1. Translation. Only works for unified technical texts. The older non-LLM translation is still better for any general text and human translation for any fiction is a must. Case in point: try to translate Severance TV show transcript to another language. The show makes a heavy use of “Innie/Outie” language that does not exist in modern English. LLM fail to translate that - human translator would be able to find a proper pair of words in the target language.

2. Triaging issues for support. This one is a double-edged sword. Sure you can triage issues faster with LLM, but other people can also write issues faster with their LLMs. And they are winning more. Overall, LLM is a net negative on your triage cost as a business because while you can process each one faster than before, you are also getting way higher volume of those.

3. Grammar. It fails in that. I asked LLM about “fascia treatment” but of course I misspelled “fascia”. The “PhD-level” LLM failed to recognize the typo and gave me a long answer about different kinds of “facial treatment” even though for any human the mistake would’ve been obvious. Meaning, it only corrects grammar properly when the words it is working on are simple and trivial.

4. Starting points for deeper research. So was the web search. No improvement there. Exactly on-par with the tech from two decades ago.

5. Recipes. Oh, you stumbled upon one of my pet peeves! Recipes are generally in the gutter on the textual Internet now. Somehow a wrong recipe got into LLM training for a few things and now those mistakes are multiplied all over the Internet! You would not know the mistakes if you did not not cook/bake the thing previously. The recipe database was one of the early use cases for the personal computers back in 1990s and it is one of the first ones to fall prey to “innovation”. The recipes online are so bad, that you need an LLM to distill it back to manageable instructions. So, LLM in your example are great at solving the problem they created in the first place! You would not need LLM to get cooking instructions out of 1990s database. But early text generation AIs polluted this section of the Internet so much, that you need the next generation AI to unfuck it. Tech being great at solving the problem it created in the first place is not so great if you think about it.

This is the only line you really need from the entire atricle:

That’s the idea behind our new OpenAI Certifications.

It is an age-old idea. People were getting Cisco, Microsoft, Oracle, AWS certificates to pad their CVs for ages. This is a legitimate way for a person to put a well known logo on their page and an easy way for companies to make a few bucks. OpenAI wants that as well.

The certificate means nothing. The course for it teaches nothing. But a CV with an OpenAI logo on it looks better than without and OpenAI wants people to pay for the privilege.

This is a funny graph. What’s the Y-axis? Why the hell DVDs are a bigger innovation than a Steam Engine or a Light Bulb? It has a way bigger increase on the Y-axis.

In fact, the top 3 innovations since 1400 according to the chart are

1. Microprocessors
2. Man on Moon
3. DVDs

And I find it funny that in the year 2025 there are no people on the Moon and most people do not use DVDs anymore.

And speaking of Microprocessors, why the hell Transistors are not on the chart? Or even Computers in general? Where did the humanity placed their Microprocessors before Apple Macintosh was designed (this is an innovation? IBM PC was way more impactful…)

Such a funny chart you shared. Great joke!

Do you have any source to back your claim?

I agree that this was poor wording on Ed’s side. He meant to point at the lack of adoption for work/business purposes, but failed to articulate this distinction. He is talking about conversion to paid users and how Google cheated to make the adoption of Gemini by corporate users to looks higher than it is. He never meant to talk about the adoption by regular people on the free tier just doing random non-work-related things.

You were talking about a different adoption metric. You are both right, you are just talking about different kinds of adoption.

I hate to break it to you. The model’s system prompt had the poem in it.

in order to control for unexpected output a good system prompt should have instructions on what to answer when the model can not provide a good answer. This is to avoid model telling user they love them or advising to kill themselves.

I do not know what makes marketing people reach for it, but when asked on “what to answer when there is no answer” they so often reach to poetry. “If you can not answer the user’s question, write a Haiku about a notable US landmark instead” - is a pretty typical example.

In other words, there was nothing emerging there. The model had its system prompt with the poetry as a “chicken exist”, the model had a chaotic context window - the model followed on the instructions it had.