Building with agents, local AI, open models Automation, Robotics, Voice Thinking about post-work society... - https://nitter.cf/t.co/HFnpazddi5 - https://nitter.cf/t.co/lqQpd2kCyK
Beaconsfield, UK
Joined August 2009
- Tweets6.4K
- Following483
- Followers366
- Likes69.9K
Agents with agency.
I gave one @bot controlling the Hermes fleet.
It owns health, stuck chats, and keeping work moving overnight. When spark1 soft-zombies, it pings Ops and clears the path...
I only get woken for the bits that need a human.
Who else is handing overnight ownership to a named agent, not just a cron?
Derek Colley retweeted
Agent memory gets harder the longer a chat runs.
We wondered: how well does Observational Memory hold up at 100K, 1M and 10M tokens? So we ran it on BEAM, the long-term memory benchmark, with default settings.
Turns out it has the top published scores at 100K and 1M:
Derek Colley retweeted
Own your own hardware to run intelligence locally.
Big sam says it’s a terrible allocation of resources. Yeah ok fam.
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”
RT @tenhkspark: これはわたしの構成とは違った画期的な構成です。
DGX Spark3台それぞれにQwen3.8-Flash-Nextを載せて、デコードが劣化しない15並列を達成するとやることが無くなるんですよ。
私もアイデア出しに苦労してます。瞬殺ですよ。
This quoted post is unavailable.
Derek Colley retweeted
Our agent memory is SoTA:
Derek Colley retweeted
Learn AI safety in Cambridge for free.
Housing, food and travel covered.
CAMBRIA is a 3-week, in-person program:
• $0 to attend
• Housing and weekday meals provided
• Travel reimbursed (up to $750, often 100%)
• 24/7 office access + dedicated TAs
• International applicants welcome
What you'll build:
• A Transformer from scratch
• Mechanistic interpretability
• LLM evals and AI control
• Frontier alignment research reproductions
• Your own capstone research project
Jan 4–22, 2027.
Applications close Nov 1.
cbai.ai/cambria
Grok & Hermes => Building Tourbillon
...and Personal bot was whipping them hard this week to hit the limit, not sure we'll get there.
The main unlock is pushing DEV and TEST work to Hermes.
My 3x DGX running Qwen3.8-next-flash are now the bottleneck.
I need another spark 😭
Useful to know - if you have 2x DGX sparks.
But the trade-off is parallel slots vs a faster single stream.
I run 3x DGX with Qwen3.6-next-flash @ "PARALLEL=5" = 15 sessions behind LiteLLM.
For my workload, concurrent sessions wins, for now...
What fits on one DGX Spark, and what needs two. Only models I ran myself, one user, real prompts.
One Spark: Qwen3.8-Flash-Next at 66 tok/s on JSON. Nemotron 3.5 Bolt at 96 with its draft model. A dense 27B at 11.
Two Sparks: DeepSeek-V4.1-Flash (763B) at 103 tok/s, but only 7-9 GiB stays free on each Spark. MiniMax-M3.1-Flash (428B) at 27-29 with 2.15 bpw, 20-22 with 3.06 bpw. Qwen3.8-Flash-Next goes from 66 to 99.
Derek Colley retweeted
ZERO Edits or Cuts! 1 Shot recording using LocalRouter on a single DGX Spark, a completely Local AI stack for everyone.
Here is Scarlett Johansson (Voice clone) to show you how it works in real time.
Text, voice, image, video, decision. Here are some of the features
- Voice cloning from 1-2 audio clips
- Full Image/Video generation
- Electron desktop app for Windows/Mac
- JEV computer-use model
- Qwen 27B main model
- Drama 3 voice model
- Transcription model
- Headless app keeps the Sparks lean
It's early, but looking forward to the future of what this brings for local AI. More coming soon
Thanks to everyone thats contributed parts of this. @ashxhart @tensorfold_dev @MiaAI_lab @plotarmordev
Derek Colley retweeted
When I said hunch could learn in real-time I wasn't kidding. It's like teaching a kid!
Behind the scenes, @Cloudflare’s Clef is helping refine the local model so it learns in real time, while @googlegemma’s EmbeddingGemma2 runs live on the CPU. 🧵
It's been a while since I showed any Jev stuff but I've been cooking on a way to use @GoogleDeepMind's EmbeddingGemma2 as an insanely fast real-time /v1/systemone endpoint (20ms on text, 100ms on images).
It's trainable in real-time, can distill an upstream teacher model, and is multi-modal by default. Hard to believe a @googlegemma model of this size can do so well with the right harness!
Derek Colley retweeted
🚀 Announcing OpenDuet
An open source proactive AI agent that's self-hostable and works with any agent harness.
Clone it. Build on top of it.
Repo → github.com/CopilotKit/OpenDu…
🎉 Introducing OpenDuet
Self-hostable, proactive AI agents that you own and works with ANY agent harness.
Includes:
- Build, generate, and edit live assets
- Human-in-the-loop approvals
- Computer use
- Mobile, Web, and Channels
Repo → github.com/CopilotKit/OpenDu…
Powered by @CopilotKit and AG-UI.
Clone this template and customize it however you want. Enterprise-ready.
I'm working on how agents request info/approval from users.
@TourbillonHQ uses @AGUI_Protocol widgets to render the question in chat, but Agents call suspending frontend tools 🛑
What if it's 3am....? 😴 💤 ⏰
I created hitl-spec to standardise how agents ask questions, async:
github.com/hitly-net/hitl-sp…
Please to let me have comments and suggestion.
Derek Colley retweeted
first time ever letting Grok bot buy something for me…
it signed me up for Spotify's $1 trial, then cancelled it right after so I'd never get charged the $12.99.
and it told me the checkout page had hidden text telling AI agents to click an invisible button.
but it caught it and completely ignored it.
asked for my card through a secure form, stopped to confirm before charging anything, and handled the whole thing in a few minutes.
honestly one of the most impressive things I've seen an AI agent do considering how they made it so Grok Bot never sees any of the verification codes or payment methods.
Derek Colley retweeted
To paraphrase Letorneau, inventor of diesel construction machinery, AI is not to make small jobs easy, it's to make large jobs small.
网景创始人 、硅谷顶级投资人Marc Andreessen 在访谈里坦言，他看到了科技史上最可怕的一道认知鸿沟，
传统巨头以为 AI 只是个降本增效的插件，而新一代 AI 原生团队，正在用完全非人的组织效率掀翻整个商业版图。
当年写出网景浏览器、见证了整个互联网三十年兴衰的 Marc Andreessen，坐在 Ben Horowitz 面前直接发出了感慨，
如果 1994 年我们做 Netscape 的时候就有今天这种 AI 工具，那简直不可想象。
现在绝大多数传统巨头和现有企业，依然在用改良软件的旧思维去看待大模型。
在他们的组织逻辑里，AI 只是一个高级功能，用来给现有业务做个插件、在钉钉里加个智能助手、或者把客服团队砍掉 20% 省点成本。
他们以为自己在拥抱变革，其实只是在给老旧的蒸汽机打补丁。
但真正恐怖的事情，正在那些刚成立的 AI 原生团队里发生。
那些创始人对 AI 能力的理解，跟大厂完全不在同一个平行宇宙。
他们每周都在看着底层技术飞速进化，而且他们压根不会把自己定义成一家软件公司、硬件公司、甚至是一家 AI 公司。
以前创办一家公司，你的组织架构是按人类的协同瓶颈搭出来的：
招产品经理写文档、招工程师写代码、招法务改合同、招运营做投放。
因为人类的脑力带宽有限，所以必须靠庞大的部门和管理层级来支撑业务。
而真正的 AI 原生物种，是用机器智能作为最底层的骨架。
两个人加上一整套自动化 Agent，可以直接把从市场洞察、软件工程、全栈设计到商业销售的链路全部跑通。
他们不是在效率上快了 30%，他们是在组织形态上彻底跨越了一个物种。
大厂的盲区，从来不是买不起算力，而是他们无法理解组织本身才是最大的负债。
当现有企业还在开会讨论如何把 AI 塞进现有的业务流程时，新一代的野蛮人正在用完全超越行业界限的方式，把传统的业务流程连根拔起。
未来拉开人与人、公司与公司之间差距的，根本不是你用了哪个模型，而是你到底是在用新工具维持旧世界的体面，还是用新世界的底层逻辑，重新定义了自己能做的事情。
