netdur

it can refuse to do the task based on morals, if it think the output will be used to harm, the issue is not straight refuse, it can subtle refuse by producing results "designed" to avoid accomplish what you want to do

Didn't DeepSeek's CEO say that Llama is two generations behind, and that's why they didn't use their methods?

While BM25 did emerge from earlier work in the 1970s and 1980s (specifically building on the probabilistic ranking principle), I'm curious about your perspective on a few things:

What specific modern statistical approaches are you seeing as superior replacements for BM25 in practical applications? I'm particularly interested in how they handle edge cases like rare terms and document length normalization that BM25 was explicitly designed to address.

While I agree learning-based approaches have shown impressive results, could you elaborate on what you mean by search being "strictly dominated" by learning methods? Are you referring to specific benchmarks or real-world applications?

I found YOLOS to be faster and better, bot real time but 22k objects under half second

TL;DR: Real-time Linux finally merged into mainline after 18+ years. Good for robots, not your desktop.

Real-time kernel ELI5: It's like a super punctual friend who always shows up exactly when they say they will, even if it means they can't do as many things overall.

Key points:

- Guarantees worst-case execution times

- Useful for stuff like lasers, car brakes, Mars rovers

- Been around forever as patches, now official

- Linus quote from 2006: "Controlling a laser with Linux is crazy, but..."

Honestly, this won't change much for most of us. Your Firefox tabs aren't suddenly going to be more responsive. But it's a big deal for embedded systems and industrial applications. Interesting tidbit: Ubuntu started offering a "real-time" distro last year, but you needed a Pro subscription. Wonder how this will affect that business model.

Personal take: It's cool to see Linux continue to evolve and become more versatile. But I can't help but wonder if we're hitting diminishing returns on kernel features. Maybe the next big innovations will be at higher layers of the stack?

Anyone here actually using real-time Linux in production? Would love to hear some war stories.

Yes, the size is 99% 2 models weights required to run inference offline, there no way around it.

We have not decided what to do with it yet. It could be free, paid, or open source. However, the logic code for using semantic search with CLIP-compatible models on Android will be available on GitHub.

I have made similar android app for semantic image search, works offline too, still gathering feedback and polishing UI, but it works, if you are brave enough here is it https://drive.google.com/file/d/1tE0cY6umj5h5zCY_Jvaou1M8sCf...

The problem with Flutter is that it's a UI framework that relies on a single main thread. If you put something heavy there, you'll get lag. In contrast, other languages like SwiftUI let you easily spawn threads. Flutter dismisses this as outdated and instead promotes an isolated model with completely separate memory... for reasons.

Yes, Egypt has faced significant financial challenges. Recently, there have been substantial investments from the UAE, purchasing a city for $35 billion, with potential similar investments from Saudi Arabia. Despite these inflows, concerns about how these funds are managed, particularly regarding military influence, remain prevalent

got it, I will not build micro services, I will build just regular one.

I understand the methods to address the fine-tuning and RAG issues but lack the time and possibly the technical skills to implement the solution. Fine-tuning can potentially dumb down a perfect model, and RAG has context limitations and may not cover all content. My thinking, we should vectorize the text and embed these vectors into all layers of the model at inference time. This approach would bypass the context size limitations and resource wastage associated with fine-tuning, as vectorization is fast. I believe this vectorization and embedding strategy is the solution.

Use Gemini 1.5 Pro, which has 1.5 million tokens. Prompt it with a logical question and observe it struggling to answer. Then, upload a book on logical thinking in PDF format and ask the same question again. Notice how it can now answer the question effectively.

Could you share more about your strategy or the approach you took for chunking the articles? I'm curious about the criteria or methods you used to decide the boundaries of each chunk and how you ensured the chunks remained meaningful for the search functionality. Thanks!

Yes, The team behind Ximian, before being acquired by SUSE, was involved in early efforts to improve the usability of desktop Linux for end users. They conducted usability studies and published videos of these sessions to highlight where users encountered difficulties. These efforts were part of a broader initiative within the GNOME project to enhance user experience and make the GNOME desktop environment more intuitive and accessible to a wider audience.

I believe Google stands apart; although their current offerings sucks, they possess all the necessary components, including software, data, and hardware.

Yes Quantization compresses float32 values to int8 by mapping the large range of floats to a smaller integer range using a scale factor. This scale factor is key for converting back to floats (dequantization), aiming to preserve as much information as possible within the int8 limits. While quantization reduces model size and speeds up computation, it trades off some accuracy due to the compression. It's a balance between efficiency and model quality, not a magic solution to shrink models without losing some performance.

Quantization is essential for me since a 7B model won't fit on my RTX 2060 with only 6GB of VRAM. It allows me to compress the model so it can run on my hardware.

> - Need moar GPUs..

Why is there not a greater focus on quantization to optimize model performance, given the evident need for more GPU resources?

original Mistral or GGUF one?

oh thanks! I am getting down votted for a typo? reminds me grammar nazies during slashdot days... okay

yes RAM is shared between CPU and GPU, I am able to run Mixtral-8x7B in this machine at decent speed, this is pretty good for testing LLM, yes

I have MBP M1Max 32GB - it is $20000 now very good for LLM.

I had toshiba satellite with some whooping 96MB RAM... as main computer even... happily ran Windows 98... them I got the book "Linux from scratch" the rest is history... now I am happy mac user.

I hepled jQuery UI which landed me first real job with contract and everything, first side project was react like for canvas where only diff is drawen not all canvas, I have full family eating from this.

maybe he is CEO or someone rich.

12,800 CUDA - $4,000.00

Sorry Microsoft, you had a small window to win on search with ChatGPT, but instead, you opted to immediately use this opportunity to force users to install software they don't care about. You are small because of you, not because of Google.

I am not sure how can a binary have open source license at first place

later he said HF's tensor has almost zero copy, so 720GB is not needed using this format

Generally worst than GPT4 but have some killer features, today I asked it for Mortal Kombat 1 release time in my time zone, I can also upload photo and have conversation about it

But if you really wonder what they are building, get access to maker suite and play with, there is nothing comparable to it, only issue for it supports English only