Search
1000 results for “quantixed”
-
@gutenberg_org British chemist #JohnDalton is the *actual* originator of #QuantumTheory. His #AtomicTheory correctly quantized *mass* a century before Planck quantized energy. 🤔
-
@gutenberg_org @wikipedia By postulating that chemical change was no more than the recombination of atoms - with fixed and *quantized* mass - #JohnDalton founded #QuantumTheory almost a century before #MaxPlanck quantized energy. 🤷♂️ #Histodons #Science #ScienceHistory
-
I decided I wanted to know what a Julia set would look like in 3D and quantized. It looks like an alien coastal town...
-
RT @[email protected] quick-and-dirty Pharo Smalltalk binding of TensorFlowLite C API. Quantized Mobilenet 1.0 244 works. #TensorFlowLite #Pharo https://t.co/hkcMm0oOSQ
-
SciPost Physics @physics
New #openaccess publication #SciPost #physics
Interplay of Kelvin-Helmholtz and superradiant instabilities of an array of quantized vortices in a two-dimensional Bose-Einstein condensate
Luca Giacomelli, Iacopo Carusotto
SciPost Phys. 14, 025 (2023)
https://scipost.org/SciPostPhys.14.2.025 -
Messing around with my new qgliss algorithm in #sndkit
A randomly generated line is used to produce a quantized melody via qgliss. this line also maps to other parameters in the patch as an expression curve.
sndkit patch:
https://paste.sr.ht/~pbatch/560654ba675a0557ea408aaf9afec85f55e6b781
qgliss:
https://pbat.ch/sndkit/qgliss/
https://git.sr.ht/~pbatch/sndkit/tree/9e8a7d543fe28813742acd06b6883e012764504a/item/dsp/qgliss.org -
Messing around with my new qgliss algorithm in #sndkit
A randomly generated line is used to produce a quantized melody via qgliss. this line also maps to other parameters in the patch as an expression curve.
sndkit patch:
https://paste.sr.ht/~pbatch/560654ba675a0557ea408aaf9afec85f55e6b781
qgliss:
https://pbat.ch/sndkit/qgliss/
https://git.sr.ht/~pbatch/sndkit/tree/9e8a7d543fe28813742acd06b6883e012764504a/item/dsp/qgliss.org -
✨ Open source RAG (Retrieval Augmented Generation) right in your browser! ✨
#SemanticFinder now offers an 𝐚𝐝𝐯𝐚𝐧𝐜𝐞𝐝 𝐜𝐡𝐚𝐭 & 𝐬𝐮𝐦𝐦𝐚𝐫𝐲 𝐟𝐞𝐚𝐭𝐮𝐫𝐞 for your search results - all in your browser.
💡There are very few capable small LLMs that offer high-quality results. Quantized LaMini-Flan-T5-783M offers good performance with 3-4s load time and >6 tokens/s after model download on an old i7.
https://do-me.github.io/SemanticFinder/
#transformers #RAG #AI #LLM #embeddings #semanticsearch #text2text #Flan #T5
-
CW: LLM use for dev (pi.dev as a slop newbie, long)
So yesterday I have actually tried https://pi.dev/ at home. In an isolated proxmox VM with a tight firewall as if dealing with dangerous bacteria 😱
And it kind of is. You can just let it install the tools it needs on Debian and all. It can modify its own configuration and tools and even write plugins for itself.
Magic yet frightening but it’s an isolated VM with only test projects. I access pi via ssh and let it do whatever it needs on its home VM. So far my laptop’s locale hasn’t been changed to zh_CN or ru_RU.
The only thing this could access on the local network was the local MLX / Ollama servers and I still felt surprised when it knew how to download a different model on my other machines via Ollama API using curl.
At the same time it feels easy to maintain control with the few set of basic read/write/bash tools it comes with. All controlled from a simple shell.
Your sessions are saved as text files and traceable and there are no hidden instructions in the prompt sent. I understand people complain about that in Claude (which I never tried). One thing I liked was asking summaries of what the fuck I was working on based on the session files. I am like that . I easily forget what I was doing even when I write the code myself.
As pi.dev (+lazypi) comes with some tools the context at startup of a new project easily goes to 18k tokens on first prompt (info on pi itself, on the tools and additional packages each add some kilobytes).
Even on further prompts speed on my local #LLMs was painful. They were ok to ask for snippets and chat locally (via OpenWebUI on another local VM). but not for so called agentic shit and rapid iteration.
So there went away my dream of seriously using pi.dev with my only local LLMs (on M1 Mac and Nvidia 3060 on PCs).
The whole idea is to save time so waiting minutes on a modification isn’t worth it, especially when trying to learn a system I don’t know. In retrospect some tests I did on the local LLM and took a full 4 minutes produced stuff that I think I wouldn’t produce in 4 minutes. But when learning to use this I jsut don’t want to wait four minutes on each test. Maybe with very careful planning it could make sense to just wait for local LLMs. And Yes the full LLMs produce better results than the quantized local versions when it gets complex. No surprise.
So I went full slopper and got a DeepSeek V4 account. It’s apparently 6x cheaper than American counterparts and most of all I’d rather be on the less worse side in the grand scheme of history. Also I am sure it’ll be working in 2027 after the US bubble pops. Yes I am a tankie in addition to a slopper now :-/
To make testing simple I asked “DeepSeek V4 flash” to build PHP sites and to configure nginx and PHP-fpm in the VM to serve it locally. The shit is fast. Much faster than me. It’s easy to feel overwhelmed by the fast pace of iteration. Remember I had planned nothing. Just went live testing and changing directions.
Summing up I went on for several hours basically reproducing a human translation web service that I used for years and now closed (icanlocalize). Basically editing strings in a SQlite db (and calling an api to translate stuff automatic) . The result is just so much pleasant (not hard icanlocalize was notably slow and confusing).
It’s a small tool only I would use (and maybe a colleague). I wouldn’t release this to the public but I can seriously use it on a local network for my own needs now.
Can’t deny that I alone would never have created something so functionally detailed with several screens, Ajax edition and handling of lot of edge cases in the same time and so easily iterated.
The machine tried to do some overly complex stuff sometimes and I was glad I actually knew how to write code (I think) to reorientate it and restructure the DB to always target simplicity.
Overall I must say there’s no reason I wouldn’t use this in the future. For such use cases. I never used this on Xcode as I don’t want to mess up my existing projects. It kind of feels like subcontracting except you can more easily cancel /rewrite stuff and iterate without waiting one day.
I am just sad I can’t use this purely locally as there’s no way the hardware I have or could reasonably get would be so fast.
pi.dev is local but I still sent all my very confidential data about this test PHP project to the CCP 🇨🇳.
That and about one dollar in token costs. It would have been less if I had understood earlier the /compact option. -
Modern #Androids pack 45+ #TOPS of #NPU compute. Instead of draining battery, a native app using #ExecuTorch and #NNAPI runs 4-bit quantized #LLMs natively, turning smartphones into a massive, parallelized compute swarm. 🧠🔋 #EdgeAI #MachineLearning #AndroidDev
Turning Android Phones Into An... -
Modern #Androids pack 45+ #TOPS of #NPU compute. Instead of draining battery, a native app using #ExecuTorch and #NNAPI runs 4-bit quantized #LLMs natively, turning smartphones into a massive, parallelized compute swarm. 🧠🔋 #EdgeAI #MachineLearning #AndroidDev
Turning Android Phones Into An... -
Doug @dougmerritt, we were discussing the status of the #Kohonen #SOM research, last week. Here are a few recent (21st Century) publications on the topic that I like, either for their undergrad accessible styles or for their advanced research ideas.
Given that this is my favourite list, it skews heavily toward DSP and DIP. But then, Kohonen did design the SOM expressly for perceptual processing of auditory and visual signals.
The idea of implementing quantised SOMs on FPGAs intrigues me, at present.
• 2001 Kohonen—SOMs 3ed
• 2001 Kiang—Extending the Kohonen SOM for Cluster Analysis
• 2001 Villmann—Exts and Mods of SOM and Apps in Remote Sensing Image Analysis
• 2002 Seiffert—SOMs: Recent Advances and Apps
• 2003 Zherebtsov—Clustering Stock Portfolios
• 2004 Bação—Intro to SOM
• 2004 Mokriš—Decreasing the Feature Space Dim by SOMs
• 2005 Guthikonda—SOM
• 2005 Huang—Exploration of Dim Reduction for Text Visualisation
• 2007 Sharma—Image Comp and Feature Extr with NN
• 2007 Villmann—Class Imaging of Hyperspectral Satellite Remote Sensing Data Using FLSOM
• 2008 Sap—Overlapping Clusters
• 2008 Skupin— Intro: SOM
• 2008 Yin—SOMs: Background, Theories, Exts, and Apps
• 2009 Campoy—Dim Reduction by SOMs that Preserve Distances in Output Space
• 2010 Dvorský—Improvements Quality of SOMs Using Dim Reduction Methods
• 2012 Kohonen—Essentials of SOM
• 2012 Asan—An Intro to SOMs
• 2014 Kohonen—MATLAB Impl and Apps of SOM
• 2015 Abdelsamea—Image Feature Classification
• 2024 Linke—SOMson: Sonification of Multi-Dim Data in SOMs
• 2025 Malik—SOMs
• 2025 Nogales—SOMs as a Way to Evaluate Optimal Strategies for Balancing Binary Class Distributionold school #AI
-
You ever just
https://archive.org/details/MIT8.04S13/MIT8_04S13_lec04_300k.mp4My alt text might be very wrong. IDK if the Pr(P) is the probability of a given value of P, or the probability density of a given vlue of P. AaAAAAAAAA
#Math #Maths #Physics #Nerd #DorkWeb #Nerds #Geek #GeekGyaru #QM #QuantumMechanics #QuantumPhysics #QuantisedMechanics #QuantisedPhysics #ItJustMeansWaves #WavesEverywhere #WeAreAllWaves
-
"Our current actions as a society seem totally disconnected from any optimized, survivable future."
J. Paul Neeley for Coda: https://longreads.com/2025/04/14/when-im-125/
#Longreads #Longevity #Biohacking #Data #Death #LifeOptimization #QuantifedSelf
-
Dự án VibeVoice vừa đạt **944 sao** trên GitHub. Phiên bản 1.8.0 đã ra mắt, bao gồm mô hình 8-bit tích hợp đầu tiên, lượng hóa động 4-bit/8-bit, và hệ thống quản lý mô hình thủ công. Mô hình mới có sẵn trên HuggingFace. Cảm ơn cộng đồng vì sự hỗ trợ!
#AI #MachineLearning #Vietnamese的各项标签 #Technology #VibeVoice #QuantizedModel #CommunitySupport
https://www.reddit.com/r/LocalLLaMA/comments/1nuxdd4/release_finally_a_working_8bit_quantized/
-
In 1803 #JohnDalton used the Laws of #Chemistry to propose the modern #AtomicTheory. It marks the birth of #QuantumTheory. (Dalton quantized mass a century before #MaxPlanck quantized energy.)
In 1914, #HenryMoseley used the X-ray emmisions of atoms to propose the quantization of #AtomicStructure.
Sadly, #Moseley was killed by a sniper a year later in the Battle of #Gallipoli in #Turkey. https://theconversation.com/how-science-lost-one-of-its-greatest-minds-in-the-trenches-of-gallipoli-45890
-
In 1803 #JohnDalton used the Laws of #Chemistry to propose the modern #AtomicTheory. It marks the birth of #QuantumTheory. (Dalton quantized mass a century before #MaxPlanck quantized energy.)
In 1914, #HenryMoseley used the X-ray emmisions of atoms to propose the quantization of #AtomicStructure.
Sadly, #Moseley was killed by a sniper a year later in the Battle of #Gallipoli in #Turkey. https://theconversation.com/how-science-lost-one-of-its-greatest-minds-in-the-trenches-of-gallipoli-45890
-
Making the most out of a small LLM
Yesterday i finally built my own #AI #server. I had a spare #Nvidia RTX 2070 with 8GB of #VRAM laying around and wanted to do this for a long time.
The problem is that most #LLMs need a lot of VRAM and i don't want to buy another #GPU just to host my own AI. Then i came across #gemma3 and #qwen3. Both of these are amazing #quantized models with stunning reasoning given that they need so less resources.
I chose
huihui_ai/qwen3-abliterated:14bsince it supports #deepthinking, #toolcalling and is pretty unrestricted. After some testing i noticed that the 8b model performs even better than the 14b variant with drastically better performance. I can't make out any quality loss there to be honest. The 14b model sneaked in chinese characters into the response very often. The 8b model on the other hand doesn't.Now i've got a very fast model with amazing reasoning (even in German) and tool calling support. The only thing left to improve is knowledge. #Firecrawl is a great tool for #webscraping and as soon as i implemented websearching, the setup was complete. At least i thought it was.
I want to make the most out of this LLM and therefore my next step is to implement a basic #webserver that exposes the same #API #endpoints as #ollama so that everywhere ollama is supported, i can point it to my python script instead. This way it feels like the model is way more capable than it actually is. I can use these advanced features everywhere without being bound to it's actual knowledge.
To improve this setup even more i will likely switch to a #mixture_of_experts architecture soon. This project is a lot of fun and i can't wait to integrate it into my homelab.
#homelab #selfhosting #privacy #ai #llm #largelanguagemodels #coding #developement
-
For the seventeenth #30MapsInAMonth are four maps revising rail electrification using Vega-lite https://vega.github.io/vega-lite/, a tool I had never used before the weekend.
As before, this uses tagged @openstreetmap railway data, but this uses a topological (TopoJSON) version, which can then be quantised to give different levels of granularity.
1/n
#Rail #Electrification #PublicTransport #Vega-lite #GreatBritain #OpenData #30dayMapChallenge
-
Great to see that #Apertus has been highlighted as a digitally independent alternative to ChatGPT yesterday! However, this needs a quick fact-check:
- The model was trained in a local data center at CSCS, the costs of the project are almost entirely covered by Swiss public institutions;
- People from all over the world have made contributions - open source LLM development is a global community of interest;
- There are groups in several countries using or fine-tuning Apertus to improve linguistic capabilities and local knowledge;
- While Apertus can run on Amazon servers, thanks to a third-party deployment script, it runs anywhere LLMs can run;
- Quantized versions are available to fit even relatively cheap consumer grade video cards (see my blog posts for details);
- The #PublicAI web interface and Apertus 8B demo runs on AWS, however the large model is hosted by CSCS as well;
- You do not need to use Google to authenticate to Public AI, it is just a convenient way to log into #OpenWebUI - if you want another provider, please suggest it;
- Apertus is not a chatbot on its own, it is a large language model that can be deployed as part of a system to provide chat services.I'll send this to the maintainers as well. Did I miss anything?
https://di.day/en/digital-switch-recipes/alternativen-zu-chatgpt-co#different-ways-to-cook
-
Building a Local-First Multi-Agent Orchestration Platform
The Problem with Cloud-Centric AI vs Local-First AI Orchestration
The cloud has long been the default stage for artificial intelligence. Frameworks such as LangChain, AutoGen, and CrewAI make it possible to orchestrate local or hosted models. However, their design still leans toward API-based, cloud-first execution. That approach works for experimentation, yet it introduces a clear weakness: dependence.
This return to autonomy echoes the early days of personal computing explored in Riding the Waves: From Home Computers to AI Orchestration, where individual control shaped innovation before the cloud era began.
From cassette tapes and floppy disks to orchestrated AI systems, computing has evolved through every wave.Every remote call carries both cost and exposure. Sensitive data must leave the machine to be processed elsewhere. Token-based billing discourages iteration. Even when using secure endpoints, developers trade autonomy for convenience. As a result, innovation is often limited by infrastructure.
A local-first approach changes that balance. It focuses on privacy, predictability, and cost control by running agents directly on local hardware. The cloud remains useful for large or complex tasks, yet local processing gives developers freedom. It does not reject connectivity; instead, it restores choice.
That principle guided the creation of a production-grade orchestration platform of roughly 3,700 lines of Python. Through seven BDD development cycles and a 96.5 percent test pass rate, it proved that a reliable system can run with zero external dependencies. Using SQLite and JSONL metrics, the same codebase coordinates multiple AI agents securely, predictably, and locally across devices.
Three-Layer Architecture of a Local-First AI Orchestration Platform
The system follows three clear layers: CLI, Orchestrator, and Registry. Each layer handles a specific function in the orchestration lifecycle.
The CLI layer, built with Typer, serves as the command surface. It offers more than twenty commands and about six hundred lines of code. Developers can initialize environments, run agents, and invoke workflows. This layer is the human-facing edge of the platform.
The Orchestrator layer, written with FastAPI, acts as the control center. It manages scheduling, routing, and task lifecycles. Its asynchronous design lets small tasks run in parallel while heavy inference jobs are handled one at a time. The main application file stays compact and easy to read.
The Registry layer defines intelligence. Eleven expert agents are declared in Pydantic configurations that describe capabilities, dependencies, and budgets. New agents can be added or updated with simple configuration changes.
FastAPI was chosen for its async speed and automatic schema generation. SQLite replaced Redis to stay aligned with the local-first approach. JSONL metrics were selected for their simplicity and transparency. As a result, commands call APIs, APIs invoke agents, and agents return results through a steady feedback loop.
These principles align with the broader ethical and security implications discussed in AI Orchestration, Security, and the Future of Work, where resilience and accountability shape the next phase of automation.
Hardware-Aware Resource Scheduling in a Local-First AI Orchestration Platform
Local-first systems must respect hardware limits. Machines differ widely: some are laptops with integrated GPUs, while others are workstation-class servers with up to 128 GB of RAM and powerful GPUs. Consequently, the orchestrator adapts through hardware-aware scheduling.
Each environment selects one of three profiles: Laptop, Workstation; or Server, defined in a simple
resources.yamlfile:profile: workstation max_agent_runs: 4 gpu_memory_limit: 16000 cpu_cores: 8
During initialization, the active profile sets concurrency gates and resource budgets. Lightweight operations run together, while heavy tasks acquire locks before execution. A dual-lock system separates general resource tracking from expensive AI calls. This method maintains parallel work without conflict.
Scheduling moves through five stages: global concurrency check, CPU allocation, GPU budgeting, codex serialization, and cleanup. Each stage keeps the system predictable and stable. Cleanup routines always release resources, even after errors.
This approach brings precision and balance to orchestration rather than experimentation.
Despite these advantages, running a local-first AI orchestration platform introduces its own constraints. The system’s performance depends directly on available hardware, and smaller machines may need to rely on compact or quantized models such as Phi or Llama variants instead of large-scale cloud models. This balance between efficiency and accuracy requires careful model selection. In addition, while workstation-class setups with 128 GB of RAM can handle concurrent agents with ease, laptops or limited servers may experience slower inference or constrained multitasking. These realities remind developers that local-first design is not about matching the cloud’s abundance, but about achieving sustainable autonomy within real hardware boundaries.
Integrating the Model Context Protocol (MCP)
While a local platform values privacy, it still needs secure communication. The Model Context Protocol (MCP) provides structured interoperability for tools that observe or influence AI workflows.
The implementation, only 254 lines of code, supports two authentication modes: simple tokens for development and shared-secret tokens for production. It runs across HTTP, WebSocket, and TCP. As a result, the system remains flexible yet secure.
Through the MCP tool system, external services can register abilities such as
memory.readormemory.write. These allow dashboards, IDEs, or bots to stream workflow events in real time. For example, a Grafana panel can show resource usage, while an IDE plugin can display agent progress.In short, MCP turns a local orchestrator into a cooperative system—connected when needed, private by default.
For a deeper exploration of how MCP enables cross-agent collaboration, see Unlocking AI Collaboration with the Model Context Protocol.
A symbolic visual of the Model Context Protocol: where developer flow, memory, and modular context converge.
DAG-Based Workflow Execution
At its heart, orchestration is dependency management. The platform models workflows as directed acyclic graphs (DAGs), where each node represents a task and edges define dependencies.
A common configuration is:
plan → (backend, frontend) → (security, qa)
The product manager agent drafts a feature plan. Backend and frontend agents work in parallel. Security and QA agents then validate results. Prompts reuse earlier outputs through simple placeholders like
{backend.result}. The queue engine runs each step, stores results, and queues the next tasks until completion.This design preserves context, improves traceability, and supports recovery from partial failure. This emphasis on context-driven execution mirrors insights from AI Agents and Large Codebases: Why Context Beats Speed Every Time.
The Three-Tier Guardrail System
Stable orchestration requires discipline. Therefore, the platform applies a three-tier guardrail system.
- Input validation filters unsafe or malformed prompts.
- Runner control manages retries and captures runtime errors.
- Output checks reject empty or inconsistent responses.
All guardrail events are logged in
guardrail_metrics.jsonlwith categories such asguardrail_block,runner_error, andvalidator_block. Developers can view them directly:python -m agents.cli.main metrics guardrail --details 5
As a result, every failure becomes visible and fixable. Silent issues disappear.
The Eleven Expert Agents
Intelligence resides in the registry of eleven expert agents. They are grouped into development, security, and infrastructure domains.
- Development:
product_manager,bdd_backend,bdd_frontend,qa - Security:
security,validator,guardrail - Infrastructure:
database,networking,web3,encryption
Each agent includes a Pydantic schema defining its role and resource limits. During startup, these definitions convert to runtime specifications. This clear separation keeps the system flexible. Moreover, every action is logged, ensuring full transparency.
Built-In Web Dashboard
Transparency should not require the cloud. Instead, the platform provides a lightweight local web dashboard with seven views: system overview, workflows, guardrails, resources, agent timeline, MCP clients, and JSON API.
Each page loads in under 100 milliseconds and refreshes automatically. It remains responsive, simple, and always available—even offline.
Context Management and Memory
Persistent context keeps intelligence coherent. The SQLite-backed memory system uses two tables:
memoryfor key-value data andhistoryfor append-only logs.Agents use REST or MCP calls to read and write context. This lets long workflows maintain state between runs. As a result, agents can recall past outputs or user preferences without external storage.
Developer Experience and Automation
Starting up is simple:
python -m agents.cli.main init --profile laptop
This single command creates all configuration files, chooses a hardware profile, and prepares directories. The CLI also scaffolds projects in five languages: Python, Go, React, PHP, and Perl. Each uses templates with variable substitution for fast setup.
With more than twenty commands and six sub-apps, Typer provides clear and self-documented interfaces. Consequently, the CLI becomes both toolkit and guide.
A BDD-Driven Development Journey
Development followed seven BDD cycles, each improving a key feature:
- MCP authentication and security
- Zero-friction initialization
- API deduplication
- Resource scheduling
- Dashboard observability
- Advanced resource tracking
- Fail-fast initialization
Each cycle used RED-GREEN-REFACTOR testing and generated living Gherkin documentation. As a result, coverage now exceeds 85 percent, keeping behavior predictable while features evolve.
A visual metaphor of how structured thinking, like Gherkin and Behavior-Driven Development, helps AI systems connect human intent with machine execution.
The importance of clear behavioral documentation aligns closely with ideas from AI, Gherkin, and the Future of Software Development: Why Behavior-Driven Development Matters.Production Readiness and Lessons Learned
The final system demonstrates production-level quality. It includes thread-safe scheduling, clear error handling, and real-time monitoring. JSONL metrics make audits simple. Configuration is idempotent and safe to repeat.
Key technical innovations include:
- Fail-fast error handling with clear fixes
- Append-only metrics for transparency
- Dual-lock control for parallel work
- Hot-swappable agent settings
- Hardware-aware scaling across profiles
Building locally highlighted several truths. Simplicity brings reliability. In addition, insight into system behavior is essential. Developer experience shapes success as much as model accuracy. Above all, privacy and control can align with capability.
The platform now runs seamlessly across laptops, workstations, and servers. Each profile is tuned to its limits, and each agent knows its role.
The Future of Local-First AI Orchestration Platforms
The local-first AI orchestration platform proves that autonomy and performance can coexist. It respects hardware, protects data, and offers hybrid flexibility. In practice, it shows that orchestration can be as private as computation itself. This serves as a foundation for tools that return control to their builders.
Next comes refinement: wider support for edge devices, stronger context management, and closer integration with ecosystems such as Claude CLI and OpenAI APIs. Although the system is already production-grade, its deeper importance lies in the idea it represents: local-first intelligence as a craft, not a slogan.
The cloud will always have its place. However, it should never be the only place. Ultimately, true orchestration begins where control is personal.
The next frontier of AI engineering will not be written in the cloud alone. It will emerge from local workstations, developer labs, and edge devices where privacy and autonomy coexist. If this vision of local-first orchestration resonates with your work or research, share your thoughts, build upon the concept, or join the discussion on how to design systems that respect both hardware and humanity. Real progress begins when we question the defaults and start building differently.
What is a local-first AI orchestration platform?
A local-first AI orchestration platform manages multiple AI agents directly on local hardware instead of relying on cloud APIs. It improves privacy, reduces cost, and increases control over performance.
How does hardware-aware scheduling improve AI orchestration?
It adapts task execution to available resources such as CPU cores and GPU memory, ensuring stability on devices ranging from laptops to 128 GB workstations.
What role does the Model Context Protocol (MCP) play?
MCP enables secure communication between agents and external tools, allowing dashboards and IDEs to interact with workflows in real time while maintaining local control.
Can local-first systems replace cloud orchestration entirely?
Not completely. The cloud remains valuable for large-scale training and inference. Local-first orchestration complements it by offering autonomy, speed, and privacy for smaller or sensitive workflows.Key Takeaways
- A local-first AI orchestration platform enhances autonomy, privacy, and cost control by running AI agents directly on local hardware.
- It features a three-layer architecture: CLI for commands, Orchestrator for task management, and Registry for defining agent intelligence.
- The platform employs hardware-aware scheduling to optimize performance based on device capabilities, such as laptops or servers.
- The Model Context Protocol (MCP) facilitates secure communication between agents and external tools while maintaining local control.
- Its future includes support for edge devices and deeper integration with existing ecosystems, emphasizing personal control over AI workflows.
-
“Scuse me while I kiss the sky”*…
In 1967, Jimi Hendrix’s manager, Chas Chandler arranged for Jimi to meet Cream…
There was a particular night when Cream allowed Jimi to join them for a jam at the Regent Street Polytechnic in central London. Meeting Clapton had been among the enticements Chandler had used to lure Hendrix to Britain: “Hendrix blew into a version of [Howlin’ Wolf’s] ‘Killing Floor’,” recalls [Tony] Garland, “and plays it at breakneck tempo, just like that – it stopped you in your tracks.” [Keith] Altham recalls Chandler going backstage after Clapton left in the middle of the song “which he had yet to master himself”; Clapton was furiously puffing on a cigarette and telling Chas: “You never told me he was that fucking good.” – source
Hendrix’s extraodinary virtuosity has, altogether justly, gotten a great deal of attention; less well noted, his incredible mastery of the technology of music making, recording, and performance. Rohan Puranik explains…
3 February 1967 is a day that belongs in the annals of music history. It’s the day that Jimi Hendrix entered London’s Olympic Studios to record a song using a new component. The song was “Purple Haze,” and the component was the Octavia guitar pedal, created for Hendrix by sound engineer Roger Mayer. The pedal was a key element of a complex chain of analog elements responsible for the final sound, including the acoustics of the studio room itself. When they sent the tapes for remastering in the United States, the sounds on it were so novel that they included an accompanying note explaining that the distortion at the end was not malfunction but intention. A few months later, Hendrix would deliver his legendary electric guitar performance at the Monterey International Pop Festival.
“Purple Haze” firmly established that an electric guitar can be used not just as a stringed instrument with built-in pickups for convenient sound amplification, but also as a full-blown wave synthesizer whose output can be manipulated at will. Modern guitarists can reproduce Hendrix’s chain using separate plug-ins in digital audio workstation software, but the magic often disappears when everything is buffered and quantized. I wanted to find out if a more systematic approach could do a better job and provide insights into how Hendrix created his groundbreaking sound.
My fascination with Hendrix’s Olympic Studios’ performance arose because there is a “Hendrix was an alien” narrative surrounding his musical innovation—that his music appeared more or less out of nowhere. I wanted to replace that narrative with an engineering-driven account that’s inspectable and reproducible—plots, models, and a signal chain from the guitar through the pedals that you can probe stage by stage…
[And probe it Puranik does– fascinatingly, stage by stage…]
… Hendrix didn’t speak in decibels and ohm values, but he collaborated with engineers who did—Mayer and Kramer—and iterated fast as a systems engineer. Reframing Hendrix as an engineer doesn’t diminish the art. It explains how one person, in under four years as a bandleader, could pull the electric guitar toward its full potential by systematically augmenting the instrument’s shortcomings for maximum expression.
“Jimi Hendrix Was a Systems Engineer,” from @spectrum.ieee.org.
See also: “The Technology of Jimi Hendrix.”
* Jimi Hendrix, “Purple Haze”
###
As we plug in, we might send well-connected birthday greetings to another wizard with wires, Geoff Tootill; he was born on this date in 1922. An electronic engineer and computer scientist, he worked (with Freddie Williams and Tom Kilburn) to design a computer memory. To that end they built the first electronic stored-program computer— the Manchester Baby— at the University of Manchester in 1948.
The Baby was not intended to be a practical computing engine, but was instead designed as a testbed for the Williams tube, the first truly random-access memory. Nonethless, Baby worked: Alan Turing moved to Manchester to use it, and the following year, it inspired the Ferranti Mark 1, the world’s first commercially available electronic general-purpose stored-program digital computer.
#computerMemory #computing #culture #electronics #engineering #GeoffTootill #history #JimHendrix #ManchesterBaby #music #Technology -
Okay, Back of the napkin math:
- There are probably 100 million sites and 1.5 billion pages worth indexing in a #search engine
- It takes about 1TB to #index 30 million pages.
- We only care about text on a page.I define a page as worth indexing if:
- It is not a FAANG site
- It has at least one referrer (no DD Web)
- It's activeSo, this means we need 40TB of fast data to make a good index for the internet. That's not "runs locally" sized, but it is nonprofit sized.
My size assumptions are basically as follows:
- #URL
- #TFIDF information
- Text #Embeddings
- SnippetWe can store an index for 30kb. So, for 40TB we can store an full internet index. That's about $500 in storage.
Access time becomes a problem. TFIDF for the whole internet can easily fit in ram. Even with #quantized embeddings, you can only fit 2 million per GB in ram.
Assuming you had enough RAM it could be fast: TF-IDF to get 100 million candidated, #FAISS to sort those, load snippets dynamically, potentially modify rank by referers etc.
6 128 MG #Framework #desktops each with 5tb HDs (plus one raspberry pi to sort the final condidates from the six machines) is enough to replace #Google. That's about $15k.
In two to three years this will be doable on a single machine for around $3k.
By the end of the decade it should be able to be run as an app on a powerful desktop
Three years after that it can run on a #laptop.
Three years after that it can run on a #cellphone.
By #2040 it's a background process on your cellphone.
-
A short preview demonstration of PALM with llama3.2 1TB as base model - A Object-Pascal-native LLM engine with inline assembler AVX2 SIMD optimizations, running on the CPU with StreamingLLM-like "endlessly" context-windowing and 8-bit quantized weights and activations (W8A8), and multithreaded/parallelized with my PasMP library. But support for 4-bit weights (for W4A8) is also on the roadmap.
-
terminology question: exact vs approximate #kNN
HNSW is approximate, brute-force exact. but what about quantized brute-force? it's both exact (brute-force) and approximate to some degree (quantized). how do you differentiate between algorithm and precision? it should be called... -
2025 is the International Year of #Quantum #Science. #IYQ Although I’m not sure why it is this year.
1803 marks the birth of quantum #chemistry. John Dalton quantized mass in his #AtomicTheory.
1900 marks the birth of quantum #physics. Max Planck quantized energy.
#ChangeMyMind (Hint: You can’t change #history.)
-
@mariapopova I believe you correctly surmise “how staggered (Emily Dickinson’s) pliant young mind must have been to learn that scientists had just proven the existence of atoms.”
But #physicists didn’t accept the #AtomicTheory until #AlbertEinstein demonstrated its verity through Brownian motion over a *century* after #JohnDalton proposed it.
Both ‘particle #chemistry ‘ and #QuantumChemistry predate their #physics counterparts by a century since Dalton quantized *mass*!
-
Q-LLL: как мы сделали LLL-редукцию наблюдаемой, управляемой и проверяемой
Мы привыкли воспринимать LLL-редукцию как «чёрный ящик»: подали целочисленный базис, получили редуцированный базис, проверили результат. Но что, если сделать процесс редукции наблюдаемым? В статье рассказываю о Q-LLL — exact-certified алгоритме семейства LLL, где классическая корректность сохраняется, но выбор редукционных действий управляется квантизированной Gram/Lovász-геометрией. Главная идея: approximate geometry observes, exact arithmetic decides, certificate proves. Q-LLL не заменяет fplll и не меняет Lovász-критерий. Вместо этого он добавляет новый слой: quantized Gram/Lovász oracle, exact gate, fair scheduler и proof-carrying certificates, которые можно независимо проверить. В статье разбираю: — почему обычного sequential LLL недостаточно для больших семейств lattice-вариантов; — что такое Lovász slack и как из него получается карта геометрических дефектов; — как работает quantized Gram/Lovász oracle; — почему approximate слой не принимает математических решений; — зачем нужны exact-сертификаты и independent verifier; — как Q-LLL становится lattice-core для nonce-observatory; — какие результаты уже получены и какие ограничения честно остаются. Это не статья про «магическую кнопку» и не claim про универсальное превосходство над fplll. Это попытка показать новый взгляд на LLL: как на управляемый, наблюдаемый и проверяемый процесс, где квантизированная геометрия направляет редукцию, а exact-арифметика остаётся источником истины.
https://habr.com/ru/articles/1031386/
#алгоритмы #криптография #ecdsa #аудит #информационная_безопасность #сигнатуры #биткойн
-
Massive black holes drag and warp the spacetime around them in extreme ways. Observing these effects firsthand is practically impossible, so physicists look for laboratory-sized analogs that behave similarly. Fluids offer one such avenue, since fluid dynamics mimics gravity if the fluid viscosity is low enough. To chase that near-zero viscosity, experimentalists turned to superfluid helium, a version of liquid helium near absolute zero that flows with virtually no viscosity. At these temperatures, vorticity in the helium shows up as quantized vortices. Normally, these tiny individual vortices repel one another, but a spinning propeller — much like the blades of a blender — draws tens of thousands of these vortices together into a giant quantum vortex.
Here superfluid helium whirls in a quantum vortex.With that much concentrated vorticity, the team saw interactions between waves and the vortex surface that directly mirrored those seen in black holes. In particular, they detail bound states and black-hole-like ringdown phenomena. Now that the apparatus is up and running, they hope to delve deeper into the mechanics of their faux-black holes. (Image credit: L. Solidoro; research credit: P. Švančara et al.; via Physics World)
https://fyfluiddynamics.com/2024/05/black-holes-in-a-blender/
#astrophysics #blackHole #fluidDynamics #physics #quantumVortex #science #superfluid #superfluidHelium #vortices #vorticity
-
New update for the slides of my talk "Run LLMs Locally": Bonsai-8B
The latest version of Llama.cpp now supports Vulkan with 1-bit quantized models like Bonsai: 8B model having 1.1 GB in size, 2.5 GB in RAM.
https://codeberg.org/thbley/talks/raw/branch/main/Run_LLMs_Locally_2026_ThomasBley.pdf
#ai #llm #llamacpp #stablediffusion #gptoss #qwen3 #glm #localai