Pinned Tweet
No API keys. No servers. Just your browser.
I built an on-device AI agent in <200 lines of code using Gemini Nano / Phi-4-mini.
It loops, calls tools, even reads your filesystem.
I built an on-device AI agent in <200 lines of code using Gemini Nano / Phi-4-mini.
It loops, calls tools, even reads your filesystem.
1
0
1
663
Ryan Seddon 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 e "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/hugging…
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 e "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/hugging…
131
602
4,437
249,549
I've switched to Zen browser from Arc but the one thing that has been throwing off is the icon messing my muscle memory when alt tabbing.
So I got my agent to make an arc like app logo with a Z and it did a really good job.
So I got my agent to make an arc like app logo with a Z and it did a really good job.
0
0
1
149
I just built some hyper personalised software to track some physio exercises while doing said physio exercises, i got it to grab a youtube video and put together a webapp with gifs and slowly increasing my reps over the 6 week recovery period. Did it all in 15mins.
0
0
0
180
Why use many token when few do trick?
Caveman is backed by @ycombinator (F26).
It started as a joke to get my AI bill under my grocery bill by having Claude speak like a Caveman. 111k GitHub stars later, it's a company.
Today: Caveman 4.0 and Caveman Cloud.
0
0
0
124
He's too scared to at you @thdxr get him.
Coding CLIs keep getting more complicated. Daemons, 3 layers of TUIs, client-server architectures.
The whole point of a CLI is to deliver intelligence with zero friction, and be unix-like.
💾 Try fx.sh. It's 6mb, starts in milliseconds, model-agnostic. It's tastefully minimal: MCP, Skills, Subagents and ACP built-in.
0
0
0
139
Has anyone else noticed chrome devtools mcp has suddenly forgot it can call webmcp tools, last week it would discover them on first load and now i have to explicitly say use webmcp tools and even then it doesn't. This is across opencode and claude desktop.
3
0
3
216
Ryan Seddon retweeted
You can now compose scenes with models from the Pardall Games library and you can share these scenes with a link.
The arena has gizmos now: position, scale, rotate models.
Open this link you are not going to regret it library.pardallgames.com/scene?p0=ids:P… #FinalFantasy #RetroGames #PS1 #PSX
The arena has gizmos now: position, scale, rotate models.
Open this link you are not going to regret it library.pardallgames.com/scene?p0=ids:P… #FinalFantasy #RetroGames #PS1 #PSX
Pet project:
1) Advanced PS1/N64 game extraction (+all animations).
2) Website where you can browse and compare models, textures, animations, FMVs, for preservation.
Already have FF7, Mega Man Legends 1 and 2 and MediEvil fully ripped.
#FF7 #FinalFantasy #FFVII #PS1 #PSX
13
198
1,772
275,285
The roadster be like
Roadster event update
We've been tracking the weather closely with local meteorologists, but given the severe conditions predicted & because this event can only be held outdoors, we've made the difficult decision to reschedule.
New date is October 15. Additional details to follow
0
0
0
179
Ryan Seddon retweeted
📦 Chrome 154 - Responsive iframe resizing 🤯
Letting an <iframe> resize based on the size of its embedded document
Haven't all widget/extension authors dreamed of this for years? 😆
Instead of using postMessage({resize})
bram.us/2026/09/23/res…
Letting an <iframe> resize based on the size of its embedded document
Haven't all widget/extension authors dreamed of this for years? 😆
Instead of using postMessage({resize})
bram.us/2026/09/23/res…
14
82
677
51,213
Ryan Seddon retweeted
kit finally finished the blog post
he covers how OpenCode 2.0 is able to modify everything about itself in real time even during running sessions
btw post whatever easter eggs you're able to find
anoma.ly/notes/opencode…
he covers how OpenCode 2.0 is able to modify everything about itself in real time even during running sessions
btw post whatever easter eggs you're able to find
anoma.ly/notes/opencode…
200
102
2,101
330,212
Jev is the new claw, picojev next.
🔥还有谁？！Jev模型刚出来没几天，NanoJev 已经把整条链路开源了，不是套一层 API，是 0.6B 并行决策模型 + 数据 + 训练管线！
NanoJev 把「状态+问题 → 完整概率分布、零 token 解码」这条路直接摊开了：
🔹底座只用 Qwen3-0.6B，一次前向就能并行批处理多个状态、多个问题
🔹动态候选 2–255 Score 三类结构化决策
🔹50×50 迷宫：244 步、36 次碰撞到达终点；贪吃蛇 12×12 吃到 27 个食物还活着
🔹模型和数据全开源，训练、评测、持久化推理服务、三栏对照演示一条龙
🔹代码规划 + 模型局部判断，不是纯聊天模型硬解游戏
仓库：github.com/TianyuCodings/…
模型 / 数据：Hugging Face 上的 C-Tianyu/NanoJev
闭源刚喊完 200 倍更快，开源已经把玩具级 System One 端到端跑通了。就问还有谁？！
#NanoJev #Jev #SystemOne #开源 #LLM #Qwen3 #AIAgent #强化学习 #RLCD
0
0
0
135
I had a network issue with xbox cloud streaming where i would get consistent stream dropping, i ended up using agents to dig into my router logs and learned a lot about @Tailscale, dhcp leases and cgnat. ryanseddon.com/hacking/agents…
0
0
0
172
Clever to use a notebook here to build the cdi, not sure if the uploading of pirated content to build it will last long though
‼️BIG NEWS ‼️
The Sonic R port for Sega Dreamcast by jnmartin has officially released! 🏁
WITH ONLINE NETWORK SUPPORT! Get your assets ready and head over to colab.research.google.com/drive/1Ss1Dzn1… so you can build your CDI today!!
Can you feel the sunshine??🌞
0
0
0
187
