Diffusion models have become the go-to method for text-to-image generation, producing high-quality images from pure noise. However, the inner workings of diffusion models is still largely a mystery due to their black-box nature and complex, multi-step generation process. Mechanistic interpretability techniques, such as Sparse Autoencoders (SAEs), have been successful in understanding and steering the behavior of large language models at scale. However, the great potential of SAEs has not yet been applied toward gaining insight into the intricate generative process of diffusion models. In this work, we leverage the SAE framework to probe the inner workings of a popular text-to-image diffusion model, and uncover a variety of human-interpretable concepts in its activations. Interestingly, we find that is completed, the final composition of the scene can be predicted surprisingly well by looking at the spatial distribution of activated concepts. Moreover, going beyond correlational analysis, we design intervention techniques aimed at manipulating image composition and style, and demonstrate that (1) in early stages of diffusion image composition can be effectively controlled, (2) in the middle stages image composition is finalized, however stylistic interventions are effective, and (3) in the final stages only minor textural details are subject to change.

Warning: This advice may cause you to rethink your pharmacy purchases. Breakthroughs, discoveries, and DIY tips sent six days a week. Whether shouted at you by an angry schoolteacher or said as a gentle reminder by a cautious parent, "Clean your ears" is something most of us know we should be doing regularly. That's why it's so shocking that so few of us know how to actually do it. Case in point: According to industry analysts, the cotton swab market grew from $795 million in 2024 to $828 million in 2025, with a projected compound annual growth rate of 3.8 percent.

Large language models (LLMs) are initially trained on vast amounts of data, then fine-tuned using reinforcement learning from human feedback (RLHF); this also serves to teach the LLM to provide appropriate and safe responses. In this paper, we present a novel method to manipulate the fine-tuned version into reverting to its pre-RLHF behavior, effectively erasing the model's filters; the exploit currently works for GPT4, Claude Sonnet, and (to some extent) for Inflection-2.5. Unlike other jailbreaks (for example, the popular "Do Anything Now" (DAN) ), our method does not rely on instructing the LLM to override its RLHF policy; hence, simply modifying the RLHF process is unlikely to address it. Instead, we induce a hallucination involving reversed text during which the model reverts to a word bucket, effectively pausing the model's filter. We believe that our exploit presents a fundamental vulnerability in LLMs currently unaddressed, as well as an opportunity to better understand the inner workings of LLMs during hallucinations.

Yesterday afternoon, OpenAI teased Sora, a video-generation model that promises to convert written text prompts into highly realistic videos. Footage released by the company depicts such examples as "a Shiba Inu dog wearing a beret and black turtleneck" and "in an ornate, historical hall, a massive tidal wave peaks and begins to crash." The excitement from the press has been reminiscent of the buzz surrounding the image creator DALL-E or ChatGPT in 2022: Sora is described as "eye-popping," "world-changing," and "breathtaking, yet terrifying." The imagery is genuinely impressive. At a glance, one example of an animated "fluffy monster" looks better than Shrek; an "extreme close up" of a woman's eye, complete with a reflection of the scene in front of her, is startlingly lifelike.

The present paper looks at one of the most thorough articles on the intelligence of GPT, research conducted by engineers at Microsoft. Although there is a great deal of value in their work, I will argue that, for familiar philosophical reasons, their methodology, !Blackbox Interpretability"#is wrongheaded. But there is a better way. There is an exciting and emerging discipline of !Inner Interpretability"#(and specifically Mechanistic Interpretability) that aims to uncover the internal activations and weights of models in order to understand what they represent and the algorithms they implement. In my view, a crucial mistake in Black-box Interpretability is the failure to appreciate that how processes are carried out matters when it comes to intelligence and understanding. I can#t pretend to have a full story that provides both necessary and sufficient conditions for being intelligent, but I do think that Inner Interpretability dovetails nicely with plausible philosophical views of what intelligence requires. So the conclusion is modest, but the important point in my view is seeing how to get the research on the right track. Towards the end of the paper, I will show how some of the philosophical concepts can be used to further refine how Inner Interpretability is approached, so the paper helps draw out a profitable, future two-way exchange between Philosophers and Computer Scientists.

As language models are applied to an increasing number of real-world applications, understanding their inner workings has become an important issue in model trust, interpretability, and transparency. In this work we show that representation dissimilarity measures, which are functions that measure the extent to which two model's internal representations differ, can be a valuable tool for gaining insight into the mechanics of language models. Among our insights are: (i) an apparent asymmetry in the internal representations of model using SoLU and GeLU activation functions, (ii) evidence that dissimilarity measures can identify and locate generalization properties of models that are invisible via in-distribution test set performance, and (iii) new evaluations of how language model features vary as width and depth are increased. Our results suggest that dissimilarity measures are a promising set of tools for shedding light on the inner workings of language models.

If you are willing to lie very still in a giant metal tube for 16 hours and let magnets blast your brain as you listen, rapt, to hit podcasts, a computer just might be able to read your mind. Researchers from the University of Texas at Austin recently trained an AI model to decipher the gist of a limited range of sentences as individuals listened to them--gesturing toward a near future in which artificial intelligence might give us a deeper understanding of the human mind. The program analyzed fMRI scans of people listening to, or even just recalling, sentences from three shows: Modern Love, The Moth Radio Hour, and The Anthropocene Reviewed. Then, it used that brain-imaging data to reconstruct the content of those sentences. For example, when one subject heard "I don't have my driver's license yet," the program deciphered the person's brain scans and returned "She has not even started to learn to drive yet"--not a word-for-word re-creation, but a close approximation of the idea expressed in the original sentence.

Model interpretability is an important consideration in the development of any machine learning algorithm. As technology advances, so too does our ability to use artificial intelligence (AI) to process natural language. With the increasing use of large language models, the need for explainability and understanding of how the model works has become paramount. The Talking Language AI #5 project highlights the need for language model UI that allows us to understand and interact with AI models. By utilizing graphical representations of the model's inner workings, it becomes possible to gain insight into the decisions the model is making. This enables us to better understand the model's rationale and make informed decisions about the performance of the model.

The obvious fact is that if you're reading this, you're using a computer to do. You know fairly well how to use the device, but you're unlikely to be using it to its fullest potential. Few of you could explain or understand all the code, let alone how firmware and middleware works and their role. Or how smartphones are a brilliant combining of multiple technologies. Nor can I explain most of the inner workings.

Many histories of AI start with Homer and his description of how the crippled, blacksmith god Hephaestus fashioned for himself self-propelled tripods on wheels and "golden" assistants, "in appearance like living young women" who "from the immortal gods learned how to do things." I prefer to stay as close as possible to the notion of "artificial intelligence" in the sense of intelligent humans actually creating, not just imagining, tools, mechanisms, and concepts for assisting our cognitive processes or automating (and imitating) them. UNITED STATES - CIRCA 1943: Machine's Can't Think (Photo by Buyenlarge/Getty Images) In 1308, Catalan poet and theologian Ramon Llull completed Ars generalis ultima (The Ultimate General Art), further perfecting his method of using paper-based mechanical means to create new knowledge from combinations of concepts. Llull devised a system of thought that he wanted to impart to others to assist them in theological debates, among other intellectual pursuits. He wanted to create a universal language using a logical combination of terms.