pilooch

No. Because the model should and will replace any piece of code. This is what happened already for other tasks, in computer vision, text (entity recognition, etc...), audio, etc.

RAG will go away, decision multimodal models / LLMs will take over. No here yet, but inevitable I believe.

The adversarial surface to the LLM remains enormous, manual cannot handle it.

So Google hasn't used an LLM to generate and test weird queries ? This is not putting the bar very high for the whole industry... There'd be so much to gain from a clean deployment... Either it hard, either it is a rush. As a machine learnist, I believe it's actually impossible, by design of the autoregressive LLM. This race may we'll be partially to the bottom.

Detection of generated images works well, as reported by lots of works on arxiv. It's not catching them all, but lots of them. One way to understand it is that DDPMs and their underlying neural network do leave patterns in the high frequencies, often invisible to the eye. Deconvolutions typically generate low noise "marbles".

Agreed. And solid, very little to no maintenance. My biggest respect to the team behing jellyfin !

Look at policy-based templating, my goto for generic AI/search algorithms with many options. May prove useful.... Or not :)

We use it professionally at work, great enough tool!

Breakthrough may well be among the recent reformulations into supervised, offline RL and related techniques. Of special interest, the decision transformer and the very emerging litterature on diffusion planning.

Generative AI basically unlocks the metaverse, since all the 3D environment can now be generated. The metaverse changes people's context and senses with generated data. Interestingly AI will paint a new world for humans to learn and experience new situations more often, more easily on a broader spectrum. Exciting !

This has been true of machine learning APIs from the early renaissance of neural networks. Computer vision APIs have been used for years to grow datasets and fine-tune models on the data.

As a matter of fact, the API/SaaS business for machine learning 'direct' model results have suffered from this since the early days. Now, what may make more sense and survive longer are full data-storage-oriented platforms, and niche products that encapsulate all underlying model calls, including to publicly available models. Sometimes glue code or even devops becomes the true solid business case of AI/ML apps ;)

The big difference is there's no full training code for llama, yet.

Consciousness might just be the brain's recurrent and unsuccessful attempt at predicting itself.

Similarly, "free will" may simply be an illusion caused by the failure of that perfect prediction of what to do/what will happen, thus leaving us with a pleasant stochasticity that we end up favoring as freedom :)

Loving pytorch but the reasoning remains strange to my ears: Python at all cost, not Julia, because of the ecosystem, well OK.

But all bullet points are about things that are easily done right now with libtorch (pytorch underlying C++ core code), and the hassle is... Python.

Well rational conclusion would be, just do everything in C++, and bind to Python. Make C++ first citizen here, since in all cases it'll be needed for performance, forever.

My take on intelligence over the past 20y has been it is high quality / efficient search of immense state spaces.

"Solving intelligence" as a famous corporation motto, might just be improving state-space search.

Humans are incredible at state space search, it's obvious as soon as you consider the potential data pointsnof any problem we face every day, from washing dishes to designing algorithms.

TikTok algorithm is like a tamgotchi with short memory. The echo chamber is actual much less than on other systems, or at least this is my actual understanding. I've been working on AI and ML for 20+y now, I'm always interested is testing the systems out there from the user side. Overall, I've been surprised by TikTok, as I did find it very enjoyable actually, and easily 'trainable'.

ML is about writing programs automatically. It is huge, like the end of programming as we know it. Think of future Blockchain algorithms, crypto functions etc as trained instead of programmed. It'll take time but ML shall supersede every programming task. Computer science is about automation, ML is the next level of autation, of programming itself. It was abput time :)

Confidence is the key, but it is sometimes out of reach, even if only momentarily.

So rationalizing is one way: about competition first, it doesn't matter (really, like... really), whatever you are doing with focus will end up different than what others do (the state space is too large), and the rest is not within your hands. No need to worry then, it's a recurrent, automatic, bad habit.

Now, overall and most useful I believe, what we are doing in tech does not matter, it'll be outdated in months, years, whatever. What matters is the people we are working and spending time with. People first, tech second.

Good luck, serenity is within reach, especially in tech, it's a matter of body and mind working well together.

Oh and exercising is fundamental, walk, run, dance, jump, ...

Torchscript is very complicated. Its internal API was not truely documented for external use last time I checked it. You can't even export to torchscript from libtorch C++, it only works from Python AFAIK.

https://developer.apple.com/documentation/accelerate/blas This runs on the AMX, so much faster than on CPU.

I've been unable to get native openblas yet through homebrew, though havent tried recently.

For deep learning inference I'm using Vulkan/moltenvk/ncnn got it to work and was able to hack around.

This works fine until a point in life at which communicating with others and answering to others, meaning you are available for them, whatever the app, is more important than anything else.

Not meaning to not boycott such apps with those you can, just giving another pov from elsewhere in life. Most of my phones are rooted and without google for years, but sometimes...

Life is short, human communication is deep and fragile. Tools do change, renew and improve, while as living humans we get older ;)

Done that since 2015. You can look at https://github.com/jolibrain/deepdetect. C++ doesn't sound ideal to many, but when your target is production, it's pretty powerful, and since c++11 probably much more comfortable than most non practitioners think. For deep learning, it is excellent for bare metal and fitting with industrial applications. Never looked back. For R&D (gans, flows, RL, ...) Python remains easier to play with.

Yes, that's 250w consumption on both sides for a one-to-one vid all. Not sure it's a good trade-off...

I did a bit with instances of an Open Source deep learning server of mine, for three years+ now. Made up to 1.5k/month, lower now but it's on my plate to revive it. It's provided side money to our service/ software business.

Keep in mind there's a 30% cut from AWS, plus if you're in the EU, you need a tier like payoneer to channel your revenues back into the EU.

Well done but be careful, TSNE has some drawbacks, well detailed here: https://distill.pub/2016/misread-tsne/

For building a more precise 2D embedding or image similarity search engine you may want to rely on a neural embedding + FAISS or Annoy instead of compressing into 2D with TSNE. When doing the later, distances don't mean much anymore.

FTR, a simple RPi4 yields ~4FPS on a DL object detection model. You may need cloud for your central hub, but neural network inference can be done locally.

Agreed. AutoML is useful, but modeling and ensuring correct hypotheses are taken into account comes first.

See https://github.com/jolibrain/deepdetect/pull/641 that uses FAISS as a backend alternative to annoy (annoy supported as well). Deletion can be implemented by removing entries from the listing db while the vector remains within the index.

Tests show that FAISS is bit better than annoy on retrieval of both small and million items indexes. It also includes ind x compression techniques that in our tests do fair very well, with very low loss on mid size 500k image indexes.

As an ML + open source developer for over 10y, MacOS support for deep learning is already long gone, Linux is the prime AI/ML OS. However, Apple and NVidia parting away is a good omen for GPU competition I believe. Whatever Apple hw comes up with, if usable outside MacOS software stack, it'd be an interesting alternative.

What would be a good source of tetrachromat images ? Would love to train a model and "see" what these women see !

vs C++-14 ? Indeed most DL is in fact C++. The Pytorch recent C++ API is a must. As professionals in this industry, my colleagues and I have switched to full C++. I'd be interested in advantages of Rust vs C++ instead of Python (which truely in terms of performances is C in the background).