A couple of new models have come out recently from Claude (Sonnet 4.8 and Fable), alongside upgrades from other major AI companies (ChatGPT has gotten much better at research). They are impressive. They are faster, better at reasoning, stronger at coding, and more reliable across a range of tasks. But it increasingly feels like the era of a new model arriving and completely changing our sense of what is possible may be behind us.
For a while, every major release felt like a step change. A new model would appear and suddenly entire categories of work seemed newly automatable. The boundary of what AI could do moved dramatically. But expectations have shifted very quickly.
I saw a joke on LinkedIn recently about a 16-year-old daughter coming to her dad and saying, 'There is no way people used to write 10-page book reports by hand, without computers and without AI. There is just no way people did that.' And the dad says, 'Honey, I wrote 10-page book reports by hand without even reading the book.'
That joke landed because it captures something true about human adaptability. We have always found ways to produce work under imperfect conditions, to improvise, to cut corners, to rely on judgment when the tools available were limited. The arrival of AI has not fundamentally changed that pattern. What it has done is shift which tasks are easy and which still require genuine effort. The shortcuts have moved, but the behaviour has not.
Now, most new model releases feel more marginal (for most tasks). They are better, often meaningfully better, but not in a way that changes everything for everyone. The improvements matter, but they are increasingly specific. In particular areas, they can still feel game-changing. But for most users, most of the time, the difference between one frontier model and the next is becoming less dramatic. This is especially important to understand because the people building these models are, overwhelmingly, software developers. And when a model becomes dramatically better at software development, it can feel to them like the entire world has changed.
For a while, every major release felt like a step change. A new model would appear and suddenly entire categories of work seemed newly automatable. The boundary of what AI could do moved dramatically. But expectations have shifted very quickly.
I saw a joke on LinkedIn recently about a 16-year-old daughter coming to her dad and saying, 'There is no way people used to write 10-page book reports by hand, without computers and without AI. There is just no way people did that.' And the dad says, 'Honey, I wrote 10-page book reports by hand without even reading the book.'
That joke landed because it captures something true about human adaptability. We have always found ways to produce work under imperfect conditions, to improvise, to cut corners, to rely on judgment when the tools available were limited. The arrival of AI has not fundamentally changed that pattern. What it has done is shift which tasks are easy and which still require genuine effort. The shortcuts have moved, but the behaviour has not.
Now, most new model releases feel more marginal (for most tasks). They are better, often meaningfully better, but not in a way that changes everything for everyone. The improvements matter, but they are increasingly specific. In particular areas, they can still feel game-changing. But for most users, most of the time, the difference between one frontier model and the next is becoming less dramatic. This is especially important to understand because the people building these models are, overwhelmingly, software developers. And when a model becomes dramatically better at software development, it can feel to them like the entire world has changed.
But software development is not the whole world.
Developers spend much of their time solving logical puzzles. They translate requirements into systems, debug problems, structure code, and reason through constraints. Current AI models are exceptionally good at that kind of work. They can write complex code, solve difficult mathematical problems, and reason through abstract logic in ways that are genuinely impressive.
But real estate is different.
In real estate, we spend our days making decisions based on judgment, experience, intuition, training, education, market context, and conversations with other people. We are not just solving logical puzzles. We are interpreting ambiguous information. We are weighing incentives. We are judging whether something feels right or wrong based on experience that is often hard to write down.
And this is where AI still struggles.
The models may be extremely intelligent in the narrow sense. They can do maths that most of us cannot do. They can solve technical problems that would take a human hours or days. They can power software development at extraordinary speed.
But they still have poor commercial judgment.
They do not naturally understand why valuations may be low in a particular market at a particular moment. They do not know whether the issue is on the buy side or the sell side. They do not know whether brokers are struggling to match buyers and sellers. They do not know whether there is market uncertainty that everyone is quietly discussing at a conference but that has not yet appeared clearly in the data.
They also do not understand your portfolio unless you explain it. They do not know how you like to build your DCFs unless you teach them. They do not know which assumptions you trust, which ones you usually challenge, and which outputs you consider unrealistic.
This is why the biggest improvements in AI output will not necessarily come from smarter models. The models are already smart enough for many tasks. In raw reasoning ability, they are already beyond what most people can do in many technical domains. The bigger gap is not intelligence. The bigger gap is context.
But real estate is different.
In real estate, we spend our days making decisions based on judgment, experience, intuition, training, education, market context, and conversations with other people. We are not just solving logical puzzles. We are interpreting ambiguous information. We are weighing incentives. We are judging whether something feels right or wrong based on experience that is often hard to write down.
And this is where AI still struggles.
The models may be extremely intelligent in the narrow sense. They can do maths that most of us cannot do. They can solve technical problems that would take a human hours or days. They can power software development at extraordinary speed.
But they still have poor commercial judgment.
They do not naturally understand why valuations may be low in a particular market at a particular moment. They do not know whether the issue is on the buy side or the sell side. They do not know whether brokers are struggling to match buyers and sellers. They do not know whether there is market uncertainty that everyone is quietly discussing at a conference but that has not yet appeared clearly in the data.
They also do not understand your portfolio unless you explain it. They do not know how you like to build your DCFs unless you teach them. They do not know which assumptions you trust, which ones you usually challenge, and which outputs you consider unrealistic.
This is why the biggest improvements in AI output will not necessarily come from smarter models. The models are already smart enough for many tasks. In raw reasoning ability, they are already beyond what most people can do in many technical domains. The bigger gap is not intelligence. The bigger gap is context.
"New" models seem to mostly emulate good human users
The recent improvements in AI models also seem to reflect this. Many of the upgrades are not necessarily about entirely new technological breakthroughs. They are about getting models to do more of what humans were previously doing around them.
Humans were adding reasoning steps. Humans were checking outputs. Humans were keeping logs of what had been done and what still needed to be done. Humans were applying quality control. Humans were running sanity checks before trusting the final answer.
Increasingly, AI is doing more of that work itself.
That is valuable. It makes the models more useful and more reliable. But it also means that the cost of using AI can increase quickly. The model is not just producing an answer. It is reasoning, checking, verifying, correcting, and sometimes re-running its own work. All of that uses compute. All of that costs tokens.
So yes, the models are getting better. But a lot of the improvement comes from delegating more of the surrounding workflow to AI, rather than from a fundamental change in what AI understands.
And what AI still does not understand, unless we explain it, is commercial context.
That is the central point for real estate.
The way to get more value from AI is not to wait for the next model release and hope that it magically understands your business. The way to get more value is to understand your own work more clearly, define your own context more explicitly, and teach AI how you want it to operate.
What matters is not just which model you use. What matters is what you tell it.
What is your investment strategy? How do you think about risk? What assumptions do you usually make? What does your portfolio look like? What kind of outputs do you trust? What mistakes do you want the model to avoid? What market context should it know before it starts producing analysis?
The more clearly you can explain those things, the more useful AI becomes.
This is also why companies are increasingly concerned about vendor lock-in. More firms we speak to are nervous about committing too heavily to one provider, whether that is OpenAI, Anthropic, or anyone else. They are also increasingly focused on cost, which links closely to Thomas's post in this month's newsletter.
Humans were adding reasoning steps. Humans were checking outputs. Humans were keeping logs of what had been done and what still needed to be done. Humans were applying quality control. Humans were running sanity checks before trusting the final answer.
Increasingly, AI is doing more of that work itself.
That is valuable. It makes the models more useful and more reliable. But it also means that the cost of using AI can increase quickly. The model is not just producing an answer. It is reasoning, checking, verifying, correcting, and sometimes re-running its own work. All of that uses compute. All of that costs tokens.
So yes, the models are getting better. But a lot of the improvement comes from delegating more of the surrounding workflow to AI, rather than from a fundamental change in what AI understands.
And what AI still does not understand, unless we explain it, is commercial context.
That is the central point for real estate.
The way to get more value from AI is not to wait for the next model release and hope that it magically understands your business. The way to get more value is to understand your own work more clearly, define your own context more explicitly, and teach AI how you want it to operate.
What matters is not just which model you use. What matters is what you tell it.
What is your investment strategy? How do you think about risk? What assumptions do you usually make? What does your portfolio look like? What kind of outputs do you trust? What mistakes do you want the model to avoid? What market context should it know before it starts producing analysis?
The more clearly you can explain those things, the more useful AI becomes.
This is also why companies are increasingly concerned about vendor lock-in. More firms we speak to are nervous about committing too heavily to one provider, whether that is OpenAI, Anthropic, or anyone else. They are also increasingly focused on cost, which links closely to Thomas's post in this month's newsletter.
Local models start making sense
At VARi, we are experimenting more and more with running open-source models locally for specific tasks. These models are now extremely capable. They may not always match the very best frontier models across every possible benchmark, but for targeted workflows they can be more than good enough.
They also offer something many companies find comforting: control.
Control over privacy. Control over cost. Control over performance. Control over infrastructure. And, importantly, a degree of durability. Once a local model is set up for a specific task, it can continue performing that task without being exposed to the same uncertainty around pricing, model changes, or vendor strategy.
That does not mean every company should immediately move everything to open-source models. Frontier models remain incredibly powerful, and for many use cases they are still the right choice.
But it does mean that the conversation is changing.
The question is no longer simply: 'Which model is the smartest?'
The better question is: 'Which model, with which context, in which workflow, under which constraints, produces the best result for this specific task?'
That is a much more useful way to think about AI in real estate.
So my conclusion this month is simple: do not get too distracted by every new model release. The progress is real and new models will continue to unlock new possibilities. But in real estate, the biggest gains will not come from treating AI as a mysterious intelligence that somehow knows your business. They will come from making your own judgment, process, assumptions, and context explicit enough that AI can work with them. The next breakthrough will not just be a smarter model. It will be a better explanation of what we already know.
They also offer something many companies find comforting: control.
Control over privacy. Control over cost. Control over performance. Control over infrastructure. And, importantly, a degree of durability. Once a local model is set up for a specific task, it can continue performing that task without being exposed to the same uncertainty around pricing, model changes, or vendor strategy.
That does not mean every company should immediately move everything to open-source models. Frontier models remain incredibly powerful, and for many use cases they are still the right choice.
But it does mean that the conversation is changing.
The question is no longer simply: 'Which model is the smartest?'
The better question is: 'Which model, with which context, in which workflow, under which constraints, produces the best result for this specific task?'
That is a much more useful way to think about AI in real estate.
So my conclusion this month is simple: do not get too distracted by every new model release. The progress is real and new models will continue to unlock new possibilities. But in real estate, the biggest gains will not come from treating AI as a mysterious intelligence that somehow knows your business. They will come from making your own judgment, process, assumptions, and context explicit enough that AI can work with them. The next breakthrough will not just be a smarter model. It will be a better explanation of what we already know.
Created with
