My stance on AI, if anybody cares.
AI is slowing down, but it was expected. There are very obvious limitations, both in terms of costs and in terms of data. To address the objection 'BUT SYNTHETIC DATA!!' quickly, no, synthetic data won't solve the issue because there are topics for which there is just not enough good quality data, and you can't generate good quality data on something that is just missing; you can only generate a better format of bad quality data, which doesn't help.
"Meh, whatever" is precisely how I feel, and honestly, it's how I've been feeling for the past few years, despite the gradual progress. We're only getting a bunch of incremental upgrades that don't actually solve real issues. Even when Sora was announced, while everyone was super hyped, I was like, whatever. What's the point if, often, there will still be hallucinations? It's just a toy. And don't get me started on AI music garbage.
For LLMs in particular, if you don't truly have a 100% reliable output, it's unusable in the long run because any output always requires human supervision. You can't have a reliable AI chatbot; you can't have a reliable AI agent. It _can_ technically be more reliable than a human _sometimes,_ but since we're talking about a machine, not a human, we expect 100% accuracy when there is an objective truth, not 80% or 90%, and not even 99%. If I ask something like 234987*34982, I want a correct answer, not a hallucinated one in which the LLM forgot to use the calculator tool. If I ask an information about something, I want correct info, not something that has a 1% chance of being completely incorrect and potentially misleading.
Therefore, the most you can have is an assistant, which is helpful sometimes and just a waste of time many other times. But an assistant for what?
- For code, sure, for small personal projects that have little to no use, it can be helpful, especially if you have limited coding knowledge. But for actual work? You always have to manually check to see if the code works as intended. Since it's often not reliable for complex applications, several serious programmers stopped using Copilot after using it for a while, as they claim it makes programming slower on average. Yes, slower, not faster.
- As a substitute for Google searching, you'd have to manually check the information because there's always the risk of hallucination (even if the probability gets smaller and smaller, it's not 0, so it makes basically no difference). So yeah, sometimes it can feel faster, but it's not really faster, because maybe you read some information it gave you that you didn't know about, and you start to wonder: _Maybe I should double-check._ And so you do a Google search anyway. Which defeats the whole purpose of using the LLM. You say Perplexity? Indeed, I'm very perplexed how would anyone use that seriously over google or any other search engine.
- As a summarizer, you have to manually check if it made up info that is not present in the original text, which means you have to know the original text anyway (= perplexity is useless). Don't make me remember how many times I had to read summaries that were full of shit invented by the LLM. This also implies LLMs aren't really appropriate to explain complex topics in a simpler way.
- To learn other languages? No. It's terrible. You can't trust LLMs for that: they mess up grammar rules, they use unnatural phrasing, and they use unnecessarily complex vocabulary. The most you can do is use them as a translator, since current machine translators stink anyway, so you may as well use something else that also stinks. But keep in mind that sometimes they will hallucinate extremely wrong translations that can give you a very wrong idea of what the actual text says. To reduce this risk, I personally use two and sometimes three different LLMs to double-check a translation. You'd be surprised to see how different the interpretations between LLMs can be, especially for Asian languages. If you think it's at least good for English, I'll let you know that I've done a C2 English grammar question a few days ago and GPT-4o got it wrong.
- To generate poems? To generate prompts for images? Sure, personal entertainment is a possible use of LLMs.
- Generate ideas for a topic, a book, a story, or whatever? You see, there's a big problem with this. LLMs tend to stick more or less always to the same kind of answer. If you ask an LLM to tell you a joke, there's a very limited amount of good and different jokes it can make. So for creativity, the thought of LLMs being creative is actually just a myth, imho. You need a considerable amount of human input before you obtain something that is truly creative with an LLM. Maybe it can be helpful in reorganizing already existing ideas and connecting with a few related ones, but existing human ideas have to exist because LLMs are pretty bad at being creative once you've explored their limits.
Don't get me wrong, I use ChatGPT multiple times a week, and that's the same for Claude. But I still consider them to be absolute garbage. They're truly helpful only sometimes, such as when I'm coding, because I'm not a good coder anyway, or when I want to translate, rewrite, or summarize some text (but I already know what the text is about, so I can manually verify and fix the output). In my experience, almost any other application outside of this is basically just dumb or very limited. It's all just empty hype.
To conclude, if you're the type of person that uses LLMs a lot every day, be careful. You're likely becoming dumber and less capable of doing anything every day.
1 - 4
This is HUGE news for the piano community. After decades of her teachings being shared primarily through workshops, institutes, and videos, we're finally getting a comprehensive written guide straight from the source. The Taubman Approach has been a game-changer for many pianists. Its focus on natural, coordinated movements and preventing harmful stress has unlocked a whole new level of freedom and expression at the piano for me. While I don't know how comprehensive this guide will be, having insights directly from Edna in book form will be an amazing reference to have. I'm sure this won't replace the need to study from other sources, but it will be illuminating to explore her philosophy and methodology in her own words.
5 - 1
Today, we mourn the loss of a true musical giant, Maurizio Pollini. Over his illustrious career, Pollini brought a unique combination of technical prowess, emotional depth, and intellectual curiosity to his performances. His interpretations of works by Chopin, Beethoven, Schubert, Schumann, Prokofiev and Stravinsky, among others, were nothing short of revelatory, and his advocacy for contemporary composers helped push the boundaries of classical music. Pollini's legacy as a pianist and musician will live on through his recordings and the countless musicians he has inspired. Rest in peace, Maestro.
8 - 0
I was listening to my own recording of the G# sharp minor fugue. Czerny indicates: quaver = 108. It feels so slow to me. It turned out that in the recording, the tempo I reached at the end wasn't the same as the tempo at the beginning. So I did a quick analysis of the tempo, and while I started at 108, most of the fugue was actually at 120. So I just recorded it again. Whatever, it happens.
0 - 0
That feel when Sorabji's music is so bad not even Eric Xi Xin Liang wants to study the whole OC
Unrelated note: an update on my progress. I got a tiny bit of motivation to work on WTC. I'm working on the A minor fugue and everything before has been recorded already. As I said elsewhere, I'm gonna publish all the remaining ones only when everything is ready, so don't necessarily expect me to finish it very quickly. But I might if I study and record fast
3 - 1
Answer to latest WW video and everything related to it: drive.google.com/file/d/1jR6cJ-IKvjaKD1nkR6l6awAMa…
*TLDR:* If you know what it is about, you know. No, I'm _not_ going to give a proper answer in a video. I do _not_ want to be affiliated with that musician or his theories anymore, so I'm not going to answer his latest video or any future videos he may or may not make about my performances. As for why my own videos have stopped for now, sorry, I guess I'm lazy.
5 - 6
Profile picture and Youtube banner changed. Just in case you were confused.
2 - 0
Coming soon on my channel:
• Brahms - 16 waltzes op. 39
• Kapustin - Piano Sonata n. 3 (live)
• Brahms - Theme and Variations, Op.18b
• Mozart-Alkan - Piano concerto K.466
Stay tuned ;)
Also, feel free to leave some suggestions in the comments (if I don't like any of the suggestions I'll continue working on Bach-Czerny, which is also coming hopefully at some point).
5 - 0
What will I upload after n. 1-30 of Czerny Op. 299bis?
2 - 5
I'm Aldo Roberto Pessolano, known artistically as Pianoth Eakòs Shaveck. My piano interpretations are born from a fusion of in-depth analysis and vibrant virtuosity. Join me in rediscovering classical narratives with a touch of the unconventional.
E-mail: pianothshaveck@gmail.com