So what does AI mean for me?

Note that AI is changing at a rapid rate. This was written in August/September 2026 so the information may have changed in since then 

AI is all but unavoidable these days 

It's becoming the assistant in your phone; all your online searches have AI based responses and every other app for your phone or computer seems to have added words like "powered by AI" into their description. They answer questions, make pictures and videos, can talk to you in fairly natural language, and some have some fairly convincing voices too. Intelligence is a stretch at this stage, but it is coming along very quickly. This will be a brief overview 

Online Large Language Models (LLMS) The most common type you'll come across. 

When most people think AI it's the LLM/chatbot that comes to mind. Names like ChatGPT, Gemini, Claude and Grok are getting to be household names and many other smaller lesser known and/or more specialised are out there.  All these models work by taking your input (questions/requests/statements) and coming up with a reply. They analyse your input and decide what it means and then come up with the most likely response. This is important because it might not be the best response. There is also a good chance that the information isn't actually accurate. But with those warnings in mind, if you check the information you are getting is correct, they can be helpful tools. The most likely response is often the correct one given the amount of training data (the information the LLM was given to prime it to be ready to respond to users). This is especially so for things like programming but some areas the systems can be quite limited. 

They all work within a context window. A single question usually is very little context while a long conversation including several documents is a larger context window. A longer or more complex conversation means that the AI can lose access to the early parts depending on how large the context window of the model you are using is. Some companies will meter the number of responses you get based on how much information needs to be considered in each response you get. This can mean you don't get as many inputs on some models/price tiers each day as each response given requires more processing. 

Many AI have tools attached to them to be able to make files, author documents, summarise email and messages, create pictures/videos etc. The list of things they can do increases rapidly. 

My experience so far with those big 4: 

nb: each major LLM company will have several models available with different capabilities 

ChatGPT (GPT 5.5): Has a wide range of knowledge, likes to be quite verbose with detailed answers and dot points. It can be tuned to a certain degree and has the cheapest low-level tier (which is still ad supported).  It's a fairly reasonable all-rounder. 

Gemini (3.6 Flash): Is very assistant like, given its integration into android phones and google smart speakers and displays as well as showing up in google searches. It can get overconfident with its answers and is prone to "sycophancy" (always telling you your ideas are clever or observations sharp). It has a large context window and is good for quick answers, but I trust its answers the least after getting many incorrect answers. Higher levels are apparently better. You can give Gemini instructions on how to respond but has a much narrower range of allowed customisations compared to Grok. 

Claude (Sonnet 5): Claude feels very measured, going back and correcting its mistakes at times in conversations. Answers feel well thought through, but it is still able to make mistakes. You can tune the way it responds between dot points and formality or being conversational 

Grok (4.5): Grok describes itself as "maximally truth seeking". It is an AI with a lot less boundaries on what it will talk about. It will match the way you speak to it and can be heavily customised to react/speak how you want. Obviously, this can be good and bad. It's probably not one I'm letting my children use but if you want something that will go on an unhinged rant with you then it may be your go to for that. It has quite a small context window in my testing so not great at longer conversations that require details to be remembered. 

I often find if one isn't getting to the right answer you can get another AI to look at that conversation and see where the first one is going wrong. 

Things to note: 

They don't think about things in-between inputs or sit with ideas; they answer in the moment based on what they know about you, giving preference to the conversation you are currently in to decide how to respond. 

They will confidently give answers that are not correct if it is what best matches the response that makes sense given their training. 

If you are using a free or low-priced tier with these systems your inputs may be used as future training data. This means you should be careful about giving it sensitive data about you as it may turn up as part of someone else's answer (it's unlikely to be tied back to you but something to think about). 

Local LLM 

Software such as LMStudio is available for your PC, Mac or phone.  They require a computer with a lot of dedicated graphics memory or a shared memory setup such as the latest Snapdragon X PCs or M1 + macs. Running locally isn't going to give you something with the power of the online models but it will give you unlimited access to it and for the privacy side it means you keep all your information on your computer. There are a range of models available this way with different strengths and can be customised further to do a more narrow set of tasks better. Generally responses will be slower and context windows will be small. 

Local Agents (such as OpenClaw) 

This is like the agentic systems below but less integrated. You can run the agent on your own computer, give it a task and let it go to work. It is important to understand here that they are not the reasoning system and will access an LLM to get information to get a task done. It will keep asking for information until it gets the task done or decides it can't be done. Accessing the LLMs this way works via something called API where you pay for "tokens". These systems can run through thousands of tokens depending on the task costing a lot more money than people realise. 

You can, however, pair OpenClaw with a local LLM (see above) and then it is just your own computer providing the responses to your own local agent. This means a less powerful overall system than using an online model, but you don't have the token costs involved. 

The Future (already starting) with Agentic systems 

Agentic Systems are different in that they are setup to monitor and automate processes. Unlike previous "if this then that" type automations, agentic systems can reason to complete a task. Imagine like a factory line that could adapt to many problems without human intervention, diagnosing problems (within the limits of its sensors and training). Not only that but it can learn from issues that happen to (hopefully) better perform its function. This will go into traditional automation as well as things like automated replies and workflow situations in offices plus more applications, many of which probably have not been thought of. 

Future Caveats 

As people get used to using these systems I think it is worth remembering that if you rely on one AI as they get better for your answers you are only going to get answers that the company who makes that AI think are OK. It is unknown how much these systems might be doctored to give certain types of information or how advertising could be so integrated that it is indistinguishable from actual information.