Expressive Limits

I've droned on about this in a few formats, but I keep stubbing my toe on it. This paper covers an approach to accelerate model capabilities by expanding the data available for re-training through the dynamic nature of Language Games.

The games in play aren't the NYT Crossword, or Scrabble - rather these are games that play with language itself. Frameworks like Socratic dialogs, debates or open ended co-creation of stories create anomalous datasets that go beyond the standard task->review based data generation.

This is the paradigm of BMAD based product creation. Role-playing and collaborative dialog through each phase of a defining and building something. Farming the labor intensive tasks out to agents that can read and write much faster than the a human; using the dialogs to extract the highest value context from the human driver.

When I've rolled through a BMAD session I've found it both incredibly useful and deceptively undermining. It's too easy to trust the output and continue. That's a weakness that I think that many people, myself foremost among them, struggle with. It's not that I think we should scrap this approach, I'm more in the - we don't really have the right tools to do this just yet camp.

Getting back the paper, I don't have the background to evaluate the math & analysis behind it, but the language of the paper itself is decidedly stacked in one direction. It's the same thing I do when I pontificate about a specific technology that I also happen to be deeply familiar with. I over index on the value and undersell the limitations.

There are interesting callouts like Expressive Limits and Information Loss - where they capture the risks of being textual only, and how that may limit the ability to include aspects of the human experience. A shocking conclusion.

What really gets me about this paper is not just that it isn't a contemplative exploration of the topic. It's also not that it's 100% written by AI according to Panagram. What sticks in my craw is that this is the messaging we keep getting about what's coming.

Satya Nadella's learning loops, which is a tidy packaged version of looping agentic harnesses still doesn't get beyond the twin challenges of automation bias, and the inability to capture individualistic voices.

I'm fully onboard with the notion that those don't matter for quite a bit of what the technology world churns out. I will be the first in line to click the button to automate building an internal admin tool, or for the love of all that is good and free; auto-generate a status email.

However for innovative value. To create something that shifts the market in a meaningful way. I can't seem to get past the limitations inherent to our current toolset. The humans and intelligence tools don't seem quite ready to go infinite.

Dirges

Here's two Dirges about the Scourge of the Spurge Surge - one AI, one human. It might matter, it might not, but I think we need to spend some time asking if it could.

The Scourge of the Spurge Surge Dirge - A poor needler's complaint without restraint

The heat doth press,
Bringing ingress.
The worming Spurge,
Begins it's surge.

Upon knees we chop,
Eyeing milky drop.
Vainly pulling,
Picking and forking.

The vile scourge,
Of its relentless surge,
Claims all foes,
As only it grows.

A Dirge for the Spurge Surge

Toll the bell, the garden weeps,
Through the soil the spurge now creeps.
A milky scourge, a weed reborn,
It surges past the trampled corn.

We pulled, we sprayed, we cursed its name,
Yet still it spreads, and still we're shamed.
Lay down the hoe, abandon hope—
The spurge has won. We cannot cope.