Why Models aren't the Answer, Labs Are.
I'm reading the tea-leaves, but the water itself is pretty dirty.
This one is a long one, but a labour of love. I’ve thought of this for a while due to my background in econ, but it definitely needs some work. I’d love to work on this further and have split it yet again into a two-parts but thank you so much for reading and subscribing. It really does mean a lot that some of you spare a few minutes to entertain my shower thoughts.
I’m not so much a fan of the capitalistic economic framework, but it does have it’s perks. Cheap data packs, accessible transport, the opportunity for me to speak with you all at little to no costs and donuts? By now, our ancestors would have labelled us as witches and burned us at the stake!
But as we find ourselves at the infancy of the AI revolution, a question needs to be asked. How do we as humans manage to adapt to the coming age? What apriori beliefs do we discard or embrace?
The reason I ask this is just a really primal emotion. Fear. For nearly ~30,000 years the human has reigned supreme. The throne has remained occupied, undisturbed and ever-powerful. Though anthropologists, archeologists, geologist and every other field offer their own reasons, a concrete fact can be established of the reign of the homo sapiens. Intellectually, we were second to none.
However, November 20, 2022 was our Hiroshima. Humanity came to terms with the fact that intellectual dominance was now no more monopolized. AI was coming for us and so we must adapt. Earth now has two intellectually - blessed creatures. The fear of being replaced, of being irrelevant has now returned. Not on a personal level, but on the scale of a species. But let’s not get so ahead of ourselves on the possible extinction of mankind. That’s for another time. But this brain-fart has allowed me to shift gears to come to terms with a renewed focus on AI, this time with a focus not on the products, but frontier labs instead. And that’s where I think we need to zone in.
But why frontier labs? Don’t the models pose a bigger threat?
To answer this, let’s first take a detour through behavioral econ. In particular, the principal-agent problem. The principal-agent problem classically explains a misalignment in incentives between a principal (the one who delegates the task) and an agent (the one who does the task). Brought about by information asymmetry, shifting priorities or self-interest, an aligned output, as described by the principal-agent problem, is never feasible.
It is thus essential, that we, the broader society acknowledge our actions as that of an unwitting, aloof principal who discharged all model development, decision-making and deployment to the agents, frontier labs. And in an attempt to take back the reins, it becomes our moral imperative to identify, align and direct incentives towards the greater Good. To do this requires us to understand the frontier labs, their motivations, their incentives, their money-trail.
‘A race to AGI’ is what you’d say, isn’t it?
You are right to some extent, but as right as saying ‘the fastest one on the track is the winner’. It’s a fact, but it’s a trivial conclusion once you understand the basic rules of the game. We’ll we need to dig deeper. To better illustrate my point, I’ll steal one of my dad’s favorite sayings — Actions speak louder than words. I’m arguing to instead shift our objective to trace not the next iteration of the o-series that OpenAI has chosen to make public but to look at their direction. To connect the dots over a period of time to better understand model development.
This lovely article from Simon Willison tracks all models from 2024 (too bad there’s nothing for 25 yet!)
Now you would accuse me of reading tea leaves, or trying to make sense of an unruly tides and I’d definitely agree. I’m not one to usually follow how many hidden-layers & which activation function is commonly used in the latest models, but in my defense, I am a voracious consumer. ChatGPT is in my blood (not these articles though!). So it is only my responsibility to try and connect the dots from a consumer-centrist approach.
I’ve played around with Claude, Gemini and GPT. I’ve tried to read their detailed system cards and tried to experiment too. And though they all seem able to perform the same task to varying degrees of success, their evolution and sister models over time paint a broader picture of how their parents think. It’s evident that Anthropic, Google and OpenAI are betting on (in my opinion) three distinct ‘interaction modules’ — approaches which might lead them to greatest profitability, ensured by companies instilling a reflex action in you to open their model the first thing in the morning.
First, let’s understand the characteristics of these interaction modules in their relation to us.
These modules have an explicit objective would be to accelarate market capture. To capture a sizeable chunk of the market in order to either a) monetize them immediately or, b) to capture future earning potential.
The second more cruel implicit objective would be the ominous attempt of re-engineering thinking, i.e. to make free thought an afterthought. The idea would be to make models a sounding board where all your thinking is automated. And trust me, it isn’t that OpenAI or Anthropic are willingly doing this. It’s an unfortunate consequence in the pursuit of an omnipotent being.
These modules are highly fungible. Anthropic could be following A, but immediately switch to B with almost little to no switching costs (partially un-true, but let’s stick to it). In fact, we see Google adopting 2-3 approaches all at the same time! OpenAI in-fact adopted a new business model altogether about a year ago.
Remember, these characteristics aren’t exhaustive but just the tip of an iceberg. And no list will ever be. Because dear reader, you and I must acknowledge what I argued for above — AI isn’t just a complex pattern recognition module. It’s an attempt to give birth to another entity capable of cognition. So though we might laugh, criticize and at times feel surprised at the sheer capabilities of AI, we must not forget that these entities will sooner or later rival us.
But returning to the interaction modules as the ultimate climax of this long-drawn article, let’s zone in on the top-three for opted by foundational research now.
The God Model
The Personal Assistant
The Jack of All Trades
While a name change is surely required, let’s stick to this. Note that these aren’t business models in the conventional sense. The ‘Aha’ moment on these modules is only when you realise that the objective right now is market capture rather than monetization. So while the God model aims to impress you doing multiple things at the same time, the Jack of All Trades offers multiple models allowing you to cater to niches. The options are numerous and the payoff, equally high.
The modules allow for a flexibility not only in price, but quantity. So a few cracked engineers might pay a steep price for The God Model in Claude Opus or Sonnet (either through Claude Code, Cursor, API acess, whatever) but this allows for a rate-limited free-plan for everyday users like me. The modules essentially allow for the unlimited breadth of access but at a tradeoff on exposure. An engineer is definitely going to value Claude 10x higher than an average home-maker or student. But ChatGPT is going to be the default for this home-maker wife/husband and students to finish assignments or develop shopping lists. I think contextualizing this, will be key to understand the thinking behind foundational model development.
I think this article has far exceeded the length I preferred so let’s visit this next week. Thank you so much for sticking with me this far. I do hope you’ll return next week too. I’m really grateful that you choose to open my mail every week :)