[HN Gopher] Brave Leo now uses Mixtral 8x7B as default
       ___________________________________________________________________
        
       Brave Leo now uses Mixtral 8x7B as default
        
       Author : b_mc2
       Score  : 209 points
       Date   : 2024-01-27 05:04 UTC (17 hours ago)
        
 (HTM) web link (brave.com)
 (TXT) w3m dump (brave.com)
        
       | rhdunn wrote:
       | If you want to run Mixtral 8x7B locally you can use llama.cpp
       | (including with any of the supporting libraries/interfaces such
       | as text-generation-webui) with
       | https://huggingface.co/TheBloke/Nous-Hermes-2-Mixtral-8x7B-S....
       | 
       | The smallest quantized version (2bit) needs 20GB of RAM (which
       | can be offloaded onto the VRAM of a decent 4090 GPU). The 4bit
       | quantized versions are the largest models that can just about fit
       | onto a 32GB system (29GB-31B). The 6bit (41GB) and 8bit (52GB)
       | models need a 64GB system. You would need multiple GPUs with
       | shared memory if you wanted to offload the higher precision
       | models to VRAM.
       | 
       | I've experimented with the 7B and 13B models, but haven't
       | experimented with these models yet, nor other larger models.
        
         | jodleif wrote:
         | I prefer koboldcpp over llama.cpp. It's easy to spilt between
         | gpu/cpu on models larger than VRAM
        
           | DrSiemer wrote:
           | Runs in Oobabooga textUi as well, if you add the llama.cpp
           | extension. Easier interface imo, plus fun stuff like coqui
           | and whisper integration.
        
           | rhdunn wrote:
           | That's interesting. It also looks like koboldcpp works better
           | with long interactions, as it only processes changed tokens.
           | I'm using llama.cpp with text-generation-webui and its OpenAI
           | compatible API. I'll have to look to see if I can use
           | koboldcpp with it.
        
             | sp332 wrote:
             | Llama.cpp has an interactive mode, but I don't think text-
             | generation-webui uses it. https://github.com/ggerganov/llam
             | a.cpp/blob/master/examples/...
        
             | jodleif wrote:
             | Indeed. Koboldcpp works fine with other UIs than the
             | bundled one.
        
           | magicalhippo wrote:
           | I've got an aging 2080Ti and Ryzen 3800X with 96GB RAM, any
           | point in trying to mess with the GPU or?
           | 
           | Haven't really been able to justify upgrading to a 4090 or
           | similar given I play so few new games these days.
        
             | htsh wrote:
             | Yes, offloading some layers to the GPU and VRAM should
             | still help. And 11gb isn't bad.
             | 
             | If you're on linux or wsl2, I would run oobabooga with
             | --verbose. Load a GGUF, start with a small number of GPU
             | layers and creep up, keeping an eye on VRAM usage.
             | 
             | If you're on windows, you can try out LM Studio and fiddle
             | with layers while you monitor VRAM usage, though windows
             | may be doing some weird stuff sharing ram.
             | 
             | Would be curious to see the diffs. Specifically if there's
             | a complexity tax in offloading that makes the CPU-alone
             | faster but in my experience with a 3060 and a mobile 3080,
             | offloading what I can makes a big diff.
        
               | macNchz wrote:
               | > Specifically if there's a complexity tax in offloading
               | that makes the CPU-alone faster
               | 
               | Anecdotal, but I played with a bunch of models recently
               | on a machine with a 16GB AMD GPU and 64GB of system
               | memory/12 core CPU. I found offloading to significantly
               | speed things up when dealing with large models, but there
               | was seemingly an inflection point as I tested models that
               | approached the limits of the system, where offloading did
               | seem to significantly slow things down vs just running on
               | the CPU.
        
             | baq wrote:
             | I had only cuda installed and it took 2 ollama shell
             | commands in WSL2 from quite literally 0 local LLM
             | experience to running mixtral fast enough on a 1070 and
             | 12700k. Go for it.
        
           | sp332 wrote:
           | Llama.cpp has --n-gpu-layers that lets you set how much of
           | the model to put on the GPU.
        
           | attentive wrote:
           | kobold bundles and runs llama.cpp. So it should be fairly the
           | same with convenient defaults.
        
         | DreamGen wrote:
         | When talking about memory requirements one also needs to
         | mention the sequence length. In case of Mixtral, which supports
         | 32000 tokens, this can be a significant chunk of the memory
         | used.
        
         | viraptor wrote:
         | And if you want better performance when talking about code, you
         | can try the dolphin-mixtral fine tuning
         | https://huggingface.co/TheBloke/dolphin-2.7-mixtral-8x7b-GGU...
        
         | thriw63748 wrote:
         | Why not normal RAM? Ryzen 5600 with 128GB DDR4 is perfectly
         | fine to run mixtral 8bit, and costs less than $1000.
         | 
         | GPUs are only needed if you can not wait 5 minutes for an
         | answer, or for training.
        
           | snowfield wrote:
           | Or if you want multiple sessions at the same time. Or if you
           | want to do anything else with your machine while it's
           | running.
           | 
           | But realistically, 5 minutes is too long. It should be
           | conversational, and for that you need at least 5 tokens per
           | second. Which your Ryzen just can't do.
        
             | MPSimmons wrote:
             | >It should be conversational, and for that you need at
             | least 5 tokens per second.
             | 
             | To be fair, a lot of people are using this for non-
             | interactive work, like batching document analysis or
             | offline processing of user generated content.
        
               | Gracana wrote:
               | In my experience, it takes some experimentation to figure
               | out a good prompt. I don't think I would have gotten very
               | far off I had to wait that long for each result.
        
               | Diti wrote:
               | This particular thread we are commenting on is about
               | Dolphin Mixtral, which is mostly used for offline code
               | completion (a la Microsoft GitHub Copilot). You don't
               | want to have to wait 5 minutes at every keystroke to get
               | code suggestions.
        
           | rhdunn wrote:
           | That was what I was referring to with the 32/64 GB systems.
        
           | irusensei wrote:
           | Why not both? Llama.cpp allows layering GGUF models between
           | GPU and CPU memory.
        
           | SkyMarshal wrote:
           | What's the bandwidth between the Ryzen and that DDR4?
        
           | brucethemoose2 wrote:
           | > perfectly fine
           | 
           | Only for very short context and responses.
           | 
           | Beyond that, the performance is painful.
        
           | juliangoldsmith wrote:
           | Not sure why you're getting downvoted. It performs decent
           | enough on my Ryzen 3600X with 64GB of RAM. It definitely
           | wouldn't be usable for production or fine-tuning, but it's
           | fine for experimenting.
        
           | dragonwriter wrote:
           | > GPUs are only needed if you can not wait 5 minutes for an
           | answer
           | 
           | Yeah, but that's generally true (or at least, "5 minutes for
           | an answer is very suboptimal", even if "can't" isn't quite
           | true) for interactive use cases, which are... a lot of LLM
           | use cases.
        
         | tarruda wrote:
         | > You would need multiple GPUs with shared memory if you wanted
         | to offload the higher precision models to VRAM.
         | 
         | Or just a powerful apple silicon machine? I've tried dolphin
         | mixtral 4bit on a 36gb ram MacBook m3, and inference is super
         | fast.
        
           | pennaMan wrote:
           | I can run 4bit on a beat up 1070 ti. GP talks about higher
           | precision models
        
             | sp332 wrote:
             | You wouldn't be able to fit the whole model into 8GB VRAM.
             | It's faster than not using a GPU at all, but most of it
             | would still be computed on the CPU.
        
               | baq wrote:
               | IME ollama ran mixtral on a 1070 fast enough.
        
           | mkesper wrote:
           | Or a Linux machine with a Ryzen using the internal GPU and
           | the unified RAM (scroll down at llama.cpp and look for ROCm).
        
             | stavros wrote:
             | What? I can do this? _Runs to the PC_
             | 
             | EDIT: I cannot, I need to install ROCm to compile with it,
             | and then install something called hipBLAS, and who knows
             | what else.
        
               | assbuttbuttass wrote:
               | OpenCL should also work on AMD cards, and is way easier
               | to install
        
               | stavros wrote:
               | Does that let me use unified memory on the GPU, though?
               | Or is it just so I can use my CPU memory?
               | 
               | EDIT: Oh, no, I have an nVidia GPU, AMD CPU.
        
               | brucethemoose2 wrote:
               | It is dead slow on integrated graphics, unfortunately.
        
             | jacooper wrote:
             | Wait ROCm support Ryzen APUs and still doesn't support
             | dedicatedly GPUs like the 6700XT?!
        
               | jacooper wrote:
               | Supports* dedicated*
        
             | KeplerBoy wrote:
             | Or a jetson orin agx (~2k$). Probably the cheapest way to
             | get an Nvidia GPU with 64 GB of RAM.
        
               | robterrell wrote:
               | Have you tried this yourself? Curious to know how well
               | this works for an LLM home lab.
        
               | kkielhofner wrote:
               | I've worked with Jetson going back to the TK1 and I
               | highly recommend you do not do this.
               | 
               | Nvidia has significant dominance in the AI space because
               | of their work on software and the overall platform.
               | 
               | With the Jetson line being the sole exception. Use it for
               | what it's for - a targeted build for an embedded/specific
               | application requiring small size and low power.
               | 
               | The software is a mess. Support for Jetson (generally) is
               | a far afterthought or not considered at all around
               | projects at Nvidia and the broader ecosystem. When it is
               | supported at all it lags behind significantly, using
               | ancient distros (Jetpack), etc. To make matters worse the
               | user base is so (relatively) tiny there are bugs and
               | strange behavior everywhere.
               | 
               | Just don't do it.
        
               | qrios wrote:
               | According to this article [1] it looks like there is no
               | complex preparation needs to run the inference on a
               | Jetson system. Should work with Mixtral too.
               | 
               | [1] https://www.hackster.io/pjdecarlo/llama-2-llms-w-
               | nvidia-jets...
        
               | KeplerBoy wrote:
               | I haven't tried it for LLMs yet, i use it for real time
               | RF processing, but I actually have one of them on my desk
               | and they are fun little devices.
               | 
               | Maybe I will try to get a 32 GB+ LLM running one of those
               | days.
        
               | fsiefken wrote:
               | I wonder what would be the cheapest way to run an LLM,
               | with the latest Ryzen integrated graphics and 64G Ram or
               | the Jetson AGX Orin 64. https://www.nvidia.com/en-
               | us/autonomous-machines/embedded-sy...
        
               | KeplerBoy wrote:
               | The Ryzen is a lot cheaper, but most likely also a fair
               | bit slower. You'd be looking at a 200$ CPU, 200$
               | Motherboard + 200$ of ddr5 ram. Throw in a case, nvme
               | drive and power supply and you're still below $1k and
               | those numbers are quite generous estimates, you could do
               | it a lot cheaper by going AM4 with DDR4 ram.
        
           | asimpleusecase wrote:
           | Did you do anything special to make that work? Is it useful?
           | Or just a toy?
        
             | EarthLaunch wrote:
             | I've been using it for 'easy' queries like syntax/parameter
             | questions, in place of ChatGPT 4. It's great for that. I am
             | using a ~48GB version.
        
             | windexh8er wrote:
             | I have a 14" MBP with an M1 Max and 64GB. The M3 won't
             | really make a difference, but the RAM, since unified, is
             | huge. I can run most models on this machine with realtime
             | performance compared to a Ryzen 7735HS and 64GB (DDR5). Now
             | I'm not saying the Ryzen setup _should_ be good, but the M1
             | architecture just makes it a much better option. I could
             | add an eGPU to the Ryzen system and it could likely do
             | better, but would also exceed the price point and
             | portability.
        
             | brucethemoose2 wrote:
             | I've had less luck with Mixtral, but I run Yi 34B finetunes
             | for general personal use, including quick queries for work.
             | 
             | Its kinda like GPT 3.5, with no internet access and
             | slightly less reliable responses, but unrestrained, _much_
             | faster and with a huge (up to 75K on my Nvidia 3090) usable
             | context.
             | 
             | Mixtral is extremely fast though, at least at a batch size
             | of 1.
        
               | simonw wrote:
               | Which Yi 34B finetunes are you using that have a 75,000
               | token length?
        
               | brucethemoose2 wrote:
               | All of the Yi 200K finetunes should support it, but you
               | have to be careful because some degrade the base model's
               | quite excellent long context performance more than
               | others. The very strong Bagel 34B DPO model, for
               | instance, basically doesn't work at long context.
               | 
               | Nous Capybara is a popular one. I personally use my own
               | merge of many models, and you can look through the
               | constituent models to see if any interest you:
               | https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-
               | megame...
               | 
               | You can't really use llama.cpp for super long context
               | btw, its just too slow and vram inefficient at the
               | moment.
        
         | chpatrick wrote:
         | `ollama run mixtral:8x7b-instruct-v0.1-q3_K_L` works fast on my
         | 3090 locally
        
         | beck5 wrote:
         | Dumb question, but how can a 32 bit number be converted to 2
         | bits and still be useful? It seems like magic.
        
           | ttoinou wrote:
           | All the 32 bits weren't necessarily used, and it's the whole
           | network itself that has to be useful. It's a tradeoff. We
           | started with very good precision to test the new method, now
           | we can optimize some parts of it
        
           | sp332 wrote:
           | The extra precision is more useful for training. Once the
           | network is optimized, it's a statistical model and only needs
           | enough precision to make good guesses. In fact, one of the
           | big papers on this also pointed out that you can drop about
           | 40% of the weights completely. I think people generally skip
           | that part because sparse matrix operations are slower, so it
           | doesn't help here.
        
             | viraptor wrote:
             | For models with dropped weights, the keyword is
             | "distilled". For example ssd-1b is a 50% size version of
             | Stable Diffusion XL (https://huggingface.co/segmind/SSD-1B)
        
               | sp332 wrote:
               | That's crazy, I've never seen one that dropped whole
               | layers from a pre-trained model. I guess that avoids the
               | sparse matrix math.
        
           | beefield wrote:
           | I think of it with this kind of analogy: the original image
           | is stored with 32 bit color scheme. You can reduce the color
           | scheme to 16 bit accuracy and still figure out pretty well
           | what the image is about. 2 bit is stretching this to a bit
           | far, basically either pixel is white or it is black, but even
           | if you lose lots of nuances in the image, in many images even
           | that gives you some idea whats going on in the image.
        
             | DougBTX wrote:
             | That's an interesting question, I wonder if there is an
             | analogy in quantisation to image dithering?
        
           | DougBTX wrote:
           | Nice graphs here:
           | https://github.com/ggerganov/llama.cpp/pull/1684
           | 
           | So for example, 2 bit version of the 30B is much worse than
           | the original, but still better than the 13B model.
           | 
           | Also, there are lots of extra details, eg, not all of the
           | weights are 2 bit, and even the 2 bit weights are higher than
           | that overall as groups of quantised weights share scale
           | factors stored elsewhere.
        
           | regularfry wrote:
           | Mixtral and others are often distributed as 16-bit floats, so
           | that chops the problem in half immediately, but then it turns
           | out that LLMs only have about four bits per parameter of
           | actual information stored. There's a lot of redundancy. The
           | ideal quantisation scheme would only throw away useless data,
           | but no quantisation scheme is perfect so they inevitably harm
           | the model somehow.
           | 
           | You've then got to remember that one thing neural networks
           | are very, very good at is being noise tolerant. In some
           | senses that's all they are - noise correction systems. The
           | inaccuracies introduced by quantisation are "just" a sort of
           | noise, so it's not surprising that they aren't fatal. It just
           | raises the noise floor and gives the model more ways to be
           | wrong.
           | 
           | Finally the thing to know is that these quantisation schemes
           | don't do a naive "chop each number down to two bits", not
           | exactly. Simplifying a bit, for each parameter in this
           | example they'd try to find a mapping from a two-bit index
           | into a four element lookup table of higher-precision values
           | such that the information destroyed by replacing the original
           | parameter by the lookup value is minimised. That mapping is
           | calculated across small blocks of parameters, rather than
           | across the entire model, so it can preserve local detail. The
           | lookup table gets stored per block, which throws the
           | compression ratio off a little.
        
           | brucethemoose2 wrote:
           | Its not really 2 bits.
           | 
           | Modern quantization schemes are almost like lossy compression
           | algorithms, and llms in particular are very "sparse" and
           | amenable to compression.
        
           | hnfong wrote:
           | This blog post might shed some light on the matter. If I'm
           | understanding it correctly, it claims there are emergent
           | features on the LLM weights that make it easier to "compress"
           | the floats into smaller bits without losing much precision.
           | 
           | https://timdettmers.com/2022/08/17/llm-int8-and-emergent-
           | fea...
           | 
           | Note that 2 bit quantization is generally regarded as too
           | aggressive. Generally 4bits+ achieves a good tradeoff, see
           | eg. https://arxiv.org/abs/2212.09720
        
         | zaat wrote:
         | Using TheBloke nous-hermes-2-mixtral-8x7b-dpo.Q5_K_M.gguf,
         | llam.cpp with oobabooga, on Ryzen 9 7950x and RTX 2070 super,
         | offloading 4 layers to the GPU.
         | 
         | 27,631MiB system memory used.
         | 
         | 7358/8192MiB usage on the GPU.
         | 
         | I'm getting 6.32 tokens/second.
         | 
         | You: You are Hunter S. Thompson, write a 450 words essay on the
         | LHC project. The writing must be in Thompson's style
         | 
         | AI: Well, buckle up, kiddos, because we're about to dive into
         | the twisted world of particle physics and the Large Hadron
         | Collider, or as I like to call it, the Doomsday Machine. This
         | beast of a thing is buried deep beneath the Swiss countryside,
         | a 17-mile-long ring of death and destruction that makes the
         | atom bomb look like a goddamn firecracker.
         | 
         | Now, let me tell you, when the Swiss government first proposed
         | this monstrosity, I thought they'd gone completely batshit
         | crazy. But then again, this is the land of Toblerone and Heidi,
         | so maybe I shouldn't be too surprised. The LHC is essentially a
         | giant, underground racetrack for subatomic particles, where
         | they're accelerated to nearly the speed of light before being
         | smashed into each other like a couple of drunken frat boys at a
         | kegger.
         | 
         | The goal of this madness? To recreate the conditions that
         | existed just after the Big Bang, when the universe was nothing
         | more than a seething cauldron of energy and chaos. And if that
         | doesn't sound like a recipe for disaster, I don't know what
         | does.
         | 
         | But let's not forget the real reason we're all here: the Higgs
         | boson, or as it's more commonly known, the God Particle. This
         | little bugger is what gives other particles their mass, and
         | without it, we'd all be floating around like a bunch of cosmic
         | tumbleweeds. The LHC was designed to find this elusive
         | particle, and in 2012, they finally did it. Sort of.
         | 
         | You see, the Higgs boson they found was a bit of a
         | disappointment. It was too light, too wimpy, and didn't behave
         | quite the way they expected it to. So now, the physicists are
         | back at the drawing board, trying to figure out what the hell
         | went wrong. And in the meantime, the LHC continues to churn out
         | more and more data, like some kind of unholy particle-smashing
         | machine.
         | 
         | But let's not forget the dark side of the LHC. There are those
         | who believe that this thing could very well be the end of the
         | world as we know it. They fear that the LHC could create a
         | black hole, or even a miniature Big Bang, right here on Earth.
         | And while the scientists assure us that these fears are
         | unfounded, I can't help but wonder if they're just whistling
         | past the graveyard.
         | 
         | So, what have we learned from all this? Well, for one thing,
         | we've discovered that the universe is a much stranger place
         | than we ever imagined. We've also learned that even the most
         | brilliant minds can be blinded by their own ambition, and that
         | sometimes, the pursuit of knowledge can lead us down some very
         | dark paths.
         | 
         | In the end, the LHC is a testament to the insatiable curiosity
         | of the human race, and our unquenchable thirst for knowledge.
         | It's a monument to the power of science, and the potential
         | consequences of playing God. And as we continue to push the
         | boundaries of what's possible, we must always remember that
         | with great power comes great responsibility. Or, as I like to
         | say, "When the going gets weird, the weird turn pro."
        
           | tacoman wrote:
           | I am using the exact same model. Ryzen 5600G w/32GB and an
           | Nvidia P40 w/24GB VRAM
           | 
           | 20/33 layers offloaded to GPU, 4K context. Uses 25GB system
           | RAM and all 24GB VRAM. 5-7 tokens per second.
        
             | zaat wrote:
             | Context is set to 32768, I didn't change it I guess that's
             | the model's default.
             | 
             | Thanks for making me feel better about investing in tht
             | motherboard + CPU + RAM upgrade and deferring the GPU
             | upgrade.
        
         | bearjaws wrote:
         | 2bit is pretty damn terrible, I don't recommend it for anything
         | serious.
        
           | ignoramous wrote:
           | At that level of quantization / distillation, smaller models
           | like _phi-2_ (q &a) and _wavecoder-6.7b_ (code-gen) might be
           | preferable over QLoRAd ones:
           | https://huggingface.co/microsoft/phi-2
           | 
           | > _2bit is pretty damn terrible_
           | 
           | Wait till you go hybrid [0] or even 1bit [1]
           | 
           | [0] https://github.com/efeslab/Atom
           | 
           | [1] https://github.com/IST-DASLab/qmoe
        
         | EVa5I7bHFq9mnYK wrote:
         | Faraday.dev has it in its selection of models now. Good for us
         | clueless Windows folks. Runs decently fast with 16gb mobile
         | 3080 gpu. Results seem better than any other free option.
        
         | MuffinFlavored wrote:
         | What differences would I measurably notice running the 2-bit
         | version vs the 4-bit version vs the 6-bit vs the 8-bit?
        
       | davikr wrote:
       | It's nice using Brave because you have Chromium's better
       | performance, without having to worry about Manifest V2 dying and
       | taking adblocking down with it. I have uBlock Origin enabled, but
       | it has barely caught anything that slipped past the browser
       | filters.
        
         | deely3 wrote:
         | Is this submarine comment?
        
           | thebruce87m wrote:
           | What is the definition of a submarine comment? Google fails
           | and ChatGPT says:
           | 
           | > A "submarine comment" on social media refers to a comment
           | that is made on an old post or thread, long after the
           | conversation has died down. This term derives from the idea
           | of a submarine which remains submerged and out of sight for
           | long periods before suddenly surfacing. In the context of
           | social media, it's when someone delves deep into someone
           | else's posts or timeline, finds an old post, and leaves a
           | comment, bringing the old post back to attention. This can
           | sometimes surprise the original poster and other
           | participants, as the conversation was thought to have been
           | concluded.
           | 
           | Which doesn't make sense in this context
        
             | mcbuilder wrote:
             | I think GP is trying to coin a term for stealth marketing
             | Hacker News comments, except the analogy doesn't really
             | make sense.
        
               | freediver wrote:
               | Seen on HN first!
        
               | deely3 wrote:
               | Definition: https://paulgraham.com/submarine.html
        
             | deely3 wrote:
             | Definition: https://paulgraham.com/submarine.html
        
         | croes wrote:
         | Brave"s support of Manifest V3 is totally dependent on Google
         | and Chrome
         | 
         | >Brave will support uBO and uMatrix so long as Google doesn't
         | remove underlying V2 code paths (which seem to be needed for
         | Chrome for enterprise support, so should stay in the Chromium
         | open source)
         | 
         | https://twitter.com/BrendanEich/status/1534893414579249152
        
           | davikr wrote:
           | Yeah, but the Brave adblocker is built-in, it's not an
           | extension.
        
             | croes wrote:
             | There are more extensions than just ad blockers
        
           | steve_rambo wrote:
           | It doesn't really matter much in practice. The built-in
           | adblocker (which does not rely on extension APIs) has
           | excellent performance, about as effective as ublock origin,
           | and supports the same filter lists.
        
             | croes wrote:
             | It does matter if you use other extensions that require
             | Manifest V2
        
             | guipsp wrote:
             | The built-in adblocker still relies on a large amount of
             | extension APIs.
        
         | charcircuit wrote:
         | MV3 doesn't prevent adblockers from existing.
        
           | rpastuszak wrote:
           | It makes them almost useless in practice.
        
             | charcircuit wrote:
             | That is a baseless statement. It doesn't make them useless
             | as they can still block ads.
        
             | HeatrayEnjoyer wrote:
             | Because the filter list is capped, right? Is there a reason
             | the Brave team cannot just remove or increase the cap?
        
               | gkbrk wrote:
               | Not just because of the filter list cap. It also reduces
               | ad blockers to static filter lists instead of powerful
               | dynamic filters.
               | 
               | MV3 makes it impossible for ad-blockers to inspect
               | requests with code and then allow/deny dynamically.
        
               | charcircuit wrote:
               | >It also reduces ad blockers to static filter lists
               | instead of powerful dynamic filters.
               | 
               | This is very outdated information and borderline
               | misinformation by representing it as how it currently
               | works. It allows for 30,000 dynamic rules and 5,000
               | session rules (session rules only persist until the
               | browser is closed).
               | 
               | >MV3 makes it impossible for ad-blockers to inspect
               | requests with code and then allow/deny dynamically.
               | 
               | Giving this ability to extensions can slow down the
               | browser for the user. These ads can still be blocked
               | through other means.
        
               | charcircuit wrote:
               | >Because the filter list is capped, right?
               | 
               | The limits are 300,000 static rules [1] + 30,000 dynamic
               | rules [2] + 5,000 session rules [3]. For reference
               | easylist is about 35k rules. The Chrome team has been
               | constantly tweaking these limits themselves and Brave
               | could set their own limits if they wish. The API is
               | designed such that extensions can query to see how many
               | rules they can use.
               | 
               | [1] https://source.chromium.org/chromium/chromium/src/+/m
               | ain:ext...
               | 
               | [2] https://source.chromium.org/chromium/chromium/src/+/m
               | ain:out...
               | 
               | [3] https://source.chromium.org/chromium/chromium/src/+/m
               | ain:out...
        
         | srid wrote:
         | > I have uBlock Origin enabled, but it has barely caught
         | anything that slipped past the browser filters.
         | 
         | I have been completely satisfied with Brave's builtin ad-
         | blocker. Does uBlock Origin catch anything in particular that
         | the builtin one does not?
        
           | rozab wrote:
           | I couldn't see Brave being able to counter, say, the recent
           | YouTube anti-adblock push as quickly as the open source
           | community did. I could see that kind of stunt becoming more
           | common as Google tries to nail the coffin shut and deny their
           | competitors this USP
        
           | SparkyMcUnicorn wrote:
           | Brave's ad blocking is compatible (and uses) the uBlock
           | Origin lists, so you shouldn't really notice a difference
           | between having uBO enabled or disabled in brave.
        
         | Diti wrote:
         | If by performance you mean browser performance, you have more
         | performance with Firefox nowadays.
         | https://news.ycombinator.com/item?id=36770883
        
           | jacooper wrote:
           | While still not being as secure as chromium and still not
           | supporting many advanced features like WebGpu.
        
           | FoodWThrow wrote:
           | I use firefox because it has the most hassle-free hardware
           | decoding in linux. However, everything basically feels better
           | with Brave, even with the same amount of plug-ins.
        
             | kristianp wrote:
             | I found the snap update notifications too annoying on
             | Ubuntu, so I tried the ppa. But it the video plugin would
             | crash. So back to Chrome for me.
        
       | firtoz wrote:
       | What are good API providers that serve mixtral? I know only octo
       | ai which seems decent but will be good to know alternatives too
        
         | jug wrote:
         | You also have Replicate:
         | https://replicate.com/mistralai/mixtral-8x7b-instruct-v0.1
        
         | htsh wrote:
         | openrouter, fireworks, together.
         | 
         | we use openrouter but have had some inconsistency with speed. i
         | hear fireworks is faster, swapping it out soon.
        
         | Tiberium wrote:
         | OpenRouter is generally a good option (already mentioned), the
         | best part is that you have a unified API for all LLMs, and the
         | pricing is the same as with the providers themselves. Although
         | for OpenAI/Anthropic models they were forced (by the respective
         | companies) to enable filtering for inputs/outputs.
        
         | firejake308 wrote:
         | I personally like Anyscale Endpoints
        
         | mark_l_watson wrote:
         | I have used both Mistral's commercial APIs and also AnyScale's
         | commercial APIs for mixtral-8-7b- both providers are easy to
         | use.
         | 
         | I also run a 3 bit quantization of mixtral-8-7b on my M2 Pro
         | 32G memory system and it is fairly quick.
         | 
         | It is great having multiple options.
        
         | Philpax wrote:
         | I've had good experiences with Together, and they have very
         | competitive pricing.
        
         | bearjaws wrote:
         | Together.ai seems to be the best, incredibly fast.
        
           | RupertEisenhart wrote:
           | These guys are much faster than openrouter, and their llama2
           | runs faster than 3.5-turbo. Amazing work.
        
         | jeanloolz wrote:
         | The creators of the model actually have their own platform
         | where you can access this model and others via API:
         | https://console.mistral.ai/
        
         | petercooper wrote:
         | Both already mentioned, but I am using Anyscale Endpoints with
         | great success, very fast and will work on ten jobs at a go out
         | of the box. Together.ai also seems to work fine in my initial
         | tests, but haven't tried it at scale yet.
        
       | fifteen1506 wrote:
       | Just checking: PDF summarization is not yet implemented, right?
        
         | 1equalsequals1 wrote:
         | The Kagi browser extension can do that, if you're a subscriber
        
       | charcircuit wrote:
       | It's interesting that they made it so you can ask LLM queries
       | right from the omnibar. I wonder if they eventually will come up
       | with some heuristic to determine if thr query should be sent
       | directly to an LLM or if the query should use the default search
       | provider.
        
       | finikytou wrote:
       | quick question I have 24GB VRAM and I need to close everything to
       | run MIXTRAL at 4 bit quant with bitsandbyte. there is no way to
       | run it at 3,5 on windows?
        
       | syntaxing wrote:
       | Interesting, I must have missed the first Leo announcement. I
       | really like how privacy conscious it is. They don't store any
       | chat record which is what I want.
        
         | Dwedit wrote:
         | There is no way to confirm that claim, just like there is no
         | way to confirm that a VPN service is "no log".
        
           | Erratic6576 wrote:
           | You gotta trust them by their word
        
           | lolinder wrote:
           | Yes, at some point if you're going to interface with other
           | humans you will eventually just have to trust their word.
           | 
           | For some people's threat models that isn't good enough, but
           | for the vast majority of people--people who aren't being
           | pursued by state intelligence agencies but who are squeamish
           | about how much data a company like Google collects--a pinky
           | promise from Brave or Mullvad is good enough.
        
             | bcye wrote:
             | I would like to think GDPR ensures this pinky promise is
             | good enough
        
             | wolverine876 wrote:
             | > For some people's threat models that isn't good enough,
             | but for the vast majority of people--people who aren't
             | being pursued by state intelligence agencies but who are
             | squeamish about how much data a company like Google
             | collects--a pinky promise from Brave or Mullvad is good
             | enough.
             | 
             | Who are you to say it's good enough (and ridicule people
             | who disagree)? We don't have too much evidence of it,
             | because they have very few options and of course most
             | people are not informed and lack the expertise to
             | understand the issues (a good situation for regulation). At
             | one point lots of people used lead paint and were fine with
             | it; they would have told us.
             | 
             | > Yes, at some point if you're going to interface with
             | other humans you will eventually just have to trust their
             | word.
             | 
             | There's technology, such as the authorization tokens used
             | by Brave, that reduces that risk. Of course, no risk can be
             | complete eliminated but that doesn't mean we shouldn't
             | reduce it.
        
               | lolinder wrote:
               | > say it's good enough (and ridicule people who
               | disagree)?
               | 
               | I'm not ridiculing anyone, I explicitly say that for some
               | people's threat models it isn't good enough.
        
               | wolverine876 wrote:
               | You said that people who agreed with you are,
               | 
               | > people who aren't being pursued by state intelligence
               | agencies
               | 
               | That implies that the only valid reason to disagree is if
               | you are pursued by state intelligence agencies. Obviously
               | it's ridiculous to think that you are.
        
               | salad-tycoon wrote:
               | Well, these days a lot of people are actually pursued.
               | You know like parents, social activists, various
               | religious movements. It's not just only the Dr. Evils of
               | the world anymore.
        
         | Tanath wrote:
         | It's foolish to trust Brave. https://www.spacebar.news/p/stop-
         | using-brave-browser
        
       | kristianpaul wrote:
       | I run Mixtral locally using ollama
        
       | andai wrote:
       | Asked Mistral 8x7B for an essay on ham. It started telling me
       | about Hamlet.
        
         | Erratic6576 wrote:
         | It must start from the beginning. Pig > piglet. Ham > Hamlet
        
       | m3kw9 wrote:
       | If you have used gpt4 and then use mistral, it's like looking at
       | a Retina display and then have to go back to a low res screen.
       | You are always thinking "but GPT4 could do this though"
        
         | mpalmer wrote:
         | Have you used mixtral?
        
       | wolverine876 wrote:
       | Kudos to Brave (for this and other privacy features):
       | 
       |  _Unlinkable subscription: If you sign up for Leo Premium, you're
       | issued unlinkable tokens that validate your subscription when
       | using Leo. This means that Brave can never connect your purchase
       | details with your usage of the product, an extra step that
       | ensures your activity is private to you and only you. The email
       | you used to create your account is unlinkable to your day-to-day
       | use of Leo, making this a uniquely private credentialing
       | experience._
        
         | quinncom wrote:
         | This is very cool, and something I'd like to integrate in my
         | own apps. Does anybody know how this works exactly, not using
         | foreign keys?
        
           | luke-stanley wrote:
           | I could guess, an "anonymous payment credential service"
           | could do something like this: 1. User completes payment for
           | the paid for service, 2. To track the payment entitlement, a
           | random, unique ID is generated by the service for the user,
           | that is not related to any of their data. 3. This ID is saved
           | in a database as a valid payment key. 4. The database records
           | IDs in shuffled batches, or with semi-random fuzzy / low
           | resolution timestamps to prevent correlation between payment
           | time and ID generation. 5. Each ID has an entitlement limit
           | or usage stopping point, ensuring it's only valid for the
           | subscribed period.
           | 
           | Another way might be Zero-Knowledge Proofs (ZKPs), but that
           | might be more complex. They might even use their BAT crypto
           | stuff for this somehow, I suppose. Whatever solution, would
           | need a fundamental solution for how to avoid correlation, I
           | think.
        
       ___________________________________________________________________
       (page generated 2024-01-27 23:01 UTC)