Episode #257

AI That Evolves: Solving the Preference Problem

Why do AI recommendations feel stuck in the past? Discover the technical hurdles of real-time learning and the future of personalized agents.

Episode Details
Published
Duration
25:28
Audio
Direct link
Pipeline
V4
TTS Engine
LLM

AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.

In the latest episode of My Weird Prompts, hosts Herman and Corn sit down in Jerusalem to tackle one of the most persistent frustrations in modern artificial intelligence: the "amnesia" of large language models (LLMs). The discussion is sparked by a query from their housemate Daniel, an engineer who is tired of the manual labor required to keep AI recommendation systems updated with his evolving personal tastes. Daniel’s dilemma serves as a springboard for a deep dive into why AI often feels like a friend who only remembers who you were years ago, and what technical breakthroughs might finally allow models to learn in real-time.

The Static Model Problem

Herman explains that the primary reason AI assistants feel "frozen" is the nature of their training. Most frontier models, such as GPT-4o or Claude 3.5 Sonnet, are trained on massive datasets and then effectively locked. This creates a "knowledge cutoff," where the model’s internal weights—the connections that form its "intelligence"—do not change based on new interactions. While these models can access the internet to find facts, they lack a native way to integrate a user’s shifting preferences into their core reasoning without being manually fed that information every single time.

The Pitfalls of RAG and "Context Rot"

To solve this, many developers currently rely on Retrieval-Augmented Generation (RAG). In a RAG setup, a user’s history is stored in a database and injected into the prompt as needed. However, Herman warns of a phenomenon he calls "context rot." As a user’s history grows, the prompt becomes cluttered with old, irrelevant data. Even with the massive context windows available in 2026, LLMs often suffer from the "lost in the middle" problem, where they struggle to prioritize recent feedback over older, potentially obsolete information. This leads to high latency, wasted tokens, and a degradation in reasoning quality.

The Barrier of Catastrophic Forgetting

A common question arises: why not just update the model’s weights every time a user provides feedback? Herman points out that this leads to "catastrophic forgetting." When a model is forced to learn highly specific new data—like a niche movie preference—it can inadvertently overwrite the patterns that allow it to perform general tasks like solving math problems or speaking different languages. The model becomes a hyper-specialized but ultimately broken tool, losing the general intelligence that makes it useful in the first place.

New Horizons: Test-Time Training (TTT)

The conversation shifts to one of the most promising areas of AI research: Test-Time Training (TTT). Unlike standard models where hidden states are static, TTT models treat these states as tiny, adaptable neural networks. When a user provides input, the model performs a small amount of "gradient descent" during the inference process itself. This allows the model to compress the context of a conversation into temporary weights. Corn likens this to "learning to play the piano in the middle of a concert." While TTT offers a way to handle massive amounts of data with constant latency, the challenge remains in making these updates permanent without triggering the aforementioned catastrophic forgetting.

The Rise of Self-Editing Memory

For a more immediate solution, Herman points to architectures like Letta (formerly MemGPT). These systems treat the AI as an operating system with tiered memory. The agent can proactively "write" to its own archival memory, essentially keeping a digital diary of user preferences. When a user expresses a dislike for a specific genre or topic, the agent updates its own notes and searches them during future interactions. This creates a self-correcting loop that feels more like a partnership than a static tool.

The Personal "Adapter" and the Strategic Flywheel

The ultimate vision discussed by Herman and Corn is the implementation of a "strategic flywheel." This involves using Parameter-Efficient Fine-Tuning (PEFT), specifically techniques like LoRA (Low-Rank Adaptation). Herman suggests a future where a user’s local model undergoes a brief, automated fine-tuning session every night. This process would take the day’s feedback and create a "personal adapter" that sits on top of a larger frontier model. By the next morning, the user has a fresh version of their AI that has "digested" their new preferences.

Conclusion: Bridging the Gap

Herman and Corn conclude that while the "holy grail" of a perfectly evolving AI is still being refined, the tools to build it are becoming increasingly accessible. By combining the massive intelligence of foundation models with local, personalized fine-tuning and self-editing memory, developers can bridge the gap between static code and the fluid nature of human taste. As we move further into 2026, the transition from AI as a tool to AI as a truly adaptive partner is no longer just a theoretical dream, but an engineering reality.

Downloads

Episode Audio

Download the full episode as an MP3 file

Download MP3
Transcript (TXT)

Plain text transcript file

Transcript (PDF)

Formatted PDF with styling

Episode #257: AI That Evolves: Solving the Preference Problem

Corn
Hey everyone, welcome back to My Weird Prompts. I am Corn, and I am sitting here in our living room in Jerusalem with my brother, looking out at a surprisingly clear afternoon. You can actually see the Judean Hills today, which is a rare treat after the dust we had last week.
Herman
And I am Herman Poppleberry, ready to dive into some serious engineering questions. It is good to be here, Corn. The air is crisp, the coffee is strong, and the prompts are getting weirder by the day.
Corn
They really are. And we have a really interesting one today from our housemate Daniel. He has been busy lately, has he not? He was telling me about that agentic United Nations project he was working on, where he had different artificial intelligence agents representing different countries. It was absolute chaos in the kitchen for three days.
Herman
I remember that. He was trying to see if they could actually negotiate a climate treaty without human intervention. It was a fascinating experiment in game theory, even if the agent representing one of the smaller nations ended up trying to buy all the carbon credits with digital goats. I think it was a hallucination triggered by a specific cultural data set he used, but it really highlighted how these agents can go off the rails when they start optimizing for local rewards.
Corn
Right, he is always pushing the boundaries of what these systems can do. But today, he is asking about something that sounds simpler on the surface but is actually a massive technical hurdle in the world of artificial intelligence. He wants to know why he cannot just have a recommendation bot, like for Netflix or YouTube, that actually learns from him in real time and stays updated without him having to manually feed it his life story every single time. He is tired of the manual sifting.
Herman
It is the classic preference problem, Corn. Daniel pointed out that if you just give a model a description of what you like, it gets boring. It gives you the same five things over and over because it is working off a static snapshot of your personality. It is like having a friend who only remembers who you were in high school and keeps trying to buy you the same brand of sneakers you liked ten years ago.
Corn
Exactly. And the core of his question is about the continuous training loop. He is looking for a continuous deployment model for artificial intelligence self-correction. He wants to know if there is a technology where the model is not just frozen in time but is actually evolving as he interacts with it. He wants a partner, not a tool.
Herman
That is the holy grail, Corn. The idea of a model that does not have a knowledge cutoff because it is constantly learning from every single interaction. But before we get to whether that exists, we should probably talk about why the current systems feel so broken for someone like Daniel who is doing high-level engineering. We are in early twenty-twenty-six, and yet we are still dealing with models that feel like they have amnesia every time you start a new session.
Corn
Well, he mentioned the knowledge cutoff. Most of these models, like leading ones such as GPT-4o or Claude 3.5 Sonnet, were trained on a massive batch of data, and then they were effectively frozen. They are like a library that stopped buying books two years ago. Even with the live search tools we have now, the internal weights of the model—the actual 'brain' of the system—do not change based on what you told it five minutes ago.
Herman
That is a perfect analogy. And while we have search tools now that let the model look things up, that is not the same as the model knowing you. Search helps with facts, but it does not help with the nuance of why you hated that one documentary even though you usually love the director. That nuance is what we call the 'latent space of taste,' and it is incredibly hard to capture in a static model.
Corn
So, Herman, let us dig into the memory side of this. Daniel mentioned storing rejections and tracking what you have already watched. In a typical setup, how is that usually handled? Because I know he has tried the standard approaches and found them lacking.
Herman
Usually, developers use what we call Retrieval-Augmented Generation, or RAG. You take the user's history, you stick it in a vector database, and when the user asks for a recommendation, you pull out the most relevant bits and shove them into the prompt. But as Daniel found out, that creates what we call context rot.
Corn
Context rot. I like that term. Is that when the prompt just gets too cluttered with old data? It sounds like a digital version of a messy desk.
Herman
Exactly. If the model has to read a list of five hundred movies you have already seen just to suggest one new one, you are wasting tokens, increasing latency, and honestly, the model starts to lose the thread. It might focus on a movie you watched three years ago instead of the one you watched last night. Plus, LLMs have a 'lost in the middle' problem where they struggle to pay attention to information buried in a massive context window. Even with the two-million token windows we see in models like Gemini 1.5 Pro, the reasoning quality degrades as you stuff more junk into the prompt.
Corn
So, if RAG is not the perfect solution for long-term, evolving preferences, what is? Daniel is asking for a continuous loop. Does a model exist that actually updates its internal weights based on a single thumbs-down on a YouTube video? Is that even theoretically possible without breaking the model?
Herman
The short answer is no, not in the way most people imagine it. If we updated the primary weights of a foundation model every time a user said they did not like a movie, we would run into a massive problem called catastrophic forgetting. It is the bane of continual learning research.
Corn
Catastrophic forgetting. That sounds like a movie title. What does it actually mean in this context? Is the model going to wake up and forget how to be an AI?
Herman
Pretty much! It means that as the model learns new, very specific information, like Daniel's preference for nineteen-seventies Italian horror films, it might literally forget how to do basic math or how to speak French. The new data overwrites the old patterns because the model is trying to optimize for the most recent feedback. It loses its general intelligence in favor of becoming a hyper-specialized, and likely very confused, movie critic.
Corn
So, it becomes a specialist in my weird movie tastes but loses its ability to help me write an email or plan a trip. That does not seem like a fair trade. But wait, I have been reading about some of the newer research lately. There is this concept of Test-Time Training, right? How does that fit into Daniel's dream?
Herman
Oh, you have been doing your homework! Yes, Test-Time Training, or TTT, is one of the most exciting developments in recent research. The idea there is that each individual input, or each test instance, defines its own little learning problem.
Corn
So, the model actually learns while it is processing your request? It is like learning to play the piano while you are in the middle of a concert?
Herman
That is a great way to put it. In a TTT model, the hidden states of the neural network are not just static vectors; they are actually tiny neural networks themselves. When the model processes your context—like your list of rejected movies—it performs a small amount of gradient descent on its own internal state during the inference process itself. It is essentially compressing the context of your conversation into a temporary set of weights that are optimized for that specific moment.
Corn
That sounds exactly like what Daniel is looking for. A model that adapts in the moment without needing a massive retraining session.
Herman
It is very close. TTT allows for constant inference latency regardless of how much context you throw at it, because it is 'learning' the context into its weights rather than just reading it over and over. But there is still a catch. Usually, those updates are temporary. Once the session ends, those weights are often discarded. It is like having a really good short-term memory that resets every morning. To make it permanent, you need a way to merge those updates back into a long-term model without causing that catastrophic forgetting we talked about.
Corn
Okay, so if the weights are hard to change continuously, what about the architecture around the model? Daniel mentioned he wants an AI bot that perfectly understands preferences without manual sifting. We live with him, and I know he has looked into MemGPT, or what they call Letta now. How does that handle the feedback loop?
Herman
Letta (formerly MemGPT) is probably the closest thing Daniel can use right now to build this. It treats the artificial intelligence model like an operating system. You have the core processor, which is the model, but then you have a tiered memory system. You have the 'core memory,' which is like RAM—always in the prompt—and then you have 'archival memory,' which is like a hard drive.
Corn
And the agent can actually write to its own hard drive? It is like the AI is keeping a diary of everything I like and dislike?
Herman
Exactly. The agent can decide, 'Hey, Daniel just said he hated that romantic comedy. I am going to call a tool to write a note in my archival memory that says Daniel hates overly sentimental plots.' The next time he asks for a recommendation, the agent proactively searches its own notes. It is self-editing memory. And in twenty-twenty-six, we are seeing agents that can maintain these personal profiles over time.
Corn
But I can hear Daniel's rebuttal already. He would say, 'Herman, that is still just a fancy version of RAG.' He wants the model to actually get smarter about his tastes, to understand the latent features of why he likes what he likes, not just read a text note that says he likes action movies. He wants that deep, intuitive understanding.
Herman
And that is where the engineering gets really deep. To do what he wants, you would need a system that does periodic, automated PEFT, which stands for Parameter-Efficient Fine-Tuning. Imagine a pipeline where every night, while Daniel is sleeping, a small script takes all of his feedback from the day, formats it into a training set, and runs a quick LoRA fine-tuning session on a small, local model like a Llama 3.1 8B.
Corn
So, he gets a fresh version of his own personal model every morning? Like a custom-tailored suit that gets adjusted every night based on how he moved during the day?
Herman
Exactly. We call this the 'strategic flywheel' or an AI feedback loop. Instead of waiting six months for a new model release from OpenAI or Anthropic, you have a local deployment that is constantly being updated in small, controlled bursts. This personal 'adapter' sits on top of the big foundation model. It is the best of both worlds: the massive intelligence of a frontier model and the hyper-specific personalization of a local fine-tune.
Corn
That sounds like a lot of compute power for just a Netflix recommendation, Herman. Is it worth it? Or is Daniel just over-engineering a solution for his Sunday night boredom?
Herman
For a single user? Maybe it is overkill. But if you are building a professional-grade assistant, or if you are Daniel and you just want to see if it is possible, it is the only way to get that true self-correction. You have to bridge the gap between the frozen foundation model and the ever-changing human. And honestly, with the hardware we have in early twenty-twenty-six, running a LoRA fine-tune on an eight-billion parameter model takes minutes, not hours. It is becoming very accessible.
Corn
You mentioned something earlier about the data integration nightmare. Daniel mentioned geo-restricted content and different platforms like YouTube and Netflix. How does an agent handle the fact that a recommendation is useless if I cannot actually watch it here in Jerusalem?
Herman
That is where 'Knowledge RAG' comes in. Instead of just using a vector database, which just looks for 'similar' things, you use a graph database like Neo4j or RelationalAI. In a graph, you can map out relationships: 'This show is on Netflix,' 'Netflix has these rights in Israel,' 'Daniel has a V-P-N that can access the U-K.'
Corn
So, it is not just 'you like sci-fi,' it is 'you like sci-fi that is currently available to you and does not require a third subscription service.'
Herman
Exactly. A graph database allows the agent to traverse these complex rules. It can see that Daniel likes movies with a specific actor, but only when they are directed by a certain person, and only when the setting is in space. That kind of relational data is what makes a recommendation feel perfect instead of just okay. And the agent can update this graph in real time. If a show leaves Netflix, the agent updates the edge in the graph, and it stops recommending it.
Corn
Okay, so if Daniel wanted to build this today, what is the actual blueprint? He has the data, he has the graph. How does he solve the frozen model problem once and for all?
Herman
He uses a hybrid, multi-model architecture. Step one: Use an agentic framework like Letta to manage the long-term state and reflection. Step two: Use a graph database to map out the actual availability and complex preferences. Step three: Use a local, small model for the daily LoRA fine-tuning. This small model acts as a 'preference filter.'
Corn
So, the big model—let us say a leading frontier model—is the one that knows everything about the world, and the small, local model is the one that knows everything about Daniel.
Herman
Exactly. When Daniel asks for a recommendation, the big model suggests twenty things that fit the general vibe. Then, the small, personalized model—which has been literally shaped by his past choices—ranks those twenty things and says, 'Out of these, Daniel will only actually like these three because I know his specific nuances regarding gritty lighting and slow-burn pacing.'
Corn
And then the agent checks the graph database to make sure those three are actually watchable in Israel.
Herman
Precisely. It is a multi-step workflow. And the final piece is the 'justification.' The agent should say, 'I am suggesting this because it has the same gritty atmosphere as that other show you liked, and it just became available in Israel this morning.' That transparency creates a new feedback loop. If the reasoning is wrong, Daniel can say, 'No, I did not like that show because it was gritty; I liked it despite the grit.'
Corn
And that feedback goes right back into the next night's fine-tuning session. It is a closed loop.
Herman
It is. And it solves the 'manual sifting' problem because the agent is doing the sifting based on a model that is literally a reflection of your past behavior. We are moving from generative artificial intelligence to agentic, learning systems. Daniel's question is right at the heart of that transition.
Corn
I wonder if this connects back to what we talked about in a previous episode, about mesh networks and local compute. If you have this personalized model, you really want it running locally for privacy reasons, right? You do not want your entire psychological profile sitting on a corporate server just so you can find a good sitcom.
Herman
Absolutely. Privacy is the biggest driver for local fine-tuning in twenty-twenty-six. If Daniel builds this, he can keep his preference weights on his own hardware here in the house. The foundation model provides the language and reasoning capabilities, but the personal 'soul' of the AI stays with him. It is much safer than giving a giant corporation a map of your subconscious tastes.
Corn
It is funny, we always end up back at the idea that the best AI experiences right now are the ones you have to build yourself, or at least heavily customize. The 'one-size-fits-all' approach is what leads to those boring, repetitive recommendations Daniel is complaining about.
Herman
It is true. True personalization requires a feedback loop that is actually listened to. Most commercial systems just use your data to sell you more stuff; they do not use it to actually get smarter for your benefit. Daniel is essentially building a 'sovereign agent' that works only for him.
Corn
I think Daniel is going to be excited to hear that. He loves a good engineering challenge. I can already see him setting up a local server in the corner of the kitchen—right next to the air fryer—to start fine-tuning his own movie critic bot. We might need to upgrade our internet again. Remember that previous episode? We are always chasing that gigabit dream.
Herman
It is a never-ending cycle, Corn. Just like the feedback loops we are talking about. But if it means I do not have to watch another documentary about ancient grain silos, I am all for it.
Corn
Hey, you love those silos! Do not pretend you do not. You spent four hours last week explaining the structural integrity of a Neolithic grain pit.
Herman
I mean, they are interesting, but a little variety would not hurt. Maybe Daniel's bot can find me a documentary about ancient irrigation systems instead. That would be a real step up.
Corn
Baby steps, Herman. Baby steps. Well, I think we have given a pretty good overview of why this is hard and where the solutions are. It is all about moving from a frozen state to a dynamic memory system, and eventually, to these tiny, frequent updates to personal models using techniques like TTT and automated LoRA.
Herman
Exactly. And for anyone else listening who is struggling with their AI feeling a bit static, just know that the industry is moving toward this lifelong learning model. It is the next big frontier. We are finally getting to the point where the AI grows with you.
Corn
It is an exciting time to be looking at this stuff. I mean, we are sitting here in early twenty-twenty-six, and things that felt like science fiction three years ago—like a model that literally learns while it talks to you—are now just a matter of choosing the right library and setting up a cron job.
Herman
It really is. I am looking forward to seeing what Daniel actually builds with this. Maybe he can finally solve the 'what should we watch for dinner' argument once and for all.
Corn
That would be the greatest engineering feat of the century. Daniel, thanks for the prompt. It really pushed us to look at the cutting edge of memory and training architectures. To all our listeners, thank you for sticking with us through the technical weeds.
Herman
Yeah, it was a great one. We know we get a bit deep into the weeds sometimes, but that is where the real fun is. If you want the surface-level stuff, there are plenty of other podcasts for that.
Corn
Exactly. If you are enjoying the show and these deep dives into the weird corners of AI and engineering, we would really appreciate it if you could leave us a review on your favorite podcast app or on Spotify. It genuinely helps other curious people find us in the algorithm.
Herman
It really does. We see every review, and it keeps us motivated to keep digging into these prompts. Even the ones about digital goats.
Corn
You can find all our past episodes, including the ones we mentioned today, at our website, myweirdprompts.com. There is a contact form there too if you want to send us a prompt of your own. We love hearing what you are working on.
Herman
And we are on Spotify, of course. Just search for My Weird Prompts. We are usually right there next to the other tech shows.
Corn
Alright, that is it for this episode. I am Corn.
Herman
And I am Herman Poppleberry.
Corn
We will see you next time. Thanks for listening to My Weird Prompts.
Herman
Until next time, keep staying curious and keep those loops closed.

This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities.

My Weird Prompts