Let’s build your great helper

Jonathan Ward ·

The great helper learns your goals over time by listening to you and observing you. The great helper then works constantly to help you achieve them, actively seeking out your feedback and learning quickly from it. Finally the great helper is loyal to you, protective of your time, resources, health, and privacy. I want an AI great helper for myself, and I want to bring AI great helpers to the world for anyone who wants one.

We’re getting glimmers of the AI great helper with the latest versions of ChatGPT and Claude, but they are still missing a lot. They only know what you tell them, their memory is superficial, they don’t actively seek feedback, the labs hold custody, and they are trained to solve standalone problems rather than further your long-term goals.

We can build something much better if we rethink the approach.

But why?

First, it is good for individuals to shape, own, and understand the technology they rely on. People should be free and sovereign. The AI agents you use should be loyal to you. You should be able to teach an AI your goals and tacit knowledge without that AI being used by the organization that owns it to disempower you.

Second, AI systems today are bottlenecked on the existence of “robust goals”. Enormous resources are going towards creating training data and environments. Outside of a few areas like math and coding, the ground truth for whether the AI agent has acted correctly exists in the mind of a human. Your goals in the context of your life are therefore extremely valuable.

If you’ve engaged deeply with these claims and disagree with them, let’s talk.

What must be built

Fundamentally, we need a low-friction way for a single person to produce enough data from their life to teach a model how to help them achieve their goals. The basic idea is “data-maxxing” - record and process as much data as possible from your life - while keeping that data under your sovereign control.

Here’s a way to visualize how the data will flow

User-Owned Intelligence Loop Your digital life is recorded by an open-source local app and stored locally on your device. It produces active, passive, and accumulated data. Data governance lets the user review and remove data before a training data pipeline uses data transformation compute to turn the raw data into signal. Model training with RL, SFT, and other methods finetunes an open weight base model such as Kimi, Qwen, DeepSeek, or GLM using user-owned or securely rented compute to create a user-owned model. The model runs private inference, and its outcomes return to active data through an online learning loop. Your Digital Life Active Data Passive Data Accumulated Data Govern Your Data Training Data Pipeline Data Transformation Compute Model Training Open Weight Models Training Compute User-Owned Model Use Your Model Inference Compute ONLINE LOOP

Your model creation, usage, and ongoing learning is enabled by combining:

  1. Your data and decision-making.
  2. The model-training and data-generating systems that I build.
  3. The open weight models and efficient inference/training stacks made by the community.

If you’re working towards the same goals and think there’s a better way to build this, let’s talk.

Who I am & Why start this now

I am a posttraining researcher with experience in human data collection, synthetic data creation, RL training, (prosaic) Alignment and launching consumer AI features.

OpenAI — Posttraining research, June 2026 – October 2020

Based on what I’ve seen, we now have the right ingredients to build this.

  1. High-quality semi-synthetic data generation is now within reach.
  2. Open-weight base models are now good enough to productively finetune.
  3. It’s now apparent that the bottleneck on agent utility in a lot of domains is how well they understand your goals, not their raw intelligence.

The primary reason for writing this post at the early stage of development is to make contact with the other people creating in this space so that we can discuss ideas together, and alpha-test each other’s tooling. It’s time to make a scene, let’s talk!

Contact info

Preferred: Email · X

Other: GitHub · LinkedIn · Substack · Google Scholar

The motivating vision of the future

In 2031, anyone who wants can own their own personal model, the compute to inference and train it, and advanced wearables and neural prostheses to steer it and for it to learn from them. They feel the inalienable satisfaction of the craftsman as they shape the AI they use with their own hands and minds. They are able to extend their idiosyncratic genius to create their own art forms and schools of thought. The world sees a flowering instead of a flattening since we have chosen to extend illegible individual minds rather than replace them with a lossy compression.

Inspiring ideas worth reading