What a cool video about how GTA3 optimized its memory usage on the PS2 (yes, I know, boring ๐Ÿค“).

It reminds me of the optimizations made on RollerCoaster Tycoon, where technical limits were part of the game design.

Now it feels like technical limits feel like a burden for creativity and not the opposite.

Its also wild that people are porting it to the Dreamcast.
#videogames #playstation #gta3 #MemoryOptimization

https://www.youtube.com/watch?v=cIbCxbrBCys

How Rockstar fit an entire city into PlayStation 2 memory

YouTube

Classic โ€“ Hacker News

Hacker News์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ ๋ชฉ๋ก์—์„œ๋Š” Rust๋กœ ์ž‘์„ฑ๋œ Unix ์˜๊ฐ์„ ๋ฐ›์€ ์ฝ”๋”ฉ ์—์ด์ „ํŠธ Zerostack, ์˜คํ”ˆ์†Œ์Šค 2.6B ํŒŒ๋ผ๋ฏธํ„ฐ ์›”๋“œ ๋ชจ๋ธ SANA-WM, ๊ทธ๋ฆฌ๊ณ  LLM ๋ฉ”๋ชจ๋ฆฌ ์ตœ์ ํ™” ์—ฐ๊ตฌ ฮด-mem ๋“ฑ AI ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์œ ์šฉํ•œ ์ตœ์‹  ๋„๊ตฌ์™€ ์—ฐ๊ตฌ๊ฐ€ ๋‹ค์ˆ˜ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ LLM ๊ด€๋ จ ๋…ผ๋ฌธ๊ณผ ์˜คํ”ˆ์†Œ์Šค ํ”„๋กœ์ ํŠธ๊ฐ€ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ์œผ๋ฉฐ, AI ์—์ด์ „ํŠธ ๊ตฌ์ถ•๊ณผ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ์ตœ์ ํ™”์— ๊ด€์‹ฌ ์žˆ๋Š” ๊ฐœ๋ฐœ์ž์—๊ฒŒ ์ฐธ๊ณ ํ•  ๋งŒํ•œ ์ •๋ณด๊ฐ€ ํ’๋ถ€ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, AI ์‹ฌ๋ฆฌํ•™ ๊ด€๋ จ ๋…ผ์˜์™€ ๋Œ€๊ทœ๋ชจ ์Šคํ† ๋ฆฌ์ง€ ์†”๋ฃจ์…˜ ๋“ฑ AI ์ธํ”„๋ผ์™€ ์—ฐ๊ด€๋œ ์ฃผ์ œ๋„ ๋‹ค๋ค„์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

https://news.ycombinator.com/classic

#rust #llm #opensource #aiagent #memoryoptimization

Classic | Hacker News

Dan McAteer (@daniel_mac8)

Meta์˜ ์ƒˆ๋กœ์šด ๋ฉ”๋ชจ๋ฆฌ ํšจ์œจํ™” ๊ธฐ๋ฒ• 'KV self-pruning'์„ ์†Œ๊ฐœํ•œ ํŠธ์œ—์ž…๋‹ˆ๋‹ค. LLM ํ•™์Šต ์ค‘ ์œ ์šฉํ•˜์ง€ ์•Š์„ ๊ฒƒ์œผ๋กœ ์˜ˆ์ธก๋˜๋Š” KV pair๋ฅผ ์„ ํƒ์ ์œผ๋กœ ์žŠ๊ฒŒ ํ•ด์„œ, ์„ฑ๋Šฅ์€ ์œ ์ง€ํ•˜๋ฉด์„œ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์„ 15~35% ์ˆ˜์ค€์œผ๋กœ ์ค„์ด๊ณ  ์ฒ˜๋ฆฌ๋Ÿ‰๋„ ๋†’์ธ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. LLM ํ•™์Šต/์ถ”๋ก  ์ธํ”„๋ผ ์ตœ์ ํ™”์— ๋ฐ”๋กœ ๊ด€์‹ฌ ๊ฐ€์งˆ ๋งŒํ•œ ์†Œ์‹์ž…๋‹ˆ๋‹ค.

https://x.com/daniel_mac8/status/2055308764547367416

#llm #memoryoptimization #meta #training #throughput

Dan McAteer (@daniel_mac8) on X

cool new memory efficiency technique from @AIatMeta. 'kv self-pruning': during training an llm is given the option to forget kv-pairs it predicts will not be useful. uses only 15 - 35% of the memory with parity on performance. as a bonus, it also increases throughput. great

X (formerly Twitter)

The understated loading design inside Transformers that saves memory

Transformers ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” PyTorch์˜ meta device๋ฅผ ํ™œ์šฉํ•ด ๋Œ€ํ˜• ๋ชจ๋ธ์„ ๋ฉ”๋ชจ๋ฆฌ ๋‘ ๋ฐฐ ์‚ฌ์šฉ ์—†์ด ํšจ์œจ์ ์œผ๋กœ ๋กœ๋”ฉํ•˜๋Š” ๋ฐฉ์‹์„ ๊ตฌํ˜„ํ–ˆ๋‹ค. meta device๋Š” ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ๋งŒ ๋ณด์œ ํ•ด ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ์„ ์ตœ์†Œํ™”ํ•˜๋ฉฐ, safetensors ์Šฌ๋ผ์ด์Šค๋ฅผ ํ†ตํ•ด ํ•„์š”ํ•œ ํ…์„œ๋งŒ ์ง€์—ฐ ๋กœ๋”ฉํ•œ๋‹ค. ๋˜ํ•œ ๋น„๋™๊ธฐ ๋ฐ ๋™๊ธฐ ๋กœ๋”ฉ ๊ฒฝ๋กœ๋ฅผ ์ƒํ™ฉ์— ๋งž๊ฒŒ ์„ ํƒํ•˜๊ณ , ๋””์Šคํฌ ์˜คํ”„๋กœ๋”ฉ์„ ์ง€์›ํ•ด ๋ฉ”๋ชจ๋ฆฌ ๋ถ€๋‹ด์„ ์ค„์ธ๋‹ค. ์ด๋Ÿฌํ•œ ์„ค๊ณ„๋Š” 70B ์ด์ƒ์˜ ๋Œ€ํ˜• ๋ชจ๋ธ๋„ ์ œํ•œ๋œ ๋ฉ”๋ชจ๋ฆฌ ํ™˜๊ฒฝ์—์„œ ํšจ๊ณผ์ ์œผ๋กœ ๋‹ค๋ฃฐ ์ˆ˜ ์žˆ๊ฒŒ ํ•œ๋‹ค.

https://www.stevhliu.com/2026/transformers-compendium-1

#transformers #pytorch #memoryoptimization #modelloading #safetensors

Transformers Compendium - Part 1

A collection of engineering details and design in Transformers.

leopardracer (@leopardracer)

16GB ๋ฉ”๋ชจ๋ฆฌ๋กœ๋Š” 35B ๋ชจ๋ธ์„ ๋Œ๋ฆฌ๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๊ธฐ์กด ์ธ์‹์„ ๋’ค์ง‘๋Š” ์„ค์ • ํ”Œ๋ž˜๊ทธ๊ฐ€ ๋“ฑ์žฅํ–ˆ๋‹ค๋Š” ๋‚ด์šฉ์ด๋‹ค. ๋Œ€ํ˜• ์–ธ์–ด๋ชจ๋ธ์˜ ๋กœ์ปฌ ์‹คํ–‰๊ณผ ๋ฉ”๋ชจ๋ฆฌ ์ตœ์ ํ™”์— ์œ ์šฉํ•œ ๊ธฐ์ˆ ์  ๊ฐœ์„ ์œผ๋กœ ๋ณด์ธ๋‹ค.

https://x.com/leopardracer/status/2043979806958596551

#llm #localai #memoryoptimization #35b #aimodel

leopardracer (@leopardracer) on X

Everyone said 16GB isnโ€™t enough for a 35B model. They were right. Until this one flag.

X (formerly Twitter)
๐ŸŽ‰ Wow, a groundbreaking realization! ๐Ÿง  Memory optimization is back in style because AI hoarders allegedly bought all the #RAM. Who knew? Next up: inventing fire! ๐Ÿ”ฅ
https://nibblestew.blogspot.com/2026/03/everything-old-is-new-again-memory.html #memoryoptimization #AIhoarders #shortage #technews #groundbreakinginnovation #HackerNews #ngated
Everything old is new again: memory optimization

At this point in history, AI sociopaths have purchased all the world's RAM in order to run their copyright infringement factories at full bl...

Everything old is new again: memory optimization

At this point in history, AI sociopaths have purchased all the world's RAM in order to run their copyright infringement factories at full bl...

Jarred Sumner (@jarredsumner)

Claude Code์˜ v2.1.47 ์—…๋ฐ์ดํŠธ์—์„œ ์žฅ์‹œ๊ฐ„ ์‹คํ–‰๋˜๋Š” ์ฝ”๋“œ ์„ธ์…˜์˜ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์ด ๊ฐ์†Œํ–ˆ๋‹ค๋Š” ๊ณต์ง€์ž…๋‹ˆ๋‹ค. ๊ฐœ์„ ์€ @cirospaciari์˜ ๊ธฐ์—ฌ๋กœ ์ด๋ฃจ์–ด์กŒ์œผ๋ฉฐ, ์‚ฌ์šฉ์ž๋Š” ๋ฌธ์ œ๋ฅผ ๊ณ„์† ๋ณด๊ณ ํ•ด ๋‹ฌ๋ผ๋Š” ์•ˆ๋‚ด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ์ž ํˆด ์„ฑ๋Šฅ ํ–ฅ์ƒ์— ๊ด€ํ•œ ์ค‘์š”ํ•œ ๋งˆ์ด๋„ˆ ์—…๋ฐ์ดํŠธ ์†Œ์‹์ž…๋‹ˆ๋‹ค.

https://x.com/jarredsumner/status/2024289291879534793

#claude #memoryoptimization #release #developertools

Jarred Sumner (@jarredsumner) on X

Long-running Claude Code sessions use less memory in v2.1.47, thanks to @cirospaciari Keep reporting issues and the team will fix

X (formerly Twitter)

Vali Neagu (@AmbsdOP)

์›๋ž˜ Gradio ๋ฒ„์ „๊ณผ ๋™์ผํ•œ API ํ˜ธ์ถœ์„ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ์œผ๋‚˜, cover ๊ธฐ๋Šฅ์—์„œ Apple ๊ธฐ๊ธฐ์—์„œ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์ด ๋†’๊ฒŒ ๋‚˜ํƒ€๋‚˜๋Š” ๋ฌธ์ œ๋ฅผ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค. ํ˜„์žฌ ๋ฉ”๋ชจ๋ฆฌ ์ตœ์ ํ™”์— ์ง‘์ค‘ ์ค‘์ด๋ฉฐ ๊ณง PR์„ ์˜ฌ๋ฆด ์˜ˆ์ •์ด๋ผ๊ณ  ์•Œ๋ ธ์Šต๋‹ˆ๋‹ค. ๊ฐœ๋ฐœ์ž์šฉ ํˆด์˜ ์„ฑ๋Šฅ ๊ฐœ์„  ๊ด€๋ จ ์ง„ํ–‰ ์ƒํ™ฉ์„ ๊ณต์œ ํ•˜๋Š” ์—…๋ฐ์ดํŠธ์ž…๋‹ˆ๋‹ค.

https://x.com/AmbsdOP/status/2019503866929164666

#gradio #memoryoptimization #apple #api #pullrequest

Vali Neagu (@AmbsdOP) on X

@joanplanas @cocktailpeanut We are doing the same API call as the original Gradio version, but I noticed some high memory usage on Apple devices for the cover feature. Right now, I'm focusing on memory optimization. I will push a PR soon.

X (formerly Twitter)

gatehouse (@imangegatehouse)

ํŠธ์œ—์€ @deepseek_ai๊ฐ€ AI ์ถ”๋ก ยทํ•™์Šต์—์„œ ๊ณ ๊ฐ€์˜ HBM(High-Bandwidth Memory) ํ•„์š”์„ฑ์„ ์ œ๊ฑฐํ•ด ๋ฉ”๋ชจ๋ฆฌ(RAM) ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ๋ฐฉ๋ฒ•์„ ์ฐพ์•˜์„ ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ DRAM ๊ฐ€๊ฒฉ์ด 10์ฃผ ๋งŒ์— 5๋ฐฐ ์ƒ์Šนํ–ˆ๋‹ค๋Š” ์ ์„ ์–ธ๊ธ‰ํ•˜๋ฉฐ ํ•˜๋“œ์›จ์–ด ๋น„์šฉ ์ ˆ๊ฐ๊ณผ ๋ฉ”๋ชจ๋ฆฌ ํ˜์‹ ์˜ ์ž ์žฌ์  ์˜ํ–ฅ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

https://x.com/imangegatehouse/status/2013167288728338722

#hbm #dram #memoryoptimization #aiinference

gatehouse (@imangegatehouse) on X

โ€œ@deepseek_ai may have found a way to solve the RAM crisis by eliminating the need for expensive HBM for AI inference and training โ€” yes, the very reason why DRAM prices went up by 5X in 10 weeksโ€ https://t.co/vPRamjORKE

X (formerly Twitter)