@yunta_tsai Not too many places offer such opportunity : felt lucky to have been in a few of them
English
Ryan Ng
24 posts
Ryan Ng
@aftermultiply
Reasoning @xAI | ex-@OpenAI | TL at Ray / Anyscale | K8s | DynamoDB
Katılım Şubat 2026
21 Takip Edilen33 Takipçiler
Pursuit of happiness, builder edition:
- Find a problem you love spending your life solving.
- Spend your life solving it.
English
I’ve been feeling a bit burnt out so I decided very last minute a few days ago to fly to scandinavia and detach a bit. As I was chilling in a random park in Copenhagen and taking this photo I overheard the couple next to me talking about world models and grounded video gen LOL
English
Excited to see people try Grok Build for web dev. Our team has put a lot of effort into improving its aesthetics, functionality, and more exciting features to be expected with recursive self-improvement loop.
It’s still early beta, and feedback is very welcome. Please try it out and let us know where we can improve.
Kilo@kilocode
Grok Build 0.1 might be one of the most underestimated AI models right now. We tested it in Kilo Code by asking it to build 5 websites from scratch. Here are the results:
English
learned this quote from 2023 is making rounds -- i actually don't think this is true anymore in 2026!
The model should be invisible.
i expect us to flip back to ux in the form of agent behavior + continual learning loops; and the alpha is in making models feel natural and as invisible as possible.
English
@CoreAutoAI “The discipline to focus on what matters before the training run, especially things like data quality and systems readiness”
English
Never understood the magic of @Tailscale until I started using it for myself
English
Best productivity hack I know is organizing your work so that you enjoy it the most
English
Interesting results
Nous Research@NousResearch
Today we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. Validated at 270M, 600M, and 3B dense scales, and at 10B-A1B MoE. The work on TST was led by @bloc97_, @gigant_theo, and @theemozilla.
English
Great blog, reminds me of the Bitter Lesson from the good old days during Macrohard @endernewton
Thinking Machines@thinkymachines
People talk, listen, watch, think, and collaborate at the same time, in real time. We've designed an AI that works with people the same way. We share our approach, early results, and a quick look at our model in action. thinkingmachines.ai/blog/interacti…
English
Ryan Ng retweetledi
Try out deepseek v4 on vLLM!
vLLM@vllm_project
🎉 Day-0 support for @deepseek_ai V4 Pro and Flash on vLLM — a new generation of DeepSeek model, purpose-built for tasks up to 1M tokens. Alongside the release, we're publishing a first-principles walkthrough of the new long-context attention and how we implemented it in vLLM. The new attention mechanism, in four moves: • Shared K/V + inverse RoPE → 2× memory savings • c4a / c128a KV compression → 4×–128× savings • DeepSeek Sparse Attention over compressed tokens • Short sliding window for locality across compression boundaries At 1M context, per-layer KV state is ~8.7× smaller than a DeepSeek V3.2-style 61-layer stack (9.62 GiB vs 83.9 GiB, bf16). fp8 attention cache + fp4 indexer cache shrink it further. vLLM side: • Unified hybrid KV cache — single logical block size (256 native positions) across all compression rates; compressor state folded into the SWA KV cache spec so prefix caching, disagg prefill, CUDA graphs and MTP reuse the same abstraction • Three page-size buckets for the full 5-way cache stack → no cross-kind fragmentation • Fused kernels: compressor + RMSNorm + RoPE + cache insert (1.4–3×), inverse RoPE + fp8 quant (2–3×), Q-norm + KV RoPE + K insert (10–20×) • Multi-stream overlap of indexer vs main-KV compression vs SWA insertion Disaggregated serving is supported out of the box and strongly recommended for best performance. Follow our recipes site for verified commands for @nvidia Blackwell (B200, B300, GB200, GB300) and Hopper (H100/H200/H20) systems. Thanks to the @deepseek_ai team for open-sourcing DeepSeek V4, and to @inferact for landing day-0 support 🤝 📝 Blog: vllm.ai/blog/deepseek-… 📖 Recipes: recipes.vllm.ai/deepseek-ai/De… 🤗 huggingface.co/deepseek-ai/De…
English
Ryan Ng retweetledi
Ryan Ng retweetledi
After an unforgettable ride, this was my last week at @xai.
When I joined in 2024, xAI was a small yet an extraordinary team. I got an amazing opportunity to build our core ML Platforms from scratch and form an ambitious, amazing team. Together we built the core systems that power frontier-model research, evaluations, human data collection, agent training, and daily productivity for every member of technical staff at xAI. It was a true privilege.
I’m deeply grateful to @elonmusk and every single person I’ve had the chance to work with across research, engineering, compute, infra, and beyond. The pace, collaboration, and shared mission to understand the universe and be truth seeking have been unmatched. Being able to contribute to data, training, and infra behind Grok-2-1212, Grok 3, Grok 4, 4.2, Aurora, Imagine and many other efforts been the highlight of my career so far.
Special thank you to my incredible team and all the brilliant people who made this ride so special. I leave more inspired than ever.
Excited for what’s next. Ad astra 🚀
English
Introducing Dex: the self-driving workspace for operators.
Dex is the first agent system with full operational context and a self-updating knowledge base.
Every datapoint from your workspace is ingested, synced, and structured into compounding context for agents to take action.
Comment "DEX" or tag @dexbythirdlayer access.
First 1,000 sign-ups get 7 days free.
After, join our rolling waitlist.
Sign up at joindex [dot] com for a fun surprise.
How dex works (threads)
English
Reasoning models think hard — but all that thinking fills up your KV cache fast.
Memento fixes this: the model compresses its own chain-of-thought mid-generation, flushing old KV entries after each block. 2-3× less peak KV cache, ~2× throughput — accuracy largely preserved.
The cool part: deciding what to remember and what to forget is a capability the model acquires through training — not something you bolt on.
Excited about where this goes — especially for agents.
Dimitris Papailiopoulos@DimitrisPapail
English
