Consider the trajectory of LLMs so far. Starting in Google’s translation research departments, following on from their groundbreaking neural network models, transformers were first and foremost a “language” model. They have grown to be able to convince people of their utility, much like the market stall owner would do pre-2000s, holding groups of up to 50 people in awe because of their fast talking, confidence and skill demoing whatever they were selling. The same works on shopping channels and more recently Youtube. Speak fast and confidently, cut all the dead air, and you got yourself the makings of a popular channel.

So for me the primary utility of LLMs are to solve language problems - translation, writing/editing, computer language use. This can be extended to solving other problems through a variety of methods, but as a baseline they are a language layer sitting above the tools people would usually use.

This ability to convert simple language into instructions or information, in any language in the dataset - computer or human - is the core innovation imho.

When you add and further train models on large sets of guidelines, now seemingly known as harnesses, you can optimise the model to know which tool to call in 1000s of use cases. This is a good start, and is why Claude, Cursor and more recently OpenClaw have been blowing so many developers’ minds. Given enough access, they can orchestrate a plan, given in plain language (as accurately as possible), to pull together tools to search for and edit all relevant “language”, be it code, or code as a representation of whatever the user would like to change or create.

I’m trying to be careful with how I put this, and not over-egg things, but I do see this harness as merely a large set of instructions over the core innovation, the language layer.

This is to remind people not to be drawn in to the intelligence trap, where just because someone is fast talking and confident, we should trust everything they say. Usually all the less so. But with the right harness in place, we can channel the energy into something useful. We can then run our checks, rewrites, storage commands for future retrieval, training and so on, but at least use the model to discuss with us and keep momentum up.

The utility is there, in the way it enables more work at a consistent level of quality, however you define that. It removes dead ends in trains of thought. It suggests best practice concrete next steps, even if sometimes generic. Who remembers writer’s block? Consistency is key to so many endeavours. Who easily gets derailed by a bad day or busy week? Using OpenClaw with Todoist via CLI (trying to be secure here) and using it to set routines and check in with me is a 2-3x productivity gain over just having access to a web-based model. At least that. I’m getting things done that were languishing, because it nags me. The language layer is perfect for that. It reaches out to me via Whatsapp, at times we agree on together.

So that’s my first point. LLMs + harness is not AGI, but it feels really useful.

Next we have the issue of bridging the gap from LLM to AGI. It would need persistent memory, and self-teaching abilities, at least, if it were to learn and execute exponentially. It looks an awful lot to me like “Skills” are the first stage of that journey.

If you’d not seen, the frontier AI companies have agreed on a few semi-standards at this point. An AGENTS.md file which sets the user-specific or project-specific rules a user would prefer to use, and more recently a Skills folder, where users work with the LLM to define the steps of a workflow (SOP) for the LLM to be able to call on and deploy at any time. It might include which tools to call, which settings to use, which formats to output, to where, etc. Rather than try to have it derive the whole workflow from first principles, humans can just tell it what they would do, and many times it can do it.

Why would the frontier AI companies admit they need users to add Skills, when they’ve spent billions training these so-called “sentient” and “AGI approaching” models? Because it’s just a language model with guidelines. So far. Perhaps the first real AGI model will include LLMs as just a language layer, while actually thinking and creatively reflecting on millions of ideas simultaneously in another kind of model. That’s all well above my pay grade, but it’s the only way I can see it happening from here. An LLM isn’t designed to think, self-learn, self-train, like the rather amazing AlphaGo did. This isn’t that kind of AI. Models will need to mix or train, but in the meantime, we keep feeding the AI company our Skills and they can build a repository of millions of human processes, to further the illusion that LLMs are working it all out themselves.

Smoke and mirrors, my friends.

I do not doubt their utility. LLM-based tools have let me code an entire automation pipeline for the industry I work in. They have let me automate many of the stressful parts of my life. They have offered clarity where I had doubt, and have given me confidence and momentum. I like the language layer + tools. But it isn’t working out from first principles how to extract specific parts of text from specific file formats, translate that with specific previous work taken into consideration through a specific matching algo and exporting and managing tags and history and versions… it’s just not going to get there in any reasonable time. But if I tell it exactly what I want it to do, which tools to use, we can get there much quicker than either of us could have alone.

That’s what’s leading me to suspect that Skills are a play for examples to train first LLMs and later AGI-adjacent models to be able to replicate humans better in the workplace. All to the benefit of the AI companies, who stand to be the next Google/Microsoft within the next decade as they dominate search, ads and now work processes.

I can only hope we all get some kind of dividend pay out from the sum total of our generous offering to the AI companies. An open model, harness, local mini-AGI - we deserve to have our pay out too. We’ve all worked very hard for this.

Disclaimer: written manually into Cursor’s agent window, to have it automatically prepped for my static site. No AI edits.