Diego Quezada

following the gradient down

[Diego Quezada @diegoquezadac]

@burkov @Typora Typora it’s really good, back in college I used to try every typing app such as Obsidian. MarkText. SimpleNotes, Caret (my favorite), etc … Recently I switched to Ulysses, you can do Markdown and LaTeX and the overall experience is so great imo.

[Diego Quezada @diegoquezadac]

@MatGuu @llama_index @UnstructuredIO Didn't know about Peslac. I just tried the free PDF parser and it's bad for extracting tables. They claim to have a "AI-Powered Data Extraction" however they don't mention anything about tables in their website.

[Diego Quezada @diegoquezadac]

LLamaParse by @llama_index is definitely something else. Best solution atm for parsing files.

I've experimented with:

1. PDF parsers
2. PDF parser + LLM
3. LLM with vision capabilities
4. Unstructured (@UnstructuredIO)

None come close to LlamaParse's premium parsing.

[BURKOV @burkov]

So, just finetuning Llama 8B boosts its performance on this incredibly hard benchmark to 45%. What a sad omission. Purely accidental, I'm sure. https://t.co/9gzPzbBVA0

[BURKOV @burkov]

So what are the real breakthroughs in AI, as opposed to fake ones like "o3 (yes, we skipped a number)"?

1. Random forest - which made decision trees strong.
2. Gradient boosting - which made decision trees even stronger while it was thought impossible.
3. Support vector machine - which made machine learning a science.
4. Factorization machine - which applied machine learning to millions of features.
5. Conditional random fields - which solved sequential predictions for the first time (as a supervised learning problem).
6. word2vec - which has shown that one can apply arithmetics to words and modify their meaning (and which opened the path to LLMs).
7. LSTM, which solved machine translation.
8. Transformer, which solved machine translation again, while it was thought impossible.
9. A very large transformer (GPT-3.5/LLM), which gave a text interface to machine learning systems.
10. GPT-4, which gave a text interface to text archives.
11. Suno, which gave a text interface to song archives.
12. DALL-E, which gives a text interface to image archives (and by extension to video archives).
13. GPT-4V, which gave language models the ability to see.

[Diego Quezada @diegoquezadac]

@hwchase17 Sounds like a never ending eval cycle. Then what, you want to evaluate the evals of the evals ?

[Diego Quezada @diegoquezadac]

@delcatta @brunoprietog he’s not even trying to hide the flex anymore 🔥

[Diego Quezada @diegoquezadac]

notebooklm is crazy good

[Andrej Karpathy @karpathy]

Oops sorry it's a new on-demand podcast on whatever source materials you give it it / link it. Generate them in Google's Notebook ML:
https://t.co/kN8GrLevOs

+ New Notebook
Link sources (whatever you want!)
Notebook guide > Deep dive conversation generate

[Andrej Karpathy @karpathy]

Oops sorry it's a new on-demand podcast on whatever source materials you give it it / link it. Generate them in Google's Notebook ML:
https://t.co/kN8GrLevOs

+ New Notebook
Link sources (whatever you want!)
Notebook guide > Deep dive conversation generate

[Diego Quezada @diegoquezadac]

@amanmathur_ @delcatta

[Diego Quezada @diegoquezadac]

@delcatta @brunoprietog bruno looking like a CEO

[François Chollet @fchollet]

When you use a tool to do a task you could have done yourself, your skills atrophy and you become dependent on the tool, giving it power over you.

This can be fine... Unless:

The task is vital, and the tool might not always be there.

You don't own the tool, giving someone else power over you.

The task is "thinking" -- surrendering your mind.

[François Chollet @fchollet]

These weakly abstract models can generalize, but only locally so. They still break down upon simple problem modifications or when encountering novel problems.

This has been true of all deep learning generated abstractions, across every domain and across all architectures.

[Diego Quezada @diegoquezadac]

@delcatta keep pushing

[Diego Quezada @diegoquezadac]

@OUIGO Je viens de recevoir un e-mail concernant l'annulation de ma réservation 8ZW966. Oigo a survendu des billets. Je ne peux pas accepter cela, il n'y a pas de billets pour le même jour, et je n'ai pas l'argent pour en acheter un nouveau, la date est plus proche et plus chère.

[Diego Quezada @diegoquezadac]

@__mharrison__ Type hints !

[Diego Quezada @diegoquezadac]

@Norbert_al @LATAM_CHI @DespegarCl Hola! Despegar no me ofreció ninguna devolución de dinero más que las tasas tributarias (100.000 app) por ende lo que hice fue cambiar la ruta (el destino) y las fechas en Despegar. Tuve que pagar una penalización (250.000 app), sin embargo, era la mejor opción.

[Diego Quezada @diegoquezadac]

@LATAM_CHI Me sorprende que como empresa no ofrezcan una verdadera solución a sus clientes para manejar complicaciones externas. Necesito cancelar/modificar el vuelo porque en Perú hay una crísis en donde ya han fallecido 46 personas. @DespegarCl me ofrece una devolución de menos del 16%.

[Diego Quezada @diegoquezadac]

@DespegarCl Tengo un vuelo a Cuzco el 1 de Febrero y, según la recomendación de las autoridades nacionales, es recomendable cancelarlo debido a la crísis social-política. ¿Qué opciones de devolución existen para estos casos especiales?

[Diego Quezada @diegoquezadac]

@DespegarCl Tengo un vuelo a Cuzco el 1 de Febrero y, según la recomendación de las autoridades nacionales, es recomendable cancelarlo debido a la crísis social-política. ¿Qué opciones de devolución existen para estos casos especiales?