Generativ AI ökar (eller sänker) din produktivitet

 AI verktyg har positiv effekt produktivitet så länge man håller sig till rätt uppgifter

Äntligen har generativ AI, eg. ChatGPT 4, funnits ute så pass länge att det börjar droppa in ambitiösa studier som fokuserar på om verktyget ger stöd eller inte 🥳

I en studie från Harvard Business School, av bl.a. Ethan Mollick (följ honom om du är nyfiken på AI och produktivitet), tar man en grupp konsulter, låter ena delen använda GPT-4 och andra delen får jobba på enligt eget huvud.

Bättre och sämre:

😁 gruppen som får använda AI presterar ca 40% bättre än den utan, detta så länge man jobbar med uppgifter som ligger inom AIs komfortzon (”within jagged frontier”)

🙁 för uppgifter utanför AIs komfortzon, blir gruppen med AI-stöd mellan 13% och 26% sämre än gruppen utan stöd.

🧐 Det gäller att ha lite koll på vad AI kan och sätta förväntningarna rätt.
Men det rör på sig…AIs komfortzon växer hela tiden så om du testade ChatGPT i början av året, är det tags att ta en ny titt.

Resultaten ger ytterligare stöd till det, numera, gamla uttrycket: AI inte kommer ta ditt jobb utan någon som använder AI kommer göra det.

Det finns helt klart risk för många bryska uppvaknanden där omvärlden har sprungit ifrån “Kodak moments”, men framför allt många möjligheter att använda detta fantastiska verktyg till att göra arbetet roligare och låta oss fokusera på det mänskliga.

Ett citat ur artikeln som stannade kvar i mig

“In our study, since AI proved surprisingly capable, it was difficult to design a task in this experiment outside the AI’s frontier where humans with high human capital doing their job would consistently outperform AI.”

🤔 AI är väldigt bra på tankearbete.

Läs artikeln för fler detaljer. Om du är har ont om tid, dela upp artikeln i 5 delar och läs lite varje dag.

ChatGPT 4 got an almost a perfect score on “högskoleprovet” (Swedish University Entrance exam) 🥳

In early January of this year I put ChatGPT (Dec 15 version) through the Swedish University Entrance exam (in Swedish), it got a normalized score of 1.4* out of 2.0, a good result but failing a bit on quantitive reasoning and reading comprehension of long texts.

Now, I challenged ChatGPT 4 with the same test and it got a score of 1.9* out of 2.0. Wow 🧠 Only 2.5% of the human participants get such a high score and it will open the doors to almost any University education in Sweden.
Similar results are seen for other benchmark exams.

ChatGPT 4, still has a way to go with some mathematical reasoning questions, however,

I consider reading comprehension as a solved problem. We can go from searching to asking and discussing. This is huge!


There are a gazillion cases where this is game changing - one of the first that comes to my mind is e-mails 📬

Say that you have a long e-mail chain and you want to catch something. Today most of us use keyword search and combine data from different e-mails in our head to get an answer. With ChatGPT 4 you can ask the question that you are thinking about and get a summarized answer taking data from all e-mails into consideration. You can even ask follow-up questions and ask to write a reply. New ways of thinking.

This will only solve situations for which the AI-model have data to leverage. The personal touch based on your experiences must be written by yourself 🤩

As for all co-workers with strong CV, ChatGPT comes with a higher salary than the old version (you must use the plus version of ChatGPT to access version 4) 💰

I’m inspired to put it through the paces in the days to come, and understand how it can help me as a project leader🚀


*Excluding all questions that were graph based and could not be input to ChatGPT, this is the most optimistic interpretation.

Is Chat GPT "smart" enough to be accepted to the university?

Let's see how well it scores on the Swedish university entrance exam ("högskoleprovet").

 

For fun, I put Chat GPT to the test and it got a score of 39 of 70 at the May 2022 exam. If excluding the diagram-based questions, the maximum score is 58, giving Chat GPT a relative score of 67%*. This translates to a normed score of 1.0 (2.0 is the maximum), which beats 70% of the participants and qualifies ChatGPT, as one of the 28 students, for the Economy Program at Örebro University. Congrats!
 
Looking a bit more into details, ChatGPT scored exceptionally well on the Swedish verbal part with a score of 25/30. I was especially impressed with the reading comprehension. It scored worse on the mathematics part, 14/28, which makes sense considering how the model was trained. It is still better than random guessing, and some mathematical "reasoning" was awesome and correct.
 
What does it all mean?
We already see amazing companies emerging using these technologies, e.g., Sana labs and Hypertype. For 2023 even more advanced language models will be released, also a model trained specifically for Swedish.
 
How will this influence my job as project leader?
 
This is a fascinating topic, and we are just at the beginning of what is likely to be the most significant technological advancements of our lifetime. I will devote as much time in 2023 as possible to explore how this can be used for (project) leaders and share my findings.
 
Let's go!
 
*Depending what assumptions are being made, the relative score is between 56% and 67%.