Published on [Permalink]
Reading time: 5 minutes
Posted in: ai • productivity • agents • open-source • privacy

The pathfinder finds the cliffs

Another week, another claimed breakthrough - this one called OpenHuman.

It lives in an avatar, has a voice and can join your Google Meet as an equal - “joins meetings, transcribes them into your Memory Tree, and can speak back into the call.” The pitch is that it is “an open-source AI assistant designed to be the memory and doer for everything you do across your tools.”

It ticks a lot of boxes. It promises all the right things. It is very tempting.

But here is the problem. When even Apple can’t get Siri to be meaningfully smarter - and that’s with billions in resources, decades of tech experience and tight hardware integration - how does a small, open-source project deliver on “personal AI super intelligence”?

It probably doesn’t. Not yet. Not fully.

OpenHuman is a real project with real momentum - around 12,000 followers on GitHub, active development and an honest “Early Beta” warning in its own documentation. Under the hood it runs a small, locally-installed AI model from Google called Gemma 3. That part is legitimate.

But a small local model and “superintelligence” do not belong in the same sentence. That model handles simple tasks well - summarising, categorising, deciding which tool to call next. Anything harder gets quietly routed to bigger, cloud-based models. If OpenHuman feels powerful, that power is mostly coming from the plumbing underneath - the integrations, the memory system and the connections to larger services - not Gemma 3 doing the heavy lifting alone.

The “privacy stays off the cloud” claim is also slippery. The documentation shows optional routing to 30+ cloud providers under a single subscription. Out of the box, the app has analytics and error-reporting tools switched on. None of that is automatically bad - but it means “private, simple and extremely powerful” is an aspiration, not a design guarantee. Worth understanding before you hand it your Gmail.

Even the originator of this agent category - OpenClaw - is struggling to scale robustly. Its creator burned through $1.3 million in AI usage fees in a single 30-day period. That is not a rounding error. It illustrates just how economically demanding these systems are at any meaningful scale - and raises real questions about whether the business models and cost structures are anywhere near figured out yet. If you are building on top of these agents or depending on them in any serious way, that fragility is your problem too.

I have been running Hermes on the side. Solid reputation, interesting approach. In practice: flaky. Not something I would go near in a production environment.

Meanwhile, Claude’s Dispatch - which lets Claude phone home to my desktop from anywhere - just quietly did something none of the others managed. Sitting in a cafe, I remembered I needed to add Apple’s WWDC to my calendar. WWDC - Apple’s Worldwide Developers Conference, the Digital Woodstock for Apple fans and developers - runs June 8-12 this year, with the keynote on Monday June 8 at 10am PT (that’s 3am AEST Tuesday June 9 for those of us setting an alarm).

Siri was clueless. Dispatch figured out the timezones, told me when the keynote was and - at my request - put it all in my calendar. No technical setup required. No error messages to Google. Just done.

The direction is clear enough. Better, more robust and more capable agents are coming. The question is when, not if.

Right now, getting OpenHuman running means executing a setup script in the terminal and then connecting it to your Gmail, Slack, GitHub and Notion. If you love that sort of thing - genuinely, carry on. But for most of us, there is nothing wrong with waiting six months for the dust to settle and the rough edges to smooth out.

The pathfinder finds the cliffs. There is nothing wrong with waiting for the map.

Update: On the day OpenHuman hit #1 on GitHub and #1 on Product Hunt simultaneously, someone unconnected to the project took the mascot image and launched a speculative cryptocurrency token on a platform called Solana; essentially creating a tradeable coin backed by nothing except the fact that people might buy it. The project team had nothing to do with it. The software is still in early beta. The hype, apparently, could not wait.

Geek speak appendix

For the technically curious.

API tokens / AI usage fees - Most AI services charge by the amount of text processed (called “tokens”). One million tokens is roughly 750,000 words. At scale, this adds up fast - OpenClaw’s $1.3M bill was a month of heavy real-world use.

Open-source - The code is published publicly so anyone can read, audit or contribute to it. Generally a good thing. It also means the quality depends heavily on the team maintaining it.

GitHub stars - A rough popularity signal on GitHub, the main platform where developers publish open-source software. 12,000 stars suggests real interest, not a hobby project.

Local model vs cloud model - A local model runs directly on your computer (private, but limited). A cloud model runs on someone else’s servers (more powerful, but your data travels there). Most “private AI” products use both.

1 billion parameter model - A way of measuring an AI model’s size and capability. 1 billion is actually quite small by modern standards - GPT-4, for comparison, is estimated to be orders of magnitude larger. Useful for lightweight tasks, not frontier reasoning.

Terminal / setup script - The terminal is a text-based interface for giving your computer instructions directly. A setup script automates a sequence of those instructions. Completely normal for developers; unfamiliar and intimidating for everyone else.

YAML files - A type of plain-text configuration file that developers use to set options. Looks simple; error messages when you get it wrong are not.

Sources:

✍️ Reply by email