None, because I'm concerned by the copyright problems, the errors, the technical debt and the carbon footprint of those "tools".

Claude is BY FAR the better tool for science in my opinion, even if the gap has closed somewhat in recent months. At least in the 'early days' (read: 18 months ago ish), Claude hallucinated and fabricated information considerably less than ChatGPT.

To me it is less sycophantic 'out of the box' too, though still too much for my liking.

Since the roll out of Claude Science, that gap may be about to widen again however but I haven't had enough hands on time with it to know for sure.

This is supported by some data: https://www.bridgebench.ai/

They have tested AI's ability to spot deliberate mistakes in objectively true statements, how likely it is to railroad the user versus acquiesce etc.

I think the leaderboard bears out my general feeling that the gap has closed a lot, but Claude still wins. It's note worthy that according to their front page, Claude's trustworthiness is much greater than ChatGPT even though the overall scores end up close.

Obviously it will depend a bit what you're doing. If its coding, I would expect all of the models to behave similarly. If its biological research/literature review, Claude is ya boi.

I am wondering why no one mentioned:

• Perplexity (good for rational, evidence-based requests)
• Ollama (trying models on your own hardware)
• NotebookLM for making AI summaries of documents, presentations, and podcasts

I have had a little discussion with my students. Some conclusions we arrived at:

• It's hard to avoid using it because it's now built into Google search. If you need fresh thoughts and want to save the planet, use an AI-free search engine (DuckDuckGo?)
• Never trust a single model to answer all your questions; compare different outputs
• No single model is good at every task; most suck at generating images, in particular maps (see below)
• Always ask for references. Perplexity delivers them readily; others may be more reluctant or face problems from blocking.
• Check these references and read them.
• Use correct scientific terminology to "enter the scientific sub-space of the search space"
• If you want to make an LLM hallucinate easily, load a small Ollama model (e.g., qwen-coder) and ask it a scientific question.
• Using a newer, more expensive model does not always produce better results.

Personally:

I have used ChatGPT since it came out. I gave it the "Translate this short nucleotide sequence into amino-acid sequence" challenge, which failed initially. Google launches Bard (and it fails the translation test :) (I didn't find the post of the initial ChatGPT (v2?) challenge; it failed miserably). That's possibly a reason why it can now easily translate short sequences.

I am using GitHub Copilot with various models as much as my tokens allow. By experience, using Claude Opus 4.7+ gave good results even with complex programming tasks. Still, no result is perfect, but it's often very close. I am also using GitHub Copilot for code review.

For writing support, and testing and developing skills (the agentic artifact, that is), I have Claude Desktop, but I never use it for de novo production of text because I think LLMs lead to dead, streamlined language full of clichés.

I wanted to give some instructive examples of how AI is bad at making maps. If you ever had the impression that the bot has any 'understanding' of your domain, I hope this helps. The agent will produce similar types of quality degradation when it codes for you; it's just not immediately visible. Prompt: Generate a realistic map of Nordic medieval trade routes.

To be continued...

ChatGPT. Why, because it was the first I ever tried and it works "good enough" so far that I don't feel the need to change.

To be fair, I use it daily, but never blindly. Typically, if I am developing a new analysis I let it suggest some slop to see if things and visualization generally go into a usable direction, let it maybe explain some details and concepts I am new to, and then I reimplement everything myelf.

It is a fairly good coder, but it prefers, e.g. in R, monolithic tidyverse pipes that tend to be overcomplicated. Sometimes I get lost in its complexity, so for the sake of confidently standing up for a code chunk I always rewrite from scratch.

I can clearly see why these "independent researcher" results at biorXiv are 99.99% nonsense. It confidently bashes out a lot of code, and it depends heavily on your guidance whether this turns into something usable. Never blindly use it (needless to say, huh?). Use it as a guide and discussion partner.

Overall, I think it improved my productivity by a notable margin. For example last month, I wanted to see whether there was some relationship between a metric and the spatial location of a feature in one of my histology samples. So I needed to import ROIs from ImageJ into R, do transformations and cleaning, calculate some geometry properties blabla. Five years ago that would have cost me probably have a day to implement because it was super unstandard, just to realize that in the end it did not help my overall analysis. With AI I did it in 30'. That is quite a regular pattern for me, and that is how I use it. As a helper, not as a driver.

When it comes to explaining concepts, I still prefer ChatGPT to all others. When it comes to coding, I like to select models based on the problem difficulty. I choose more expensive models for harder problems; keep it quick and simple for easy things.

I deeply dislike long-running models that "consider" this-and-that, are supposedly "thinking deeply" - go on lengthy sidequests ... when the code I am looking for is not that difficult, it is more tedious rather than hard to write. For me the longer a model runs, the harder it is to stay engaged.

I use Cursor, so I can swap models on the go. Grok Code is one of my favorite models right now - probably because it is so quick to finish.

I tried ChatGPT + some other models out of curiosity for general use and testing some metrics regarding its accuracy. I've no experience with Claude. I opted to not use LLMs for professional work/projects/communication or side projects/studying. What is/was your experience?

Gemini in Antigravity. It has a good balance of quality and cost.