the divinity in you has awoken the divinity in me. I wish greater wealth, happiness and prosperity for you.
Joined October 2018
- Tweets7.4K
- Following3K
- Followers236
- Likes40.4K
Crypto Lelouch retweeted
Get it here:
github.com/MiaAI-Lab/DeepSee…
Crypto Lelouch retweeted
20 days ago we built a system that writes GPU code. Today its kernels outperform the state of the art. We just open-sourced 9 megakernels, which outperform NVIDIA’s stack on DeepSeek and run self-driving, robotics, protein folding and world models 2–3× faster than the current best systems. Every model runs in a single launch instead of thousands, and most of our kernels were generated in under an hour.🧵
We invented a recursive megakernel generator called Hyperkernel, the most complete solution to megakernels and GPU efficiency.
Hyperkernel-generated megakernels achieve state of the art performance on 11 different domains: inference (DeepSeek, Kimi and GLM), pretraining (Qwen), hardware fault recovery, robotics (π₀-FAST, ACT) diffusion and real time world models (Matrix-Game), full self driving (OpenPilot), voice models (Moshi), genomics (Evo 2) and protein folding (ESMFold).
All of this in a single kernel launch.
For context, Hyperkernel generated fully optimized megakernels for robotics, genomics and voice in under 50 minutes, and doubled the efficiency for >5 models generally using only about 60-100 lines of code per megakernel. This speaks to the unparalleled standard for code quality in the automated generations.
Our first release today open sources the megakernels themselves. Over the next couple weeks, we will release the entire Hyperkernel stack, along with the complete serving stack for Kimi and GLM (20-30% inference efficiency gains across all concurrencies).
Out of the box GPU machinery performs terribly for unorthodox research. We built Hyperkernel only 20 days ago for internal use; but given its multi-modality, we have decided to open source it.
Creative research should not be bottlenecked by the lack of dedicated infra teams. Hyperkernels offer a one stop shop infra stack for all research. We would love for the open source community to add to it :)
GitHub repo: github.com/Hyperstitioncc/Hy…
NOTE: We primarily use NVIDIA B300s, but our kernel generation extends to other hardware (B200s, AMD, TPUs).
1/9
We absolutely crush NVIDIA and achieve the best low latency inference for DeepSeek v4.1 Flash.
Compared with NVIDIA’s published vLLM results, Hyperkernel delivers:
• 36.7% higher per-user generation speed at concurrency 1.
• 28.2% higher per-user speed while serving two users, versus NVIDIA serving one.
• 15% higher per user speed while serving four users, versus NVIDIA serving one.
• 90.6% higher per-user generation speed at concurrency 4, compared with the highest published C4 AgentX interactivity for TP4 B300s.
• 28.5% higher per-user generation speed at concurrency 8, compared with NVIDIA's published result.
• 36% faster median per-request decode speed.
Crypto Lelouch retweeted
438 tok/s at 16 users DeepSeek V4.1 Flash on 4x DGX Spark, now open source:
80 chat / 138 code single stream
438 tok/s at 16 users
1M context, 10k tok/s prefill
image support soon
github.com/Picom-code/pico-d…
Crypto Lelouch retweeted
Dataset: huggingface.co/datasets/tenc…
Paper: arxiv.org/pdf/2611.02902
Crypto Lelouch retweeted
Tencent just released the TailorRig dataset on Hugging Face
39,073 rigged humanoid characters with skeletons and skinning weights - perfect for training auto-rigging and garment-aware skinning.
Crypto Lelouch retweeted
We open-sourced LivLiv Player, a 2.5D web character player, ~34 KB gzipped, with lip sync and webcam face tracking. Easy to embed in your site or app.
We also open-sourced our 21 featured characters, carefully made with @livliv_ai , are ready to use.
github.com/livliv-dev/player
Crypto Lelouch retweeted
Kimi K3.1 Pro open-weights are out.
Weights ~1.5TB
huggingface.co/moonshotai/Ki…
Crypto Lelouch retweeted
To save money, I gave Opus 5.5 a cheap DeepSeek V4.1-Flash-1017 to write the code... but the accuracy dropped and the bill went up 🫠
Opus 5.5 alone: 81% of tasks, $1.20 per task
Opus + DeepSeek V4.1-Flash-1017: 67%, $1.29 per task
DeepSeek’s Super-agent post tested scheduling on simulated coding (not real tasks). I tested 7 team setups on 12 real, hard ML coding tasks.
> Telling Opus “your tokens cost 20x more” instead of banning it from coding: the worker wrote 94% of the code, same 67%, and the cost per task dropped to $0.71
> Teams aren’t cheap: Opus re-caches its whole context after every long worker call, and that costs more than the worker itself
> 71% of team failures are tasks the worker also fails alone
Thread ↓
Crypto Lelouch retweeted
Introducing: Phone Harness MCP 🔌
Any agent that supports MCP can now control a real iPhone or Android in the cloud:
> works in Claude, ChatGPT, Cursor, Codex, and more as a custom connector
> nothing to install, just paste a URL and sign in
> sees the screen, taps, types, and swipes like you do
Available in phone-harness 0.3.3.
github.com/phone-harness/mcp
Crypto Lelouch retweeted
PAYWALL HACK:
Carousel paywalls absolutely destroy single-screen paywalls.
They allow you to tell a structured story rather than cramming too many features into one screen (e.g., Slide 1: Core benefit → Slide 2: Social proof → Slide 3: Pricing options)
Here's a great example:
Crypto Lelouch retweeted
Your Harness. You Define It.
Meet Creator Mode in DeepSeek Harness.
Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer.
What will you create?
Crypto Lelouch retweeted
done, it's open source now ✅
iOS phone farm - register real iPhones, watch + control them live in the browser, schedule tiktok on a postgres-backed scheduler.
free, self-hosted, apache-2.0.
- git: github.com/Git-Agni/prod-FAR…
- DIY steps: gethandler.ai/ios-farm
Crypto Lelouch retweeted
GLM-5.3-Flash on two DGX Sparks with NVIDIA's
FP4-Local weights: TensorFold is up to 1.5x faster
than vLLM for one user, and level at four.
One user, real prompts, tok/s (TensorFold vs vLLM):
JSON 51.5 vs 50.3
code 47.1 vs 41.0
rewrite 52.0 vs 40.7
explain 42.0 vs 33.3
story 36.3 vs 24.6
A 128K prompt reads in 91 s, and turn 2
comes from the cache in 1.3 s.
This is the stock TensorFold release, no patches.
My first PR to @ashxhart's TensorFold landed
this week, a small one. Happy to keep helping improve it.
GLM-5.3-Flash on two DGX Sparks with @NVIDIAAI's FP4-Local weights (TP=2): 31 tok/s for one user, 55 for four.
Decode, total tok/s (thinking off, 200-token prompts, 150 tokens out):
1 user 31.2
2 users 45.3
4 users 54.7
4 users at once is the server limit (max-num-seqs 4).
An 8k-token prompt reads in 3.6 s.
Engine: vLLM 8244d1c (eugr/spark-vllm-b12x nightly-20261015), the model card's command with a few changes. The important one: PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True on both nodes (from Bizuayeu's report in discussion #7 on the HF card). Without it, free memory fell to 4 GB right after the weights loaded. With it, the low point was 12 GB during the load and about 20 GB stays free while serving.
Crypto Lelouch retweeted
Abliteration, on demand ⚡️
FlashAlign abliterates per request, on the weights you already serve. No permanent changes.
By default, only the users, channels and tools you authorize get it. Everything else stays stock.
First step toward dynamic real-time alignment.
GitHub 👇
Crypto Lelouch retweeted
RIP paid AI memory tools.
MemPalace just hit 49K stars for one reason: it outperforms everything, runs locally, costs nothing, and never sends your data anywhere.
Most AI memory systems work like a junk drawer. You throw everything in and hope semantic search finds the right thing. MemPalace organizes memory into a structured palace where people and projects get wings, topics get rooms, and original text lives in drawers that can be searched individually without crawling the whole corpus.
The benchmark numbers are what shocked me. 96.6% retrieval recall on LongMemEval with zero API calls, zero LLM, zero cloud involvement. Add the hybrid pipeline and it hits 98.4% on questions the system had never seen. Every result is fully reproducible from the repo.
→ Verbatim storage, nothing gets summarized or rewritten ever
→ 29 MCP tools for palace reads, knowledge graph operations, cross-wing navigation, and agent diaries
→ Auto-save hooks for Claude Code keep memory updated before context compression hits
→ Python 3.9 plus ChromaDB is all you need, around 300MB disk and zero API keys
49K stars. MIT License. 100% Opensource.
github.com/MemPalace/mempala…
Crypto Lelouch retweeted
This is why I think AI will become significantly cheaper and more capable over time rather than more expensive and capable.
arxiv.org/pdf/2609.19969
