In today’s free post, I look at why LLMs will persist in spite of their flaws.
If you like what you read here and you’re not yet subscribed, consider subscribing. Most of my posts are free though deeper dives tend to be paid.
The case against LLMs is easy to make and even easier to posture around. They hallucinate. They forget things. They can’t reason. They fail at math, logic, planning, and anything that looks like actual cognition. They consume electricity like a mukbang consumes food, and they require an industrial-scale capital stack to deploy meaningfully.
And yet: they will persist. That’s not optimism. It’s just path dependence.
The Core Critique
Let’s steelman the strongest arguments against LLMs as a long-term substrate for AI:
They don’t reason. LLMs don’t have a world model. They don’t bind variables. They don’t reason over symbolic structures. What they produce is fluency, not knowledge. Which is why they hallucinate, contradict themselves, and routinely fail tasks that any halfway competent symbolic system could handle.
They don’t remember. Even with a 100,000-token context window, LLMs are still high-bandwidth amnesiacs. They don’t track narrative arcs or causal chains. They can’t meaningfully reflect or re-evaluate across time without external scaffolding.
They don’t plan. These are not agents. They don’t have goals, memory, or agency. Every time you invoke an LLM, you’re getting a brand new, stateless entity. Agentic behavior must be bolted on from the outside using hard-coded loops, vector stores, and function calls.
They don’t align. You can’t really steer them. RLHF is a kludge. Guardrails are porous. Prompt injection and jailbreaks are trivial. Interpretability remains a speculative frontier. If you're trying to build something you can trust, LLMs require duct tape, prayers, and constant monitoring.
They don’t scale efficiently. They’re compute hogs. Training is expensive. Inference is expensive. The economics only work for a few hyperscale players sitting on energy subsidies and GPU monopolies. Everyone else is underwater.
So yes, LLMs are structurally, economicaly, and cognitively flawed.
And yet like mushrooms and cockroaches they will persist.
Why the Critique Misses the Point
Technology is not a meritocracy of ideas. It is an accretion of capital, tooling, and lock-in. LLMs may be suboptimal, but their limitations are no longer the relevant point. It is more helpful to ask: Are they entrenched enough to persist, and evolve into something better?
And on this question, the answer is emphatically yes. Let’s walk through the structural reasons why.
1. Path Dependence Is Everything
History is not full of optimal technologies. It’s full of dominant ones. x86, JavaScript, SQL, Excel, TCP/IP: none of these are ideal. They are durable. Why? Because everything from training pipelines to developer tooling to regulatory frameworks to educational curricula congeals around early winners.
LLMs are now an early winner. The flywheel has started: model weight distribution, ecosystem investment, enterprise deployment, retraining cycles, API integration, fine-tuning frameworks, prompt engineering playbooks. You don’t throw this away. You build on it.
2. Capital Formation Is Destiny
Once tens of billions have been deployed into an architectural paradigm, it gains the kind of gravitational pull that no better idea can easily escape. Entire venture funds, chip fabs, cloud stacks, and workforce upskilling programs are being built around LLMs. That capital is not patient enough to wait for a better architecture. It will evolve the current one until it works.
3. Hybridization Beats Replacement
LLMs won’t be replaced. They will be absorbed into larger hybrid stacks:
LLM + retrieval-augmented generation (RAG)
LLM + tool-use (function calling, code exec)
LLM + memory (episodic + long-term state)
LLM + symbolic reasoning (external solvers, planners, agents)
LLM + multimodality (vision, speech, video, embodiment)
The result isn’t a pure architecture. It’s a kludge. But a powerful one. Just like the modern browser. Just like the JVM. Just like Kubernetes. Clean design dies in the trenches of economic reality.
4. Better Architectures Lack a Theory of Power
It’s easy to say, “we need something better than LLMs.” But what exactly? Symbolic reasoning? It’s brittle and unscalable. Neurosymbolic systems? Promising, but early. Embodied cognition? Great, once you solve robotics and perception at scale.
The alternatives aren’t obviously better. They’re just differently broken, and far less mature. Meanwhile, LLMs continue to improve. Context windows are expanding. Tool use is stabilizing. Memory integration is underway. Scaffolding is becoming systematic. Alignment remains hard, but nothing else is easier.
So no, there is no obvious replacement. There is only accumulation and adaptation.
5. The Game Is Already Rewriting Itself Around LLMs
Whatever else you think of them, LLMs are already rewiring workflows in content, marketing, customer support, software engineering, and legal. Not always well. Not always efficiently. But persistently.
And when workflows change, everything downstream, including hiring, training, vendor selection, and compensation models, shifts with it. That’s the real moat. Not just the model weights, but the org charts and decision trees that quietly rearrange themselves around the new substrate.
Final Thought
The right critique of LLMs is not that they're flawed. Of course they’re flawed. The right critique is: How far can we push this flawed paradigm before it breaks, and how much can we graft onto it before it becomes something else entirely?
The next wave of intelligence won’t come from a clean-slate revolution. It’ll come from a layered mess, where flawed components become functional through relentless stacking, scaffolding, and augmentation.
LLMs are not optimal. But they are embedded. And that makes them, for now, inevitable.
Coda
If you enjoy this newsletter, consider sharing it with a colleague.
Most posts are public. Some are paywalled.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
Thank you for stating the critiques so well.
Sorry to disagree with your conclusions.
If it were true that a critical mass of real-life workflows had moved to LLM dependence, you would be right. But it’s not true. The very flaws you identified make it not so.
There’s a lot of hype. A lot of noise. But name any business process at scale in the world that has converted. There ain’t any. Not even the ones supposedly in the sweet spot of these models, like coding.
Hadoop. The Metaverse. Yesterday’s false paradigm shifts. It happens that sometimes it doesn’t happen.