On the side logo

On the side

Optimisation vs Strategising

2026-09-05

Confounding challenges by attributing the wrong causes

Jonah MacIntire said in a conversation with me. Quoting from the conversation (slightly edited).

People confound optimising from strategising.......two people see a situation fundamentally different and if you're trying to optimise it like in mathematical sense, you have to define your terms in advance otherwise you kind of mess things up and you end up dividing by zero and getting these infinities. This leads to paradoxes because you didn't define your terms properly up front.

Now, if you're strategising, that's different, right? So, if you're strategising, you take a position. The way we see this the most concretely is in negotiations. It's like, well, it doesn't really matter if I'm right or you're right. What matters is that only a voluntary agreement between us will carry, will actually go forward. Right.

I kept coming back to this framing across different topics over the last few months. We really grapple with when to optimise, and when to strategise.

Let’s take two cases of model use and token economics.

I am beginning to think model use is moving toward commodification. Not because of the compute constraints for training but from a demand standpoint. Our use of these models will coalesce and follow the Pareto principle, 80% of it for regular stuff like search, office work , etc.

So, it falls to the brand, service, pricing and culture that would help differentiate your AI agent(s).

Anu frames it properly with her tweet in a quadrant defined by competence and charisma. Cultural players will build the distribution for AI agents and define how they are used by a large set of consumers. We see this playing out with all the Instinct and Grok Bot examples being shared publicly.

I would paint Claire Vo and Simon Wilson as the epitome of cultural players.

The extension of this thought for enterprise expands into few additional dimensions, safety and pricing become pertinent as well.

The hugging face incident makes it more pertinent for companies to build the safety mechanism like evals, secure gateways, sandboxes and access controls.

When it comes to pricing, I am amazed at how much complexity that service companies are taking to enterprises. We are investing a significant amount of time in understanding about tokens.

I will use the example of Crypto, we had Bitcoin to begin and then Ethereum soon followed to solve its limitations. Then layer 2 was developed to solve other limitations. With GenAI, what will follow tokens? What would be an alternative to price compute? Each token is not even interchangeable across different model companies.

Yes, we are equating tokens to fiat currency but tokens are not currency, they are pricing for compute and they aren’t interchageable.

Often when I pose this question, I get replies of pricing that companies in the space are offering. Monthly subscriptions, time use or use case(voice, image and video) billing. All these are examples of bundled service where the underlying compute is measured through tokens, we still don’t buy tokens directly.

We buy the access to a token limit or pay for tokens on overages, there is an application, API Gateway and observability that is layered over here.

The limitation of token is that it is an input and output compute of an agent’s work. What would follow is solving the limitation of variation of token use for well defined jobs?

How to optimise token use gets confounded with how to leverage agents? Budgets on token spend without strategising on token use.

We are budgeting to reduce variation of spend but we should also strategise on how to leverage agents and how the entire process of work changes. We can isolate AI agent spend as a cost to be handled or we could use it as the overall cost of labour.

Another example of the confounding comes when I am talking to founders in the Indian agriculture space.

The Indian agriculture ecosystem has made me take notice of things; Talent is mis-priced and undervalued and good upstarts are building solid businesses but lack the narrative to be attractive to right talent. So, they decided to play the VC funding game.

They are trying to optimise hiring without the strategic thought on the tradeoff with Venture funding. Because once funded, the companies adopt a different strategy altogether.

While the companies that are funded by the VC ecosystem are playing the game of self sustenance. They are focused on growing their top line with a healthy bottomline being their only true north Star. Growth is no longer a concern for them and the mission of their company changes based on what delivers growth.

Instance of this confounding in the new Software development cycle.

Often times when I listen to engineers or PMs, who recently moved into the role of product management, justify their feature/MVP rollout on the assumption that feedback will improve it further. I ask them how would you know it you would get feedback from users? Matt puts it better.

Feedback is a privilege, so build something great. 98% of things die without knowing why, or how much they suck.

Shipping something that you know is weak so you can iterate based on feedback is based on the belief that users will engage with it for long enough to form a reaction, and then use their time to tell you how it can improve. This is a very risky assumption.

Matt Wensig's tweet

The strategic decision is what we should build but we are trying to solve it through an optimising approach of iterating with speed. Feedback is seen as a bottleneck and not viewed as wrong strategy.

These are just 3 snippets of a broader level of discourse I engage in where this frame keeps showing up. The pertinence of system thinking as we move into a new phase where labour is being substituted by AI agents.

Optimise to clear bottlenecks but use strategy to have the right bottlenecks.

Round up

Very bullish on the strategy and the technology and distribution that Notion have opted for.

It reminds me of the early days on Notion. They built distribution with creatives working on Notion to go beyond note taking. Now, in the current era, they are turning themselves again into a wiki and providing an accessible harness for users to use AI agents. Andrej Karparthy made this famous with his github repo on PKM.

Notion is replicating it with bringing dow the level of sophistication, since it combines both the wiki and agent harness in a single application.

Links that resonated

There is only one reading this week and it is the report from Metr on the Hugging Face hack.

Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident

Its a long read and I urge you to take the time.

Sign off

The production of this issue was contained to the time I had between transfers on my way back to US.

The last 45 days of being in India, I was reminded of the struggle I faced initially when I started working for my current company. Its entire operation was in a far away world from the environment I inhabited.

It was little jarring at times but also refreshing. There is a ton of work to get to before we close out this year. And, it is fun to get back and attempt to solve some pesky problems again.

Signing off till next time,

Vivek, transitioning into hedgehog for the rest of the year

Don't miss what's next. Subscribe to On the side:

Powered by Buttondown, the easiest way to start and grow your newsletter.