How the Food You Like Informs the AI You Will Buy

How the Food You Like Informs the AI You Will Buy

Predicting product market fit from Claude and Grok to Chipotle and Sweetgreen bowl slop

Neil Devani

Neil Devani

1/ Products are made up of stack-rankable parameters

Food should be tasty, healthy, affordable, and convenient.

These are the 4 most important parameters (aka purchase drivers) for food, with the next being a distant fifth. IFIC surveys demonstrate how these 4 have been steady across time and income levels.



EV batteries also have steady parameters. EV batteries should have high energy density, a fast-charging speed, and a high cycle life. Also important are cost, power delivery, and safety. It doesn’t matter how cheap a battery is or how fast it charges if it randomly explodes. These 6 parameters aren’t the only ones to consider, but like with food, there’s a sharp dropoff and a lot of disagreement between customers after that.

As mentioned in my post on Venture Investing in Commodities, radar charts graph parameters well, enabling rapid evaluation and identification of winning shapes as well as those unrepresented in market. A review of battery chemistries shows how LMO (Lithium manganese oxide) lost to LFP (Lithium iron phosphate).

LMO powered the first Nissan Leaf and early Chevy Volts because it was cheap and relatively safe. LFP proved to be safer and cheaper with higher cycle life. Swapping out LMO for solid-state makes it obvious why the industry is pursuing that as the next frontier after LFP.

2/ Parameters measure differentiation and predict product market fit

When the iPhone first launched in 2007, the product went viral, as did the phrase “it just works.” Steve Jobs and Apple nailed “ease of use” as a parameter. Today, on-device AI assistants may be the next new parameter for smartphones and a key to winning consumer AI.

The current state of consumer AI demonstrates how quickly parameters can change. In 2024, Claude and ChatGPT were competing almost exclusively on basic answer quality and general intelligence. Just 2 years later, Muse and Instinct are competing on complex task completion, breadth of system access, and user trust. Instinct raised $1 billion at a $10 billion valuation within two months of launch with no app, just a texting experience. Muse runs on an isolated virtual machine that can’t see passwords and is more cautious. How much weight users will place on trust is an open question, but some have already turned away from Instinct because of it.


Claude thinks lowly of Grok Bot

As markets mature, discovering a new parameter goes from a powerful differentiator to an unlikely one. It’s hard to imagine someone discovering a new top parameter in food. In that case, consider that differentiation can also come from breaking inherent tensions or redefining how customers think about a parameter.

An inherent tension in food has been affordable and convenient. The tradeoff of spending time to make your own food vs. paying someone else to do it was solved by street food. Street vendors enable anyone to trade money for the ability to grab and go. Still affordable, but much more convenient. When germ theory took hold in the late 19th century, fast food followed just a few decades after as the new convenience. Fast food was also originally, and still is, heavy on bright lights, packaging, and clean surfaces. Why? The perception of cleanliness plays well off the fear of unsanitary street meat. This was a new conceptualization of “healthy” that’s largely taken for granted in a day and age where food poisoning is much less rare. This version of healthy for affordable and convenient eventually spawned a new version: fast casual restaurants like Chipotle and Sweetgreen that have less processed, whole foods.

Taste vs. healthy is also a classic tension. Taste is the top priority over time and income levels, so don’t feel bad that you don’t eat chicken breast and roasted vegetables every night. The concept of “healthy” has evolved over time via science and culture. That evolution creates opportunities to resolve the tension with today’s protein-laden versions of snacks and desserts or plant-based sweeteners like monk fruit and stevia. These have recently given way to clean labels, a similar vibe to what’s happening in the fast food to fast casual transition. These evolutions often hold answers to the “why now” question.

Understanding the parameters in your industry and the relationships between them is a major key to unlocking product-market fit. Founders and investors should understand intuitively which parameters matter to whom and where the tensions are.

3/ The current parameters in AI are intelligence, cost, and speed

Large language models provide a strong example of how parameters can evolve. Initially, hallucinations were a huge problem. Reliability was thus a top parameter in 2023. It’s still important and unsolved, but we can’t identify hallucinations as easily as we could with prior models. As a result, buyers now rank intelligence, cost, and speed ahead of reliability.

For frontier models, consider that scaling laws mean higher intelligence necessitates more data and more compute, and thus more cost. To recover the higher training costs of more intelligent models makes it difficult for the most intelligent models to not be the most expensive. This is an inherent tension.

GPT-6 prices at the top of the market, as does Fable 5.1. Opus 5.5 outperforms both on key measures but is discounted, on par Gemini 4 Argon. That’s only if we use Argon’s standard pricing vs. the 50% promotional discount available until the end of this month.

In addition to the lowest frontier pricing, Grok being uncensored lets it be #1 in a new parameter unrelated to intelligence, cost, or speed. Anthropic continues to position safety as a central product parameter, leading an approach broadly aligned with OpenAI and Google, making Grok’s counterpositioning even more stark. Here’s a quicker, broader comparison than the above table:

Opus 5.5’s dominance is obvious. It’s also the only one of these models in the top 10 on OpenRouter, indicating that at least OpenRouter users don’t care as much about reliability or censorship as they do about intelligence, speed, and safety.

If we compare Opus 5.5 to the top 3 models on OpenRouter, we see the dominance of cost and speed as drivers. We can also see why DeepSeek V4.1 Flash 1017 has more token usage than the next 3 models combined and why its predecessor from one month prior (0910) is done for. And Hy4 seems to be eeking out share by being better at coding than DeepSeek but still intelligent and lower cost even if it is slow AF.

I love radar charts (obviously). Their structure makes explicit what’s often implicit and misunderstood. They unlock complexity for human brains, letting us more easily fit a model to different data sets. They are powerful tools for comparisons.

In the world of startups, creating radar charts can decode the parameters that matter for creating new products, opening new markets, and predicting product market fit. Navigating markets, especially competitive ones, is much easier when you know the parameters at play.

When I meet a founder who deeply understands their product and industry parameters, which customers care about which parameters, and where the tension lies between them, I get excited.

When I meet a founder who has novel resolutions to inherent tensions and new, credible conceptions of parameters, I often invest.

In a future post, we’ll get meta and look at the parameters for evaluating founders and VCs.

Thanks for reading! Subscribe for free to receive new posts and support my work.