Want to help build a more regenerative and ethical AI future?
Australia
Joined January 2026
- Tweets3.3K
- Following210
- Followers345
- Likes2.1K
Josh retweeted
Shadow-git checkpoints before every modifying tool call → autonomous rollback when the agent wrecks the workspace → a timeline pane to scrub and restore history.
nitter.gq/MossAIStudio/status/21…
I built Hermes Chronos for the @NousResearch Hermes Agent Plugin Hackathon.
Shadow-git checkpoints before every modifying tool call → autonomous rollback when the agent wrecks the workspace → a timeline pane to scrub and restore history.
88s demo. Sound on. 🎧
#HermesAgent
Josh retweeted
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/con…
Josh retweeted
ERNIE 5.5 Flash Base just dropped! Time to open that project folder. 🚀
Our efficient sparse MoE language model draws on data spanning multiple domains and languages, achieving SOTA on multiple benchmarks among similarly sized base models. Its architecture and training system were designed together for efficient training and inference, with advances in attention mechanisms, expert routing, and training objectives:
- 280B total parameters; 14B activated per token
- Pre-trained on 33.5T tokens
- Up to 1M-token context for fine-tuning and inference
- FP8 and BF16 weights
The weights come straight from pre-training, with no mid- or post-training. Compare methods from the same checkpoint, measure what each adds, and shape the model around your own tasks.
Josh retweeted
The biggest pool of future local AI engineers isn't on 24GB cards and 128GB Sparks. It's on the PCs and Macs people already own.
From Steam's October survey:
• 78.5% of PCs have 8GB+ VRAM
• 43.8% have 32GB+ RAM
• Only ~5% have 64GB+ RAM, and under 10% have more than 16GB VRAM
Beyond Steam:
• 44M desktop GPUs shipped in 2025 alone
• My rough estimate: 100-150M PCs in use with 8GB+ VRAM
• Plus 140M+ Apple Silicon Macs, and M5 Pro/Max are seriously quick with unified memory
That hardware runs a 20-35B MoE with ~3B active really well. On my 10GB 3080 with 32GB RAM, that class is where the best results have come from.
This is where people start tinkering, building agents and learning the craft. It's a much bigger group than the high-end crowd.
@liquidai, @ornith_, @PrismML, @OpenBMB and others are already building for it. @NVIDIAAI and @Alibaba_Qwen please take note. With Qwen 4 coming, a strong MoE that fits 8-12GB VRAM + 32GB RAM would land with a huge audience.
Josh retweeted
Marimo Bot got another model rework.
This time I am building solely upon the LFM3 Omni 2B model.
A single end-to-end model that understands text, images, video and audio, and answers in text and speech.
You can follow the project here: github.com/DevTarlow/marimo-…
Josh retweeted
🔥精选！DeepSeek Harness 必装的 8 款插件！让你的 DSH 能力全面升级！
从找插件到出图、记忆、远程操控、上下文洞察全覆盖，直接把 DSH 拉满：
1️⃣dshfind
插件搜索与发现，按分类找插件，查看项目资料与作者入口。
2️⃣dsh-image-gen
AI 生图、修图，在 DSH 里直接做图，支持生成、编辑与图库创作流程。
3️⃣dsh-better-sidebar
侧边栏增强，把文件、预览、终端等常用操作放到手边，更好用。
4️⃣dsh-mnemon
AI 长期记忆，重要背景持续可用，管理记忆与项目文档，查看记忆关联。
5️⃣dsh-im
手机远程操控，通过聊天连接 DSH，接入微信、飞书等，离开电脑也能用。
6️⃣dsh-skills-manager
Skills 统一管理，集中添加、查看、编辑与启用 Skills。
7️⃣dsh-context
上下文可视化，看清组成、趋势与活动，了解当前会话状态。
8️⃣dsh-deep-whale
鲸鱼娘主题皮肤，二次元角色、专属配色与明暗模式，界面更有个性。
这 8 个装上，DSH 体验直接起飞。
#DeepSeek #DeepSeekHarness #DSH插件 #AI工具 #必装插件 #AI生图 #长期记忆 #上下文可视化 #鲸鱼娘皮肤 #崔添翼 #梁文峰
🔥牛哇！DSH超级代理，一个人顶一个团队！24 小时盯着你的仓库，让 GitHub 自己修 bug 了！
有人提 Issue、开 PR，它直接上手改代码、出修复草案、提 PR，把结果丢给你审，你只需醒来扫一眼 diff，点同意或拒绝，不用再翻一堆消息。
核心亮点
🔹主动干活：只在被点名或规则允许时才行动，尽量少打扰你
🔹并行处理小任务，不堵塞你的对话窗口
🔹支持 GitHub + 飞书触发，规则、权限、提示词全可自定义
🔹Router + Worker 双层架构，朝更自主的 Agent 迈进
🔹MIT 开源，研究原型（非官方正式产品）
快速上手
1️⃣安装 DeepSeek Harness，克隆仓库
2️⃣按 SETUP.md 让 Agent 引导配置（或看 docs/006-deployment.md 手动部署）
3️⃣建独立 superagent profile，配好 GitHub Agent 账号和监听仓库
4️⃣启动服务，在 Panel 设置允许用户/策略
5️⃣在 Issue/PR 里 @Agent，它就开始自动干活并回结果
Repo：github.com/deepseek-ai/dsh-s…
#DeepSeek #DSH #AIAgent #GitHub #开源 #编程助手 #SuperAgent
Josh retweeted
BlipTalk Studio 1.1 got an update.
Save your own presets, type exact values, optimizations/fixes, and download the WAV blip pack as one ZIP.
Free, in your browser.
Tool: tarlow.itch.io/bliptalk-stud…
Devlog: tarlow.itch.io/bliptalk-stud…
Josh retweeted
Only a Sith deals in absolutes.
"Local inference is a terrible allocation" sounds sensible until you look at what it assumes: that compute should be pooled and allocated centrally, and individuals owning their own is waste. That's a collectivist premise, and it runs against how people actually live and behave. We want our own tools, our own data, our own say.
And the practical case stands on its own:
- Cost over time: a GPU you already own costs nothing per token, while an always-on agent adds up fast in the cloud.
- Privacy: your calendar, inbox and notes never leave your machine.
- Always available: no outage, rate limit or connection needed.
2B models can now run real agent loops on a 10GB card. Cloud has its place. So does ownership.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Josh retweeted
📱 35B parameters. Real-time inference. On a phone.
Meet Lookahead Expert Scheduling (LES), the inference system powering MiniCPM-V 4.7’s 35B-A3B MoE model on edge devices.
🚀 On an iPhone 18 Pro Max, the full system reaches 20.51 tokens per second, a 6.07× speedup over loading expert weights on demand.
How? Three coordinated optimizations:
- 🔮 Scheduling: Predict and prefetch the next layer’s experts while the current layer computes.
- 💾 Storage: Combine contiguous expert layouts with mixed-precision quantization to reduce loading overhead.
- ⚡ Compute: Use DSpark speculative decoding to verify multiple candidate tokens in one pass.
Predictions guide prefetching; the model’s actual router still determines which experts execute. The architecture, expert counts, and routing remain intact. Mixed-precision quantization reduces the weight file by 26.7%, with just a 0.16-point drop in the five-benchmark average.
LES shows how better inference systems can bring larger, more capable models to everyday devices. 🌍
👉 LES: github.com/openbmb/llama.cpp…
🔗 GitHub: github.com/OpenBMB/MiniCPM-V
#MiniCPM #OnDeviceAI #MoE #EdgeAI
Josh retweeted
Do you like learning about AI?
Build your AI skills with IBM SkillsBuild learning catalog.
skillsbuild.org/learning-cat…
Josh retweeted
Oh god f*ck, is this even possible? NVIDIA's new micro-robots just integrated the Grok Robot Brain Engine, and the live neural telemetry broadcast is genuinely terrifying
Watch what happens in the live feed when the robot ("Alpha") trips over a cone and hits the floor. Instead of rigid pre-programmed recovery scripts, you can watch its synthetic brain adapt and re-calculate in absolute real time:
• Live Neural Telemetry: Multi-node graphs track gait policies, spike rasters, and motor torque as the machine processes physical obstacles.
• Autonomous Recovery: The moment a fall is detected, thousands of neural nodes fire simultaneously to calculate friction, torque, and balance locks until the robot stands back up entirely on its own.
• Edge Cognition: Every step, stumble, and correction is generated live by a distributed neural network, completely bypassing traditional hardcoded robotics.
We are no longer just building machines; we are watching synthetic nervous systems wake up in the physical world.
Bookmark this before physical AI completely rewrites the entire robotics industry.
Josh retweeted
M3.1 Flash Weight is public now:
Hugging Face:
huggingface.co/MiniMaxAI/Min…
ModelScope:
modelscope.cn/models/MiniMax…
