The State of AI - Models, Economics, and the Builder's Dilemma

September 9, 2026

As newer AI is making things easier and easier to make applications, there seems to be revived value added to testing, where the pressure of applications needing to be tested has been a value proposition. Having things like GPT-6 (Astra) especially being able to extract value on this makes this much better.

I am expecting other companies to work on similar models, especially low-latency, small, very fast models like Gemini or even much smaller active parameter models to speed up models for the same use case might be better.

Also, previous to Astra, there was just not enough high quality data to train on for making computer use reliable for RL . With Astra, this is changing, and I will not be surprised if others are able to distill its capabilities quickly, as computer use should not rely more on Chain of Thought (reasoning) but more so reliant on output, which makes smaller model refinement better.

Monolithic vs. Specialized Models

The current take on LLMs has been not to split models, but if an Astra sized model is required for computer use, that's not as valuable as having a much smaller model being able to do computer use with great reasoning much better. So I do expect, again, a split for models doing different tasks to come back. Especially models which are reasoning focused to break and find ideas being tuned, and coding focused and tool focused should make the entire field much better.

Current economic alignment with model providers doesn't incentivize this, as still a lot of subsidy is there. When this needs to be pruned, I expect model base training and Reinforcement Learning would be done, then capabilityspecific RL would be there especially splitting models to:

Prompting (orchestrator and getting or trying new ideas),
Coding (generally most apps and tasks can be done with code, including robotics though a different need of a world model is required, so not as much),
Interaction (being able to do tasks on a computer without [MCP by using a computer with a UI) and test as well as use apps like human using ui to do things.

This would be split, I would expect, especially as this can help in reducing the cost of running models.

For example, having a Luna or Sonnet model RL'd for all these tasks, and the general talking model becomes Sol or Opus, would fit my view currently.

While we are currently in financial abundance, especially with subscriptions on Codex or Claude Code having much higher limits, I generally expect this to not be so much the case over time. Also, other companies might split and solely try to focus on one part of this task. Though big models which can do all things are still very valuable, a large number of tasks can be offloaded that still is my current perspective.

Especially because we are seeing more and more inference and more compute being poured into inference instead of training with an even larger ratio, this option, on top of hardware optimized for inference (or even models' hot path weights baked into accelerators), is something my thoughts align with.

Economic Alignment and Human Accountability

The economics of new things and capabilities has changed a large part of the industry, and as OpenAI stated, economically valuable work is extracted. Though this might not mean all is lost, as there need to be people still who can be held accountable and check in on details.

Requirements of such people, and how much they can review, and the scope of tasks is increased while quantity is decreased, providing a two-way bleed where other economists' warnings of companies not having any money from consumers going into a cycle becomes more true.

We are already expecting to see CEOs and business owners feeling like it's a dream come true, as AI agents can work for cheaper (provided in current plans on subsidy) 24/7 with very little pushback on rights and the complexity of hiring humans.

But as more and more people are replaced, I would expect we would see major deflation on the rise as people don't have money to spend, causing a debacle where essential parts of livelihood go through inflation while non-necessities go to deflation (my personal take on how the current economy is moving).

As of September 9th, 2026, in Nepal and in India, but also throughout the world, we are noticing major inflation in many things, including:

Food (caused by multiple ongoing wars),
Hardware (caused by AI inference and the overall AI rush),
Basic amenities are currently increasing due to crude price jumps and so many things reliant on natural gas and crude (disrupted by war).

Builder Burnout and AI Fatigue

Though AI has had many advancements in recent times, it also has caused many people to face more and more pressure. As many people realize they can build what they want and giving freedom and fate's knot to be given into their hands: if you fail, then it's your fault, as tools to do anything (almost) are available, it seems.

And this is making builders have the idea that the current time is the best time to make, as in the future we might not have such cheap AI, as competition or an AI bubble crash might cause it to be more expensive, while yesterday's models were weaker. Especially as many are pointing out, this is a digital gold rush, and most are realizing it really is a gold rush and try to capitalize, but forgetting gold only holds value cause it's limited.

Economics of Builders & The SaaS Sell-off

As there are more and more people building things, products, and ideas, less and less people are buying things, leaving builders with no customers. And this seems to be the case especially where people are selling advanced AI tools created using AI, as others just try to clone what parts are required as a minimum for each person, leaving barely any room for anyone to buy any digital product.

This is also portrayed in many memes and reels, and shown as technically illiterate people asking to create in-house Claude AI or AWS. But this is a valid point which technical people point to the impossibility for many things, but many smaller apps which rely on you not being able to create a complex production-ready app for all are becoming more and more viable.

Even for a 100-person company, for their basic requirement, paying for an app with a 100-user license vs. creating and patching a smaller in-house app in 1 month is seen as currently economical. This is the main reason for the major SaaS sell-off earlier this year.

And this not only seems like a fair point, but it is becoming easier and tools are getting better. While later this might require 1 person to create while a 100-person workforce shrinks to 20–30, simplifying at each level causes the next round to shrink the company, and there is barely any idea which is able to save this.