Huberman’s episode on MDMA is out.

I finally learned to resolve the git conflicts correctly I think. Apparently, besides knowing the meaning of merge and rebase and what exactly do they do it requires human / social knowledge about who from your team did exactly what (if reading diffs is not straightforward) as well as having the responsibility to decide which version is the correct one. If you don’t have all three, don’t resolve the conflicts, ask your senior, it may not end well.

And yes, read the messages. Always read the fucking messages. And ask ChatGPT/stackoverflow what they mean.

Prompt engineering is quite approachable to learn if you understand the laws of formal logic.

I think this skill is going to cost a shitload of money.

And I am in the middle of it 🤗

When I chose natural language processing in winter 2017 because at the time I was learning many languages and self studying applied linguistics it was a unpopular obscure field with Jurafsky’s Speech and Language Processing book with crazy mathematical methods (transducers, automata, maximum entropy models, context-free grammars etc). Bag of words and TF-IDF!

In 2019 in the uni when BERT was already out we still were studying all that stuff!

I got the lowest grade (7.5) in my diploma at Introduction to Human Language Technologies’ exam because I did something wrong in a hidden markov chain problem for part-of-speech tagging. Now I think about it, it seems absolutely insane.

The spring of 2022 I was running around looking for a new job because I wanted more money for personal reasons. The only people that were interested in me were machine translation (as in my current job at that time) people, guys who parsed HR stuff (job positions etc.), a company with a symbolic rules (!!!) system for FAQ support and some practical task people like news toxicity detection. Obscure (again)! Cheap! Let’s be honest, absolutely useless. In spring of 2022. A year ago.

I started my current job I’m in love with on 7 of November 2022.

On November 30 of 2022 ChatGPT launched.

I can’t say we can do anything in applied natural language processing. But we can do A LOT.

And we understand what is going on.

Sorry for being so technical, I just want to show you the contrast:

before: so many super complicated methods different and tailored for every NLP method, every NLP task its own field and job specialization

vs.

now: ChatGPT/GPT-4 etc. that does it all and then all the other (infinite) new applications

Yet another hilarious thread on AI models’ peculiar blunders.

https://twitter.com/greg16676935420/status/1669364945224359937?s=46&t=tJA3XyJ6UXEb-9Si1PfeJQ

I rather like this one.

(I’m home in on Friday night down with an extremely tenacious flu so feel free to mute/unsubscribe, I may keep shitposting.)

Ahahahahaha

Also the dataset itself is janky. Yet again the AGI is a near miss.

Nobody:
Absolutely nobody:
David on Sunday afternoon:

(He’s writing me the math of his new idea of transformers matrices decomposition in harmonic polynomials. Of course he doesn’t expect me to read it today nor react in any way but it is funny. Sometimes I get curious, open the messages and get sucked in in the middle of the weekend.)