yayr

each billion parameters using 16 bit floats requires around 2 GB of GPU or TPU RAM. ChatGPT is expected to have around 1000 billion. Good open source LLMs have around 7-20 billion currently. Consumer GPUs currently max out at 24 GB. You can now quantize the model to e.g. 4 bits instead of 32 per parameter and do other compressions, but still there is quite a limit what you can do with 24 GB of RAM. The Apple unified memory approach may be a path forward to increase that... so one box gives you access to the small models, for a GPT4 like model you'd need (for inference and if you had the model and tools) probably 100 of those 4090s or 25 of H100 with 96 GBs I guess to fit in 2 TB of model data.

that's interesting. Could that be a use case for microwave drilling?

e.g. with https://www.quaise.energy

Can someone explain or point towards an explanation, what the performance impact of such a query is? How does it scale with number of edges or depth of recursion? Is there a special smart way of indexing this to be efficient?

to contribute to the architecture of solutions is hard indeed... However, to contribute small everyone in most countries can, simply by choosing what and how to consume and by voting, wherever that is possible

One assumption could be that there are certain products/services that have age related regulations and Paypal needs to comply. Maybe, if your product or supplier wasn't on that list but you still got asked for a birthdate there was a misconfiguration in that regulation rule set...

The OpenAI business model seems to become an ecosystem play. Plugins, apis etc. They are new to it, but have a super strong partner with MicroSoft and strong startup experience for it in the leadership team. We will see if Google is able to play that well and not only release research and basic features.

nice, maybe you can make the initializing prompts in

https://github.com/enricoros/nextjs-chatgpt-app/blob/main/pa...

transparent to the user and even changeable by them.

Here are some of my favorite features:

- structure the content around conversations, so a thread can be easily put to rest picked up later while keeping context, while a different thread can be picked up in the meantime

- tag conversations, e.g. like google mails

- involve multiple AI engines in one conversation, in case the chatbot has different engines available

- be device agnostic, so I can switch a conversation e.g. from laptop to phone and back

- involve other humans in a conversation

it is a result of trying to align the AI model after it has been trained using self-supervised learning with an addition human feedback step. OpenAI has a model / process called InstructLM or InstructGPT

see e.g.

https://arxiv.org/pdf/2212.10496.pdf

https://arxiv.org/pdf/2203.02155.pdf

also an approach to solve the alignment problem.

https://en.wikipedia.org/wiki/AI_alignment

Google has had a huge history of poor product execution. So long it was not strongly apparent in ads and also not so much in search but many other areas. Now they again show this poor product execution and lack of customer focus in AI. The reason that they struggle is not that they don't have smart brains in their org, but that they are not able to translate that power into customer value and imagine and iterate effectively towards user space. Let's see if they can get the curve...

good point, probably a final version should have a pipeline like 1. cluster song segments into styles and 2. search for each cluster or only the main cluster.

What would be a good NN architecture for the first step?

it is not the typical "song radio" approach that would provide more a popularity / artist similarity based playlist but a real "similar song" google like search. I think that is valuable to discover new music, especially to find unknown artists. But if someone were to embed it into streaming clients that would need to be a different feature too, probably more search related.

is there any open source reference implementation for such a vm that is dynamically generated based on JS/TS?

well, the ETH project has many sides, one is technology, another is coin distribution (especially with a proof of stake model). That does not imply a scam, but also is a problem for long-term stability

see e.g. here https://cointelegraph.com/news/64-of-staked-eth-controlled-b....

or https://www.bitcoin.com/get-started/how-was-eth-initially-di...

well, the binary system is underrated as a general purpose tool. you could start by teaching kids that with five fingers you cannot only count to five but e.g. to 31...

It's great! I just could not figure out how to zoom the chart / coordinate system with the mouse / touchpad. pinching and modifier keys did not work.

great insight, but one point I miss is that to play games with cash-flow you got to have it in the first place.

Some ideas really benefit from investment before a product is marketable (i.e. for research, production). Kickstarter etc. provide a great tool to generate that for consumer oriented products before the product is launched. There are many markets where this is not so simple or not possible.

I would not be surprised if the counting one day will be reset to the invention of the computer. That would make us now 87 AC (anno computeri) assuming we count the invention from 1936.

This is an absurd situation. I wonder how many "promises" made by Elon have been broken regarding how he wants to manage Twitter and make it a better place.

E.g. what about transparency here? Where is the explanation why this account was suspended?

I am not sure, that with Elon in the hot seat the trust of the public into Twitter may be restored anytime soon... very sad.

would it be an approach to use separate color channels for the freq amplitude and freq phase in the same picture? Maybe the network then has a better way of learning the relationships and there would be no need for the postprocessing to generate a phase.

Can it run PyTorch, TF etc?

universe is a relative word, and typically treats us at being at the center (which in the whole of reality is somewhat unlikely). So it only describes that part of reality observable by us. 8,5 bn ly refers to roughly half distance to all of the observable events in that observable reality at the time these events occured. Since 3D space presumably expands, it would now take much more time for any event in our observable "universe" to reach from one end of the observation limit to the other...

by the way - if you want to go down further that rabbit hole a good place to start is to search for "levels of multiverse"

even though currently just one state is implementing that in Germany, that decision shines out. France has recently imposed a similar ban.

I don't understand, why this service cannot be provided from an EU data center. Every hyperscaler has these, MS also.

> Physical and behavioural changes have also been found in people: testosterone and dopamine production is increased and more risks are taken.

...

> Up to one-third of humans might be chronically infected.

That is certainly an interesting research vector...

I thought that was already obvious when Google started working on cars ;-)

those guys are also working on it atm :-) https://github.com/lstein/stable-diffusion/pull/179

looks like there is an easier path using metal shaders: https://dev.to/craigmorten/setting-up-stable-diffusion-for-m...

and https://github.com/magnusviri/stable-diffusion/tree/apple-si...

then - how far away are we from having it on M1/M2 Macs, at least with regular processing? openvino may be one path I suppose: https://github.com/openvinotoolkit/openvino/issues/11554

I wonder how usual it is for bigger tech companies to give a majority of their staff wide productive access?

> For context, to provide the electricity demand for kansas would require about 35% of the total geothermal flux into kansas. New Jersey's electricity demand is 7 times higher than what its geothermal heat flux could provide. For the whole US, electricity consumption is about 75% of geothermal heat flux if you ignore variations like the yellowstone hotspot.

It would be great if you can provide a bit more details on those assumptions. Thanks