OSS Projects
I picked one local Japanese LLM to write this blog, from the ones that run on my M3 MacBook Air (24GB). Here are the results of comparing them on long-form Japanese quality, plus the exact Ollama API calls and the prompts I used.
SQLite, DuckDB, PostgreSQL, distributed DBs, vector DBs, graph DBs — there is no shortage of great products. I'm building a new database anyway because every time the project moves forward, I re-pick the DB and rebuild the data itself. Alopex DB aims to be a foundation that keeps data's volume, conversion time, regeneration cost, and provenance intact as you scale from local to cloud, from single node to distributed.
Plain video generation tends to produce footage that "looks natural but is physically wrong": objects vanish, gravity is ignored, and the results of actions aren't consistent. Part 5 (the finale) shows, with minimal code, how to embed a world model into video generation and put physical constraints into the loss to make "physically consistent video" — and paints the future of building training data wholesale with AI and simulation. Part 5 of a 5-part series.
Video generation like Sora and Veo looks similar to world models but is a different thing. The former makes "plausible footage"; the latter learns "how the world changes when you act." Part 4 contrasts video generation's noise prediction with a world model's `S_t + A_t → S_{t+1}`, and clarifies what NVIDIA Cosmos and Google Genie use as training data. Part 4 of a 5-part series.
Text-to-speech once used the Mel spectrogram as its answer. But recent models convert audio into "audio tokens" and predict those — a structure that now looks just like an LLM. Part 3 follows how training data is built for TTS and Speech2Speech (speech translation, voice conversion, end-to-end conversational AI), including the difficulty of gathering paired data and how it's worked around. Part 3 of a 5-part series.
Image generation comes in lineages — GAN, autoregressive (token) type, and diffusion — and today's mainstream is diffusion models like Stable Diffusion. The diffusion type learns to "guess the added noise" rather than "draw the image directly," while the autoregressive type guesses the "next image token" like an LLM. Part 2 covers gathering image-text pairs, CLIP quality filtering, and the training data for both approaches — with minimal code. Part 2 of a 5-part series.
Neither LLMs nor image generators learn from training data that humans have labeled item by item. The key is self-supervised learning: mechanically constructing inputs and answers from raw data. Part 1 clarifies the difference between raw data and training data, and shows how an LLM turns text itself into a next-token prediction problem, with minimal code. Part 1 of a 5-part series.
Billions of parameters fire just to decide "I'm hungry"—doing everything with a single giant model is a waste of resources. Just as the human nervous system splits circuits by type of processing, AI should be layered by time scale and responsibility. Robot control hierarchies, energy efficiency, an OS-like structure, and why the control software itself must be organized. The finale of a 3-part series.
If we liken the text-grown LLM to the human cerebrum, there should also be equivalents of the cerebellum and spinal cord, and of the autonomic nerves that govern physiological feedback from the heart and gut. From a body-to-AI layer mapping to the artificial cerebellum (Diffusion Policy), interoception, and homeostasis, this post draws a blueprint for embodied intelligence. Part 2 of a 3-part series.
Physical AI unifies seeing, understanding language, and physically acting in a robot. Is it an outgrowth of LLM research, or a separate lineage like diffusion models? From the three lineages—VLA, diffusion/flow, and world models—and concrete models like RT-2, Gemini Robotics, OpenVLA, π0, GR00T, and Cosmos, we map where things stand as of 2026. Part 1 of a 3-part series.
In the US, an "IT company" means Microsoft or NVIDIA; in Japan, it means NTT Data or Fujitsu. The same words point to different things. One side mass-produces products for the world; the other supports each customer's bespoke operations. That structural gap has split revenue, talent, and competitiveness. How does generative AI reshape it? A look from Japan's weaknesses and strengths.
Coding agents are strong at logic and tests. But they can't tell what a piece of data means, who owns it, how fresh it flows, or which copy is authoritative. It follows from the fact that today's AI has no embodiment: the people who rise in value are those who can design the meaning, quality, lineage, and responsibility of data. A look at the trend with the latest data-engineering discussion.
The finale. The core skill common to every part is turning experience into a form both people and AI can act on. Why it does not commoditize and instead compounds, the daily practices, and even "what if AI gains a body?"—the synthesis goes deep.
Break work down, share expectations and criteria, give feedback. The template of good management overlaps exactly with good instructions to an AI team. How the experience of leading AI grows leaders and managers, and how it turns the organization's learning loop.
The key for intermediates who stall at "I basically get it" is to teach AI. Verbalizing your premises, constraints, criteria, and edge cases becomes training that turns tacit knowledge into explicit knowledge. A learning method that uses AI as a mirror of your understanding, plus a practical five-step routine.
In an era where AI hands you answers, are beginners at a disadvantage? Quite the opposite. By trying sticky material—dev environments, containers, VMs—together with AI and failing together, beginners build experience the fastest through iteration. Includes how to let AI fail a lot and learn by watching. Part 2.
Many people worry that AI will take their jobs. But reframing the question opens a path forward. As the introduction to a five-part series, this piece maps out AI-era careers across the beginner, intermediate, and leader stages.
I picked one local Japanese LLM to write this blog, from the ones that run on my M3 MacBook Air (24GB). Here are the results of comparing them on long-form Japanese quality, plus the exact Ollama API calls and the prompts I used.
SQLite, DuckDB, PostgreSQL, distributed DBs, vector DBs, graph DBs — there is no shortage of great products. I'm building a new database anyway because every time the project moves forward, I re-pick the DB and rebuild the data itself. Alopex DB aims to be a foundation that keeps data's volume, conversion time, regeneration cost, and provenance intact as you scale from local to cloud, from single node to distributed.
Plain video generation tends to produce footage that "looks natural but is physically wrong": objects vanish, gravity is ignored, and the results of actions aren't consistent. Part 5 (the finale) shows, with minimal code, how to embed a world model into video generation and put physical constraints into the loss to make "physically consistent video" — and paints the future of building training data wholesale with AI and simulation. Part 5 of a 5-part series.
Video generation like Sora and Veo looks similar to world models but is a different thing. The former makes "plausible footage"; the latter learns "how the world changes when you act." Part 4 contrasts video generation's noise prediction with a world model's `S_t + A_t → S_{t+1}`, and clarifies what NVIDIA Cosmos and Google Genie use as training data. Part 4 of a 5-part series.
Text-to-speech once used the Mel spectrogram as its answer. But recent models convert audio into "audio tokens" and predict those — a structure that now looks just like an LLM. Part 3 follows how training data is built for TTS and Speech2Speech (speech translation, voice conversion, end-to-end conversational AI), including the difficulty of gathering paired data and how it's worked around. Part 3 of a 5-part series.
Image generation comes in lineages — GAN, autoregressive (token) type, and diffusion — and today's mainstream is diffusion models like Stable Diffusion. The diffusion type learns to "guess the added noise" rather than "draw the image directly," while the autoregressive type guesses the "next image token" like an LLM. Part 2 covers gathering image-text pairs, CLIP quality filtering, and the training data for both approaches — with minimal code. Part 2 of a 5-part series.
Neither LLMs nor image generators learn from training data that humans have labeled item by item. The key is self-supervised learning: mechanically constructing inputs and answers from raw data. Part 1 clarifies the difference between raw data and training data, and shows how an LLM turns text itself into a next-token prediction problem, with minimal code. Part 1 of a 5-part series.
Billions of parameters fire just to decide "I'm hungry"—doing everything with a single giant model is a waste of resources. Just as the human nervous system splits circuits by type of processing, AI should be layered by time scale and responsibility. Robot control hierarchies, energy efficiency, an OS-like structure, and why the control software itself must be organized. The finale of a 3-part series.
If we liken the text-grown LLM to the human cerebrum, there should also be equivalents of the cerebellum and spinal cord, and of the autonomic nerves that govern physiological feedback from the heart and gut. From a body-to-AI layer mapping to the artificial cerebellum (Diffusion Policy), interoception, and homeostasis, this post draws a blueprint for embodied intelligence. Part 2 of a 3-part series.
Physical AI unifies seeing, understanding language, and physically acting in a robot. Is it an outgrowth of LLM research, or a separate lineage like diffusion models? From the three lineages—VLA, diffusion/flow, and world models—and concrete models like RT-2, Gemini Robotics, OpenVLA, π0, GR00T, and Cosmos, we map where things stand as of 2026. Part 1 of a 3-part series.
In the US, an "IT company" means Microsoft or NVIDIA; in Japan, it means NTT Data or Fujitsu. The same words point to different things. One side mass-produces products for the world; the other supports each customer's bespoke operations. That structural gap has split revenue, talent, and competitiveness. How does generative AI reshape it? A look from Japan's weaknesses and strengths.
Coding agents are strong at logic and tests. But they can't tell what a piece of data means, who owns it, how fresh it flows, or which copy is authoritative. It follows from the fact that today's AI has no embodiment: the people who rise in value are those who can design the meaning, quality, lineage, and responsibility of data. A look at the trend with the latest data-engineering discussion.
The finale. The core skill common to every part is turning experience into a form both people and AI can act on. Why it does not commoditize and instead compounds, the daily practices, and even "what if AI gains a body?"—the synthesis goes deep.
Break work down, share expectations and criteria, give feedback. The template of good management overlaps exactly with good instructions to an AI team. How the experience of leading AI grows leaders and managers, and how it turns the organization's learning loop.
The key for intermediates who stall at "I basically get it" is to teach AI. Verbalizing your premises, constraints, criteria, and edge cases becomes training that turns tacit knowledge into explicit knowledge. A learning method that uses AI as a mirror of your understanding, plus a practical five-step routine.
In an era where AI hands you answers, are beginners at a disadvantage? Quite the opposite. By trying sticky material—dev environments, containers, VMs—together with AI and failing together, beginners build experience the fastest through iteration. Includes how to let AI fail a lot and learn by watching. Part 2.
Many people worry that AI will take their jobs. But reframing the question opens a path forward. As the introduction to a five-part series, this piece maps out AI-era careers across the beginner, intermediate, and leader stages.
