frag

Great job! Did you finetune your own Laya for the snake game or what?

that's not a local LLM. If it's local, it doesn't matter in this case. Laya is a System 1 "AI", namely works like a classifier, given a state and questions, it shoots probabilities for each. I publish an episode tomorrow about Laya and Typesafe AI on https://www.youtube.com/@DataScienceatHome

Stay tuned ;)

of course. I am the author of https://www.youtube.com/@DataScienceatHome

nothing is permanent... and i wonder if they actually delete your account (of course not)

Ah fun enough I wrote a post just about that https://open.substack.com/pub/defragzone/p/ceo-is-obsolete?r...

who are you? police? lol not sharing my notes. If you wanna read, cool. If you don't, cooler.

Thank you for noticing! I’ll keep writing my thoughts and express them in clearer English with the tools I (we) currently have. One day, when Italian becomes a global lingua franca, I’ll wow you even more with my fluency :D

Thanks for reading!

You are very wrong. I polish my non-native english with Ollama and deepseek. Content is my own. Like on my podcast (datascienceathome.com) that exists way before GPT was even a name.

Welcome to cognitive capitalism https://defragzone.substack.com/p/welcome-to-cognitive-capit...

Took 6 months to migrate to Rust almost 90% of our commercial stack at amethix.com, 4 years ago. Since then we all sleep better.

Exactly. What I consider a patch and definitely a symptomatic solution is "solved" via agents that search the web (e.g. asking for the weather forecast of this year - in that case the LLM cannot know the year I am referring to, if not via a web search). Generally speaking, LLMs lack direct temporal awareness. Standard models do not model the flow of time unless explicitly. Some models can encode a model of time when trained on sequential video data and rely on external encoders to provide temporal structure. But that is a very narrow application (video in this example). That cannot be considered a generic form of awarenes of time as a concept through which facts can change.

true that

and we also see with what consequences

Fair points, and I appreciate the depth of your critique. You’re right—broad appeals to "moral evolution" can feel hollow without actionable solutions or a deeper analysis of the mechanics involved. My intention wasn’t to provide an exhaustive exploration but to spark a discussion, which your comment does beautifully.

Just to clarify: I don’t consider myself a Democrat (in the context of US elections). In fact, if I were American, I definitely wouldn’t have voted for Harris.

I referenced Trump and Meloni as case studies simply because they’re among the most recent examples. That said, I firmly believe two things:

1) The masses have an incredibly short memory. 2) It’s not just the right that engages in this kind of manipulation.

link is broken

Exactly. The tests are concerned with integrating the ML component into the broader picture of requirements and engineering. Engineers are not really interested in why the statistics of a model is failing. And should never be (that's why testing model internals in TFML is an anti-pattern)

I guess moving to Matrix?

Interesting post. Maybe to complete about adversarial examples in medicine https://medium.com/fitchain/attacking-deep-learning-models-3...

even for large projects?

voice upload is not working :(

If it's about an AI startup here are some tips https://medium.com/@frag/how-to-fail-with-artificial-intelli...

I want to cry... remembering the old good time of C programming... ohhhh

what AWS is giving for 1 year is really really minimal for a PoC or testing stuff. A small website would already get into trouble

interesting

lol... how about forking from github and train a model on Elon. I hate his euhmm thing so badly!! :D

the neural network is just like a baby brain. If it is fed with Keith Jarret, yes it will also learn to deal with that

the entire project is on github. In the folder data/ there are some samples to "see" it in action. Otherwise you have to train it on your voice and apply to whatever sound you like

Exactly! With just 5 epochs and some hours on a low budget GPU I got 81% accuracy. Not bad at all, considering that no knowledge of MFC & Co. is required.

That's right. This is not state of the art. LSTM is good for sequences and sequential data (like audio). But with the approach hereby described, prediction can be done in parallel and asynchronously. Which is something ;)