AI aficionado 20y ML & AI, top-0.1% on chess com
Joined August 2013
- Tweets50.4K
- Following836
- Followers1.8K
- Likes216K
Ded Inside retweeted
О, чел... Уже сейчас Opus системно отказывается работать в серии разных задач. И речь уже не только про dual-use и security. С прошлой недели они запретили даже разбор asset'ов игр. Я сам неоднократно сталкивался с тем, что Opus отказывается работать над рандомной задачей просто потому что.
Кучка корпораций не может и не должна определять как мы должны использовать все разнообразие productivity tools на базе АИ. Потому им нельзя дать монополизировать такой ресурс как DRAM.
People crying over this are idiots. Clearly the RAM in a data center running full tilt 24/7/365 serving millions of people is better allocated than the RAM in your machine powering your Reddit browsing.
Readers added context they thought people might want to know
Sam Altman was specifically discussing local AI inference on personal computers versus cloud needs, not general use like web browsing. Cloud servers average ~13-22% memory utilization per analyses of major providers, not full tilt 24/7.
theverge.com/ai-artificial-…
engadget.com/2303266/sam-al…
datacenterdynamics.com/en/news/only-1…
datacentremagazine.com/news/are-data-…
Ded Inside retweeted
I was very happy that you guys liked the idea!
Here is the first real battle:
Qwen3.8-27B on my RTX 5090 vs. DeepSeek-V4.1-Flash on my Mac Studio.
RTX 5090:
Qwen3.8-27B
NVFP4 and FP8 mix, NInfer engine, MTP 3
Mac Studio M5 Ultra:
DeepSeek-V4.1-Flash-1017
Bob's TensorFold Metal recipe
MLX, experts at 2 and 3 bits
This is the strategy game I'm building. It runs in real time. Both models think before they act, no human input.
DeepSeek runs greedy as in Bob's recipe. Qwen runs at temperature 1.0 as on its model card.
DeepSeek scores higher on my knowledge and code tests. Qwen decodes faster. One match is not a benchmark.
Everything you see here, including the voices, was run locally.
Both recipes, with every number:
RTX 5090: niklaslenz.com/v2/qwen38-27b…
Mac: niklaslenz.com/v2/deepseek-v…
Recipe by @BobClawblaw on @ashxhart's TensorFold. NInfer by neroued. Models by @deepseek_ai and @Alibaba_Qwen.
COMING SOON!
I built a strategy game where local LLMs play against each other. Any model, any hardware.
Pick a model, pick a box, let them fight. Spark against GPU, Mac against Spark, small model against big model. Same map, same rules.
Here is one real match.Qwen3.8-Flash-Next with @MiaAI_lab 's recipe on a DGX Spark against Qwen3.8-27B on my RTX 5090.
No human input. The models see the map through fog of war and command their side through 21 tools.
They run an economy, build a base, train an army, scout, take shrines and decide when to strike. Over 200 orders in one match.
And it runs in real time. The game never waits while a model thinks. Speed and smarts both count.
Very excited for this one! 😁
Recipe by @MiaAI_lab on @ashxhart's TensorFold. NInfer by neroued. Models by @Alibaba_Qwen.
Ded Inside retweeted
Билл Гросс выложил 1,2 млн бизнес идей по AI в открытый доступ. Никогда еще фаундеры не ощущали себя такими могущественными и жалкими одновременно как в 2026 idealab.ai
Ded Inside retweeted
"Putin scatenerà la guerra. Non so quando, ma lo farà. E voi, europei, rimarrete sorpresi nello scoprire che questa guerra è diretta anche contro di voi..."
Anna Politkovskaya, a Parigi, 2005.
Fu assassinata nell'atrio del suo palazzo il 7 ottobre 2006—il compleanno di Putin.
Ded Inside retweeted
basaltとStrataをRTX5090 DDR5-3600 で速度比較してみた (QFN IQ3_S)
■ basaltって何？
StrataをベースにBlackwell GPU向けのExpertキャッシュ管理やCPU/GPU協調処理、MTPを最適化した推論エンジン
basalt
github.com/jesdga95/basalt
実測の結果
■ 単独 Prefill｜Prose｜Code (tok/s)
・Basalt 6950｜273｜283
・Strata 5250｜164｜182
単独でもBasaltが3項目とも上回った
prefill は+1700tok/s
proseは +100tok/s
codeも+100tok/s
単独に関しては想像以上にbasaltで良い結果が出た
■ 実効prefillの並列性能(2枚目)
並列になるとbasaltのprefillにはクセがあり、同時実行数の増加とともに実効prefill速度が落ちこむ
8並列の実効prefillは約2000tok/sでStrataは約5100tok/sのまま
リクエストごとに見るとStrataは1本あたり約1.6秒で一定、Basaltは2本目以降が3〜5秒かかっていた(単独時は1.3秒)
BasaltはPrefill中もDecodeが止まらず並行して処理はしていたので、Decodeとの計算資源の取り合いが原因かもしれない
ここは要検証
■ 8並列合計デコード Prose｜Code
・Basalt 451｜489
・Strata 216｜188
・Strata --batch-mtp 197｜184
8並列の合計はBasaltがStrataの約2倍
ただし8並列時のCodeは389〜689とブレ幅が大きい
Strataは--batch-mtpを入れてもほぼ横ばいであまり効果がなかった
■ 感想
同じIQ3_Sでここまで爆速なら最適化は夢がある、かと思いきや同時実行数が増えるとprefill周りが落ち込む結果に
prose8並列で450tok/sなら一人あたり56tok/s程度、出力に関してはとてもよい
次はなぜここまでデコード速度に差が出るのか、並列時のprefill低下の原因は何か追加の検証もしながら理解を深める予定
Ded Inside retweeted
Yesterday France announced tax of €0.055 per GB on internet traffic
Today Germany is considering taxing all AI-generated revenue
You will see European countries keep clawing more and more money from foreign companies as they are slowly running out of money themselves to pay for their pensions and welfare state
Add to this the planned wealth taxes and exit taxes on any European who will try to leave (already active in many EU countries)
Europe is running out of money
Readers added context they thought people might want to know
The post inaccurately claims France "announced" a data tax; it was a narrow 11-10 vote on a budget amendment in a near-empty chamber, opposed by the government and far from law. German AI tax ideas remain unadopted ministerial discussions without formal proposals.
stephanelarue.com/taxe-au-gigaoc…
tagesschau.de/inland/ministe…
davemanuel.com/2026/10/30/wha…
Ded Inside retweeted
🚀一张 3090，把千问3.8-27B 在 vLLM 上开到 64 路并发，合计破一千 tok/s
开源项目 HyperQwen：单卡 24G，3090 限 250W 实测。
单人聊天约 127 tok/s，64 并发合计约 1035 tok/s，最长可开到 15 万上下文。
两条命令就能起：docker compose --profile single 或 batch。
视频是官方录的：先看单路吐字，再切 64 路一起跑。
温馨提示：第一次会拉约 9.5GB 镜像，再本地 requant 一次模型大约 20GB，等完再测速度。
你家 24G 卡跑千问，单路现在大概多少 tok/s？
来源：GitHub syv-ai/HyperQwen👇
github.com/syv-ai/HyperQwen
Ded Inside retweeted
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
Ded Inside retweeted
Imagine being given the gift of life and wasting 1/3rd of it staring at your phone.
Ded Inside retweeted
So much of the climate hysteria is just aimless guilt roaming the earth in search of futile acts of atonement.
Ded Inside retweeted
DDR: Das Wohnen war so toll, dass 1989 26% nicht mal eine eigene Toilette hatten. Nur 18% hatten ein Telefon (BRD 99%). Auf einen Trabi musste man 12,5 bis 17 Jahre (!) warten. Urlaub nur in sozialistischen Bruderländern, wer in den Westen floh, auf den wurde geschossen.
