←
TheDavidTai4,878 posts
TheDavidTai avatar

TheDavidTai

@TheDavidTai

Founder, hacker, stochastic tool. @microsoft escapee. Currently @eigenlabs. Checkout my github actions emulator: notdownhub.com

◷ Joined September 2018▧ 477 Photos and videos
TheDavidTai avatar
TheDavidTai@TheDavidTai·Oct 29

EXL3 has one of the best ratios of accuracy per bit and now its also one of the fastest format for M5 and NAX enabled Macs! TurboEXL3 is a fused kernel library tested on Qwen 3.8 Flash. It can running the model up to 140TPS decode using full EXL3 conversion by @turboderp_ with equal prefill performance. My TurboEXL3 library is faster than the standard Q4 block based quants. It achieves this through fusing the encoding and decoding kernels directly into each batched GEMM and GEMV calculation. The accuracy gains from EXL3 then boost the token drafting which offsets any performance losses from EXL3. You can choose EXL3 at 5.05 bits for nearly lossless quality compared to BF16. Choose EXL3 @ 3.05 bits for a version with the accuracy of NVFP4 with 33% smaller size. Available soon on mlx-serve. See the attached article for full technical details. The benchmarks were run with thinking low and greedy drafting.

1:00
Adina Yakup avatar
⟳ TheDavidTai retweeted
Adina Yakup@AdinaYakup·4h

Tencent just released Hy4 flash 🔥 huggingface.co/collections/t… - 220B / A12B - 1M context - Custom license - Native text & image understanding - 20x lower cost per task vs Hy4 - FP8 & GGUF available

Post attachment

Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base. 224B (12B active), 1M context, natively multimodal. Fast, efficient and built for everyday productivity. It's yours now. And stay tuned for Hy4 coming up next! 🤗: huggingface.co/tencent/Hy4-f… 🐙: github.com/Tencent-Hunyuan/H… 📄: hy.tencent.ai/research/hy4-f… Released under the Tencent Hy License. See full terms here: huggingface.co/tencent/Hy4-f…

Post attachment
rohit avatar
⟳ TheDavidTai retweeted
rohit@krishnanrohit·19h

I dug into this a bit. It looks like we're at highs, but not unprecedented highs for major systems failures. We're seeing alarming headlines every day, but a lot of it is recency bias.

Post attachment

Is it possible we're hearing about this stuff more because people are more jumpy about cyberattacks?

Chris Kearney avatar
⟳ TheDavidTai retweeted
Chris Kearney@SiliconForested·11h

🎉 Introducing SiliconForest: a new open operating system built on Darwin and XNU, the open-source core of macOS, running natively on an M1 MacBook Pro. Here’s the interesting part: it runs unmodified Mac apps. This first version runs Apple’s Terminal and Activity Monitor, copied straight from a macOS install, plus the stock Mac build of Firefox. Working today on a 13" M1 MacBook Pro: GPU acceleration (Vulkan 1.4), hardware video decoding, Wi-Fi, sound, keyboard and trackpad, and the Touch Bar. It’s Apple’s open-source kernel, with the parts Apple doesn’t publish reimplemented from scratch. I’ve also brought my Wayland compositor, Chonkstep, over to Darwin, and the results so far are fantastic. The idea: a free, hackable OS that can run native Mac apps. It’s early. Big apps like Logic Pro and Photoshop need a lot more (Metal, for a start), but that’s the direction. It stands on great work: PureDarwin (which we fork), the Asahi Linux GPU and hardware drivers, Omacom Mesa, and the Finch frameworks project. Source is coming soon. Right now it runs on one machine, the 13" M1 MacBook Pro. If you’d like to help bring it to other M-series Macs, get in touch. github.com/chriskearney/sili…

1:10
TheDavidTai avatar
TheDavidTai@TheDavidTai·10h

That's quite impressive. Feels like Microsoft is selling close to cost.

Microsoft have just made a comeback to challenge frontier models. Performance is close, but cost is a win.

Post attachment
TheDavidTai avatar
TheDavidTai@TheDavidTai·13h

OTOH its Sol 6.2 at home.

Pay attention here the situation is crazier than I thought, GPT Sol 6.2 at low is much better than deepseek max a quarter of the price :) Someone might say no not really it probably isn't at real tasks okay bro Sol 6.2 on high makes the gap unbelievably huge and still cheaper than fu*king Deepseek. The only reason to use deepseek now is running it locally, the API is so dead at this pricing.

Post attachment
TheDavidTai avatar
TheDavidTai@TheDavidTai·14h

We might live in a time line where Jeff Bezo's defective engines might be handicapping every one of SpaceX's major US competitors.

ULA Posted an update confirming what we'd all guessed, Vulcan failed because of low booster performance. Centaur operated correctly but was too low and slow, and the boosters which had caused problems on previous flights operated correctly.

Post attachment
TheDavidTai avatar
TheDavidTai@TheDavidTai·14h

I think this is Claude getting smarter and being able to handle task scheduling in a single agent. This does not work as well on other models. The harnesses were a crutch for needing to have a dedicated agent and session to organize work because the agent couldn't do that. Fable was the first one that could really do tasking well but not its in Opus 5.5.

This by Anthropic seems both right & a condemnation of many “agentic solutions” & harnesses being built at companies. Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all claude.com/resources/article…

Post attachmentPost attachment
OpenAI avatar
⟳ TheDavidTai retweeted
OpenAI@OpenAI·19h

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
SuSu_酥酥👅 avatar
⟳ TheDavidTai retweeted
SuSu_酥酥👅@NFT_Chen·Nov 17

🤨炸了!新美国开源模型 Maxwell-V1-Flash竟然碾压 Qwen3.6-35B-A3B!从零训练,不是Qwen微调! 35B总参/3B激活,Apache 2.0开源,权重仅~73.6GB,轻松塞进单卡128GB DGX Spark。 性能亮点: 🔹Terminal-Bench 58.65(Qwen 47.42)|SWE-bench Verified 65.4(Qwen 63.4) 🔹BrowseComp 76.3(Qwen 47.0)|Tau3-Bench 67.37,Agentic搜索直接起飞 🔹256K长上下文RULER 93.35(Qwen Base 85.75),Block Sparse Attention提速6.7x 🔹数学推理/多语言/指令跟随全面领先Qwen3.6-35B-A3B、Gemma等同类 架构亮点:256专家/12激活 + 4残差流mHC + Multi-token Prediction + 26T多语言预训练。 本地部署新选择,Agent/Coding/长文一把梭。 GitHub技术报告:github.com/maxwell-model/Max… HF权重:huggingface.co/maxwell-model… #MaxwellV1Flash #OpenSourceLLM #MoE #本地大模型 #AIAgent #Qwen挑战者 #美国开源

Post attachment

🤯牛逼!开源AI又赢麻了!Qwen3.8 27B、DeepSeek V4.1 Flash 这些本地量化版模型,现在普通人现在也能超方便调用了! Hugging Face 和 OpenRouter 正式联手:
✅ 在HF直接用OpenRouter超稳推理
✅ 开源模型权重一键直达
✅ Agent调用记录还能存回HF,方便微调分享 以前搞AI像打零和战争,现在开源靠协作把门槛打下来了! 更多协作,更少内卷,普通人也能玩转顶级模型 #OpenSourceAI #HuggingFace #OpenRouter #Qwen #DeepSeek #GLM x.com/nft_chen/status/212209… x.com/OpenRouter/status/2122…

NVIDIA Newsroom avatar
⟳ TheDavidTai retweeted
NVIDIA Newsroom@nvidianewsroom·Nov 17

The NVIDIA Board of Directors has established the National Security and Export Control Oversight Committee. Read the release: nvda.ws/4eErJ5F

🥨 Richard avatar
⟳ TheDavidTai retweeted
🥨 Richard@pretzelhds·Nov 17

Did you know that running your Mac for more than 50 days without a reboot will fuck with your internet connection? Yes, really. There is a 32-bit integer timing your TCP connections and if you let it overflow MacOS will run into all kinds of fucky issues with its TCP stack. For example I was unable to upload things bigger than 4MB. I was wondering why my Claude sessions started to hang seemingly at random. The fix? Literally just reboot your machine. That's it. What the fuck photon.codes/blog/we-found-a…

Post attachment