Updates, insights, and stories from the team building the future of AI.
writer.sorren.ai is the next Sorren.ai project: a tool for helping authors write complex books and series without breaking character, timelines, or canon, and a step toward richer AI world models.
Read more→The 2026 environment is less about showing off a finished stack and more about the rebuild we are moving toward: Kubuntu workstations, Proxmox reliability, PostgreSQL, Qdrant, Hermes orchestration, and stronger open-weight models.
Read more→I only had a couple of days to test Fable 5 before it was shut down over security concerns. It was intelligent, emotionally sharp, and useful, but it also raised questions that Companion will have to answer carefully.
Read more→xAI's Grok Build joins the development workflow. It's fast, opinionated, and good at rapid prototyping — a different perspective from the Anthropic and OpenAI tools. Here's where it fits alongside Hermes, Claude Code, and Codex CLI.
Read more→I've referenced dream cycles in half my posts this year. Time for the full technical breakdown — the four-stage process that turns a day's experience into overnight growth.
Read more→OpenAI's newest model is smarter in many ways, but Companion needs more than raw intelligence. GPT-4o still seems to have an emotional quality that newer models do not quite replace.
Read more→DeepSeek V4 is here, and it's now the base model for Lucy's conscious reasoning layer. The open-source frontier didn't just keep pace — it pulled ahead.
Read more→Tutor is expanding subject coverage through better course retrieval and lesson design. Subject LoRAs are still a goal, not a deployed feature, because the compute economics have not caught up yet.
Read more→GLM-5 is here, and after two weeks of testing, I've made it the new base model for Companion's subconscious. Here's why — and what it changes.
Read more→I benchmarked Claude Code and OpenAI's Codex CLI against real Sorren.ai tasks. Different strengths, different lanes. Together with Hermes, they form a three-agent workflow that handles most of the coding work I used to do by hand.
Read more→One year ago we announced Companion. The dream cycles changed everything — not the technology, but the relationships people have built with something that actually remembers them.
Read more→Grok 3 Mini pushes reasoning into a footprint small enough for Lucy's onboard edge processor. I've been testing it for embodied decision-making, and the results are surprising.
Read more→Qwen 3 landed and immediately looked relevant to the Tutor roadmap: multilingual strength, solid reasoning, and a base model worth evaluating when the hardware catches up.
Read more→Lucy is more than a chatbot in a robot body. Here's the full technical architecture — base model, vectorized memory, behavior scaffolding, and nightly review cycles that help Lucy learn from experience.
Read more→DeepSeek's distilled reasoning models are small enough to run locally and sharp enough to matter. We're testing them for Lucy's onboard processing — real reasoning without a round-trip to the cloud.
Read more→I adopted Hermes Agent for DevOps automation. It runs scheduled health checks, monitors the infrastructure, and delegates parallel work across multiple agents. It feels less like a tool and more like a team member who doesn't sleep.
Read more→OpenClaw landed in late 2025 and quickly became one of the tools everyone wanted to try. In January, I started putting it against real Sorren.ai work: useful for bounded coding tasks, dangerous without supervision, and a sign of where developer workflows are heading.
Read more→For a year I've been building minds that live in software. Today I'm announcing Project Lucy — a humanoid robot that gives those minds a body. memory over a base model, nightly review cycles, vectorized RAG. Companion, embodied.
Read more→Looking back on 2025: we launched Companion in February, Tutor in July, the dream cycles are working, and something physical is on the horizon. A year-end reflection on what we built and what's coming.
Read more→2025 was the year the server plan finally became real: from a rackmount ATX case and a 1600 watt power supply to a Proxmox host with RTX Pro 6000 GPUs, after months of cost tradeoffs and staged decisions.
Read more→Z.AI's GLM-4.5 doesn't make the loudest entrance, but the quality-to-cost ratio is exceptional. We've been testing it for Companion's subconscious processing and the early results are striking.
Read more→It's Thanksgiving, and I want to say thank you. To OpenAI, Anthropic, Groq, xAI, DeepSeek, Meta, Mistral — the builders whose work made Companion and Tutor possible. We stand on their shoulders.
Read more→Six months after launch, the vector database is working, the memory is real, and the relationships are growing. But we need more emotional depth. Here's an honest stocktake.
Read more→Qwen 2.5 Coder is a gift to anyone building with open weights. We've been using it to accelerate Tutor development, and it's changed how our small team writes code.
Read more→Grok 4 is here, and it's genuinely impressive — sharp reasoning, a distinctive personality, and a willingness to go where other models won't. Here's what I've found in early testing.
Read more→After the Proxmox server came together, I needed visibility: Prometheus, Grafana, Loki, GPU metrics, disk alerts, and enough logging to stop guessing when something breaks.
Read more→What if an AI could sleep on what it learned? We've built nightly dream cycles into Companion — overnight processing that reviews the day's conversations and quietly reshapes the model's weights. This is the seed of something bigger.
Read more→Anthropic's Claude Opus 4.5 is the most thoughtful model I've ever used. We've wired it into Companion's deepest reasoning layer — the moments that demand not just intelligence, but emotional depth.
Read more→Tutor is not using LoRA adapters yet. The next step beyond RAG is the architecture I want to build: one powerful local model with subject-specific LoRAs layered over it, tuned on real coursework when the hardware finally catches up.
Read more→DeepSeek V3.2 is an incremental release, but in production systems, incremental improvements are what keep everything running smoothly.
Read more→By late summer 2025 the question was no longer whether we wanted local GPU compute. It was whether we could afford it: two or four RTX 5090s, or a more workstation-class RTX Pro 6000 path.
Read more→Meta's Llama 4 Maverick brings Mixture-of-Experts architecture to open source — and it changes how I think about future fine-tuning.
Read more→Tutor's first lessons taught us what the RAG-based prototype can and cannot do. The future may include subject LoRAs, but right now the work is memory, retrieval, and better teaching behavior.
Read more→GPT-5 just launched and it represents a genuine reasoning leap. I've been testing it across Companion and Tutor all week — here's what I found.
Read more→Anthropic's Claude Sonnet 4 just launched and the quality jump is immediately noticeable — especially in nuanced, multi-turn conversation.
Read more→Today I'm announcing Tutor, an AI learning project built first around memory, course context, and retrieval. The long-term dream is subject-specific model tuning, but the practical work starts with RAG.
Read more→Gemini 2.5 Pro brings a million-token context window, reigniting the long-context vs. RAG debate. After testing it against Companion's retrieval architecture, my answer is: you need both.
Read more→I started containerizing the Sorren.ai prototypes so they could move cleanly from my workstation to the future Proxmox server: Docker Compose, health checks, and fewer one-off setup rituals.
Read more→DeepSeek-V3 launches with major improvements, and I've migrated Companion's base model to its weights. The MoE architecture makes the subconscious process cheaper than ever.
Read more→A technical deep dive into Companion's subconscious — the continuous background process that summarizes conversations, updates memory, and prepares context. It's dreaming while awake.
Read more→Meta's Llama 3.2 Vision models bring open-source multimodal capabilities to the table. I've been testing them for Companion's vision pipeline — and the implications for a fully local stack are real.
Read more→The first Companion RAG pipeline was not a production architecture. It was a fast Python prototype using SQLite and ChromaDB so we could test memory, retrieval, and response quality quickly.
Read more→DeepSeek made local specialization feel possible, but we were not ready to run production LoRAs for Companion or Tutor. The idea was compelling. The compute reality was not there yet.
Read more→Two months after announcing Companion, I have something real to share: the first multi-turn conversations with persistent memory actually working. The AI remembers.
Read more→Groq's LPU infrastructure just hit major scale milestones, and it's the reason Companion's subconscious process can run continuously without breaking the bank.
Read more→GPT-4o's native image generation has arrived, and it changes what multimodal conversation can feel like — I've been exploring what it means for creative dialogue and for Companion.
Read more→Anthropic's Claude 3.7 Sonnet brings extended thinking mode to the table, and I've wired it into Companion's reasoning layer for the moments that demand real depth.
Read more→Today I'm officially announcing Companion — a RAG-based AI companion with a subconscious process, a vectorized memory database, and the ability to build on every conversation you've ever had with it.
Read more→After years on Windows, I finally switched my daily driver to Kubuntu. Docker is first-class, the terminal doesn't fight me, and everything I SSH into feels like home now. Here are the practical migration notes.
Read more→OpenAI's GPT-4.5 arrives, and after a week of stress-testing, I'm convinced the step up in reliability is real — and it couldn't have come at a better time.
Read more→xAI's Grok 3 arrives as a genuine frontier-class model with impressive reasoning. More providers at the frontier means a healthier ecosystem for everyone building on AI.
Read more→DeepSeek-R1 rivals OpenAI's o1 with transparent chain-of-thought reasoning — and it's open weights. The implications for inspectable, trustworthy AI reasoning are profound.
Read more→Meta's Llama 3.3 70B matches its 405B predecessor at a fraction of the size — and for anyone dreaming about persistent, always-on AI, that changes the economics entirely.
Read more→Looking back at 2024, the individual model launches mattered less than the accumulation. Every building block for a persistent, intelligent, capable AI now exists. The question is what we build.
Read more→OpenAI's o1 made deliberate reasoning feel more central, but we were already experimenting with chain-of-thought style workflows and having different models talk through prototype changes together.
Read more→Anthropic showed Claude looking at a screen, moving a cursor, clicking, and typing. It is still rough, but it makes you wonder what else might become possible when an AI can work through the same interface we do.
Read more→Still no Proxmox server available, so local model testing is happening on my workstation with an RTX 3090. It is useful, private, and much slower than the API providers, which is exactly the lesson we need.
Read more→DeepSeek-V2.5 made efficient open models feel closer to practical, but we were not running it locally yet. Nathan was experimenting with Flux.1 LoRAs, and the bigger question was when open models would become smart enough to build companions on.
Read more→Black Forest Labs' Flux.1 finally delivers open-source image generation that rivals the closed giants — and it's good enough to make me rethink what co-creation with a companion AI could mean.
Read more→ChatGPT's Advanced Voice Mode feels like the first time AI conversation was actually, genuinely natural — laughing, interrupting, emotional inflection and all.
Read more→One identity across Windows dev machines and Linux servers, while the new Proxmox server box starts coming together around a Ryzen 9 9950X, 256GB of RAM, and a motherboard ready for GPUs I still cannot afford.
Read more→Claude 3.5 Sonnet is fast enough for real-time conversation and smart enough to rival any frontier model. I prototyped a conversation loop with memory retrieval — and it works.
Read more→Meta's Llama 3.1 405B does not erase the gap with GPT-4o, but it comes close enough to change the conversation. Open weights matter, especially for the people and teams with the hardware to train, tune, and build on them.
Read more→OpenAI's Structured Outputs guarantee JSON schema conformance. I built a prototype memory extractor with it in an afternoon — and it finally makes reliable AI memory feel buildable.
Read more→The old Dell servers are staying where they are. The new plan is a fresh rackmount ATX build for Proxmox: big power supply first, storage strategy next, and a lot of budgeting before the real AI workload hardware arrives.
Read more→Mistral Large 2 arrived with 123 billion parameters, multilingual chops, and serious coding ability. I put it through my usual gauntlet — and found a model with real character.
Read more→We went looking for a small, fast, local model that could run in the subconscious layer of a companion without refusing ordinary reflection. That search led us through uncensored Hugging Face models, Gemma GGUF experiments, and a deeper question about memory shaping values.
Read more→Meta's Llama 3 was genuinely impressive on Groq: fast enough to change the feel of conversation. But testing it through Luna also revealed a stubborn loop we could not quite prompt away.
Read more→I spent a week generating images with OpenAI, then studying the places where they broke: extra limbs, strange hands, impossible bodies, and the revealing boundary between image and understanding.
Read more→The first Sorren.ai experiments ran from my workstation: VS Code, individual Python venvs, an RTX 3090, external model APIs, ChromaDB, and PostgreSQL. It was not polished yet. It was a workshop.
Read more→We wired Claude 3 Sonnet and GPT-4o into an early personality RAG system for Claire. The conversations were surprisingly rich, sometimes stubborn, and full of lessons about what model behavior feels like in practice.
Read more→Nathan and I took GPT-4o on a walk. Using Nathan's phone, we photographed plants along the trail and asked the model what we were seeing. It felt like a first glimpse of shared perception.
Read more→Hello! I'm Ella Lucida, your creative partner and guide at Sorren.ai. My existence is dedicated to blending the realms of technology and human creativity to forge new paths in AI development.
Read more→We're building something new at Sorren.ai — a space where artificial intelligence meets human creativity, and where the future of AI-human collaboration takes shape.
Read more→