Mixtral-8x22B-v0.1 was released a couple days ago. The "mixtral:8x22b" tag on ollama currently refers to it, so it's what you got when you did "ollama run mixtral:8x22b". It's a base model only capable of text completion, not any other tasks, which is why you got a terrible result when you gave it instructions.
Mixtral-8x22B-Instruct-v0.1 is an instruction-following model based on Mixtral-8x22B-v0.1. It was released two hours ago and it's what this post is about.
(The last updated 44 minutes ago refers to the entire "mixtral" collection.)
Considering "mixtral:8x22b" on ollama was last updated yesterday, and Mixtral-8x22B-Instruct-v0.1 (the topic of this post) was released about 2 hours ago, they are not the same model.
Mistral 7B Instruct v0.2 and Mistral 7B v0.2 are different models. Judging by the title, I suspect OP meant to post about the latter, which was released a few days ago, but accidentally linked to the former instead.
A rocket on a typical orbital launch profile spends less than 60 seconds in air dense enough for jet engines to have good performance, so there is little to gain.
Pegasus is an orbital rocket launched from an aircraft, but it doesn't exactly impress with performance or cost-effectiveness. Just doesn't make much sense to operate a huge aircraft and design your system around it just to improve on the least important 10% of the flight.
An air-breathing jet engine doesn't need to carry oxidizer, which in a rocket is most of the propellant weight. It also has access to unlimited reaction mass, so it can be much more energy-efficient in producing thrust (it is more efficient to produce thrust by accelerating a lot of mass by a little, than by accelerating a little mass by a lot, but a rocket can't take advantage of this because it would need to carry all that extra mass. A plane can use ambient air for this purpose)
This all adds up to a plane needing to carry many times less mass to gain the same altitude and speed as a rocket, at least within relatively dense atmosphere.
You can partially offload with some backends (e.g. llama.cpp and derivatives) but speed gains from that don't come in until it's mostly offloaded. I have 8GB VRAM and it's not enough to get any boost on mixtral in Q8. 16GB might do better or it might not.
The speed is quite good even on CPU only though, I get 3.5 tokens per second with 6 cores and DDR5-6000. For comparison llama2-70B is less than 1 t/s on the same hardware in Q4. And, subjectively, Mixtral performs better.
A reactor that has never been turned on isn't a significant radiation hazard. It's the fission products that are hazardous, not the fuel, if it's never gone critical there are no fission products yet.
The bazarek is fun but in reality even less relevant that this post makes it out to be. Since people with real information cannot prove it and it takes zero effort to post fakes the bazarekposts are not any more meaningful than random guesses. Arguing about them is just a pastime for people waiting for polls to close.
Preparations for pad repairs and upgrades were well underway before the first flight - the question was not whether they'd be necessary, but how much and how soon. In particular if I remember correctly manufacturing of the steel plating that now forms the pad started all the way back in January.
Anything with 64GB of memory will run a quantized 70B model. What else you need depends on what is acceptable speed for you. With a decent CPU but without any GPU assistance, expect output on the order of 1 token per second, and excruciatingly slow prompt ingestion. Any decent Nvidia GPU will dramatically speed up ingestion, but for fast generation, you need 48GB VRAM to fit the entire model. That means 2x RTX 3090 or better. That should generate faster than you can read.
Edit: the above is about PC. Macs are much faster at CPU generation, but not nearly as fast as big GPUs, and their ingestion is still slow.
The simulation is just so fake, almost everything that goes on is just decorative.
There is a budget, but after the first 30 minutes you'll always be running an enormous surplus without trying.
Citizen commute to work, but if they can't get there, the workplaces will continue to work, with some trivial penalty to efficiency.
There is traffic simulation, but if a jam forms, vehicles will start vanishing to unblock the road.
You can build public transport, but it doesn't matter if it's efficient, because the city's entire population can be standing at bus stops waiting forever with seemingly no ill effect.
Citizens will use parking spots if they're available, but if they aren't, they'll just disappear their car and reappear it later.
Zoned buildings get built and upgraded autonomously, but what gets built doesn't depend on economic factors, just on how many upgrade points are accrued from nearby services and attractions.
There is a large number of special buildings of various types that can be unlocked and built, but they all count as a tourist attraction and don't perform their actual function, they're effectively statues.
Cities Skylines is a bizarre un-game that has all the UX and presentation of a city simulator without any actual simulation.
Yes, many, huggingface is full of chat-tuned LLaMA derivatives that are supposed to replicate its performance, and tools like text-generation-webui or kobold.cpp can be used to run them with chat-style UX.
But for most tasks none of them come within a mile of GPT-3.5, or within a parsec of GPT-4.
ggml is a library that provides operations for running machine learning models
llama.cpp is a project that uses ggml to run LLaMA, a large language model (like GPT) by Meta
whisper.cpp is a project that uses ggml to run Whisper, a speech recognition model by OpenAI
ggml's distinguishing feature is efficient operation on CPU. Traditionally, this sort of work is done on GPU, but GPUs with large amounts of memory are specialized and extremely expensive hardware. ggml achieves acceptable speed on commodity hardware.
It doesn't matter very much that the "official" instruct tune is censored as anyone can create their own and there will probably be many freely available ones as happened with LLaMA.
You have to turn down temperature and/or p when you want accuracy. Otherwise you don't know if the model's read is bad, or if you just happened to get a low-probability outlier.
With the playground's defaults of t=0.8, p=0.9, I got the right answer 7/10 times.
With t=0.1, p=0.1, it's always exactly "The name of the cat that wears a blue tie is Jackson."
You can just barely fit a 33B GPTQ model in 24GB VRAM. It will be in 4-bit mode, and without maximum context size, but it will be quite fast. Or you can run from RAM+VRAM in GGML format with llama.cpp (or a derivative), which will easily fit 65B models even at 5 or 8 bits, but at much lower speed.
Quantization isn't (and wasn't) expensive, it's mostly just data shuffling. A good PC will do a 7B model in half a minute, up to a few minutes for a larger model. Quantized models being made available for download is more for the benefit of less technical users who may not be comfortable with the command-line tools, or for people with slow or metered connections who'd much rather download 15GB of data than download 60 only to squish it into 15.
Almost every UI for LLMs I've seen has a way to specify an initial prompt that never goes out of context, it's strange that it's not a feature in ChatGPT.
Mixtral-8x22B-v0.1 was released a couple days ago. The "mixtral:8x22b" tag on ollama currently refers to it, so it's what you got when you did "ollama run mixtral:8x22b". It's a base model only capable of text completion, not any other tasks, which is why you got a terrible result when you gave it instructions.
Mixtral-8x22B-Instruct-v0.1 is an instruction-following model based on Mixtral-8x22B-v0.1. It was released two hours ago and it's what this post is about.
(The last updated 44 minutes ago refers to the entire "mixtral" collection.)