web analytics

AI – The Good, the Bad and the Ugly: Part 5

Anthropic and Pentagon

I have been planning to write a longer post on what AI I am using, but I have been too busy to produce a satisfactory post. Also, my approach has been changing recently, as I am exploring new avenues like Claud Code and Cowork.

But, given the recent news about Anthropic’s conflict with Pentagon, I thought I would write in support of Claude.

I use non-genAI, Grammarly, all the time. It has been a game-changer for me, as a non-native English speaker. I also use Claude to evaluate some of my writings.

For tasks I would previously search on Google, I use Kagi (which is non-AI) for short searches, and I use Perplexity to find answers to more substantial questions. For scientific literature search, I use Undermind, Claude, Perplexity, and Scite. I still use Google Scholar, but there are better alternatives.

For more serious tasks, I use Claude, with great caution. For very serious stuff, I use the same prompt in several AIs and then compare the results.

I do not use Grok, CoPilot, Meta, or Gemini.

I used ChatGPT when it was the only genAI available, but stopped. I appreciate OpenAI’s dominance in the market and the quality of their product, but I am not happy with their principles (or the lack thereof).

I have always found Altman’s ideas distasteful. There is something in the way they ruthlessly pursue their goals regardless of costs. All of that, and much more, apply to Grok, which I never used and I do not plan to use at all.

Why Claude? Apart from the fact that they are leading among the frontier AI, and the Code and Cowork promises are impressive, they have been much more open about the AI advantages and disadvantages, and about AI problems. I have read some of their publications, and I always found their approach refreshing. Similarly, I like Aravind Srinivas, who is the guy behind Perplexity, and their app has been a lifesaver at times, so I can forgive some of their faults.

So, yes, they are all sharks, but some are nice sharks…

AI – The Good, the Bad and the Ugly: Part 4

How not to use ChatGPT when applying for jobs

I occasionally receive very nice letters from scientists seeking to join my group at the University of Strathclyde. Very often, the e-mails come from people whose first (or even second or third) language is not English. I understand that they might want to use any available resources to help in crafting what possibly could be the most important letter in their career.

And yes, both the non-generative and generative AIs might be very helpful. Starting with non-generative AI tools, Grammarly is an excellent tool, worth every penny (although not very cheap). E-mail clients (Outlook, Gmail) come with spelling and grammar tools, and so do Word and other editors (like LibreOffice).

The advent of generative AI has changed the way we look at editing text. It is now easy to give it some tasks – for example, a summary – and, voila, it will craft a pretty decent text.

It is probably the task that gen-AIs are best suited to: Given some text, find what the “best sounding” continuation or expansion of it is. This is exactly what they are designed for and trained for: to use statistical regression to predict what the “most likely” text will be.

And, they are actually very good at it and getting better. I just received an email from someone who appears to be a genuine student from a Global South country, seeking to apply for a PhD position with me. The letter starts with a pretty good summary of my research, although using phrases that I probably would not write myself.

Maybe I am too modest – or too English (after 33 years living in the UK) – but I would not necessarily describe my work as “groundbreaking“. I would be careful in saying “Your expertise in (…) aligns perfectly with my research interests”, or that somebody can be “particularly intrigued” by my research in three quite different areas.

Such an exaggerated tone is very typical for genAI output. Unless explicitly prompted not to (which some more academic-focused tools do), the genAI tools tend to write very eager, over-enthusiastic and well-pleasing texts. They are designed to please you, because that means you will interact with them more.

But, I can buy it; people in different cultures express themselves differently. Even Poles or Scots speak differently from English, and I have experience in dealing with people from India, Thailand and even China. So, so far, I am giving the applicant the benefit of the doubt.

And then, comes this paragraph (screenshot of the original e-mail where I marked the relevant parts in red):

Ouch. So, clearly, the person did not even bother to read the text before pressing “send”!

This is an unforgivable error, and the sender gets this reply:

I must point out that your email contains several placeholder phrases such as “[relevant skills or experiences]” and “[software/tools you are proficient in, e.g., R, MATLAB, Python]” that have not been filled in. Using generative AI like ChatGPT to write personal letters, and not bothering to read the text and fill in the gaps left by the software, as shown below, is considered highly unprofessional and disqualifies the applicant.