdecross We Finns love death metal. Taylor Swift is not as popular as it's in the USA"
X Link @SebAaltonen 2025-10-28T10:55Z 46.1K followers, XXX engagements

"I plug in a single 40Gbps USB-C cable between my computer and my monitor in the morning. Monitor has USB-hub connecting mouse and keyboard. Single USB-C cable carries the 4K 240Hz HDR image charges laptop at 65W and connects keyboard and mouse. Don't want a cable hell"
X Link @SebAaltonen 2025-10-28T14:13Z 46.1K followers, 217K engagements

"Going to skip M5 Max. It's more powerful than my M3 Max but this time I am doing the upgrade for me. Last time M1-M3 Max halved the compile times = business reason = save money. M3 Max is almost perfect. The only problem is display latency. M5 Max doesn't fix it. M6 does"
X Link @SebAaltonen 2025-10-28T15:15Z 46.1K followers, 35.9K engagements

"I had the urge to write my own graphics API too. I understand :)"
X Link @SebAaltonen 2025-10-28T16:05Z 46.1K followers, 7858 engagements

"@Bearspacegoat @Tymscar Yeah. There's several comparisons around and the standard authors did proper scientific research. There's no visible difference except in some artificial test patterns that are designed to break it"
X Link @SebAaltonen 2025-10-28T17:55Z 46.1K followers, XXX engagements

"Made a X week refactoring to port our UI rendering to use our new RHI. Was using GLES2-style wrapper before. UI rendering still is using a global RenderPass object. Using Codex (AI) to refactor away the global. I instructed it to pass the RenderPass object around. Example:"
X Link @SebAaltonen 2025-10-29T08:40Z 46.1K followers, 4762 engagements

"This refactoring is MUCH simpler than the one I did before. That X week refactoring had much more global state and iffy dependencies. AI is basically a glorified search-and-replace here. But it can do more complex transforms. This is monkey work so using AI is very helpful"
X Link @SebAaltonen 2025-10-29T08:42Z 46.1K followers, 2276 engagements

"Of course I will review and test every change the AI does. This refactoring will touch over XXX files. Mostly single line changes for passing the RenderPass around in function signatures. It's nice to have a better search-and-replace. Similarly AI is a better Google search too"
X Link @SebAaltonen 2025-10-29T08:43Z 46.1K followers, 2052 engagements

"We were missing instructions in our file about our project structure. Added more info to avoid AI modifying external libs in the future. This also seems to speed up the AI. Finds code faster as it doesn't need to scan all the data/build/libs folders"
X Link @SebAaltonen 2025-10-29T08:45Z 46.1K followers, 3557 engagements

"@KimmyOverflow The downsides: X. Docking station is only compatible with that laptop. When you upgrade your laptop you need to buy a new dock. X. Docking station takes table space. A single USB-C cable to the monitor takes no space and you don't need to replace it when you upgrade your laptop"
X Link @SebAaltonen 2025-10-29T10:04Z 46.1K followers, XXX engagements

"@badguy974 There's not many EU companies worth investing into. ARM and ASML are of course super important for the whole AI industry. TSMC uses ASML machines for producing all their chips for Nvidia/AMD/Apple. And ARM instruction set used in all phones tablets and Nvidia server CPUs"
X Link @SebAaltonen 2025-10-29T21:47Z 46.1K followers, XXX engagements