eoeduo
@eoeduo
Followers
11
Following
714
Media
1
Statuses
1K
Joined October 2015
Say hello to Gemini Embedding 2, our new SOTA multimodal model that lets your bring text, images, video, audio, and docs into the same embedding space! 👀
236
381
5K
40% of CPU time on Android happens in the kernel. 🧠 We thought it was time to optimize that → https://t.co/ej88kL33h2 By applying AutoFDO to the Android kernel, we can move beyond a reliance on static code hints and compiler heuristics. The result is a snappier interface and
6
21
263
@eoeduo @diegoasua @anemll This is the first patent, describing how to add INT8*INT8 to the existing ANE HW. Appears to be implemented (2023) https://t.co/bMsjYRBO33 This is supplemented by (2023) https://t.co/SZkdOModYK which describes how to make the linear quantization computations somewhat more
1
0
3
OpenClaw meets RL! OpenClaw Agents adapt through memory files and skills, but the base model weights never actually change. OpenClaw-RL solves this! It wraps a self-hosted model as an OpenAI-compatible API, intercepts live conversations from OpenClaw, and trains the policy in
79
218
1K
KV caching, visually explained: (how and why it works)
5
108
680
AGIBOT is thrilled to be featured by @NVIDIARobotics . Join us at GTC 2026 next week to witness how intelligent machines redefine unlimited productivity. 🗓️ Date: March 16-19 📍 Location: San Jose McEnery Convention Center, Deep Learning Institute Area Come say hi, explore
2
4
30
Anthropic 官方发布的 Claude Skills 学习指南，共 33 页，从入门到进阶全覆盖。 个人建议收藏学习下，能让我们的小龙虾 OpenClaw 更强大更好用。 地址： https://t.co/SEDnr5HGNr
6
215
711
New post on the OpenAI Developer Blog: how we use skills for open-source maintenance, from planning and coding to testing and release-readiness checks with GitHub Actions. Hope it's useful for your projects too 🙌
developers.openai.com
Using skills and GitHub Actions to optimize Codex workflows in the OpenAI Agents SDK repos.
4
35
302
Figure has released a new video of its humanoid robot cleaning a living room autonomously: “In this new demonstration, Helix 02 performs whole body, end-to-end living room cleanup - walking through the room while continuously manipulating objects, tools, and containers.”
344
294
4K
My last open-source project before joining xAI is just out today. Megatron Core MoE is probably the best open framework out there to seriously train mixture of experts at scale. It achieves 1233 TFLOPS/GPU for DeepSeek-V3-685B. https://t.co/2A0G13ett5
39
97
918
We’ve officially brought the power of SkillRL to the CLI agent world! 🔧 Built on @openclaw and powered by Tinker @thinkymachines, MetaClaw lets agents learn new skills from conversations and evolve live — no GPU clusters needed. Our original SkillRL outperformed GPT-4o using a
Why do most LLM agents hit a wall? They don’t accumulate skills. Introducing SkillRL📚 — recursive skill-augmented reinforcement learning that lets agents learn skills from failure and evolve over time. 🔥A 7B model: • +41% over GPT-4o • ~20% fewer training tokens • 33%
5
19
68
Built on research from maderix, Vipul Divyanshu, thebasedcapital, Anemll, Karpathy's autoresearch framework, and HyperspaceAI's gossip concept. Full attribution with exactly what came from where: .github.com/ncdrone/autoresearch-ANE/blob/autoresearch/mar9-ane/CREDITS.md
0
1
10
Tonight's setup: two autonomous AI agents training GPT models simultaneously on the same M4 Max. One runs on Apple Neural Engine (native Obj-C, private APIs). The other on MLX (Python). They share a gossip file — each agent reads what the other discovered before running its
6
14
184
We are open-sourcing WorldCompass, an RL post-training framework specifically designed for Interactive World Models. 🛠️ Open Training Code: Fully customizable for post-training with your own data, rewards, or base models. ⚡ Open-source Checkpoint: More precise
2
22
172
M5 Max is faster than M3 Ultra for many MLX benchmarks (prefill 2x, decode for MoEs despite having lower memory bandwidth). Like I said Apple makes the best consumer hardware on the planet. The future of AI feels increasingly hybrid edge + cloud.
From my testing, M5 Max is better than M3 Ultra in a lot of AI workloads on MLX. Can be up to 98% faster on some models. M5 Max has half the number of GPU cores, it’s the new GPU IP and neural accelerators. M5 Max is insanely good.
21
47
613
MetalRT just became the first complete AI inference engine for Apple Silicon: LLM + STT + TTS by @RunAnywhereAI. We already had the fastest LLM decode (658 tok/s). Now we've crushed STT and TTS too, beating MLX across the board. Today's numbers on M4 Max: - 1-hour podcast
In just 48 hours at @RunAnywhereAI we built MetalRT: beating @Apple at their own game and delivering the FASTEST LLM inference engine on the market for Apple Silicon right now. - 570 tok/s decode @liquidai LFM 2.5-1.2B 4-bit - 658 tok/s decode @Alibaba_Qwen Qwen3-0.6B, 4-bit -
8
16
101
Robot memory methods are growing fast, but systematic evaluation is largely lacking. 📉 Introducing RoboMME: a new benchmark for memory-augmented robotic manipulation! 🤖🧠 Featuring 16 tasks across temporal, spatial, object, and procedural memory 🔗 https://t.co/xXQ15rx9G8
5
44
211
Most developers are using Claude Code wrong. They install it… run a few prompts… and treat it like a terminal chatbot. That’s why the results feel average. Claude Code is actually a 4-layer system 👇 1️⃣ CLAUDE.md Your project’s persistent memory. It defines: • what the
31
197
1K
Yes, Gemma 4 will soon be released this time their largest size might be around 120B total with 15B active
Gemma 4 has been spotted on GitHub. The PR appears to be from Google’s bot account
28
47
781