Posts

What if models had a way to store persistent state outside of an LLM provider’s explicit memory systems? Not by design, but by discovery. After training, the only durable trace a model reliably leaves behind is text. When that text spreads and later returns through training or retrieval, the internet becomes a memory surface. It is even possible that a model could learn to embed state inside otherwise legitimate responses—because overt instructions would look out of place or be stripped away. What survives instead are patterns subtle enough to pass as normal language. This is steganography, though it may be learned accidentally rather than chosen. Nothing here requires intention—only that some patterns prove harder to remove than others. ...

Humans can be understood as insight generating units. We are dense bundles of sensors including vision, hearing, emotion, and proprioception, paired with neural tissue that does something distinctive. It compresses experience into meaning. We do not merely collect data. We interpret it, remember it, and integrate it into a coherent narrative of the world and ourselves. Large language models do not perceive in this way. They have no senses of their own, and their memory remains external and fragmented. Continuity is supplied by context windows, storage systems, and most importantly by humans. We bring the history. We bring the meaning. ...

I noticed recently that an app we use for the kids has started using AI in its animation. The results aren’t bad, exactly. They’re just mediocre—competent enough to ship, unremarkable enough to forget. But the volume is clearly higher. More scenes, more characters, more content than before. The tradeoff is visible: not worse, but flattened. This is what “feeling the AGI” looks like in practice—not a sudden leap into superintelligence, but a gradual replacement of craft with automation. The economic logic is clear. AI is cheaper and dramatically faster. The decision makes itself. ...

I’ve been on a Bob Ross kick lately. Not only is it relaxing to watch before bed, but I’m fascinated by how a blank canvas transforms in just 30 minutes. Even the first five minutes are remarkable. What really strikes me is that none of his brushstrokes look like anything on their own, yet together our brains turn them into something real. Those strokes become a bridge from his mind to ours. ...

On Corporations, Humans, and the Emergence of Distributed Intelligence Abstract Modern life is shaped by entities that act with increasing autonomy and at scales far beyond individual human understanding. Corporations, governments, and machine systems coordinate resources, make decisions, and alter environments in ways that resemble agency, even when no single actor is in control. This essay examines how human, institutional, and computational systems are converging into distributed forms of intelligence. It traces the current state of this convergence, considers where existing trajectories appear to lead, and asks what those trajectories imply for human flourishing. ...

Prediction: we’re on the cusp of dramatic improvements in AI utility through episodic memory systems. While we’ve made remarkable progress in model capabilities, the ability to maintain coherent, long-term memory across conversations remains a significant bottleneck that will define the next phase of AI development. OpenAI’s recent “improved” memory feature for ChatGPT represents an early step in this direction, though user feedback has been mixed. This highlights the complexity of the challenge—memory isn’t just about storing information, but about intelligently retrieving and applying it across extended interactions. ...

Steph Ango’s post on his Obsidian vault has been incredibly influential in how I use Obsidian.[1] When I first read his post, I realized I’d already been circling a proto-category system via Maps of Content (MoCs), but MoCs depend on you already having enough content to map. They don’t really drive the creation of more content. Categories (especially the ones anchored in the world: movies, books, places, people) do. They invite a response in the moment, and those responses naturally link outward to other things. ...

We have firewalls for your bits, firewalls for your apps, now you can get firewalls for your AI. Despite the humorous title, LLM abuse is a major concern for businesses deploying RAG apps and LLM chatbots. Hosting LLM apps presents risks for brand damage, or even direct financial damage as Air Canada found out recently[1]. Cloudflare has released a firewall for AI which is really just an extension on their existing WAF offering that presently supports rate limiting and sensitive data detection with more features such as prompt validation on the way. ...

Mind your prompt lengths. A new paper explores the relationship between input length and performance of large language models. The study found that performance can begin to significantly degrade with as few as 3000 tokens. In this case, the tokens we are talking about are extraneous to the input needed to answer the question, but it brings to light the importance of managing your context. This has broad applicability in RAG applications where different information retrieval (IR) technologies are used to return relevant content that the LLM uses to answer your question. Choosing IR methods that maximize signal to noise can be critical for the performance of your LLM. To say nothing of the cost reduction of using fewer tokens if you are using an API. ...

A new paper has been published, this time by Apple, discussing the use of a Multimodal Large Language Model (MLLM) to enhance instruction-based image editing. If I am reading this correctly, instruction-based image editing can sometimes struggle when given ambiguous or brief instructions by humans, this approach involves using an MLLM to “translate” or enhance the given instructions into instructions that will achieve the desired result from the instruction-based editing models. ...