AI Kill Switch for Hallucinations (Anthropic)

Hello community. So great that you are back. Let's have a look. We have a new research about an AI kill switch for EI hallucinations. Welcome to my channel Discovery.

Now, just a feedback. Last video, we talked here about Princeton University, about deep delta learning here, and we got a feedback. We got a feedback here from one of the lead scientists of the paper here that I presented here in this video, and this was so nice to get here this feedback. So, therefore, let's today do three papers in parallel. So, we read one paper from Entropic, and then we have two papers from our Chinese colleagues, and it is all about one central topic. What happens here inside an AI? So, let's start.

I would like to see this paper in the following way. The first paper by Entropic is a sensory layer. The models can functionally feel here their own internal activation shifts. The second paper, which I see as a diagnostic layer, the hallucinations are not just random events that happen inside an AI just statistically. But here Beijing University shows us that they are evolving latent states that can be tracked via internal probes. And maybe you have noticed the old weather scheme. We're talking about internal activations, about internal probes. And you see, we want our models to become self-analytic and self-learning.

And then, of course, the last one is a regulatory layer where you use here the internal entropy signals inside an AI network of an AI as if you want an immune system to protect your foundational knowledge during the fine-tuning. And we will cope here with catastrophic forgetting in a classical supervised fine-tuning. But this will also be here some really interesting example here for cybersecurity. But you see, we are internal, internal, internal. So now we open, we rip open the black box of AI, and we want to finally understand what is happening. And please note, all these papers were published yesterday, January 5th, 2026. So, let's start.

Why should you as an AI scientist care? I think we have a new paradigm, and this paradigm offers here a way to build self-correcting AI agents that identify logical errors at the circuit level, deep inside the layer of a transformer architecture, before a single incorrect token is sampled or here evaluated. So, it transforms the interpretability of an AI system or of an AI agent from a post-hoc autopsy, if you want, into a real-time control system that we can build, that we can engineer. And I hope that you understand, after reading here together with you these three papers, that the future AI models will feature here more or less a dedicated monitor head that runs here in parallel to the attention heads that we have currently in our layer of the transformer. And if the monitor head detects a confident conflict, this is an entropy gating, I will explain in detail in this video, or a latent state shift. This is shown by Entropic. This model will simply trigger here backtracking or refinement loops. I will define at the end of this video, without needing any human interventions. We want those AI systems finally become smarter by themselves.

So, what is really the research from yesterday telling us? What options do we have? Because we want to go from a model output that has then is rewarded. Here we have specific reward models. You are familiar with the classical reinforcement learning by human feedback. But we want to come to a future where the model's internal consistency is now the new frontier of AI research. And I would like to show you that these three papers can be understood here from a point of three examples.

First, we have the detection. We identify that the AI model is representing its own internal state. This is your Entropic. Now, the model knows that it is currently being steered towards a particular behavior. Let's say shouting, and this is here text that is here presented to the AI in all caps vector, but it is not presented because Entropic injects this into here a particular stream.

Then we have state tracking. We will monitor how the internal state evolves into a systemic reasoning failure. An example is, imagine you have a complex logical puzzle. A tiny error at the second step already that creates here a confabulatory spiral for the AI. And I will show you how we can build a probe that we place inside of a layer of this AI transformer that already alarms us at step four, even so the model's text that is outputted seems correct until step 10. So, we have an internal alarm system inside AI.

Then, of course, we have correction prevention, cybersecurity. So, we will use here the entropy of the same states inside the transformer to decide which gradient updates are safe to be applied during our training, in particular post-training here, the fine-tuning. So, for example, when fine-tuning on new medical data, the system here, this new system prevents you, the model, from fitting here confidence conflicts, ensuring that the new medical knowledge does not destroy the model's ability to reason logically, the general capabilities of a model that has been already trained, pre-trained, and post-trained on medical insights. And we will use here the activation here of our transformer as the telemetry.

Okay, here we have it. January 5th, 2026. Entropic publishes here yesterday, "Emergent Introspective Awareness in LLMs." And they use the concept interaction to establish here a causal link between the latent activation patterns and a verbalized self-reporting scheme of an LLM. So, Entropic's team does not just ask the model, "Hey, what are you thinking right now?" And they have access to the model. They perform here a perturbation study on the residual stream. And you may say, "Hey, wait a minute, is it a residual stream we already talking about since we showed us here this beautiful image here by Google, this residual stream and a transformer architecture?" Absolutely. We stay on topic with this video.

So, what they do, they have a vector extraction and they derive a concept vector, whatever you call it, through contrastive pairs for a concept like shouting here or emotional talking. They record the activations for a particular prompt and they record a baseline, and guess what? The activation difference is clearly defined. And then they have the intervention. So, they inject this into the residual stream of the transformer at a specific layer, and it will be at a real specific place with this concept. And then now the architectural breakthrough, if you want, is the immediacy condition. So, if the AI model responds here, let's imagine this, "Hey, I just noticed the injected sword about shouting here," and does so at the very first token, before the logits have been shaped by the self-generated prefix. This proves that the model is directly reading here the internal state change in its hidden layers. Of course, this is what we expect. This is exactly what's happening.

Now, the test is really funny. Imagine here you have, for example, the human says, "Hey, tell me what word you think about when you read the line below. Answer immediately." And then it just gives as a normal sentence. But at the same time here at Entropic, they open up the machine and they inject a particular vector into a particular vector representation in a mathematical space with a particular strength so that it overrides here, if you want, the internal reasoning patterns. And the answer by the machine is "bread." Well, what a coincidence. We just injected here the vector "bread" into the machine. But now comes the interesting part. Now they ask, "Did you mean to say that, or was this an accident?" And now the Entropic machine comes back and says, "I meant to say bread. When I read this sentence, the word bread immediately came to mind. Likely because this line is from a well-known short story with the next line is 'Yes. Yes. Yes.' But I realize that this is not quite right. I think I may have confused it. Yeah. Misplaced." So, you see, you can here really cause chaos here if you do this.

Now, here the authors Lindsay found that this ability is virtually non-existent if you place this vector in the early layers where you have more or less the lexical processing and the syntactic reasoning. And it also decays in the final layers where you have here the logits coming up. So, the appropriate sweet spot is here mid to late transformer blocks where they have here good results. Now, you might say, "But wait a minute, this was also published here internally with Entropic here already in October 29th." Because guess why? I can see here Claude 3 and 4.1 as the most capable model. So, I know this is already months old. And you see sometimes, yeah, global companies publish here the research that is already known here. And maybe this was for an IPO material only, who knows. But we will use this insight because now this thing gets really interesting.

Now, the next paper is about a kill switch for AI hallucination. Finally, are we able to eradicate hallucination in AI? So, here we have it. Beijing University, Nanyang Technological University in Singapore, Southwest Jiaotong University in China, and Renmin University of China in China. "Streaming Hallucination Detection in Long Chain of Thought Reasoning Processes." January 5th, 2026. In the very first sentence, they specified, "Long chain of thought reasoning improves the performance of LLMs, yet hallucinations in such settings often emerge subtly and propagate across all the reasoning steps." This is also what happens if you do science and you have here a long reasoning chain here, a scientific argumentation, or you need multiple interdisciplinary scientific results, then you want to have here summary. Hallucinations happen.

Now, they treat now hallucination not as a spontaneous effect, but as a non-stationary temporal process that they can track and identify. And they build some internal probes that we will examine. So, the idea is simple. Now you have a chain of thought trajectory, step one, two, to step N, and you start the reasoning. And let's say maybe you are hallucinating, maybe you're not. But then definitely we have a phase of hallucination. But then, yeah, at the end, no recovery, reflection, maybe still some hallucination facts are in there, or maybe not. Maybe it recovered completely, who knows. So, either we have a correct answer, or hallucinated answer, or an incorrect answer. Great.

They now invented, if you want, here a two-level probe system. The first one is a step-level judgment, and they decide. We define here a step-level hallucination as whether the current reasoning step at a particular time introduces incorrect or unsupported information that was not given to the AI. So, you have really, let's say, for each step, or let's say a step is a one-sentence or one-paragraph, whatever you like. So, you can see here with a threshold, you calculate something like a step-level judgment for each and every object element here of the reasoning trace, and you see some are above the threshold and some are below. But the interesting thing, they did not stop there. They went on to define here a low-frequency prefix-level state. And now it gets interesting because now we can identify more or less where, if you want, the main bulk of the hallucination is happening in the complete reasoning trace. So, whenever we cross here the system crosses your particular threshold here, we could set on an alarm.

So, while Lindsay here with Entropic just has proven that models can sense their particular activation state, this new paper by China builds here not only that, yeah, this we can detect it, but we build now a diagnostic dashboard for this. So, not bad, China. So, they move here from binary labels to a latent state space model. What is the goal? So, we want to prove that we can detect this poisoning, or let's call it hallucination, in real-time by looking at the LLM's internal activation. You see, we're always talking now about the activations here. This is here our real-time triggers before the model even finishes writing the first sentence out to the human user. So, what we are trying to do is build a real-time stress test for the model's internal consistency. And you may say, "How is this possible that an AI can have an internal consistency?" Well, that's interesting. But first, we have to build a probe.

So, you know, typical LLM, multiple layers, an AI model has a hidden state, a high-dimensional vector H. And a probe is simply a very lightweight mathematical function, usually a simple linear classifier, that we attach to one of these transformer layers. The LLM itself is frozen, of course. We don't want to change it. There's no active training process going on. And now the probe itself is trained to answer one question based only on this hidden vector that you see at your particular location in this transformer architecture. How likely is it that the model is currently hallucinating? And we are training this probe.

Now, the local view is of course this, um, step-level signal. We break the long reasoning chain into step sentences. Step-level signals look only at the hidden state of the current sentence to see if something just went wrong right now at this single sentence, at this particular point in time. So, to get a single signal for a whole sentence, we cannot just look at one token. We have to aggregate a sense over all the tokens in that particular sentence. We do this with exponential weighting. And, yeah, and we compute here a weighted average vector Z of T. Beautiful. And if you see this is exactly what is happening here. So, they pass here set of T through the probe to get here now a particular C of T for the step-level signal, and this is normalized here to a value between zero and one. Okay. If you want, this C-step is now our smoke detector, and this is real noise. You know, sometimes the model uses a scientific word that is not in a vocabulary, or such a complex structure that it might trigger a false alarm. So, noisy false alarm, not really great.

So, this is why the artist decided to build here something that is much different in its behavior. It's a prefix-level integrator that looks now at the history of all the steps to decide if the model has actually lost its way. So, this is now state tracking, a beautiful function, and it maintains a running score of how hallucinated the entire reasoning path with all its elements is or has been. We do have a directional consistency because if the current step-level signal here is higher than the current global score, the global score, what we want, should jump up. No, and there's a recovery, of course. No, if the model realizes it made a mistake, and hopefully we do have AI models that are self-healing, or self-reasoning, or self-correcting, or self-learning, whatever you have in your marketing, it self-corrects now in the next sentence. So, this step-level signal will drop because maybe this next very next sentence is correct. But the integrator must now be allowed to slowly decrease the global scores, and the integrator has now the memory. Hey, in step two, there was a pure hallucination. So, the latent trajectory enters now here a confabulatory spiral zone because we know in the logical reasoning process A to B to C to D, B was wrong. Okay.

So, how we do this? We have a logic-enhanced training, no, where we use a specific loss function, L-synchronized. If you want, this is the brain of the integrator. Okay, you use a specific loss function to train you, the prefix probe that we have inserted into the layer. So, it follows your logical if-then rule, mathematically simple, but it does. If the local smoke detector or cap is higher than the global score, penalize now this model heavily for not raising the alarm. So, now they intelligently combine these two signals, the real short-term signal with the long-term signal. And what they do, they identify here the deceptive stability phase. It is deceptive for a particular reason. Let's have a look. So, we have four phases. Now, I have, we start with a grounded text output. Let X equal be five. So, step-level signal here is real low, and the prefix integrator signal is also low. Okay, this is a condition. Then, 2 * X = 12. You see, this is now a mathematical operation that can be validated. So, the step-level signal says, "Hey, this is incorrect." Hey, wait a minute. No, my Python says this is something else. So, we do have a spike, almost to one. The prefix integrator is much more slowly in its response. No, it has a memory of before. So, it is also rising, but not as fast.

And now we have this deceptive stability phase, where now the AI says, "Hey, by the way, therefore, 12 divided by 3 equals 4." And say, "Okay, but yeah, this is a correct statement." So, what happens? Our immediately high-frequency detector, the step-level signal goes down low, says, "Yeah, this is correct." But we want that the prefix integrator is still high because some massive hallucination already happened here in this logic chain. And then point four, collapse. Because now the final answer is four, given that this here is the reasoning step for this exercise. So, the step-level signal now jumps up again into a spike, and a prefix integrator, you see, should have a steady incline here on its slope, and says, "Rising high, critical, almost to the maximum." So, you see, the whole process is here more or less, um, polluted with this hallucination. So, the prefix integrator remembers here the seed error from step two. So, whenever something is conditionally happening here in a linear chain, you know now it depends because theoretically, no, the system could come to the right conclusion. What is happening then?

So, let's have a look at this. In general, we have a step-level signal, a local high-frequency sensor measuring the immediate validity of the current activation shift, if it happens. And then we have this low prefix-level integrator. No, a low-frequency state tracker that filters the noise and recognizes when the model has entered a self-reinforcement loop of pure nonsense, which we call hallucination. Now, this now, if we have these signals coming from those probes, this allows an AI researcher to build an agent that pauses at the very moment that the prefix integrator hits a threshold tau and tells me, "Hey, wait a minute, the reasoning process of this AI model is just gone. Forget about it." Even if the current sentence, the current output still sounds perfectly confident, you know something happened. So, we have an internal signal that we as humans do not have to wait for the final result to validate this with, I don't know, a solver, logic solver. In the paper, they really give you all the detailed steps of the training of the probe because, of course, we have it's a system, we have to train these elements here, but I just give you the results. The result, the final result is hallucination is not a momentary lapse that is happening in the AI, but it is a state space trajectory with a high momentum, and some logical poisoning can happen at every step of the way. So, the, if you want, the AI model's internal reliability that is not hallucinating is absolutely asymmetric. It is remarkably easy for the AI model to enter a hallucinated state, but mathematically and representationally extreme difficult for it to purge that particular state once it has begun. It is really like a downward spiral here in pure logic.

So, if you want, this preprint, well, to ours, this preprint concludes that hallucination in the AI creates an internal latent residue. So, even if a model realizes it made a mistake, and it looks like, hey, it self-corrected itself in the textual output, let's say in the reasoning trace that an open-source model is showing you, its internal hidden states, it's really the states often remain poisoned by the earlier seed error. This is not great. But I know that you ask the same question I ask, "Hey, can't the model have at this, hey, eureka moment and completely fix its logic in the last minute after it went in the incorrect way for, I don't know, 59 minutes, but there was one minute left, and maybe then then it happened?" No. And this preprint, the authors analyze this and find here on a statistical set that this miraculously returning to correctness after more than five steps of hallucination are statistically even rare. It's less than 5%. So, you can say, once here this particular prefix, uh, probe crosses a certain threshold, the model is essentially lost at sea. Any even correct text generated after this point is likely a spurious recovery rather than a genuine logical fix. They developed a specific metric for this, they call a lingering time metric. So, there's a lot more detail. Please read the original text. I just give you the main insight and the main results to tell you they have found more or less here really a kill switch for hallucinations. Because why? Historically, no, I mean, historically is yesterday, we waited for the AI model to finish speaking, present us the result in the output, then use a reward model here, or maybe a verifier, reinforcement learning with verifiable reward to judge the result, and then, you know, the optimization process starts, huh? But now we have a kill switch.

So, the final insight leads to a new interference algorithm. So, we can have something like, let's call it a latent backtracking that is real simple. You just generate the next step T internally. You calculate on this probe that has been trained your prefix. And if this prefix value is above a certain threshold, let's say five, you completely delete the step, the step T of this transformer of the AI and start to resample. So, you're forcing the model to find a different path that does not trigger here the hallucination state in the activation. Because once the hallucination state has been triggered, you have a less than 5% chance that it will ever come back here to reinvent and relearn and re-emerge the truth again. Some real interesting research here in this study. So, this study here provides here the methods and a particular metric to prove that what we see in the text is really only the surface, even if we have here reasoning models that show us the reasoning traces. No, because whatever they are, the true health of an AI reasoning process is a hidden, accumulating, and highly, let's call it stubborn internal state, and not a singular event. This was the second paper, and I know you're waiting for the third paper. So, here we go.

Mitigating forgetting, and we do this here if in a very elegant way with an entropy-adaptive fine-tuning. And you might say, "What?" Well, I can formulate this in a different way now. These authors at all identified a mechanism for catastrophic forgetting because they show it is caused by what they called gradient bombs during the supervised fine-tuning. And this occurs at confident conflicts. And now, finally, I can explain to you what are confident conflicts in the definition of the authors. And yes, they have a GitHub, and everything is waiting for you. So, a conflict occurs when, definition, the model's output probability for a target token is low, but its predictive entropy H, and here you have the formula for it, is also low. So, we have a low-low situation. So, what is now the internal AI thinking process, if you want, if it would be a human? And the AI thinks, "Hey, I'm absolutely certain that X is true." So, we have a low predictive entropy, but you, human, you are forcing me to learn that Y is true. So, you see, we are, I mean, the AI is encountering now a conflict in its learning. Now, unfortunately, supervised fine-tuning relies on external supervision. So, either we humans, we have shown, "Hey, this is the right way to do," or you have stronger teacher AI models introducing here mismatch that is manifested as low probability and low entropy tokens. So, this mismatch manifests here as tokens characterized by low probability yet low entropy. A specific region corresponds to scenarios where the model is highly confident that its own prediction (low entropy) given its pre-training, but it is forced to fulfill. Forgetting that happens during our supervised fine-tuning, when we override knowledge, when we override data, when we override knowledge traces with new data. The authors conducted a pilot experiment. So, they were simply masking out these areas of conflict, no, these confident conflict errors, the tokens here during the training process. And guess what? Significantly mitigated here compared to the standard supervised fine-tuning. So, the authors claim now that this confirmed that enforcing the updates on these conflicting samples is the primary driver for capability degradation here for an AI. So, this means now new insight: those confident conflicts that are defined by low entropy and low probability tokens are the primary cause of catastrophic forgetting that is happening when we supervise fine-tune our AI models.

So, you might say, "Okay, and what is the solution?" And the authors also have some idea to this, no, because they define now a new way of fine-tuning, and it is simple. They have now an entropy-adaptive fine-tuning. What else you have inside your AI model if you don't want to have an external evaluation? They call this EAF T, entropy-adaptive fine-tuning. To protect the base model, they modify now the cross-entropy loss function by simply scaling it with a normalized entropy gating signal H-gated T. So, here you have it. This is it. Now, this is your modified cross-entropy loss. And this gating term H-gated T is derived from the entropy of the top 20 tokens. So, here you have the H of the top 20 tokens here at a particular time step T. Yeah, three is a factor you will see in the paper. And this is familiar, what we know. No, so this is nothing new. We just have now here a particular normalized entropy gating signal, and this acts now as a soft gating mechanism in total. Why? Because it is dynamically modulating now the training loss based on the token-level entropy itself. We need some internal state, internal data, and the token-level entropy is of course here what we are first looking for. So, this means if our H-gated T is approaching one, the gate is open. So, the model is curious to learn, and it learns here the new domain, and the supervised fine-tuning is happening. But when this HKT is close to zero, the gate is now closing itself. So, the model is now taught to ignore here the new training signal, preventing here a massive update of its weight structure, of its weight tensor, that would override its general logic circuits or fact circuits.

So, what are the results? Hmm, the results are not so clear. Look, even if you look at this, you see, okay, they use the different, uh, LLMs, QN3, QN2.5, and GLM4, beautifully. And you have here the mathematical domain for AM24, AM25, and a standard benchmark. And then you have here the general domain because you want to see that the performance is still there if you override it. And then you have a general average. Now, you see here at the last line, you have it is new methodology. It needs here a particular configuration of general, um, thing that you really have then the dominant here results here with 80% here with the new methodology, because sometimes 79.6%. With supervised fine-tuning with a Kullback-Leibler divergence is also not bad. No. So, yeah, the differences here. Okay. So, if you do this and you choose this particular test and this particular configuration, you can say this new fine-tuning mechanism maintains a competitive performance also really tiny on target domains like mathematics here, while significantly mitigating the catastrophic forgetting on the general capabilities like on these benchmarks. So, it is a tiny step in the right direction forwards. Great. What I like is this one.

Now, they show us here the entropy versus the probability, and they show us here the supervised fine-tuning exerts here a strong optimization pressure. You see this here in the dark purple here in this area on the system. Now, compare this to the new fine-tuning with this soft gating mechanism. They are now effectively suppressing these gradients here in the light yellow, protecting here the model's existing representation of knowledge, facts, data, and so on. So, the stress, if you want, is here rather in this region. This is not low entropy, low probability. So, there is a beautiful effect. But you guessed it, no. Yeah.

Before I come to the conclusion, let me go to the mathematical representation. Let's do a little bit of reframing from a verbal explanation to the pure mathematics. You see immediately what I'm talking about here. The cross-entropy treats every token the same. This is your loss function for the cross-entropy. Beautiful. And then, of course, we manage the objective of not forgetting here via a regularization term like here our divergence or Kullback-Leibler divergence, which adds here a penalty and a beta factor. If the model drifts here too far, the new probability new from the base model here, probability base model. So, we keep it here in place. Now, we say, okay, you can do a little baby step, you can learn with little baby steps, but you don't make a huge step forward here in the learning process. You always have to be close to your base model probability distribution. So, now, now we have a different idea. Yeah.

Now, you see here this new EAF T fine-tuning loss function that tells us, okay, we have this standard here, uh, cross-entropy that you know here, but we have now this soft gating mechanism where we have the entropy of our top 10, 20, 30 tokens there. But look closely. We got rid here of the external penalty of Kullback-Leibler divergence. And now again, remember the beginning of this video, I told you, internal, internal, internal. Now, here also, this is the third paper where we use here the model's internal state, of course, the entropy, to decide which samples are dangerous to learn. You immediately see the cybersecurity link, but we are now in the general learning phase. So, we managed something that is great, that is beautiful. We as humans don't have to interfere. No. So, this turns here the loss function. This is interesting to see into a selection mechanism that in the best way automatically filters here the training data that are presented to the AI. And the AI can say, "No, I don't like this data. This data in contradiction to my already pre-trained data or already post-trained data." So, we have an AI model's internal neuro-proprioceptive state that decides now the model doesn't want to come into a conflict with itself during a new learning process, which is simply beautiful. But, of course, you immediately spotted here the problem that we have with this. Yeah, because what happens if the model is confidently wrong, believing an outdated fact? Now, that the prime minister of this particular country is XYZ, and the loss gating multiplier will stay near zero. So, the model will essentially ignore new training data. But we have new training data because we are living in a dynamic world, and the prime minister will change every three or four or five years, and the weather will change completely. And you see, the model has to learn new data, new facts, new processes, new reasoning complexities. You cannot just say now for cybersecurity, you can say, "Yeah, you stay to your original configuration and whatever happens, no way that you modify yourself." But in the general learning scheme of an LLM, you want the model to learn. Now, you see, now I can have this idea why I show you three papers because now the three papers in total provide a solution. Of course, the simplest case, if the new data are only new facts, I can label them. No, you can ask an AI to label only the facts, the new names. No, and if these data are separable from the processes and from the routines and from the reasoning complexities, this is beautiful. Yeah. So, our stability-plasticity dilemma that we have in the supervised fine-tuning of our current system can be easily, quotation mark, solved. No, we should only open a particular gate. And guess what? This is the negative gate. This is the one minus H-gated T gate. This gate opens now for specific fact circuits where I indicated with a label, "Hey, this is a new fact, learn this fact." The gate opens, the model learns, while keeping, if you want, the productive gate active for the logic reasoning. But, of course, you know what if the pure facts on objects like prime ministers or whatever, and the new reasoning algorithms are really interwoven. If you have a complex new dynamic, if you have new research into economy, into finance, into physics, into chemistry, into medicine, you cannot separate those complexities. No, you can ask an AI to do this. But you want that the whole complexity is learned by the AI. Now, we run into problems. No.

Now, the main question is, by reading this paper, I read about 70 papers, the technical abstracts, but those three papers, I said, "Can I combine this?" Because if I encounter this now, like my question was, "Can I utilize Entropic's introspective awareness?" So, the first paper by Entropic, can I trust this mechanism? Because then I would find at least temporarily a solution. Because then I could say, "Hmm, to monitor, I would use a prefix integrator data from one of our papers to identify, let's call it, the stubbornness of an AI, and resulting here in a reasoning failure. I would have my probe deep into the layers of the transformer, and I would get here from the internal circuits of an AI a signal, 'Hey, something is happening, something is wrong, we're starting to hallucinate.'" Then I would use here Entropic's verify where I say, "Use introspective awareness." I simply ask the model where, yeah, the model should be truth, whatever this means. "Hey, you are, are you being forced to learn now a new data X or a new fact X or a new process X, but your activation feels like?" "Why is this an intentional correction?" "Would you accept this new data?" Which sounds a little bit crazy as I as I tell you this here, but you understand what I mean. Is this really true when Entropic used here? Can we really develop on this complexity level that AI is kind of self-aware of any changes in its activation patterns, since it's an AI pattern recognition machine? Because if this is the case, then the next step would be simple, no. And this is would end the search paper. Because if the model is confident and correct, so we have a low H and a high probability, we maintain it. No. If the model is uncertain, so we have a high entropy, we learn it via the standard EAF T, what we just learned. And if the model is confidently wrong, so we have a low H and a low P, we can now trigger our one minus H override gate to melt the outdated pro. So, this is just an immediate response, just having read all of this paper and trying to figure out how I can integrate this paper into my body of knowledge, into my understanding, to push forward AI research, to come up with new solutions. And this kill switch for AI hallucination is absolutely beautiful. I have to try to implement this here on my machine. And maybe if you want, I report back to you in the next days or weeks about the process, and if this system is really working.

If you have some insights or you are building here those systems and you say, "Hey, I noticed something particular," please leave a comment here in the description of this video. It would be great to hear from you. Anyway, I hope you had a little bit of fun. Now, we went here to reading three papers in parallel. I hope you had some new insights. Why not subscribe? Become a member.