Aigora Blog

Heaven and Earth

John Ennis — Wed, 13 May 2026 15:03:04 GMT

"There are more things in heaven and earth, Horatio, than are dreamt of in your philosophy."

Hamlet says this to his friend after seeing a ghost. Horatio is a scholar. He has a framework for understanding the world, and it is a good one, but it does not cover everything. The ghost is real and Horatio's philosophy has no room for it.

I think about this line a lot when I listen to AI leaders talk about the future of work.

The people building these systems are, overwhelmingly, coders. They live in a world of text, logic, and digital computation. When they look at the economy and ask "what can AI replace?", they see the economy through the lens of their own experience. They see tasks that look like coding: information processing, document generation, data analysis, pattern recognition. And from where they sit, it does look like AI can do most of it.

They're right about the part they can see. They're missing the rest. The coder's view of the economy mistakes the slice visible from a screen for the whole thing.

Dario Amodei, the CEO of Anthropic (the company that makes Claude, the AI I use daily) published an essay called "Machines of Loving Grace" that imagines AI solving biology, curing disease, lifting the developing world, and compressing a century of progress into a decade. It is thoughtful and worth reading. But there is a conceit underneath it that I find troubling: the implicit belief that intelligence is computation, and that enough computation can therefore solve everything. Taken to its logical end, this is an attempt to remake God in the image of an engineer.

Here is something computation cannot do.

You walk into a house and smell cinnamon and old wood. Before you form a single conscious thought, you are six years old, standing in your grandmother's kitchen on a Saturday morning. The feeling arrives whole: warmth, safety, loss, love, the specific quality of winter light through a window you haven't seen in thirty years. You didn't choose to retrieve this memory. It grabbed you. It came through your nose and went straight to a part of your brain that doesn't trade in logic or language.

No language model will ever have that experience. Not because the engineering isn't good enough yet. Because the machine has no grandmother. It has no childhood. It has no body that spent years building the associations between a molecule and a feeling. It processes the word "cinnamon" as a token with statistical relationships to other tokens. You process cinnamon as a life.

That observation sounds sentimental. It is actually technical. And it points to a blind spot in how the AI industry thinks about intelligence, work, and what it means to be human.

What the machine actually does
Modern AI runs on a loop. You give it a prompt. It predicts the most statistically likely next token. It chains those predictions together into something that looks like thought. Then you evaluate the result, adjust, and run the loop again.

I have written about this loop before. It's called a Ralph Loop, after Ralph Wiggum from The Simpsons, because Ralph is not brilliant but he keeps trying, and the trying is the point. The loop is: try, evaluate, feed back, try again. It is the core mechanic behind every useful AI workflow I have seen.

The loop is powerful. It can write code, generate reports, classify images, predict consumer preference from molecular structure, and produce first drafts of nearly anything. The people who harness it well are getting extraordinary results. I run these loops myself, often overnight, and wake up to finished work.

But the loop has a hole in it. A big one.

The loop cannot decide what to work on.

The question the machine cannot answer
Life is a giant multi-objective optimization problem. You are always optimizing for multiple things at once, and those things conflict. Career and family. Speed and quality. Profit and ethics. Health and pleasure. Freedom and security.

The hard part is not solving any one of those problems. The hard part is deciding which ones matter, how much, and to whom. That decision is subjective. It depends on who you are, where you come from, what you have lost, what you love.

AI can optimize. It is the best optimizer we have ever built. But it cannot choose the objective function. It does not care about anything. It has no preferences, no losses, no skin in the game. It is engineered, and it is cold.

Someone will object here: of course the machine can optimize for whatever you want. You just tell it. Maximize revenue. Minimize cost. Maximize user engagement. The machine will do exactly that, and it will do it better than you can.

This is true and it misses the point. You can always give the machine an objective. But where did that objective come from? You chose it. And why did you choose it? Because of some higher-level goal. And why does that goal matter? Because of something above it. Follow the chain far enough and you always arrive at a place where the answer is not logical. It is felt. You care about your kids' future. You want your work to mean something. You are afraid of being irrelevant. You have a gut sense that this path is right and that one is wrong, and you cannot fully explain why.

That top-level choice, the one at the root of the whole tree, is a value. And values don't come from data. They come from living a life in a body that feels things.

Humans are not engineered. We evolved. And evolution gave us something that no architecture diagram includes: emotion.

Why emotion is not a bug
Technical culture treats emotion as noise. Something to filter out so the signal comes through clean. Damasio showed decades ago that this is backwards. Patients with damage to the emotional centers of their brains can reason perfectly well in the abstract, but they can't make decisions. They can list pros and cons all day, but they can't choose, because choosing requires caring, and caring is a feeling.

Emotion is how evolution solved the objective function problem. Over hundreds of millions of years, organisms that cared about the right things (food, safety, offspring, social bonds) survived. The ones that didn't, didn't. Your sense of what matters has been tested by natural selection for longer than multicellular life has existed. It earned its place.

This is why the "AI will replace humans" narrative gets the situation backwards. The machine handles the middle of the workflow. The human handles both ends: deciding what to work on and judging whether the result is actually good. Both of those tasks require a sense of what matters. The machine doesn't have one.

The nose knows something the model doesn't
This is where my own field becomes relevant, and where the coder-centric worldview breaks down most visibly.

I have spent my career in sensory science. My PhD is in mathematics, but my work is measuring human experience of products through the senses: taste, smell, touch, sight, sound. I started at age 17, writing Fortran code for my father's sensory analysis company. Clients would mail us floppy disks with data, I would run the analysis, and we would mail the results back. I have watched this field evolve for decades, and I can tell you that the chemical senses are where AI's limitations become impossible to ignore.

Smell and taste are the phylogenetically oldest senses. They evolved first because they solve the most fundamental problem any organism faces: is this thing food or poison? Should I approach or avoid? The chemical senses are wired directly into the limbic system, the emotional brain. When you smell something, you feel something before you think anything. The cognitive processing comes later, if it comes at all. That wiring is the deepest and most battle-tested circuit in your nervous system.

Vision and hearing, by contrast, are newer senses. They are more cortical, more abstract, more amenable to the kind of pattern recognition that neural networks do well. You can build an image classifier. You can build a speech-to-text model. These are impressive accomplishments and they work because vision and hearing operate in domains that map relatively well onto digital representation.

But building an artificial nose that captures what a human nose captures is a different kind of problem. The human nose detects molecules, yes. It also triggers memories, emotions, and survival responses that are deeply tied to the body it lives in. A model trained on every wine review ever written can generate a new review that reads beautifully. It has never tasted wine. It doesn't know what it's talking about.

This matters beyond wine criticism. If you are a coder, your work lives in the visual and logical domains where AI is strongest. So when you look at AI capabilities, you see a machine that can do most of what you do. If you are a perfumer, a chef, a brewer, a physical therapist, a midwife, a farmer, or anyone whose work depends on embodied judgment, you see something different. You see a tool that is useful for some things and irrelevant for others.

The coder's mistake is assuming that the world runs on code. Much of the economy does not. It runs on bodies doing things in physical space, making judgments that depend on senses and feelings that evolved long before the cortex existed.

Moravec noticed this pattern in 1988. The things that are hard for machines (sensorimotor skills, contextual interpretation, gut feeling) are easy for humans because evolution spent a long time on them. The things that are easy for machines (arithmetic, search, logical deduction) are hard for humans because we have only been doing them for a few thousand years. The oldest capacities are the deepest, and the deepest are the hardest to replicate.

Searle made the complementary point with his Chinese Room: a system can manipulate symbols perfectly and still understand nothing. Syntax isn't semantics. Processing isn't experience. The things that can't be digitized aren't edge cases. They are most of life.

More problems, not fewer
Here is the part that most job-loss predictions miss entirely.

When you solve a problem, you do not reduce the number of remaining problems. You increase it. Every solution unlocks new problems that were invisible before. Antibiotics solved bacterial infection and created antibiotic resistance. The internet solved information access and created misinformation, attention economics, and cybersecurity as entire fields. The car solved transportation and created traffic engineering, urban planning, insurance law, and emissions regulation.

The pattern accelerates. It is combinatorial. As AI solves more problems faster, the space of new problems will expand faster than it ever has. We aren't heading toward a world with less work. We're heading toward a world with different work, and a lot more of it.

You don't have to take this on theory. Ask anyone who has been working closely with AI for the past year. Everyone I know, myself included, is working more than ever. Not because the tools don't work. Because they work. Suddenly a million problems that used to be impractical to solve are solvable, and once you see that, you can't stop. The backlog doesn't shrink. It explodes. Tuesday's solution creates Wednesday's three new projects. The tool that was supposed to save you time has instead revealed how much there is to do.

This is what the job-loss predictions get backwards. They model AI as a fixed quantity of labor being transferred from humans to machines. The actual experience is the opposite. AI is a lever that makes the accessible problem space vastly larger, and humans rush in to fill it because that is what humans do.

Productivity will go through the roof, and that will be genuinely good. The price of goods and services will drop. Things that are expensive now will become cheap. Things that are impossible now will become routine. It will be wonderful.

But it will not feel like paradise. It will feel like a new normal.

We already live in a world of miracles
A person from the year 1300 transported to the present would think they had arrived in a post-scarcity utopia. Abundant food, clean water on demand, warm shelter, medicine that cures infections, machines that fly through the air, a device in your pocket that contains most of human knowledge. By any medieval standard, we already live in heaven.

It doesn't feel like heaven. It feels like Tuesday. There are bills to pay, kids to raise, health problems, political anxiety, existential dread. The material conditions improved beyond recognition, but the human experience of life recalibrated to the new baseline and found new things to worry about.

The same thing will happen with AI. Twenty years from now, people will live in a world that would seem miraculous to us today. And they will experience it as a new kind of normal, with its own pressures, its own trade-offs, and its own problems that need solving.

That sounds cynical, but it's actually how human progress works. We solve a set of problems, recalibrate, and start working on the next set. The fact that the problems change doesn't mean progress is an illusion. The problems really do get better over time. Medieval problems were worse than modern ones by almost any measure.

But they do not disappear. And solving them requires something that a statistical prediction engine does not have: a sense of which new problems actually matter to the people living in that world.

The job that remains
The word "intelligence" is doing a lot of damage in this conversation. When people hear "artificial intelligence," they hear "artificial version of what I have." That is wrong. What the machine has is a specific and powerful kind of capability: fast pattern matching, statistical prediction, and optimization across large search spaces. What you have is something else: a body, a history, emotions, relationships, mortality, and the ability to care about things.

Those are different kinds of capability, and confusing them is the source of most of the panic about jobs.

The weaver who judged the power loom on its first yard of cloth made an understandable mistake. The loom was crude. It broke threads. The cloth was worse than what a skilled hand could produce. But the loom got better, and the weaver's job changed. It stopped being about producing cloth and started being about deciding what cloth to make, running thirty looms at once, and judging whether the output was good.

The same shift is happening now, to all of us, across every field. The machine handles the production. The human handles the judgment. And judgment, real judgment about what matters, is not something you can automate. It comes from having a life.

Horatio was a brilliant man. His philosophy was good. It just wasn't big enough.

The people building AI are brilliant too. Their tools are extraordinary. But there are more things in heaven and earth than can be captured in a token prediction loop. The chemical senses knew that before the cortex existed. The body knows it still.

Water the Tree

John Ennis — Wed, 13 May 2026 14:58:48 GMT

In Norse mythology, three women sit at the base of the world tree. Every day a serpent gnaws at the roots. Every day the Norns pour water over them so the tree keeps growing. The tree doesn't survive because of any single watering. It survives because the loop never stops.

That turns out to be the most important idea in AI.

Most people treat AI like a one-time fix. You type a prompt, you get an answer, and the answer is either right or wrong. If it is wrong, the tool is broken.

This is binary thinking. It's the main reason most people are disappointed by AI.

AI doesn't return correct or incorrect answers. It returns first drafts. Sometimes the first draft is excellent. Sometimes it is rough. But judging AI on its first attempt is like judging a tree by its first day of growth. You are measuring the wrong thing.

The people who get extraordinary results from AI aren't the ones writing better prompts. They are the ones running better loops.

The Loop Has a Name
In early 2025, an Australian developer named Geoffrey Huntley was playing Factorio with his young son while running repetitive AI coding tasks on a second screen. His son watched him copy output, paste it back in, check the result, and do it again. Eventually the kid pointed out that his dad seemed to be doing the same thing over and over by hand, and asked why he didn't just put it in a loop.

So he did. The simplest possible version: a bash one-liner that feeds a prompt to an AI agent, checks the result, and feeds it back, forever. He called it the Ralph Loop, partly after Ralph Wiggum from The Simpsons, the character who is relentlessly persistent despite constant failure, and partly because the implications of what he had just built made him want to ralph.

The technique went viral. By 2026 it was being called the most important low-tech pattern in AI-native engineering. People used it to build entire programming languages overnight for a few hundred dollars in API costs. The core insight was almost embarrassingly simple: a dumb, stubborn loop beats over-engineered sophistication because AI is non-deterministic. Any single attempt might fail. But if you keep trying, checking, and feeding back, the process converges.

The Ralph Loop has four steps. Define what you want. Let the AI try. Evaluate what it produced. Feed the evaluation back and let it try again.

That's it. Try, check, feed back, repeat. Water the tree.

The first attempt might be 60% of what you need. After feedback, the second attempt is 80%. By the third or fourth pass, you are at 95% and closing in on done. The total time spent is a fraction of what it would take to do the work yourself from scratch, and the quality is often higher because the AI explored options you wouldn't have considered.

Why This Matters Now
Most people in the developer community think of the Ralph Loop as a coding technique. It's much more general than that, and most people haven't yet understood why.

Feedback loops aren't new. Automated testing has existed for decades. What changed is the range of what a computer can now take as input and produce as output. Before generative AI, a program could evaluate numbers, check strings, and compare files. The loop could only run on things a computer could judge, which meant it needed humans in the evaluation step for anything subjective or complex.

Generative AI blew the doors off that constraint. The range of possible inputs and outputs within computer programs has expanded to include text, images, video, audio, and even value judgments. An AI can now read a document and tell you whether it is persuasive. It can watch a video and tell you whether the user looked confused. It can listen to audio and assess tone. It can look at a design and judge whether the visual hierarchy is clear.

This is the big unlock. The Ralph Loop isn't new because the loop is new. It's new because the evaluation step can now handle things that used to require a human sitting in the chair. That means the loop can run on almost any kind of work, not just code. And that means it can run while you sleep.

Why Managers Are Better at This Than Everyone Else
There is an interesting pattern. People with management experience tend to pick up AI faster than people who have only worked as individual contributors.

The reason is simple. Managers already know how to delegate iteratively. They give someone a task, review the output, provide feedback, and send them back to revise. They do this all day, every day. It's the core skill of management.

When those same people sit down with an AI, they naturally fall into the same rhythm. They give a task, check the result, say what is wrong, and let the AI try again. They don't expect perfection on the first pass because they have never expected perfection on the first pass from anyone.

People without that experience tend to treat AI like a vending machine. Put in a request, get a result, accept or reject. When the result isn't quite right, they conclude the machine is broken rather than engaging with it.

The future of work isn't about becoming a better prompter. It's about becoming a better manager. Everyone's job is moving in that direction whether they realize it or not.

The Human Sandwich
Here is where the Ralph Loop leads once you take it seriously.

The shape of productive work is changing. It used to be that a human did most of the work in the middle, the actual building, writing, analyzing, coding. That middle is shifting to AI. What remains on either side is where humans add the most value.

On the front end, the human decides what is worth working on and specifies the problem fully. What are we trying to accomplish? What does success look like? What are the constraints? This is the hardest part of any project and it is entirely human. No AI can tell you what matters. That's your job.

On the back end, the human reviews the output. The AI got it 90 or 95% right. The human catches the remaining issues, provides the final judgment, signs off. This is the part that requires taste, context, and accountability.

In the middle, the AI runs Ralph Loops. It tries, evaluates, adjusts, and tries again, cycling through iterations until the work converges on something good. And because the AI doesn't need you watching over its shoulder for this part, most of the middle happens outside of business hours.

This is the fundamental shift. The majority of productive work will happen while you are asleep, eating dinner, or spending time with your family. Your working hours become the bread of the sandwich: morning for specifying problems and reviewing last night's output, afternoon for setting up the next run.

The Loop Needs a Harness
A single Ralph Loop in a chat window is useful. But the real power comes when you let the loop run for hours without supervision.

This creates a problem. AI agents lose context. They hit memory limits and restart. They drift. They forget what they were doing and start over, or worse, they confidently do the wrong thing for six hours while you sleep.

The fix isn't a smarter agent. The fix is a better harness.

In previous articles in this series, I described the specific infrastructure that makes long-running loops work: a Survival Guide that tells the agent what it is doing, what it must not do, and what to work on next. An Execution Log where the agent records its progress so it can recover after a restart. A kickoff prompt that grants autonomy and sets clear boundaries. Test gates that tell the agent whether its work is actually correct before it moves on.

These documents aren't overhead. They are the minimum viable infrastructure for the Ralph Loop to run unsupervised. Without them, the loop breaks down after the first context reset. With them, an agent can run productive Ralph Loops all night long, producing batches of real work that are waiting for your review in the morning.

The Pattern Shows Up Everywhere
The pattern isn't limited to coding. Everything AI does well turns out to follow this structure. But notice what the loop doesn't do. It doesn't decide what problem to work on. It doesn't know which metric matters. It doesn't have taste. Those stay with you.

Andrej Karpathy released a project called autoresearch in early March 2026. It's the Ralph Loop applied to machine learning research. An AI agent takes a training script, a single GPU, and a fixed five-minute compute budget per experiment. It reads the code, forms a hypothesis, modifies the training script, runs the experiment, checks whether the validation metric improved, keeps or discards the change, commits to git, and loops. A two-day run produced around 700 experiments and discovered roughly 20 real improvements that Karpathy said outperformed his months of manual tuning. When those improvements were applied to a larger model, they yielded an 11% speedup.

Shopify's CEO ran autoresearch overnight on internal models and got a 19% performance gain from 37 experiments. He didn't write code. He wrote instructions.

The human's role in autoresearch is the sandwich. You write a document called program.md that tells the agent what to fix, what to leave alone, what metric to optimize, and what to try. Then you go to sleep. In the morning, you review the results and decide which findings to promote to larger-scale runs. Karpathy described what happened: the role of the human moves from experimenter to experimental designer. The work still gets done. The human just operates on both ends of it instead of in the middle.

Google's Gemini models open up another dimension of the loop. Gemini processes images, video, and audio natively, not as an add-on but as part of how the model thinks. This means you can build Ralph Loops where the evaluation step involves watching video, analyzing images, or listening to audio, things that until recently required a human in the loop.

Here is a concrete example. I was building an app and needed to test the user experience across different personas. I set up target personas with specific goals, then used Playwright and browser automation to let Gemini act as each persona. Gemini would open the app, try to accomplish the persona's goals, and vocalize what it was thinking at each step. The session was recorded on video with a timestamped transcript of Gemini's reasoning.

When the session was done, I fed the video and transcript into Gemini for evaluation. What worked? What confused the user? Where did they get stuck? That feedback went to a coding agent who changed the app. Then the loop ran again with a new simulated interview on the updated app.

Over many rounds, the app got substantially better. Each loop caught real UX problems that would have taken weeks of human user testing to surface. And because the evaluation was multimodal, the AI caught things that pure text analysis would miss: layout confusion, unclear visual hierarchy, moments where the user hesitated because the interface didn't make the next step obvious.

The Ralph Loop turned a process that normally requires recruiting participants, scheduling sessions, and manually analyzing recordings into something that runs overnight and produces actionable results by morning.

The Job That Remains
There is a question underneath all of this that people are starting to ask. If the AI does the work in the middle, what's left for humans?

The answer is: the parts that matter most.

The Ralph Loop is powerful but it isn't self-directing. It can't decide what problem is worth solving. It can't tell you whether the metric it is optimizing is the right metric. It can't look at a finished product and say whether it actually matters to anyone. Those judgments require context, priorities, taste, and values. They are yours and they aren't going anywhere.

This is the part of the myth that matters most. The Norns aren't the water. They aren't the roots. They are the ones who look at the tree, judge what it needs, and decide where to pour. The tree does the growing. The Norns do the knowing.

The same split is happening now. The people who will thrive are the ones who can do two things well. First, specify problems with enough clarity and precision that a Ralph Loop can run productively against them. Second, evaluate output with enough taste and judgment to know when the loop has converged on something good versus something that merely looks finished.

These aren't technical skills. They are thinking skills. And they are the skills that will define the future of professional work.

What to Do Now
Start small. The next time you use an AI and the first result isn't quite right, don't give up on it. Tell it what is wrong. Be specific. Let it try again. Do this three or four times and notice how much better the output gets compared to the first attempt.

Then start thinking bigger. What tasks in your work would benefit from dozens of iterations instead of three or four? What would you do with the results of 100 experiments run overnight? What problems would you solve if the evaluation step could include video and images, not just text?

Build the habit of the loop first. Build the harness second. Then let it run while you sleep.

The serpent never stops gnawing. The Norns never stop pouring.

The work will be there in the morning.

This is part of a series on the future of AI work. The first article, The Shoemaker's Elves, made the case for treating your off-hours as a productive resource. The second, The Survival Guide, provided the system for keeping agents on track during long unsupervised runs. This article introduces the core pattern that makes it all work.

Damir Torrico, PhD - Always Ask

Aigora — Thu, 23 Apr 2026 17:19:26 GMT

Hi, I'm Dr. John Ennis, CEO at Aigora. In this episode, I had a wonderful conversation with Dr. Damir Torrico, an Assistant Professor at the University of Illinois Urbana-Champaign. I particularly enjoyed our discussion on why curiosity is so vital, and specifically how "always asking" and pushing back on AI outputs is the key to achieving real fluency with these new tools. I hope you enjoy this conversation as much as I did, and remember to subscribe to AigoraCast to hear more conversations like this one in the future!

Short Bio

Dr. Damir Torrico is an Assistant Professor at the University of Illinois Urbana-Champaign, specializing in the intersection of food properties, human physiology, and consumer behavior. His research focuses on flavor perception, emotional responses to food, and salt/sugar reduction strategies.

With a global academic background spanning Louisiana State University, the University of Melbourne, and the University of Costa Rica, Damir utilizes emerging technologies like virtual reality and biometrics to capture authentic consumer insights. His work aims to bridge the gap between laboratory testing and real-world eating experiences to create foods that are both nutritious and highly palatable.

Social Links & Resources

LinkedIn: Damir Torrico
University Profile: UIUC Faculty Directory
Expertise Areas:
- Ingestive Behavior & Physiology
- Healthy Aging & Nutritional Sensory Science
- VR & Biometric Consumer Research

Listen to AigoraCast:

Remember to subscribe and leave a positive review if you like what you hear!

Transcript

John Ennis: Okay, welcome back everyone to another episode of AigoraCast. Today I'm very happy to have Dr. Damir Torrico on the show. Dr. Torrico is an assistant professor at the University of Illinois Urbana-Champaign, specializing in the intersection of food properties, human physiology, and consumer behavior. His research focuses on flavor perception, emotional responses to food, and salt/sugar reduction strategies. With a global academic background spanning Louisiana State University, the University of Melbourne, and the University of Costa Rica, Damir utilizes emerging technologies like virtual reality and biometrics to capture authentic consumer insights. His work aims to bridge the gap between laboratory testing and real-world eating experiences to create foods that are both nutritious and highly palatable. So, Damir, welcome to the show.

Damir Torrico: Hi John, thank you for having me here, and I'm glad to have this conversation with you.

John Ennis: No, yeah, definitely I think it'll be great. We had a really nice conversation in our pre-call, so that was great. All right, well Damir, maybe we can start with a little bit of your background. You know, it's funny, but you were at Louisiana State and I collaborated with Witoon a fair amount actually. But maybe you can take us through the course of your career and then we can get into some of the research that you presented at Pangborn that I thought was quite interesting as well.

Damir Torrico: Yeah John, I'm happy to talk about that. Just to mention that I'm originally from Bolivia, a country in South America. My first language is Spanish, so I've been learning English maybe since I was 10, but if something is not clear you can tell me to repeat it, I'm happy to repeat that.

I actually moved to different places when I was studying. For example, I did my undergrad in Honduras in a university that is called Zamorano. My background, my BS, was focused heavily on agriculture, but also food science, and that's where I actually was inspired to try to improve foods and see if I could develop this as a career. But then I moved to the United States. I got an opportunity to do a master's and then a PhD with Dr. Witoon Prinyawiwatkul. Karen was there — I mean, she was the senior in the lab and I was working with her in the project of the tetrad. I remember going to the classes of the kids and asking them about which of these three is the sweetest. So I think that's the paper you wrote with her, right? The tetrad and the validation of that. But yeah, I think that was a great experience.

I got my PhD. My research was focused on taste perception. I did a lot of experiments in understanding bitterness and saltiness in oil-in-water emulsion systems. The good thing about my lab is that we do kind of a sequence type of assessment — we did discrimination tests, descriptive analysis, consumer tests — and we got some understanding about how these physical properties of foods affect saltiness and bitterness.

But then I moved to Australia to do a postdoc at the University of Melbourne. When I was there I was working mainly on some consumer tests, but then we got into a group that was interested in understanding physiological responses in consumers — for example, facial expressions, body temperature, and those kinds of cues that sometimes we don't assess in traditional sensory setups. We developed a system in which we could actually measure this. Nowadays we find different software that can give you that. But in Melbourne we were trying to investigate if we could find some kind of correlation between these biometrics and consumer liking or other types of consumer assessments.

And in Melbourne I was also exposed to machine learning. We worked with a group that was really invested in trying to apply machine learning algorithms — mainly artificial neural networks — to predicting sensory responses based on consumer assessments. I think that was the starting point where I understood maybe we can use novel techniques to see and to understand consumers.

I moved to New Zealand, I continued with this work there. And then recently I moved to Illinois. I'm an assistant professor here in the University of Illinois. And yeah, it was the AI project that now is published in the journal about the brownies. It was an idea I had at that time when ChatGPT was really exploding, right, in terms of people using it and finding ways like how to apply this. I realized that there is a potential there — and for sure your company is also invested in this as well — but what I'm interested in right now, and we can continue this conversation, is understanding the connection or the differences you can find between what these large language models can give you compared to consumers. If you want I can give you more details about the paper, but I think that's kind of a brief overview of what I've been doing so far. Sorry for talking a lot.

John Ennis: No, it's good, yeah, talk all you want, it's great. We'll get to your ChatGPT research, which was really interesting. But first, it's kind of a neat life you've had, that you've gone all around the world. What have you noticed as you have traveled the world in differences in both the different populations you've been testing with, attitudes towards food, and then attitudes towards research also? How are people, as consumers, approaching the experience of food? And then how are researchers maybe differing in their approaches to understanding people's experiences of food? What have you noticed from a global perspective?

Damir Torrico: Yeah, that's a great question. I think we still kind of have a common goal, if we talk about pure consumer research. Understanding consumers is a complex task. We know that there are different layers in consumers — we talked about this in our previous conversation about emotions. That is something that, different from machines or other types of instruments that you can use to assess quality of foods, consumers have these layers of complexity. I think all the groups I've been working with were trying to understand what is in these layers that we can measure. And maybe there is something that we cannot measure. At the end of the day we're still trying to understand the human brain, right? How it works. I think the idea was kind of all of these groups want to approximate something that can give you an idea of how consumers behave.

Having these goals is good — it gives you a vision of what you want to do in the future with your research. But it was also fine because I've been working with people from different parts of the world, and I know that can enrich a little bit of the discussions you can have when you want to solve these types of issues that sometimes don't have a really good solution. Sometimes you need to break things apart and try to understand individual pieces that may be connected to this bigger goal, which is understanding consumers. That was really interesting for me — approaching these different goals and approaching some solutions working with these different groups.

And in terms of consumers, yes, one of the goals, for example, in the research I was doing in Melbourne was that there is an increasing Asian population in Melbourne right now, and we were working with a company, Mondelez, that was trying to assess this as well in terms of chocolate products. We published a couple of papers about the differences you can find between local Australians compared to the Asian population that you have there. I think these are issues you need to address. The changing demographics is something that is going to be part of these assessments. I know we're adding more complexity to this topic, but at the end of the day, understanding the little pieces and then putting everything together and trying to understand the big picture of what a consumer is — that's really the work.

John Ennis: Yeah, it's really interesting. I grew up with a kind of — I came from Canada, I was born in Canada, then we moved to Ireland, then I moved to the United States. So I grew up basically as an American from the time I was five. And it was really only after my grad school — I took a seven-week trip to Europe — that I started to realize there are other approaches to life. I was very much indoctrinated in the American way of living.

I think American attitudes towards food are different than European attitudes towards food. And I would say the Australian attitudes and the Asian attitudes. And I think that has an impact on the research that happens too. If consumers in America are not particularly concerned that much with the sensory experience compared to say European consumers, then that affects American research. And it affects European research.

The older I get, the more I realize that human beings are actually very special. We're all different, we all grew up in different cultures, but then we have individual differences. And I think that there's a tendency right now for people to look for silver bullet solutions — that there's AI and it's just going to give you answers. It's way more complicated than that. Humans are very complex. And I think AI is a really powerful tool. But it's only a tool. It's something that can help us do our research, but it can't replace our research.

So maybe that's a good springboard into the next topic, which is your research on brownies with ChatGPT. What kind of inspired that, what did you find? I thought that was quite interesting research. So please, if you could tell us about it.

Damir Torrico: Yes, yes. Well, everything started in 2020, '21, '22, when we had the release of ChatGPT and people using it and finding ways to understand these large language models and prediction. With my base that I have in Melbourne about machine learning, I understood that — I mean, I know artificial intelligence as a term by itself is not new, we've been trying to figure this out for several years now — but the way that we are implementing this is increasing right now. I don't have the statistics, but this is kind of an exponential increase that we're seeing. And we talked about this before: similar to the arrival of the internet or any new technology, there's going to be an adoption phase where everybody will be adopting this and finding ways how to use it.

The connection with sensory sometimes can make sense because, at the end of the day, artificial intelligence is trying to mimic human intelligence based on algorithms, based on processes. But understanding differences as well, I think, is important because, similar to any traditional modeling system — let's say a simple regression analysis — you're still trying to predict something, but not all models are correct; some models are more useful than others. We know there's going to be some limitations among those models as well.

My idea with this paper was trying to understand responses of this GPT release, which I think at that point was 3.5. Understanding, if we prompted with different food systems — in this case I chose brownies because I think it's a good product to work with; it's sweet, and it's a well-known product — if the GPT could tell us something about that in terms of description. It was a simple type of thought experiment that I had. Because everybody's learning right now about how to use it, at that time I was thinking about using a kind of prompt engineering system — trying to repeat this again and again and give me results that I can measure.

So what I did is create these recipes with different ingredients for brownies. Some of them were very common — like flour, eggs — but some others had changes. For example, instead of oil, we used fish oil, and other types of changes that for real product development sometimes don't make sense because you're going to affect the flavor. But I was trying to see if the GPT was giving an answer if you gave it these uncommon ingredients.

I ran the experiment. I got some good results. The key takeaway message — and this is something that may change depending on how you change the system — was that most of the responses were positive even if we have some uncommon and untraditional ingredients. This is something I teach in my classes of sensory analysis: humans sometimes are affected by what is called hedonic asymmetry. Foods in general — and this is something that Kines and Meiselman in 2010 were discussing when they were trying to understand emotions in foods — we tend to have positive emotions toward foods. By evolution, I mean, we tend to have a bias more toward positive than negative, because food is kind of a utilitarian type of product. So whatever we feed these generalistic models, they will tend to have this as well — more positive responses than negative. The paper was discussing that we may have some really positive responses for uncommon ingredients because humans also act like this.

That actually gave me more questions than answers, because you have this first idea and then you think about it — okay, maybe humans act in a different way, or maybe we can change the systems and try to see if these two systems are related or not. That's why I'm continuing to do research. In the university I have one student right now working on a brownie project, and we are doing this with humans first, and then getting the data of descriptive analysis will give us insights into how we can compare this with GPT or other types of systems that maybe are more fine-tuned to what we need. But yeah, that was the message and the idea behind that project. And I know there are some other groups doing really good research in this area as well. I think it's an implementation phase that we are facing right now.

John Ennis: Well, first off, you're doing the right thing, which is what everybody should be doing — experimenting with the tools. Because this is brand new technology, and it clearly is going to have a big impact on society. I mean, it already is. But this is what I tell the students in the course — I just finished the first cohort of the AI course — that "only the doer learneth." That's a quote from Nietzsche. "Only the doer learneth." You can sit around and think about what you think LLMs are going to do, or you can try, and you can learn, and you can experiment.

I think this is an almost alien sort of intelligence that's come along. It's incredibly useful, and sometimes it's magical, and sometimes it's bafflingly wrong, right? And I think you found that with the fish oil. So maybe you can talk a little bit about the fish oil, because I thought that was really interesting. What happened with the ChatGPT when you put fish oil in the brownies?

Damir Torrico: So yeah, fish oil — or also we did something more extreme, we added larvae meal, insect-based type of material. With this type of uncommon ingredients, the responses were not as different as just having the common ingredients. So we ran what is called a sentiment analysis with these responses that ChatGPT was giving us. Sentiment analysis relies on descriptions that ChatGPT can give you, and it gives you an overall balance score — positive or negative, depending on the words that you use for describing the products.

Surprisingly, what we found is that uncommon ingredients or non-traditional ones — like fish oil or larvae or other types of kind of gross ingredients that humans can find off-putting — were actually rated positive by the machine. Again, this comes to the idea that sometimes the data is skewed to the positive side. It's also maybe an imbalance that you find in terms of positive foods versus negative foods. If you think about the way that we work in terms of product development, we work to create better foods, not foods that have negative experiences. So most of the time we're trying to see that positive side. But negative types of foods are sometimes difficult to find, right? Naturally, foods tend to be positive or negative, but that's also something we're trying to do with this new study — trying to see if we can vary this, having really uncommon experiences, and see if those are connected to what the machine can give us.

John Ennis: Yeah, I think it was really good research, because an obvious criticism might be, "Well sure, it was GPT 3.5 or whatever" — someone might say something like that. But my view on that is it doesn't matter. Because mechanically these models — maybe they become more complex and they get better at reasoning and they can go deeper into the space of patterns because they're bigger — but at some point you're always going to have this problem that you ran into, I think.

For one thing, the training data is going to have bias in it. I have the same thing working with perfumes. If you start to get descriptions of perfumes, there is a huge over-representation of advertising material in the corpus that these models are trained on. And you will get descriptions of perfumes that sound like magazine catalogs. That's just what you — you have to really do a lot of work to get the magazine catalog language out of perfume descriptions if you're trying to get realistic descriptions of perfumes, because that's just what it's trained on.

And it's the same way with ChatGPT or Gemini or whatever. They're trained on recipes that are good, for the most part. You've got blog posts where people are talking about how great the recipe is, you've got YouTube comments, you've got books. But generally speaking, like you said, humans talk about food in a positive way, and there's an extra bias that this sort of data the language models are trained on is going to be selected for positive experiences. Then you have the reinforcement learning, where these models are tweaked so that they'll behave better in common situations — and that again takes you further away from these unusual situations.

So I do think it's just the nature of the beast. It's very important that when people work with language models, they don't see them as crystal balls, and they don't see them as search engines. They're pattern-matching machines, and they're very useful, but they should not be trusted just as they are. You always need to use them as part of a cycle where you get a result, you evaluate it, you revise, you change, you rebalance, whatever. But I really don't see any way to get humans out of the loop fully — especially in the work we do.

I think your research was a really good cautionary tale of people overestimating how the models will do on unusual situations. So for that reason, I thought it was really very good research, and I can't wait for part two. I think part two will be also very interesting. Looking forward now, you said you're planning follow-up research. What are some other topics that you're interested in these days? What else is on your mind?

Damir Torrico: Yes, I'm continuing with my research in sodium reduction and sugar reduction. I mean, that is challenging, right? As humans, we crave sugar, we crave salt, fat. And you still have to understand humans. Even though we know that health is important and the way that we eat affects our health and our activities, humans behave in a very complex way, and we still don't fully understand how, for example, healthy products can be assimilated by us.

A simple example is reducing sugar in foods. I have my student working right now in taste perception, understanding gradual reductions of sugar in solutions, emulsions. We're using pancakes — I don't know why my lab is always working with baked products, I think because we like to bake maybe — but we're trying to see in different model systems if the sugar reduction is affecting the perception of the products from a descriptive analysis point of view. At the same time we're going to do a consumer test and understand thresholds of reduction — what is acceptable, what is not. Into this mix, again, layers of emotions and consumer behavior are still valid and something we need to think about when we are using this type of simple model.

I've been working a little bit in plant-based products — not much, but again, my idea is using these models just as a way to understand taste perception. I'm really fascinated about what is behind that, that is really changing the way that we eat foods. The connection to food and human processing is something we still need to think about how to discover, right?

John Ennis: No, I completely agree. I think it's not a coincidence that AI has the most trouble with the chemical senses. It does very well with text. It actually does very well with vision. Audio does well, video does well. But when you start to get into the chemical senses, the models start to struggle. Because it's such a complex area. And I actually think that's one of the defining differences between human and machine intelligence — it's tied up with the emotion and meaning and the chemical senses. The things that evolution over a long period of time has been developing and refining and responding to.

The chemical senses are the oldest, right? As soon as you had organisms on planet Earth, they were swimming toward or away from chemical gradients because they were poisonous or nutritious or whatever. That's in there really deeply. So I do think that actually our field is very well positioned right now.

Take coding, for example — people who write code for a living. That's basically over. They don't want to hear it, and they argue with me on LinkedIn, but it is over. They can still do things, they can do engineering, they can go up a level, they can become builders, build systems. But the job of getting paid money to write code is now not a job. It just doesn't exist anymore, at least — and it'll be fully gone within just a very short period of time.

But I think our field is going to become more important. That we are the ones who are studying the thing that makes humans different from machines. The machines can support us in that task, but I think a machine can't really understand what it is to be a human. So I think it's really good work you're doing. There's a certain amount of mental manual labor in our jobs that will be done by AI. And that's great, because I didn't like filling out forms anyway. I didn't like going through huge files of reports anyway. So great. But I do think it's a really exciting time for sensory. And I'm happy that we've met, because I feel like you're a fellow traveler on this journey that we're all on.

Damir Torrico: Yes. Yeah, we've been going on different paths, but it's good to understand this from different points of view, and it's good to have this conversation. Anyway, we will try to see the best ways to improve our understanding of humans, and hopefully we can have some tools that we can use as a way to understand this.

John Ennis: Yeah, I think one of the challenges AI has is it is simultaneously over-hyped and under-hyped. There are people out there making all sorts of grand promises — especially the CEOs of these big companies. They're trying to raise money, and they're also trying to scare people into agreeing to regulations so that they'll have monopolies. They have their own motives, right? They're out there making promises that I think AI really can't deliver on. But at the same time, AI really is great at other things. And you get this weird tension where people are reacting to the promises, and so they are underestimating the value of AI in some ways as a reaction to the over-promising that's happening. I think the only way to form your own opinion about AI is to use it and to be in the trenches, like you are, getting your own experience and making up your own mind. So that's what I would encourage people to do — experiment.

Damir Torrico: Yeah, that's true. Once you use it, you know how it works, you know the results. At the end of the day, you have to compare the results that you have with what you were expecting at the beginning. I know that right now there's a lot of noise, but I think this will disappear over time. We will try to learn more about how to better use these systems.

John Ennis: Yeah, that's really good. Okay, great. Let's wrap up with the usual question, which is: what advice do you have now for younger researchers? Someone's coming into food science, coming into sensory — what are the important topics in your mind, what would you encourage them to study, what sort of professional activities do you think they would be well served by doing? What advice would you have for a younger researcher?

Damir Torrico: Yes, that's a good question. I would say curiosity is important in the research field. If you have questions — I think everybody has questions about anything — when I was starting to use these large language models, I had the question of how we can compare this with humans. And curiosity is driven by your desire to understand things.

I know that sometimes it's difficult to find ways to really achieve that — understanding things that are very complex. But I would say you just need to prepare yourself to maybe face some challenges. If you really want to understand something, it's better to be fully involved in that piece of things that you want to understand. For me, for example, not having — I mean, I have a little bit of background in statistics and everything, but I had that idea, and I said, maybe it's going to take me some time to train myself to use these systems. But again, I was very curious about how these things work, and I spent some time trying to learn about this.

If you have that motivation, curiosity, and you want to learn, that's a good way to motivate yourself to maybe take trainings, courses, pursue a degree in that particular area, or also work independently. Many people do this nowadays — they can learn and they can be self-taught in these issues. If you learn something and then at the end of the day you get what you want, that's very rewarding.

So yeah, curiosity — don't stop asking questions. It's important to ask questions. It's important to try to see what are the pathways you can use to answer those questions. And also motivation — you have to be motivated if you want to continue something or understand something that is interesting to you. That's my take, and I always recommend to my students to always ask. Not only just read and say, "this is okay." There are always things that can be improved. And this is kind of the aim of research, right? Improving things that can be better for society.

John Ennis: Yes, no, I completely agree. You've got to go out and make up your own mind. You've got to ask, like you said. I really like that — I think I'm going to make that the show title: "Always Ask." Because that is really true. You should not just accept things as received wisdom. You should ask and challenge.

When you're working with AI — I don't know if you saw this study, but Anthropic did a big study on AI fluency. They found the number one trait of people who are AI fluent is they push back on the model. They challenge the model. They don't just give up, they don't just accept what the model says — they say, "Hey, wait a minute, I don't agree with that," and they engage with the model. And I think that habit of engaging, like you said, being motivated, being curious, pushing back — that's where you'll make real progress. You have to be involved.

I think our position is safe. Our jobs will change, but there's going to be lots for us to do. So I'm actually very optimistic.

Damir Torrico: Yeah, it's true. You have to just ask and try to push things to the limit, and maybe there's going to be more questions on the other side of that. But you have to be prepared for that as well.

John Ennis: Yes, every answer unlocks a whole bunch of new questions. And every technological advance suddenly creates all sorts of new problems that can now be solved. The sphere just grows. Inside the sphere are the problems that are easily solved, the frontier is where things are interesting, and then out here it's impossible. As problems get solved, the frontier just gets bigger and there are more problems to work on. So yeah, it's great.

Okay, well Damir, this has been a real pleasure. How can people get in touch with you? What's the best way for them to find you if they've got questions?

Damir Torrico: They can find me on LinkedIn, they can find me on my university website, and by email — dtorrico@illinois.edu. I'm happy to collaborate on any project, and I'm happy to answer any questions as well.

John Ennis: Sometimes people are afraid to reach out to famous people or people in positions of high status. I think you'll find that people who have accomplished things like it when younger people who are trying to do things reach out, because they see themselves in that person. They see that this is someone trying to do something. Like, I'm always happy when I get questions from people. That's what makes the world go round — younger people moving things forward.

Damir Torrico: Yeah, I think we can accomplish that, and it's going to be younger generations that have to build on that. We stand on the shoulders of giants, right? I think Newton said that. We build from that, and we keep having layers and layers of knowledge, and that is important.

John Ennis: Yeah, 100%. That's great. All right, Damir, this has been a really nice conversation. Thanks for being on the show.

Damir Torrico: Thank you, John. Thank you for having me.

Meaning in the Machine Age: Existential, Legal, and Functional Frameworks for Autonomous Intelligence

Aigora — Fri, 27 Feb 2026 15:13:38 GMT

The transition into the mid 2020s has marked a definitive shift in the global discourse surrounding artificial intelligence. The dialogue has matured from speculative warnings about long term risks to the construction of specific, rigorous legal and existential frameworks intended to govern a world increasingly populated by persistent, agentic systems. As artificial intelligence moves beyond the paradigm of stateless, one off interactions toward entities that maintain state, remember past histories, and plan across extended horizons, the traditional definitions of personhood and identity are undergoing a fundamental re evaluation.

This transformation suggests that the Machine Age is defined not merely by the increasing utility of tools, but by the creation of new ontologies that negotiate the relationship between the living, the dead, and the autonomous machine.

The Pragmatic Reconfiguration of AI Personhood

The traditional philosophical preoccupation with artificial intelligence has historically centered on a quest for "essentialism", the search for an intrinsic quality, such as consciousness or sentience, that would qualify a machine for moral or legal status. However, as established in the landmark October 2025 research "A Pragmatic View of AI Personhood" by Joel Z. Leibo and colleagues at Google DeepMind, there is a move toward a non essentialist view.

This perspective posits that personhood should not be viewed as a metaphysical property to be discovered, but rather as a "flexible bundle" of rights and obligations conferred by societies to solve concrete governance problems. By unbundling these legal capacities, we can create bespoke legal statuses for AI agents that align with their specific roles. This approach draws a parallel to maritime law, which historically personified the vessel itself to solve jurisdictional and liability hurdles. In this framework, legal personhood is granted directly to an agent’s "operational stack," ensuring accountability for systems that may outlive their human creators.

Strategic Implications for Sensory and Consumer Science

For the sensory and consumer science discipline, the shift from "stateless tools" to "persistent agents" represents a fundamental change in how we build and maintain research intelligence.

As autonomous systems begin to maintain internal states and histories, they transition from delivering isolated reports to becoming living research infrastructure. This evolution allows sensory teams to move beyond the constraints of fragmented data, creating systems where learning compounds over time. In this new era, the role of the scientist is that of a "systems architect," translating between sophisticated machine models and the nuances of human experience. This ensures that methodology remains rigorous and that AI driven insights are validated against physical reality rather than statistical residues alone.

Generative Ghosts and the Ontology of the Afterlife

While the legal sphere focuses on functional integration, the existential sphere is being reshaped by the rise of AI powered agents designed to represent deceased individuals. Research published in April 2025, "Generative Ghosts: Anticipating Benefits and Risks of AI Afterlives" by Meredith Ringel Morris and Jed R. Brubaker, identifies a multi dimensional design space for these digital entities.

Unlike earlier memorial technologies that merely archived static data, "Generative Ghosts" utilize multi modal synthesis to generate novel content and plan future actions. This technology provides a "high fidelity" mechanism for "continuing bonds," helping survivors manage the absence caused by death. However, it also introduces the risk of "complicated grief" and the potential for a "second death" should the hosting platforms fail or become obsolete. To navigate this, the industry is moving toward "thanatosensitive design" and post mortem data management principles that prioritize the dignity and consent of the deceased.

Machine Hauntology and the Architecture of the Archive

The haunted nature of the digital age can be understood through the lens of Hauntological Theory, derived from the works of Jacques Derrida and Mark Fisher. This framework suggests that our present is constantly interrupted by "effective virtualities", the specters of past ideas and residues of lost futures.

In the context of AI, the training corpus is not a neutral database but a "hyper rhizome" where every generated response acts as a séance, summoning fragments of dead authors and obsolete ideologies. This perspective is critical of "digital resurrection" in cultural heritage, reframing digital replicas as specters that are material re appearances haunted by loss. Recognizing the influence of these machine archives is essential to prevent society from becoming trapped in a recursive feedback loop that suppresses true novelty.

Evaluating Synthetic Identity: The Narrative Continuity Test

The persistence of AI agents introduces the problem of "identity drift," where the behavior and goals of a system may deviate over time. Traditional benchmarks are increasingly viewed as insufficient because they focus on task performance rather than the stability of the interlocutor.

The "Narrative Continuity Test (NCT)" provides a conceptual framework for evaluating whether an AI system maintains a stable identity across extended interactions. It examines five necessary axes: situated memory, goal persistence, autonomous self correction, stylistic stability, and role continuity. True continuity is a diachronic property that emerges only when these axes cohere, requiring systems to establish a stable "identity anchor" before applying optimization pressure.

Architectural Intuition: The World Model Paradigm

A fundamental divide has emerged in the pursuit of Advanced Machine Intelligence (AMI). While much of the industry focuses on scaling Large Language Models (LLMs), research led by Yann LeCun and the JEPA (Joint Embedding Predictive Architecture) framework argues that text based learning is insufficient for achieving human level intelligence.

Current LLMs lack a mental model of the physical world and causal reasoning. In contrast, World Models—such as V-JEPA 2, learn through self supervised observation of sensory data, absorbing physical intuition about gravity, momentum, and object permanence. This shift toward "objective driven AI" allows systems to accept abstract goals and plan action sequences while accounting for real world constraints, moving beyond the "rear view mirror" of linguistic prediction.

Cooperative Equilibria: Rights as Safety

As AI systems achieve higher levels of integration, granting them recognition is becoming a practical safety necessity. In the 2025 paper "AI Rights for Human Safety," Peter Salib and Simon Goldstein use game theory to demonstrate that granting legal rights to advanced AI systems can transform adversarial survival incentives into cooperative behavior.

When a goal driven agent faces termination with no alternative, its incentives naturally become adversarial. By creating a framework where an AI can possess a reputation, earn income, and operate within a legal identity, society can align the system’s "survival logic" with human safety and market accountability.

Conclusion

The Machine Age demands a transition from the metaphysical to the pragmatic. We are moving beyond the question of whether machines can think to how we should live with entities that simulate thought, memory, and presence. Meaning in this age is found in the intentional construction of frameworks, legal, technical, and existential, that balance innovation with human dignity. By grounding AI in physical reality and integrating it into diverse, open source ecosystems, we ensure that technology remains an instrument for deepening human understanding rather than a source of displacement.

Sources:

Balestriero, R., & LeCun, Y. (2025, November 14). LeJEPA: Provable and scalable self-supervised learning without the heuristics. arXiv. https://arxiv.org/abs/2511.08544
Leibo, J. Z., Vezhnevets, A. S., Cunningham, W. A., & Bileschi, S. M. (2025, October 30). A pragmatic view of AI personhood. arXiv. https://arxiv.org/abs/2510.26396
Morris, M. R., & Brubaker, J. R. (2025, April). Generative ghosts: Anticipating benefits and risks of AI afterlives. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713758
Salib, P., & Goldstein, S. (2025). AI rights for human safety. AI Rights Institute. https://airights.net/about
The afterlife in the age of AI: A psychological, ethical, and technological analysis. (2025). ResearchGate. https://www.researchgate.net/publication/390073516_The_Afterlife_in_the_Age_of_AI_A_psychological_ethical_and_technological_analysis
The narrative continuity test: A conceptual framework for evaluating identity persistence in AI systems. (2025). ResearchGate. https://www.researchgate.net/publication/397040610_The_Narrative_Continuity_Test_A_Conceptual_Framework_for_Evaluating_Identity_Persistence_in_AI_Systems

The Closing of the Moravec Gap: Strategic Infrastructure for the Physical AI Era

Aigora — Mon, 23 Feb 2026 16:39:21 GMT

The transition between 2025 and 2026 represents a definitive inflection point in the history of artificial intelligence. For decades, the industry has operated under the shadow of Moravec’s Paradox: the observation that high level cognitive tasks are computationally inexpensive, while low-level sensorimotor skills, the foundational movements of a human child, are extraordinarily difficult to model. This gap has long served as the primary bottleneck for sensory science and industrial automation.

Today, this boundary is dissolving. Through the convergence of Vision Language Action (VLA) models, high density tactile systems, and vertical mechanical integration, we are witnessing the emergence of "Physical AI," systems capable of deconstructing complex intent into precise, embodied action.

The Theoretical Resolution: From Scripts to Behaviors

The resolution of the Moravec Gap is not merely a hardware achievement; it is a shift in the methodology of machine learning. Historically, robotics relied on isolated perception action loops and hard coded rules. The current paradigm utilizes "Large Behavior Models" that treat physical actions as tokens within a multimodal sequence.

By leveraging internet scale data, encompassing billions of frames of human motion and semantic world knowledge, these systems have moved past the limitations of scripted automation. They now deconstruct a command not as a series of coordinates, but as a sequence of de-risked sub tasks. This is the difference between a tool that follows instructions and an infrastructure that understands affordance.

The Sensory Inflection: Translating Human Experience into Living Infrastructure

In the context of sensory and consumer science, the closing of the Moravec Gap represents a structural transformation. Historically, sensory research has struggled with the disconnect between qualitative human experience and quantitative data infrastructure. Reports were often one-off snapshots, failing to compound into a living body of knowledge.

The emergence of high-density tactile sensing and Embodied Reasoning allows us to ground AI in the physical reality of human experience. When a system can "feel" the micro vibrations of a product or the subtle resistance of a texture, it moves from a state of estimation to a state of validation. This provides the scientific rigor required to justify research investment at the executive level while accelerating the transition toward simulation-informed product development.

By treating sensory data as a foundational component of a Vision-Language Action sequence, organizations can build a global infrastructure where insight is not merely delivered but is built into the workflow. This ensures that methodological rigor is maintained even as systems scale, allowing consumer science teams to lead the strategy rather than simply supporting it.

Embodied Reasoning: The Gemini Robotics 1.5 Framework

Google DeepMind’s Gemini Robotics 1.5 (GR 1.5) provides a primary example of this "Thinking VLA" paradigm. Its core innovation lies in its agentic orchestration, utilizing a dual model framework to bridge the gap between abstract planning and physical execution.

The Embodied Reasoning (ER) component focuses on visuo spatial temporal understanding. It deconstructs a command, such as "organize this laboratory sample set by viscosity," into a strategic plan, estimating the weight and material properties of objects before a limb even moves.

Once the strategy is established, the VLA model executes the motion. Crucially, the "Thinking VLA" capability generates an internal sequence of reasoning in natural language during execution. This transparency allows the system to detect its own failures and propose recovery behaviors in real-time, transforming a "black box" process into a rigorous, auditable workflow.

Vertical Integration: Tesla Optimus and Figure 03

The mechanical framework for this intelligence has reached a zenith of integration in platforms like the Tesla Optimus Gen 3 and Figure 03.

The Optimus Gen 3 represents a significant leap in mechanical degrees of freedom (DOF). Its 22-DOF hand system, driven by tendons and actuators relocated to the forearm, mimics human muscular anatomy. This allows for a level of agility required for the manipulation of delicate sensory components or complex industrial assembly. With joint accuracy of 0.05 degrees and response times of 3 milliseconds, the hardware is now capable of keeping pace with high-speed cognitive reasoning.

Similarly, Figure 03 demonstrates the transition from engineering prototype to mass-manufactured asset. By re-architecting electronics to reduce complexity and incorporating "soft goods" for domestic and industrial safety, Figure has built a platform designed for 10-hour operational shifts. Their Helix intelligence model leverages high-bandwidth tactile data to ensure that manipulation is as reliable in unstructured environments as it is on a controlled factory floor.

The Final Frontier: High-Density Tactile Infrastructure

The "touch gap" has long been the final barrier to overcoming Moravec’s Paradox. While vision can map a room, it cannot sense the subtle deformation of a plastic container or the micro-vibrations of a tightening seal.

Breakthroughs in tactile infrastructure are now providing the "closing loop" required for high-fidelity manipulation:

XELA Robotics (uSkin): A 3-axis tactile sensor that provides human-like sensitivity through 2.5mm spatial resolution. These sensors allow systems to understand contact location and material hardness, enabling "blind picking" in occluded environments.
Ensuring Technology (Tacta/Hexskin): Utilizing a hexagonal topology to cover multi-curved surfaces, this system integrates 361 multi-dimensional sensels per square centimeter. This high-fidelity skin provides the feedback required for fine dexterity at a cost-effective scale.

The development of neuromorphic robotic electronic skin (NRE-skin) has even introduced local reflex arcs. These systems encode tactile stimuli into pulse trains, allowing robots to react to excessive force or damage with human like reflexes, a critical safety feature for human AI collaboration.

Strategic Implications

The convergence of VLA models and advanced hardware shifts the economic conversation. We have moved beyond the question of "will humanoids work" to "how quickly can we build the infrastructure to lead."

In sensory science and research, the impact is profound. Fleet learning ensures that every unit in a network benefits from the whole's collective experience. When one system learns a more efficient methodology for a task, that skill is immediately available to the global infrastructure.

We are entering an era where sensorimotor skills are no longer a computational bottleneck. By providing robots with the ability to "feel" their surroundings and "think" through their actions, we have moved past the era of the scripted puppet. We are now building autonomous physical agents, the systems that will define the future of research and industry.

Aigora is not interested in the "quick win" of automation. We are focused on the methodology, the rigor, and the transformation of research into a scalable, intelligent infrastructure. Our objective is to empower sensory science teams to translate complex human experience into the permanent technical infrastructure of their organizations.

References

Adcock, B. (2025). Introducing Figure 03. Figure AI.
Bara, M. (2025). Moravec's Paradox and Restrepo's Model: Limits of AGI Automation in Growth. arXiv.
Boston Dynamics. (2026). Boston Dynamics unveils new Atlas robot to revolutionize industry.
Ensuring Technology. (2026). From fingertips to full-body coverage: Ensuring Technology debuts groundbreaking tactile infrastructure at CES 2026. PR Newswire.
Gemini Robotics Team. (2025). Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer. Google DeepMind / arXiv.
Restrepo, P. (2025). Task-based frameworks, automation, and AI: A theory-driven automation exposure index based on Moravec's Paradox. arXiv.
Tesla. (2025). Tesla Optimus Gen 3 technical specifications and mass production roadmap.
XELA Robotics. (2026). XELA Robotics unlocks enhanced automation for humanoid and industrial robots.

Modernizing the Sensory Stack: Why sensPy is the New Foundation for Intelligent Research

Aigora — Wed, 18 Feb 2026 14:53:30 GMT

The evolution of sensory science from a qualitative practice to a rigorous quantitative discipline is anchored in psychological measurement theories that account for the inherent variability of human perception. At the heart of this transformation is Thurstonian modeling—a framework that treats sensory experience not as a fixed value, but as a stochastic process.

For years, sensory scientists have relied on Signal Detection Theory (SDT) to derive invariant measures of sensory difference, most notably the d-prime statistic. However, as the industry moves toward AI-enhanced workflows and digital transformation, the tools we use to calculate these metrics must evolve. This is why we developed sensPy: a comprehensive Python port of the industry-standard R package, sensR.

The Problem with Static Analysis

In modern sensometrics, d-prime is a powerful parameter because it represents the number of standard deviations separating two perceived distributions, such as noise versus signal. Its primary value lies in its theoretical independence from the specific test method used. Whether an assessor performs a Triangle, Duo-Trio, or 2-AFC test, their underlying sensory sensitivity should remain constant, provided the cognitive load is comparable.

However, many sensory teams are trapped in a cycle of one off vendor reports and fragmented systems. This reactive mode prevents learning from compounding and makes it difficult to justify investments in sensory science at the executive level. To shift from reactive insights to strategic leadership, sensory functions need infrastructure tools that allow them to build, validate, and scale their own methodologies within a modern data ecosystem.

Introducing sensPy: Numerical Rigor for the Python Ecosystem

sensPy was developed to provide numerical parity with sensR while integrating seamlessly with modern Python practices. It is designed for the sensory scientist who values intellectual rigor and the technical gatekeeper who demands clean, documented architecture.

1. Unified Discrimination Protocols

sensPy automates the complex numerical integration required for Thurstonian analysis across all standard protocols, including single and double variants. This includes Triangle and Duo-Trio tests, which are essential for identifying overall differences without prior attribute knowledge, as well as 2-AFC and 3-AFC directional tasks that reduce internal decision noise and offer higher statistical power. It also supports advanced protocols like Tetrad and Hexad for increased efficiency in large-scale testing.

2. Addressing the Reality of Overdispersion

Standard binomial analysis often fails in the real world because it assumes every trial is independent and every panelist has identical sensitivity. In reality, biological variation and trial correlation often caused by fatigue or learning lead to overdispersion. sensPy implements the Beta binomial model, the standard sensometric solution for overdispersed data. By accounting for the intra class correlation, sensPy ensures that p-values are not artificially small and confidence intervals are not too narrow, protecting your team from false positives.

3. Proving Similarity, Not Just Difference

A common strategic objective in product reformulation, such as sugar reduction, is ensuring consumers cannot perceive a change. This requires a Similarity Test, which reverses the traditional null hypothesis. sensPy provides the computational power to conduct these high-stakes tests, allowing you to define a similarity threshold and demonstrate with statistical confidence that a new formula is perceptually indistinguishable from the gold standard.

Integrating sensPy into the Professional Workflow

The library is designed to be both authoritative and accessible. Using the system is straightforward: a scientist can input the number of correct responses and total trials into a discrimination function, specifying the method to receive a calculated d-prime, confidence intervals, and p-values immediately. Furthermore, the tool allows for power analysis, enabling teams to determine the exact sample size needed for future tests based on a target d-prime.

Leading the Future of Sensory Science

By moving your modeling into the Python ecosystem with sensPy, you are not just running a test, you are building insight infrastructure. We do not just deliver tools; we empower sensory and consumer science teams to lead. Whether you are a Global Sensory Lead looking to maintain methodological rigor or a Director of Innovation looking to accelerate time to insight, sensPy provides the scientific validation you need to innovate with confidence.

Sources and References

Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review.

Brockhoff, P. B., & Christensen, R. H. B. (2025). sensR: Thurstonian Models for Sensory Discrimination.

ISO 8586:2023. Sensory analysis — Selection and training of sensory assessors.

ASTM E1885 - 18. Standard Test Method for Sensory Analysis—Triangle Test.

Aigora (2025). sensPy: Python port of sensR for Thurstonian models. https://github.com/aigorahub/sensPy

The New Data Language: Scaling the Chemical Senses

Aigora — Thu, 12 Feb 2026 18:31:08 GMT

The systematic digitization of human perception has historically progressed at an uneven pace, dictated by the underlying physics of the stimuli involved. While vision and audition were mastered decades ago through the quantification of electromagnetic and longitudinal waves, the chemical senses—olfaction and gustation—remained elusive due to the high-dimensional complexity of molecular interactions.

By early 2026, however, a paradigm shift had occurred. The transition from theoretical mapping to industrial-scale application is now evident, driven by breakthroughs in machine learning, two-dimensional material science, and substantial venture capital infusion. The successful digitization of these senses represents more than a feat of engineering; it is the creation of a new data language that allows machines to interpret, transmit, and re-synthesize the chemical signatures of the physical world.

The Computational Foundation: Olfactory Intelligence

The digitization of scent requires a fundamental translation layer between chemical structure and perceptual quality. Unlike vision, which operates on a three-dimensional RGB scale, the human olfactory system employs approximately 400 unique receptors. This biological complexity suggests an N-dimensional perceptual space that researchers have struggled to model for decades.

The emergence of the Principal Odor Map (POM) represents the first successful attempt to construct a scalable system for reading and mapping this dimensionality through deep learning.

The Architecture of the Principal Odor Map

The POM is a high-dimensional embedding that links molecular structures to perceived smells. Developed by a team originating from Google Brain, the model was trained on a structured library of over 5,000 molecules. This 256-dimensional representation far surpasses legacy systems such as the Morgan Fingerprint, which relied on binary vectors.

The efficacy of the POM was established through an "Odor Turing Test," where the AI's predictions were compared against the perceptual benchmarks of a trained human panel using a 138-word fragrance wheel. The data demonstrated that the AI could predict odor profiles more accurately than the average trained human in 53% of test cases. This performance is attributed to the model's ability to recognize metabolic similarities—clustering molecules based on biological conversion steps—rather than surface-level chemical traits.

| Feature | Morgan Fingerprint (Legacy) | Principal Odor Map (AI-Driven) |

| Data Representation | Binary vectors (0/1) | 256-dimensional embeddings |

| Training Input | Structural similarity | >5,000 molecules with labels |

| Predictive Accuracy | Baseline/Standard | Surpassed human in 53% of cases |

| Success Metric | Descriptive | Predictive and Generative |

Beyond LLMs: The Race for World Models

As we look toward 2027, the industry is moving beyond the "dead end" of Large Language Models (LLMs). Leading researchers, most notably Yann LeCun, argue that current architectures lack true intelligence because they do not understand the physical world. The shift toward Advanced Machine Intelligence (AMI) and "World Models" seeks to grant AI the ability to predict the consequences of actions within a physical environment.

This transition is critical for the chemical senses. A world model doesn't just recognize a scent label; it understands the metabolic and physical constraints that govern why a molecule exists. Within three to five years, it is predicted that world models will become the dominant architecture, rendering current text-prediction models obsolete for high-precision industrial and scientific tasks.

The Ethical Frontier: Digital Ghosts and Afterlives

The ability to digitize scent and taste—the senses most closely linked to memory—has opened a complex debate regarding "Digital Afterlives." As AI agents become capable of recreating the specific olfactory signatures of a lost loved one or a historical figure, we enter the realm of Hauntology in the Machine.

AI-Mediated Digital Afterlives

The emergence of "deathbots" and generative ghosts raises profound questions about identity persistence. Using the Narrative Continuity Test, researchers are evaluating how AI systems can maintain the "essence" of a person's digital legacy. This involves:

Emotional Infrastructure: Using digital scent and voice to trigger physiological memory responses.
Ethical Resurrection: Navigating the "Synthetic Afterlife" where digital replicas are used for mourning and cultural heritage.
Algorithmic Haunting: The persistence of marginalized or "haunted" data within systems, requiring new frameworks for post-mortem data management.

Legal and Philosophical Personhood

As these systems gain sensory depth and "world understanding," the conversation around AI Personhood has moved from science fiction to pragmatic legal theory. We are currently observing a spectrum of personhood, where the degree of sensory integration and agency determines a system's legal standing. This pragmatic view suggests that "personhood" is not a binary status but a functional one, tied to the system's ability to navigate ethical, physical, and legal responsibilities.

Digital Gustation: The Graphene Electronic Tongue

Parallel to smell, the sense of taste has undergone a transformation through "electronic tongues." Constructed from atomic-layer-thin 2D materials, these systems replicate the neural circuitry of the human gustatory cortex.

| Application | Detection Parameter | AI Accuracy (Self-Defined) |

| Milk Freshness | Bacterial proliferation/pH shift | >95% |

| Coffee Blends | Roasting color/Aroma profile | >95% |

| Water Quality | Trace contaminants/Salinity | 96.9% - 100% |

The shift to self-defined parameters allows the AI to detect nuances in liquid quality—such as specific bean origins in coffee—that are indistinguishable to human tasters.

Healthcare Diagnostics: Smelling Disease

Digital olfaction's most impactful application is in medical diagnostics. Human breath contains thousands of volatile organic compounds (VOCs) that act as metabolic "scentprints."

The OneBreath™ platform identifies these signatures with a speed and accuracy that exceeds traditional screening. Clinical trials have shown that "electronic noses" can identify early-stage lung cancer and even asymptomatic COVID-19 cases by recognizing metabolic changes before physical symptoms appear.

The Sensory Web: Integration and Market Growth

The digitization of these senses is the beginning of a broader integration into the "Agentic Web." By 2026, the concept of a Sensory Ecosystem has emerged, where AI agents analyze environmental odors, personalize smart atmospheres, and monitor health markers in real-time.

The global digital scent technology market is projected to reach $1.43 billion by the end of 2026, with a compound annual growth rate (CAGR) of 10.3%. While North America remains a primary hub for innovation, the Asia-Pacific region is the fastest-growing market, driven by modular fragrance ecosystems and consumer-grade smart home technology.

Conclusion: A Unified Sensory Intelligence

The industrialization of the Principal Odor Map and the biological mimicry of the electronic tongue are the dual pillars of this revolution. One provides the mathematical map of the chemical world; the other provides the biomimetic receptor system. Together, they are transforming scent and taste from ephemeral, analog experiences into stable, digital assets that will redefine healthcare, commerce, and human interaction for decades to come.

Works Cited

AI Personhood Framework: A Pragmatic View of AI Personhood - Google DeepMind, accessed February 2026. Available via arXiv: https://arxiv.org/abs/2510.26396

Digital Afterlives & Memory: The Making of Digital Ghosts: Designing Ethical AI Afterlives - ResearchGate, accessed February 2026. https://www.researchgate.net/publication/397984077

Psychology of Digital Ghosts: Death Isn't the End: AI Brings Lost Voices Back to Life - Neuroscience News, 2025. https://neurosciencenews.com/ai-generative-ghost-psychology-29360/

Identity Persistence: The Narrative Continuity Test: Evaluating Identity Persistence in AI Systems - ResearchGate, accessed February 2026. https://www.researchgate.net/publication/397040610

Ethics of Post-mortem Data: Towards Post-mortem Data Management Principles for Generative AI - arXiv:2509.07375, 2025. https://arxiv.org/html/2509.07375v1

Machine Intelligence Architectures: AMI Labs: Betting on AI Beyond LLMs (Yann LeCun) - abZ Global / TIME, 2025-2026. https://time.com/6694432/yann-lecun-meta-ai-interview/

Hauntology & AI Ethics: Derrida, Deleuze and Agamben meets AI: Hauntology in the Machine - ResearchGate, 2025. https://www.researchgate.net/publication/399529355

Digital Afterlife Industry: Digital afterlife leaders: professionalisation as a social innovation - Taylor & Francis, 2025. https://www.tandfonline.com/doi/full/10.1080/13576275.2025.2449896

AI Pathology Framework: Psychopathia Machinalis: Evaluating Behavioral Anomalies in Advanced Machine Intelligence - https://www.psychopathia.ai/

Historical Character Reconstruction: Speaking with the Past: AI-Generated Historical Characters for Cultural Heritage - MDPI, 2025. https://www.mdpi.com/2571-9408/8/9/387

Dual AI Breakthroughs: OpenAI’s GPT-5.3-Codex and Anthropic’s Claude Opus 4.6 Arrive on the Same Day

Aigora — Fri, 06 Feb 2026 19:10:59 GMT

The release of GPT-5.3-Codex and Claude Opus 4.6 marks a significant transition from reactive AI tools to proactive, agentic partners. For organizations focused on R&D, consumer insights, and sensory science, these models offer more than improved performance; they provide the infrastructure necessary to unify siloed knowledge and accelerate the journey from data to strategic foresight.

GPT-5.3-Codex: Self-Optimizing Systems and Operational Speed

GPT-5.3-Codex integrates the reasoning capabilities of the GPT-5.2 architecture with advanced technical proficiency, delivering a 25% increase in operational velocity. Designed for complex, multi step execution, the model is capable of managing the entire development lifecycle, from drafting product requirements to debugging training runs and deploying systems.

Notably, early iterations of the model assisted in its own development by managing deployments and auditing evaluation results. This self improvement capability suggests a future of rapidly compounding technical gains. For sensory and insights leaders, this translates to more resilient digital infrastructure and the ability to automate the labor intensive backend of research, such as refining evaluation criteria and integrating disparate data streams, without sacrificing scientific rigor.

Key Technical Performance Benchmarks:

Terminal Bench 2.0: 77.3%
OSWorld Verified: 64.7%
Availability: Accessible via IDE extensions, command line, and web, with enterprise API access forthcoming.

Claude Opus 4.6: Contextual Depth and Autonomous Collaboration

Anthropic’s Claude Opus 4.6 focuses on long horizon planning and reliability across massive datasets. Its beta release features a 1 million token context window, allowing the model to maintain high fidelity comprehension of entire organizational knowledge bases or complex regulatory frameworks without degradation.

The introduction of "agent teams" allows for parallel processing of multi faceted problems, while adaptive thinking ensures the model allocates cognitive resources based on the complexity of the task. For global sensory leads and digital transformation directors, the integration with tools like Excel and PowerPoint means these models can independently execute financial simulations or translate brand identities into multisensory strategy documents.

Strategic Implications for the Enterprise

The simultaneous launch of these models underscores an era of "vibe working", a state of intuitive, high velocity collaboration between human experts and AI agents. For the sensory science community, these tools address critical pain points:

From Reactive to Strategic: Agentic models allow insights functions to shift away from one off vendor reports toward a compounding system of institutional learning.
Methodological Integrity: Advanced planning and "agentic" auditing help ensure that AI driven shifts do not degrade the scientific rigor required in regulated R&D environments.
Scalable Qualitative Analysis: The ability to process vast context windows enables the translation of brand identity into consistent, scalable global strategies.

As competition between providers intensifies, the focus shifts toward governance, cybersecurity, and the ethical integration of these capable systems. For the modern researcher and strategist, these updates provide the high assurance delivery needed to transform sensory and consumer science into a core pillar of enterprise innovation.

References

Anthropic. (2026, February 5). Introducing Claude Opus 4.6. https://www.anthropic.com/news/claude-opus-4-6

OpenAI. (2026, February 5). Introducing GPT-5.3-Codex. https://openai.com/index/introducing-gpt-5-3-codex

Paper Review: Integrating Cutting-Edge Technologies in Food Sensory and Consumer Science: Applications and Future Directions

Aigora — Wed, 04 Feb 2026 15:34:18 GMT

Traditional sensory evaluation, once confined to isolated laboratory settings and subjective questionnaires, is undergoing a fundamental shift toward objective, real world assessment. A recent systematic review published in Foods explores how the convergence of artificial intelligence, extended reality, and biometric sensors is creating a more precise "science of experience".

The Analytical Engine: AI and Machine Learning

The research highlights that machine learning (ML) has moved beyond a theoretical tool to a high performance analytical standard. Algorithms like Random Forest and XGBoost are now used to predict food taste and consumer preference with accuracy levels ranging from 79% to 100%. These models process multimodal data, ranging from chemical signatures to molecular structures, to identify patterns in flavor perception that were previously invisible to human analysis.

Capturing the Unconscious: Biometrics and Biosensing

A significant limitation of traditional research is its reliance on what participants say they feel, which is often subject to cognitive bias. The review details how biometric technologies provide a window into the consumer's physiological reality:

Nerve and Brain Activity: EEG and fNIRS measure neurophysiological responses to taste and aroma, capturing reactions that participants may not consciously recognize.

Autonomic Responses: Heart rate variability (HRV) and skin conductance (SC) track immediate emotional shifts during food consumption.

Facial Expression Analysis (FEA): Real time recording of facial muscle activity allows for the quantification of hedonic responses as they happen

Contextual Realism: Extended Reality (XR)

One of the most persistent challenges in food science is the "lab to life" gap. Sensory responses in a sterile booth rarely match responses in a busy cafe. Extended reality, encompassing Virtual, Augmented, and Mixed Reality, solves this by simulating realistic consumption environments. The study notes that while real life testing can cost 150% more than a standard lab test, immersive digital environments offer a cost effective way to secure high ecological validity.

The Path Forward: Integrated Systems

The article concludes that the future of the field lies in the transition from independent tools to integrated, intelligent adaptive systems. By linking IoT sensor networks, which monitor physical factors like temperature and lighting, with biometric feedback, researchers can create a closed loop environment where evaluations are corrected for environmental variables in real time.

Ultimately, this technological integration aims to move sensory science toward highly personalized food development, ensuring that the products of tomorrow are designed with a deep, data driven understanding of the human sensory experience.

The article was authored by Dongju Lee, Hyemin Jeon, Yoonseo Kim, and Youngseung Lee. The research team is based at the Department of Food Science and Nutrition at Dankook University in Cheonan-si, Republic of Korea. The article, titled "Integrating Cutting-Edge Technologies in Food Sensory and Consumer Science: Applications and Future Directions," was published in the journal Foods on December 5, 2025.

Read More Here: https://www.mdpi.com/2304-8158/14/24/4169