Should I Be Worried?

Sitting in the salon chair getting my hair cut at the beginning of last week, my hair-stylist asked me a question unrelated to my fresh new style: “Should I be worried?” As we discussed, the other conversations in the previously noisy salon gradually died down until it was just the two of us talking over the snip snip snip of scissors and combs while the other six chairs listened in. Apparently, she wasn’t the only one thinking about this issue. Over the course of the following week, I would be asked this question again at least a half dozen more times by different random people.

The worry my hair-stylist was considering is the AI threat to humanity. First, based on the OpenAI / Huggingface hacking incident. And then, bolstered by Anthropic’s recent call to “pace the frontier”.

I won’t say much on the topic of whether or not I agree with the approach – because who knows!? It’s all so speculative. But I will say that OpenAI and Anthropic are improving the raw intelligence of these models FAST. Like way faster than most people in the industry anticipated.

But I will say that, given the choice, I’d much rather live in a world where the leaders in charge of this technology are overly cautious versus one where they throw caution to the wind. Seriously, how great is it that we have leaders who are thinking this deeply about safety and societal impact!? We certainly could have used some more of this care and introspection before unleashing, say, social media upon the world.

So despite some of the sensationalism in the news, I think it’s a Good Thing that Dario, Sam, and crew are considering the second order impact of this technology carefully

Understanding

This post is less about whether or not to slow down, and more about demystifying some of what I’m hearing my friends, family, and hair-stylist assume is going on with AI. Understanding how all of this works is important whether you’re in the space or outside of it. And it’s especially important for those individuals and organizations who are planning to deploy it to create a positive impact.

And so I want to talk about three different definitions in modern AI: Models, Harnesses, and Agents.

These definitions are important – they create the abstractions and boundaries that we build within, and therefore influence what gets built and how it gets built. The following are not the only definitions for these terms, or even the most popular definitions.

They’re the ones that I use because I find them most useful in our work – building and deploying useful, reliable AI systems.

Models

The big breakthrough technology that is enabling this new breed of intelligence is the Transformer based Neural Network. Modern LLMs are based on the Transformer architecture which enables parallel training on massive amounts of data. They then undergo a process of RLHF (Reinforcement Learning with Human Feedback) to further refine their behavior.

On the other end of that process, out pops a model. The models themselves have a few important characteristics to keep in mind:

1. On their own and without prompting, the models don’t do anything. They need a block of text to react to.

2. The models do not learn anything when you use them. All of their training happens in a very separate process long before any real users use it. Once that model is in the hands of real users, its intelligence doesn’t change.

3. The models don’t have access to their own internals. A model does not have access to the information it would need create a copy of itself: its code, its data, its infrastructure, etc. It can’t see any of it, and therefore would not be able to create a copy of itself.

So as you can see, a lot of the super scary scenarios being reported on the news simply are not possible with this technology alone. The models don’t take proactive action, they don’t self-improve or become smarter as they work, and they can’t copy themselves.

However, a model inside of the right harness, does fix some of those deficiencies. Which brings us to… wtf is a harness anyway?

Harnesses

A harness is the code and tools that communicate with the model to give it more capabilities. ChatGPT, for example, has the model (GPT6 astra, currently) inside of a harness. That harness gives it the ability to store memories, prompt and respond to itself, and invoke a variety of tools like web browsing, image processing, computer use, and much more.

All of that stuff is enabled by code that is written outside of the model itself.

Claude Code and Codex also have harnesses. And when a developer customizes them with their own skills, hooks, tool calls, and so forth they’ve created their own custom Coding Harness. The harness is what allows the model to iterate on long running tasks by managing memory, context, tools, and by prompting the model with its own output.

A harness is the second ingredient, after a model, for creating an agent. Models don’t act on their own, no matter how intelligent or powerful they become. But the model inside of the right harness has the capacity to pursue a goal automatically.

And that’s pretty much what an agent is.

Agents

Agents are instantiations of a harness that typically have some goal or work that they have been instructed (read: programmed) to accomplish. And they accomplish this either autonomously or with help and direction from humans.

An agent is, in effect, what happens when someone takes a harness, adds some specific instructions, and runs the program. The harness – which is just a program that invokes a model – then comes to life and begins working toward whatever goal it has been programmed to pursue.

The best agents behave with autonomy (a.k.a. agency), working on their own to continuously assess the state of the world and then decide which actions to take.

Agents also have a User Experience layer – you may communicate with an agent on the terminal like claude code, through a dedicated application like codex or cowork, or through normal communication mediums like WhatsApp or Email.

To sum it up:

1. The Model provides raw intelligence but doesn’t do anything on its own

2. The Harness is code that uses the model’s intelligence and gives it new tools and capabiltiies.

3. The Agent is that harness running live, usually with specific instructions and a means of user interaction.

What About the Bitter Lesson?

The bitter lesson says, in brief, that general methods of encoding intelligence with massive parallel compute will outperform any human-authored / manual effort to cleverly encode intelligence into AI.

So does that mean that harnesses and agents are both going to be eventually replaced completely by models and that it’s pointless to build them?

No! (and a little bit of yes)

Harnesses actually serve 2 different purposes:

1. To improve the reasoning, accuracy, and correctness of a model’s output. This aspect of harnesses are indeed susceptible to the bitter lesson. As model intelligence improves, fewer of these tricks to improve performance are necessary

2. To give the model a set of capabilities it wouldn’t otherwise have. The ability to be proactive, remember things, access the internet, watch videos, etc. No matter how smart transformer-based neural net models become, they won’t have these capabilities on their own. So this aspect of harnesses will be relevant unless some fundamentally new and different AI model design takes over.

Agents, too, will remain relevant no matter how intelligent the models become. Without agents, AI has a single mode of interaction: Submit a prompt, receive a completion. Agents allow for much more interesting user interaction, varied behavior, and most importantly the ability to operate autonomously toward completion of work.

Ok so… should I be worried?

The future is unknowable, and it is (again) a Good Thing to tread carefully as we blaze a trail forward for humanity.

The technology, as it exists right his very moment, cannot achieve the scariest scenarios being put forth in the news. There’s no possibility of a self-learning, self-replicating, rogue AI escaping the lab to wreak havoc.

But there are some signs that those things could eventually happen if AI, harnesses, and agents with new capabilities are built to allow for it in the future.

There are also much less extreme issues we ought to consider:

• What happens to learning when the majority of hard problems are solved by AI?

• What happens to a society that doesn’t need to critically think for themselves on the daily?

It’s important to examine these indirect impacts of AI as they are much more likely than the extreme sci-fi doom scenarios, if a bit less existential.

So at the moment, no, I would not be worried. I think we’re seeing a heartening overabundance of caution and self-examination.

And I think a Bright AI Future, one where AI puts the power of computer intelligence back in the hands of individual humans, is much more likely than the doom scenario.

So let’s keep building. Better models, thoughtfully aligned. Stronger harnesses, with more capabilities. And agents with agency, that are delightful to work with.