Solid core + capability-based sandboxes + LLMs = Users with superpowers
MemGPT
Memory-GPT (MemGPT) - Towards LLMs as Operating Systems - Teach LLMs to manage their own memory for unbounded context!
Build agents, not pipelines
There are only two ways to use LLMs in a computer program: as part of a pipeline, or as an agent. In other words, either you express the control flow of the program in code, or you give a LLM tools and allow it to manage the control flow itself1.
Magicoder: Empowering Code Generation with OSS-Instruct
We introduce Magicoder, a series of fully open-source (code, weights, and data) Large Language Models (LLMs) for code that significantly closes the gap with top code models while having no more...
Pony
An open-source, object-oriented, actor-model, capabilities-secure, high-performance programming language.
Sprites - Stateful sandboxes
Persistent, hardware-isolated execution environments for arbitrary code. Run AI agents, untrusted code, and more in secure sandbox environments with checkpoint & restore.
Guardian Angels: LLM Personalization for Productivity and Security
I propose an approach for highly personalized LLMs, for near-future productivity gains and personal info/cybersecurity against increasingly powerful LLMs: they should, in the spirit of uploading, try to emulate the user’s values and preferences in order to amplify the principal—not replace them. I discuss a package of techniques and proposals to accomplish such ‘guardian angels’; dynamic evaluation of LLMs combined with active learning and elicitation and heavy inner-monologue search/data-augmentation.
Can LLMs Be Computers? | Percepta
We build a computer inside a transformer — executing arbitrary C programs for millions of steps with exponentially faster inference via 2D attention heads.
psf (@psf@oldbytes.space)
In their article ["Malleable Software"](https://www.inkandswitch.com/essay/malleable-software), Ink & Switch list various ways that modern software applications fail to serve the user: - Software is too rigid: users can’t tailor it to suit their workflow, except by entering the role of programmers. - Software applications are overspecialized "avocado slicers", not general tools like "knives", so their lack of tailorability is keenly felt. - Software doesn’t provide a reasonable learning curve from beginner to expert; in other words, time spent learning an application doesn’t unlock tailorability, instead the end user must learn an unrelated skill like a markup/scripting language or even the tool’s implementation language. - Software applications are increasingly siloed, so sharing data between apps, let alone orchestrating cross-app workflows, is overly difficult. These are real problems. From what I can tell, Ink & Switch are software researchers, and view the problems from a highly technical perspective, focusing on novel infrastructure that enables end-user programmability. They pose questions such as "how do we ease users into programming?" and "instead of shipping a dead, static binary, can we ship a living image containing programming tools allowing it to be modified in arbitrary ways while still running?" These ideas are technologically exquisite, but they fail to provide a gentle learning curve. (cont'd) 🧵
‘Doom Loop’: OpenAI and Microsoft Admits LLMs Are Destroying the Web and Built on Theft
"Millions of people around the world will soon consider large models ‘hoovering up’ all their work to be an astonishing theft of unprecedented proportions."
Artificial
An LLM is a computer program. We should talk about it like a computer program.
Systems design 3: LLMs and the semantic revolution
Long ago in the 1990s when I was in high school, my chemistry+physics teacher pulled me aside. "Avery, you know how the Internet works, righ...
Restate - Build innately resilient distributed apps
Restate is a lightweight runtime that lets developers build innately resilient distributed apps without the complexity tax.
Dancing mad with sandboxing
Kefka is a Go-native shell sandbox with coreutils, Python via WebAssembly, and more. Learn the works of madness that went into making this happen!
How LLMs Actually Work
A from-the-ground-up walkthrough of how modern LLMs work, from tokens to transformer blocks to the next-token loop
Fight fire with pedantry
Over the last two years, I’ve heard a bunch of speculation about how to design a programming language specifically for an LLM. It usually centers around typechecking, formal verifiability, or context engineering. It’s a fun thought experiment, but always felt aimless. First, my experience has been that designing for humans is the best way to design for an LLM, as long as you restrict yourself to text. Second, being in-distribution is so critical that there’s no point unless you’re also training a frontier coding model.