Hi all(emaal)!
Recently I wrote a small document on ‘What to know about AI when starting to experiment with it‘ for an acquaintance working as a lawyer in Singapore, thought I might as well share it here - maybe some of you can also get some value out of it. There are a few example prompts near the bottom, which you might like perusing, and one ““technique”” I don’t read about at all but find myself increasingly amazed by — the ability of the latest models to read between the lines, i.e., reading subtext (even your own subtext as reflection tool).
I’ve also been surprised by how easy it has become to get “local LLMs” (AI chatbots that work purely on your own device, no internet needed, no worries about putting in sensitive information) working on linux/macOS via the combination of Ollama and Open WebUI. Sounds a bit technical, may do a write up about it later for those interested. (For the more nerdy folk reading this: it’s just 3 CLI commands and a docker install away, with one easy CLI call per model to download new ones. More involved for Windows.)
Also enough happening on the AI front wrt developments, I will write a wrap-up of the current state later.
But now without further ado:
AI - pointers for gaining a feeling with the technology
This is a small document outlaying, in no particular order and by no means exhaustively, important themes and information to keep in mind when first approaching the latest in AI developments from a non-technical perspective.
Some basic information in short:
When people talk about AI, they largely mean systems build upon Large Language Models (LLM), which are AI models that are trained to generate text that likely sounds to be correct, based on patterns it has learned from data it has been given. Similar techniques are also used for generating images, sound, etc. (which we thus call Generative AI, and if AI models combine multiples of these modalities we additionally label them as Multi-Modal)
We do not understand these models well enough to be able to remove information that has been put into the model. This means that anything that has been put in, literally cannot be removed as of this moment. BE VERY CAREFUL WITH SENSITIVE INFORMATION, I cannot stress this enough. Officially, via paid access information is not stored for most AI models, but trust that as far as your trust the organisations behind them.. There are ways to ensure privacy through 'running AI models locally', but this is a bit more involved and technical (can tell you more about that if interested, can imagine that's important wrt confidentiality clauses).
One model can be better at something than a different model and worse at something else, and it also depends on your personal taste. I use Claude for coding/programming, and Gemini for most other things currently. Experiment and see what works best for you.
See these AI models as interns. Their output is only as good as the instruction you give them! If you want them to produce a certain format of output, explicitly state this. If you want them to consider certain contexts, arguments, perspectives, or think in a certain way, add this to the instruction (prompt) you give to the model. And just like the output of an intern, always check and validate whether what they give you is actually proper, mistakes can slip in at the most random of moments.
It is incredibly important to use proper benchmarks to see if an AI model will actually be of use to you. Does it actually save you time, does it increase your enjoyment in work, does it affect the quality of your work, are your ideas now better...
It helps to view work as a collection of tasks, and experiment with individual tasks to see where an AI model could help you be more efficient and where it does not make any sense (yet).
There are multiple types of AI models (LLMs), you can roughly classify them as such:
'Classic' models: regular AI models (LLMs), good for simple to slightly more complex tasks, generally works better with clear single-purpose tasks.
Reasoning models: """thinks""" before it generates answers, uses more energy and is slower, higher chance of solving tasks with multiple elements.
Deep research models: use web-search before reasoning over the gathered sources to generate answers, very slow and uses loads of energy, better at creating comprehensive reports on a singular complex topic.
Agents: specialised AI models that are able to act autonomously in a limited way in the real world (e.g., send an email, post a note to a program, put an event in your calendar)
Agent networks: multiple specialised AI models that are tied together to autonomously achieve a complex task (e.g., periodically perform a search on relevant materials online, then making a summary, sending that summary to an assistant by email with an overview of the content such that he/she can verify it before sending it out as a newsletter)
An agent can be nothing more than a classic model that is given a very specific instruction ('System prompt') in the background that it always checks before it looks at your instruction/prompt. There are easy ways to build these agents for yourself nowadays for tasks that you perform routinely, thus removing the need for writing the same instruction every time.
Some AI models have access to web-search tools and can give you references/sources you can use to validate their output, which can be very important within certain topics.
Examples: perplexity, gemini, chatgpt with web search enabled.
Generally, from both an ecological and an economic-temporal perspective, use the smallest/weakest model that is able to execute your instruction. You first need to build intuition in order to be able to determine this on the fly, but have this rule in the back of your mind as you experiment with the technology.
Small overview of current AI (LLM) model names and respective companies, non exhaustive, with the current best models named from smaller (weaker) to bigger (better):
Gemini (Google): gemini-2.5-flash < gemini-2.5-pro [Reasoning models with web-search]
Claude (Anthropic): claude-4-sonnett < claude-4-opus [Classic models that can also enable reasoning]
ChatGPT/GPT/O3 (OpenAI): GPT4o < GPT4.1 < GPT4.5 ~< O3 [Classic models to reasoning models]
LlaMA (Meta): Llama 4 Scout < Llama 4 Maverick < Llama 4 Behemoth [Classic models]
DeepSeek (DeepSeek): DeepSeekV3 < DeepSeekR1 [Classic model ('V3') to reasoning model ('R1')]
...
I personally use Claude-3.7-sonnet for my programming needs, perplexity for additional web search, and Gemini-2.5-Pro for my semi-complex questions. (Very) easy questions I find are often not worthwhile to ask a chatbot, as a quick search online is faster than writing an entire prompt to get the output I want.
I use Poe as a platform to have access to pretty much all models for one fee, there are loads of similar platforms out there. I haven't really bothered doing an exhaustive search for alternatives yet and comparing them.
Some prompting techniques to test:
For all examples, I also explicitly share my instruction at the start on how i make this example, as that is I think also useful to see.
ask the model to read between the lines, it may surprise you at how good it is at reading subtext (even with your own subtexts, as a reflection tool)
whilst using the AI models replace sensitive information with placeholders. e.g, instead of inputting someone's social security number, replace it with {social security number}, including the {} braces.
use the AI models to formulate arguments both in favour and against your work.
ask the AI model to remove all mentions of X, convert British English to American English, format information into a table (or .csv file), create headers appropriate for a long piece of text, and perform other small but tedious edits to textual content.
Some closing thoughts, from my perspective as an AI specialist at a university:
Looking at the capabilities of current AI technologies based on language models, and even pessimistic perspectives on the growth of these systems, AI programs are set to significantly disrupt and replace entry-level jobs at practically all firms and sectors in white-collar (and even blue-collar) work. Many students graduating within the coming years will not be taught yet how to properly gain value from AI systems in their tasks, as university programs are slow to adapt to the realities of the new work field. This can, and likely will, create tension between efficiency-based metrics asking for the automation of entry-level tasks, and longer-term benefits of a company to integrate and nurture new talents and teach them the ropes of their new role (something they are generally not taught at uni). It's important for organisations to consider how to balance these efficiency drives with the integration of new talent for the long-term prosperity of the organisation itself as for larger society -- especially so long as university programs have not yet adapted to the new reality of work.