Old Bottlenecks New Bottles

The future is made of the same stuff as the past.

Electronic computers enabled the first data practitioners to flood the firm with information. As computation got cheaper, so did data. The Internet increased the interconnectivity of data, which meant we got to curate, generate, transform and publish data about data inter-connectivity. LLMs are an accelerant.

The job titles evolve: database analyst, analyst, data analyst, business intelligence analyst, web analyst, digital analyst, data scientist, data engineer and now, agentic analytics engineer and agentic analytics analyst.

The same stock problems persist: data quality, numeracy, insight and actionability. These are all the polite framings within organizational life. Beneath each term are billions of emails echo’ing the same words since the origin of the memo. A person born in 1400 has just as much trouble understanding exponents as a person born in 2000. Compounding isn’t intuitive. Data has always been dirty, in multiple formats, and fragmented. Insight has always been scarce. Actionability always constrained by authority, real and imagined.

The same stock solutions permeate around data management: centralize it into a mart, a warehouse, a lake, a lakehouse; interlink it with a graph, a mesh, a fabric, a blockchain. These are all polite framings of systematically integrative integrations. If only it were in one place, we’d know everything. Well, a frontier LLM Model has centralized quite a bit of knowledge into one place. I’m skeptical of claims that we know everything. It sure does enhance confidence though, doesn’t it?

In spite of all of these new solutions, the grand bottleneck, OODA-Lag, persists.

That’s OODA for Observe Orient Decide and Act. And lag as in the temporal lag, the massive delay from when data is first observed to when it is acted upon. A key root cause of OODA-Lag is attention. The collective brains within a firm can only process so much. An organization can only have so many number one top priorities (this appears to top out at 15 number one priorities in the most disciplined of orgs!). And there is a perception, an incorrect one, that there is only so much authority, so much power, to go around.

An agent only has so much attention, context and authority.

Even the mighty agentic swarm appears to run into issues with perspective coordination.

This gap in authority constrains experience accumulation, which slows learning, with causes OODA-Lag to increase.

Software engineers are finally, at long last, feeling the effects of OODA-Lag. Data is coming at them faster than they can pay attention to it. All the bottlenecks in software engineering that were there prior to LLM-assisted code generation will persist long afterwards.

Now I hear many of you already, across the vast oceans of time and space, that the solution lies in Test-Driven-Development (TTD), so we’re going to deploy a swarm of agents to write TTD’s, and then they’ll write TTD’s to check the TTD’s. It’ll be TTD’s all the way down.

I also hear Good’s thesis in this: once a superintelligence self-improves itself recursively, assuming we can retain alignment perhaps, we’d all be left behind. It’s juice for the Gooder Agents meme. Maybe OODA-Lag is a key reason why AI isn’t yet building a better version of itself.

If you listen closely you can hear the ghosts of a million library scientists whisper something from beyond the grave. What’s that?

What’s that they’re saying?

Ontology?

Ontology they say?

If only there was a new bottle.

The New Bottles

New bottles are great because they of enable new mixes. A lot happens when the past is funnelled into the future. Lots of spills. Lots of new concoctions. New terms to cloud the tag cloud. Think of the synergies we’ll integrate holistically.

Some people have been giving their new agentic intelligences the authority to wipe out their hard drives. That’s exciting. It’s a tough lesson. Some people are giving them signing authority, the ability to make financial decisions which bind their humans to an obligation. And isn’t that interesting?

So the response to misjudgement are probably going to mirror the past. Engineers will engineer a system of surveillance and oversight designed to reduce the authority of agents to make irreversible decisions. They’ll create enormous agentic bureaucracies, where different LLM’s will be tasked with writing and reviewing different forms before other agents are allowed to go ahead and execute on a plan to fill in the form. The effect on OODA-Lag is predictable.

But you know what will fix OODA-Lag? More agents which will intermediate the disagreements between different parts of the synthetic bureaucracy, and alert a human that’s in the loop! The human will get an email. Yeah, an email, that they’ll have to reply to. Or maybe a voice memo. Or, maybe, wouldn’t it be cool if they host a standup meeting with the human, and each agent will go over where they’re blocked, and ask for help from the human. Or maybe the human will just be prompted to jiggle their mouse at it.

I’ll predict that in most places, the relative deflation in token cost will enable ever spiralling agents on agents on agents. It seems likely that in the short run, 90% of projects will end in fractal synthetic bureaucracies. This will be a recurrent pattern until people, or agents, learn of a better way.

So what would that better way be?

The New Bottle

Technology is an amplification of intent.

Intent is multi-dimensional monkey paw.

Intent, expressed in vague specifications riddled with contradictions and conflict, produce outcomes that are generally reflective of those specifications. One can improve the odds by writing and reflecting on better specifications. It won’t necessarily solve the alignment problem entirely, but it will make it easier to deduce, manage attention, and improve the system over time.

Better intent, better purpose, better outcomes.

Using AI to understand ones own intent could be a massive improvement. It might loosen the neck. That’ll make it better for some.

And that’s a new bottle.