🇻🇳🇻🇳🇻🇳 Threads • https://xitter.cf/t.co/qINX2Pmvwt
Joined December 2015
- Tweets6.7K
- Following537
- Followers248
- Likes7K
Phan Hoàng 🇻🇳 retweeted
Qwen Concept 27B: we rewrote 9 weight matrices in Qwen3.8-27B. No gradient steps, no training data.
It thinks less and scores higher: GPQA Diamond +4.3 pts // LiveCodeBench v6 +13.0 // 27–54% fewer thinking tokens.
Open weights and methodology paper below↓
Weights and working paper are released. As we run more robustness checks, we also discover that this model edit unexpectedly adds +13% performance on coding tasks.
concept.country/research/wei…
Phan Hoàng 🇻🇳 retweeted
We are releasing a new package from @roboflow, meet reid!
It turns a crop of an object into an embedding, so you can tell objects apart, like this player across clips of a game.
Can't wait to see what you build with it!
github.com/roboflow/re-ID
Phan Hoàng 🇻🇳 retweeted
GRPO problem explained very simply
Phan Hoàng 🇻🇳 retweeted
🔥Significant step forward for speculative decoding! 🔥
Tencent just released DSpine for Qwen3.8-27B and it beats DFlash2 by any category!
huggingface.co/tencent/Qwen3…
Phan Hoàng 🇻🇳 retweeted
Anthropic just broke down the components of agent memory
the formula:
agent memory = transcripts + notes + dreaming + wikis + retrieval
past experience → correct mistakes → connect knowledge → next session
this is must-learn ↓
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
Phan Hoàng 🇻🇳 retweeted
Video editors are cooked. Claude Opus 5.5 just learned to steal any video you show it...
hyperframes just dropped /reverse-any-video. plus 10,000+ music tracks, sfx, 75,000+ stock images and tts. free
this is the 3D + video skill stack i pair it with:
VIDEO
┣ hyperframes · write HTML, get a video. UPDATED nov 4
┃ ┗ github.com/heygen-com/hyperf…
┣ remotion skills · claude cuts real videos in react
┃ ┗ github.com/remotion-dev/skil…
┗ openmontage · a full agentic video studio, 12 pipelines
┗ github.com/calesthio/OpenMon…
MOTION
┣ gsap skills · official easing, timelines, scroll motion
┃ ┗ github.com/greensock/gsap-sk…
┣ video-shotcraft · cinematic product promos, 124 shot recipes
┃ ┗ github.com/Vincentwei1021/vi…
┗ motionmaxxing · turns AI slop into motion-designer films
┗ github.com/Tejashmakwana/mot…
BLENDER
┣ blender mcp · claude drives blender. whole scenes
┃ ┗ github.com/ahujasid/mcp-for-…
┣ blender-skills · 94 skills, blockout → engine export
┃ ┗ github.com/arjun988/blender-…
┗ dream-loop · blender + image gen + a critic that scores every render
┗ github.com/achimala/dream-lo…
3D WEB
┣ threejs-game-skills · playable three.js games, AAA look
┃ ┗ github.com/majidmanzarpour/t…
┣ anatomy · isometric exploded machines, SVG + WebGL
┃ ┗ github.com/wheresryan22/anat…
┗ 3dviz-pro-max · idea in, three.js or blender scene out
┗ github.com/viettranx/3dviz-p…
update hyperframes first:
npx hyperframes skills update
every card in the video is a real 3D scene I rendered for that skill. same little claude, one blender file
Phan Hoàng 🇻🇳 retweeted
This is f*cking insane... Andrej Karpathy. 8 years at OpenAI and Tesla.
Last week he packed everything he knows into one free 2-hour lecture.
Agents → Loops → Harness → Graphs
People burn $15k on bootcamps that cover less than this.
2 hours feels like a lot right now.
Don't let it scroll past your feed.
Watch it today, then save the full graph engineering guide below ↓
Phan Hoàng 🇻🇳 retweeted
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over.
And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams.
So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super happy to finally release "The Ultimate MoE Training Handbook", free and open, organized around the questions training teams hit:
- how many experts, how big, shared or not?
- why is half my step spent in all-to-all, and how do I hide it?
- why is my run 25x slower with no error in the logs?
- softmax, sigmoid or sqrt(softplus) router, and why keep it in fp32?
- aux loss or DeepSeek's loss-free bias?
- upcycle a dense checkpoint or train from scratch?
- can I serve with fewer experts per token than I trained with?
⚠️ Warning: it's long (~70k words, budget a few days, sorry) but with 55 interactive figures, code you can run and numbers from our own runs.
Read it here: huggingface.co/spaces/huggin…
Phan Hoàng 🇻🇳 retweeted
This week we were blessed by:
1. @Nouamanetazi dropped MoE training handbook: huggingface.co/spaces/huggin…
2. @yb2698 started training harness adapters: huggingface.co/blog/harness-…
3. @joelniklaus released the harness leaderboard: huggingface.co/spaces/joelni…
Have a lovely weekend y'all. 🤗
Phan Hoàng 🇻🇳 retweeted
I beat the ten percent challenge in 4 hours 🙀
I compressed Qwen2.5-7B to 9.4% of its size (15.2 GB → 1.43 GB) with no weight training.
On a locked exam it scored 0.799 vs the original's 0.7675, and still passes after learning from user feedback. It follows all 5 challenge rules. 🚀
Built for the @tokenbender 10% challenge.
Proof + code: github.com/Bobliuuu/ten-perc…
demo and explanation below 🧵
I would like my 1K USD 😗
i’ll sweeten the pot.
1k USD prize if you can meet this challenge on a commercial class task, above 0.5B params, in one week.
github.com/DeusExperiments/t…
Phan Hoàng 🇻🇳 retweeted
We tested a lot on this: 20 films, simple to complex, from UGC to talking heads to motion graphics. Our reverse-any-video vs REA-Video, two experiments:
1. Analysis: cuts, editing, how each shot was made
2. Full 1:1 remake: same model, same keys, isolated containers
Experiment 1, analysis (hand-labelled cuts), ours vs REA-Video:
• Cuts found: 108/108 vs 59/108
• Precision: 96% vs 89%
• Time: 81 s vs 1,595 s, ~20× faster
• Its misses: talking heads, heavy edits
• It called live footage AI in 3 of 8
Experiment 2, full remake, ours vs REA-Video:
• SSIM: 0.96 vs 0.78
Bottom line:
• Simple UGC: both great
• Talking heads with motion graphics, B-roll and heavy editing, and launch videos with complex motion graphics: we do better
• REA-Video remakes slightly faster and cheaper
example:
REA-Video，一个逆向视频的开源项目。
装好后把视频丢给 Agent，它会自动拆出每个镜头，分析画面是实拍、AI 生成还是代码动效。
还能反推出每个镜头的Prompt、运镜方式、剪辑节奏、字幕和音效，最后生成一份带关键帧截图的完整复刻方案。
相当于给它一条成片，帮你倒推出一份详细的制作教程。
免费开源，支持本地视频分析。
GitHub：github.com/JuneYaooo/rea-vid…
Phan Hoàng 🇻🇳 retweeted
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
Phan Hoàng 🇻🇳 retweeted
Today we release Mirage: A 22 MB visual-language model that runs scene descriptions + object detection + OCR + segmentation + spatial relations analysis + extractions + decisions, in a single pass.
It can do all that in 74 ms on Samsung Galaxy S26 NPU or 107 ms on its CPU.
Mirage has a 7.8M encoder + a 40.8M decoder at 2bits, and describes images in English, German, French, Spanish, Italian, Dutch and Polish.
We trade general language capacity, to compete top small VLMs today (SmolVLM2-500M, LFM2.5-450M and Florence2-230m) on these tasks.
On 256 VQAv2 questions, feeding Mirage's structured output and full trace to text-only Qwen3-0.6B raised its score from 9.3 to 31.3 out of 100, without fine-tuning either model.
Read more and benchmark it yourself here:
cactuscompute.com/blog/mirag…
Phan Hoàng 🇻🇳 retweeted
Harness adapters training is now live!
We are training LoRA adapters on LFM2.5 / LFM3 from @liquidai and the recent Twil-9B model from @thewebAI using RL to make them more efficient and accurate on agentic tasks!
You can follow the training going live here: harness-hub-harness-adapters…
huggingface.co/blog/harness-…
Phan Hoàng 🇻🇳 retweeted
We're releasing RealCUA, the first computer-use dataset of open-ended work in evolving environments.
600+ hours (and growing!) of real multi-objective work with interleaved tasks across thousands of steps in realistic desktops.
11-hour sample & access: data.fig.inc/realcua
Phan Hoàng 🇻🇳 retweeted
how we’re building decision models at @liquidai
(aka: “how we reverse-engineered Jev”)
谢谢 @aiDotEngineer shanghai for having me.
here is the thread version of my talk 🧵
Phan Hoàng 🇻🇳 retweeted
一个适合在个人电脑上学习微调+RL的小项目，虽然很小，但是是一个很好的学习案例
微调+RL Qwen3-1.7B
kevin-wu.ca/blog/qwn/
github.com/kevinyvv/qwen-3-1…
Phan Hoàng 🇻🇳 retweeted
How are Decision Models Trained?
Today, we are publishing the Drex Technical Report for the developer community! 🚀
Discover details on how the State-of-the-art Open-Source decision models are trained with RLAF.
Build more at: console.nace.ai/ (try for free)
P.S. Drex Max 1.1 and Drex for Finance are closed sourced and designed for enterprise use-cases
Phan Hoàng 🇻🇳 retweeted
Hermit v0.2.0 is out. 🐚🦀
1. Harness evolving with your coding plan: 𝐧𝐨 𝐀𝐏𝐈 𝐤𝐞𝐲 𝐧𝐞𝐞𝐝𝐞𝐝.
2. Now supporting 𝟏𝟐 𝐡𝐚𝐫𝐧𝐞𝐬𝐬 𝐞𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐦𝐞𝐭𝐡𝐨𝐝𝐬: Prompt-only, GEPA, SkillOpt, MWM, Harness², Continual Harness, ReHarness, Meta-Harness, ACE, AHE, RRSI, EvoSafeHarness. (More to come!)
3. A new agent abstraction that makes adding methods much easier.
𝐆𝐢𝐭𝐇𝐮𝐛: github.com/Infini-AI-Lab/her…
𝐎𝐫𝐢𝐠𝐢𝐧𝐚𝐥 𝐑𝐞𝐥𝐞𝐚𝐬𝐞 𝐀𝐫𝐭𝐢𝐜𝐥𝐞: View article
