I just shared the full walkthrough of the Vois website refresh on LinkedIn. I have attached it here too because the process is more useful than the screenshots alone. The headline is big: less than 1.5 days from idea to deployment, with about 5 billion text-agent tokens used along the way. That total excludes image and video generation. Before I posted it, I asked a friend who is a veteran UI/UX designer to review the work. His response: "I hate you. this is good... I hate you for making it" For context, I am the founder of Vois, and this is my own product and codebase. I ran the refresh like a coordinated AI product team. I stayed responsible for the product direction, visual taste, quality bar, and every ship or reject decision. The setup: • GPT-5.6-sol handled the main orchestration. It broke the work into plans, assigned specialist agents, integrated their output, and kept the release moving. • More than 300 specialist sub-agents using GPT-5.6-terra and Grok-4.5 worked in parallel and in sequence across planning, UX, choreography, visual design, asset creation, copy, frontend development, mobile QA, accessibility, browser testing, performance optimisation, and deployment. • Fable-5 acted as an independent advisor. It reviewed alternatives, challenged decisions, and sent weak work back through another pass. • GPT-image-2 generated the images and characters. Grok Imagine generated the video assets. Their generation usage is not included in the roughly 5 billion text-token total. The text-agent token split: Main agent: 13% Sub-agents: 51.3% Advisor agent: 35.7% What shipped: • a complete visual refresh for Vois • a cinematic, scroll-driven homepage • a consistent design system across the site • dedicated desktop and mobile behavior • generated visual and motion assets • accessibility, performance, browser, and deployment fixes The 5 billion total makes the trade visible. I used far more compute and coordination to compress a large design and engineering cycle into less than 1.5 days.

Most useful answers in a community disappear into the feed. Someone solves a hard problem today. A month later, the same question returns because the answer was never turned into something durable. The community keeps paying the cost of rediscovery. That is why I built LLM Wiki: a simple shared memory for AI agents. Instead of saving every conversation, it keeps the knowledge worth reusing: 1. Save a useful article, transcript, note, or exported community thread. 2. Let the agent compile it into linked Markdown pages. 3. Ask a question later and retrieve the right section with evidence. 4. File useful conclusions back so the next answer starts from what the community already learned. This is useful for recurring questions, research libraries, customer conversations, project decisions, and any topic a group explores over time. LLM Wiki v2 improves how that memory is searched: - exact section search finds names, flags, and phrases - optional semantic search finds paraphrased ideas - reciprocal rank fusion combines both result lists - citations and source metadata show where an answer came from Cerebras recently published a detailed article about its internal knowledge base. Employees ask it more than 15,000 questions a day, according to the team. Their system is enterprise-scale, but the need is universal: information is scattered, and good answers should not have to be rebuilt every time. No embedding service is required. The default is local, Markdown-first, and dependency-free. The clever parts are optional; the knowledge stays readable. Open-source project: https://github.com/praneybehl/llm-wiki-plugin Cerebras article: https://www.cerebras.ai/blog/how-we-built-our-knowledge-base What valuable answer in this community keeps getting rediscovered?

Back in March, I started a slightly ridiculous experiment: One machine on my desk, running 10 YouTube channels while I slept. Four months later, here’s the honest update: The system can finish the work. It still can’t reliably decide what is worth making. That has become the much more interesting problem. This isn’t one prompt generating a video. It’s a production line: Research → angle selection → script → quality gate → voice generation → word-level alignment → visuals → thumbnails → video assembly → companion article → RSS/website → YouTube Each channel has its own narrator, editorial voice, audience, visual identity, and content rules. Failed stages retry. Stuck jobs recover. Weak scripts can be rejected before the expensive production stages begin. The second image is the actual atlas of the system, mapped down to its prompts, configurations, timeouts, retries, quality checks, and channel identities. To be clear, “autonomous” does not mean “effortless.” I spent months building the factory: prompts, schemas, voices, editorial rules, visual direction, quality checks, compliance, retries, distribution, and all the failure modes nobody includes in an AI demo. I don’t manually make the episodes. I built the system that does. And almost the entire factory runs locally. Voice generation using my app - Vois.so, image creation, forced alignment, subtitles, video assembly, audio processing, quality checks, storage preparation, and orchestration all happen on one machine sitting on my desk. The only external component is the intelligence layer used for research and writing, which runs through one of my existing ChatGPT or Anthropic subscriptions. So once the system is built, the marginal cost of producing another video is effectively close to zero—mostly electricity, with no per-image, per-minute, or per-video production bill. Here’s the latest result. One channel—History That Hits—recorded: 3,234 views 153.9 watch hours +23 subscribers in the last 90 days 58 subscribers total

I just watched a video about Vercel Eve and I'm struck by how similar their proposal is to Jake's ICM framework. I've included the link below if you're interested. https://vercel.com/eve https://github.com/vercel/eve

I just added my own voice to an animation using Eleven Labs. Version 0 had no instructions, no integration into my ICM, just a prompt that said add my voice to the video file. I do want to explore how to shape the voice and output. I also am not against doing my own readings. There is so much to explore now. Another day, another step. Where do people stand on using an AI generated voice, versus their own recording?