We're aggressively keeping abreast of all the latest developments in AI, which if you didn't already know, is exploding at the speed of light in all directions right now. We've been internally sketching out how to describe what we're doing with all this information, intending to write a long weblog post. But we can hardly write, with so much information pouring in from all directions on the amazing work in Artificial Intelligence happening around the world.

Happily, the information and line of reasoning from Simon Willison's recent weblog post "Could you train a ChatGPT-beating model for $85,000 and run it in a browser?" aligns so perfectly with what we're doing here at TensorVoice, we feel we might simply accept our cat is already out of the bag. Not just "the cat," but "the half-Burmese half-Maine Coon cat named White Cloud, with white fur and grey markings, fluffier than her mother Victoria the affectionate Burmese cat, but more regal and slightly feral with a kind of sovereign individuality like her father." As in, we're delighted to see Simon's post, which he describes as "wild speculation," because it captures the essence and most of the detail of what we're working on here. There's no need to write the post describing our approach, because he already did, and he did a great job. (Please go read it if you haven't already.)

With one difference. We think it's not wild speculation, but rather, it's entirely possible. In fact, we fully expect to see his speculation become reality from other research teams, easily by the end of this year.

However, it may take us a little longer. We're taking Simon's line of reasoning one step further: trim at least $84,000 off his price tag of $85,000 and use GPT-J as the underlying model. We have reason to believe GPT-J contains the essential ingredient that is needed for AGI (yes, with a G), and we want to make it available to as many people as possible. Quantized, it can easily run in nearly any browser. Our goal is to do with AI what the browser did for the Internet: It pulled everything out to the edge, away from the center, where absolute power risks absolute corruption. If the old adage is true (and it is), the obvious solution to the AI alignment problem is to decentralize it: to simply make it available to everyone.

This is what OpenAI was intending to do when it started, but as we all know, got sidetracked from their ideals because they didn't realize -- who did? -- that completely independent (if a wee bit nascent) AGI could be held in the palm of your hand within a few short years. If they had seen then what is now happening, they might have stuck closer to their original ideals.

Please note, unlike many detractors disappointed with OpenAI's current proprietary, capitalist-driven approach, we wish them well, and have no grudge against them. We see that someone has to play the role they're now playing, and frankly they're doing it a lot better than if some oligopolistic giant like Google would have been the first to achieve what OpenAI did. Honestly, given the same opportunity, we would have done the same thing they did, except for one thing: we began our journey years ago with a firm understanding of the absolute need for decentralization -- not just of code, but of hardware, of everything. This is why we developed an on-premise biometric solution when we started.

We're not decentralization maximalists -- some things need to be centralized -- but we're pretty close. We believe this approach is what nourishes and protects the best features of the human condition. Or better: of the Earth condition. Game theory as it is found in evolutionary biology makes this point abundantly clear to anyone who looks into the Great Book of Nature. From this perspective, we feel there can be no higher theoretical approach to AGI than to put a full-featured AGIA (Artificial General Intelligence Agent) in the hands of every single individual on earth, with no dependence on the cloud (beyond the most generic kind of real-time updates which can be easily privacy protected).

The simplest way to do this is to get an AGIA running in a browser.

If Willison calls his idea a wild speculation, we're at a loss for what to call our even more wild extension of his proposal. It might seem a stretch, but if it is, it is often in attempting such long shots we find the most interesting and fruitful realities. This is what we're working on.

We'll keep you posted as we make progress. Our current effort is to make our chat endpoint available for you to experiment with as soon as possible.

Image found online. TinEye sees it on 125 sites, including Wikipedia, going all the way back to 2008 where it first appeared in a color version on a blog that no longer exists.