[DRAFT] Everything Is a Context Problem
Harrison found me at noon under the beautiful Florida sun, slumping and sulking through dozens of HelpScout threads. Enumerated in the way a studying PhD would escalate the mortifications of their apprenticeship before finally dissolving in front of the thesis board. I was grumpy. Was it because I had just finished clearing the inbox for the seventh day in a row? Of course it was.
We already tried the obvious. Shut down inbox on weekends. Expand our documentation for customers. Tried the AI support bots others were hawking. On paper, these were all more than reasonable. Real companies, real technology.
None of them worked.
The bots weren't dumb. They just didn't know anything. They read our in-house notes and guessed at the answer. A customer would write in asking about if their July payment was recorded. The rented product would send back a generic article about our billing practices. The same one it sent to everyone. It didn't know who the customer was. Nor their history with our company. Nor the specifics of their billing setup. A simple:
if question = billing, send billing.md
Good article. But at the wrong time. Which means it was the wrong article. A few more rounds of this, and I wanted to leave the labyrinth.
JetBlue dot com, here we go: prepare to be bounced around for an hour or two, without an answer in sight. First it asks for confirmation codes. Then what the issue is. Then instead of cancelling the flight, sends a link on how I can use the bot to cancel a flight. How is this possible? The LLM models are better than ever. They should see my issues and solve the problem. But no matter what I try, I always request to speak with a person.
Companies seem to insist on shelling out millions of dollars for these AI solutions. Generic products that nervously avoid real problems. Frustrating customers and somehow creating more low-value work for the employees. How did we get here?
§When Genius Models Fail
We are way past the excuse of intelligence. These models appear to have performance comparable to a PhD. 91% on GPQA Diamond, against a 65% human-expert baseline. They are solving open math and computer science problems. Industry-ignorant brainiacs. Do you feel like ChatGPT or Claude Cowork is a graduate student? I certainly don't. Brainpower can't be the problem.
We improved our own support metrics, reducing resolution time by 88% and replies to resolve by 46%. All from a custom solution. Yet, we struggled for over a year to correctly identify how to guide AI to work.
These are organizational failures.
An LLM is a PhD hire on their first day in the field. Brilliant. Can spin a thesis with the best of them. Pointed in the wrong direction, they will hurt your business. Not because they can't help, but because you're asking them to solve PhD-level problems with pre-school level tooling.
Context and tooling become the new bottleneck.
I've been tracking where our tools fail. With over 2,000 historical errors to choose from since March across 57,000 actions, 73% are through one incorrecly calling our database [1].
The models can't own these failures- they lie with me. For neglecting to provide the proper context. Same flop JetBlue's bots are making, at a different altitude.
§A Model for Value
If context is the scarce input, the question becomes who ends up owning it.
AI capabilities are commoditizing and converging. It is not clear that any Frontier Lab will have the right to snare significant value going forward:
Domain, system, and individual knowledge is boutique. Broad intuitions can only travel so far. Universal tools lose usefulness the farther they travel from your articles of incorporation. Real value must be secured at the company level. Institutional knowledge, compounding and reinforcing itself throughout a company.
Generic tooling won't work. Similar to asking somebody in another PhD program what you should research. Without background on your specific business model and challenges, you'll get a middling answer. The alternative is bespoke — scaffolding tailored to you and the terrible problem you wish to conquer. Indexing to intelligence instead of context is a failure-mode.
The assumption that these models are good enough to just "figure it out" is a poor one. Integration costs will collapse once the model crosses a threshold. Yet they will never touch zero. You will have to build custom harnesses. Doing it now compounds institutional knowledge and earned secrets. Incorporating them into your scaffolding will amplify over time. And the kicker is, as the models get better, so will the harness.
As the models graduate after defending their thesis, context will always remain an issue. Super intelligence may be good at solving problems. ASI will look like a swarm brute-forcing to create context. Understand what the issue is, and tackle it with vigor. Yet, will always be slower than knowing the problem to point at from the beginning. Even Skynet will need onboarding.
Footnotes: [1] LLMs can provide you with reams of data on what they are doing. This allows for faster feedback cycles leading to exponential improvement. Reviewing agent threads and doing basic blocking and tackling will lead to impressive gains in a short period of time.