causalmodels

Reinforcement learning from human feedback [1]

[1] https://en.wikipedia.org/wiki/Reinforcement_learning_from_hu...

Fiascos like this display neither experience nor intelligence. This whole saga was a colossal failure on the part of the previous board.

Zero chance OpenAI ends up seeing any more compute if they merge.

I don't agree with sex work, but I will absolutely defend the rights of people to get paid for their labor.

A gig worker or freelancer who sets up an LLC is essentially no different than an individual. They should have protections.

Isn't the EU currently considering making open source ai developers legally liable for their models?

They’re not posts, they’re native ads.

This is quite harsh for a repo explicitly created to experiment with non-Unix OS ideas. It is an experiment, it doesn't have to e useful.

Personally, I enjoy when people post things like this as it helps me learn.

This is a decent place to get started https://arxiv.org/pdf/math/0009118.pdf

You can extend some of the thinking in this paper to cover action spaces for deep RL models to really have some fun!

I’ve used algebraic topology for certain robotics tasks when statistical saftey wasn’t good enough and when trying to generalize techniques to learn on graphs.

The results are spectacular, but the application set is indeed very limited.

My undergrad research was on CW complexes and I did spend the following decade trying find applications for it in ML. It has worked out approximately twice, but my god is it powerful when it naturally fits a problem space.

Hoverboards.

> What I hear is that training AI on AI outputs is fruitless. To me, thats as good as test as any: when you can use at scale an AI product to inform the training of an unrelated AI system I will be impressed.

The phi-1 team did this a few weeks ago.

When did you graduate? None of my humanities courses required handwriting and that was ten years ago. Plus all non-intro humanities courses were either take home exams over a few days or final papers.

My parents had the same idea but instead of smart phone use, it was religion. It wasn't so successful for them.

Hard disagree. Phi-1 seems to be a harbinger of what is to come. However i think there is a plausible argument to be made that their training process approximates a kind of distillation

It's not an either or. We're going to leverage the web trained LLMs to bootstrap the specialist models via combination of training token quality classifiers and synthetic data generation. Phi-1 is a pretty good example of this.

> Rule of thumb is that you need ~20 tokens per parameter.

That rule of thumb is wrong. The chinchilla paper has it anywhere between 1 and 100 tokens per parameter.

His solution is a global regulatory regime to ban new large training runs. The tools required to accomplish this are, IMO, out of the question but I will give Yud credit for being honest about them while others who share his viewpoint try to hide the ball.

Funny to see someone call for wide scale cooperation to stop training LLMs but can't seem to get people to cooperate on the embargo.

Prescreening? yes. Ad targeting that looks an awful lot like prescreening? No.

IMO the ethnic quotas would be the most unsavory aspects to an American audience.

I'm going to assume you know how to stand up and manage a distributed training cluster as a simplifying assumption. Note this is an aggressive assumption.

You would need to replicate the preprocessing steps. Replicating these steps is going to be tricky as they are not described in detail.Then you would need to implement the model using xformers [1]. Using xformers is going to save you a lot of compute spend. You will need to manually implement the backwards pass to reduce recomputation of expensive activations.

The model was trained using 2048 A100 GPUs with 80GBs of VRAM. A single 8 A100 GPU machine from Lambda Cloud costs $12.00/hr [2]. The team from meta used 256 such machines giving you a per day cost of $73,728. It takes 21 days to train this model. The upfront lower bound cost estimate of doing this is [(12.00 * 24) * 21 * 256) = ] $1,548,288 dollars assuming everything goes smoothly and your model doesn't bite it during training. You may be able to negotiate bulk pricing for these types of workloads.

That dollar value is just for the compute resources alone. Given the compute costs required you will probably also want a team composed of ML Ops engineers to monitor the training cluster and research scientists to help you with the preprocessing and model pipelines.

[1] https://github.com/facebookresearch/xformers [2] https://lambdalabs.com/service/gpu-cloud

edit: these costs are for the 65B parameter model.

Last week I went to an urgent care clinic to get treated for strep throat. Many of the steps did not occur as you describe. Here is what happened.

1. I signed in using an automated terminal. A tablet which scanned my ID and insurance card. I sit down in the waiting room. A few minutes later a nurse comes to get me.

2. The nurse takes my temperature and swabs my throat for a rapid test which is then done in front of me. The nurse leaves the room while the test is reacting. The interaction takes about five minutes.

3. A NP came in to look at the test and my throat. The only interaction I had with the NP was when they asked me if I thought I had strep to which I replied "yes". The NP diagnosed me with strep and wrote me a script for amoxicillin in less than five minutes.

4. I left the urgent care clinic without seeing the clerk at checkout. The entire visit including waiting to see a medical professional took less than 30 minutes.

5. Ten minutes later I pick up my course of antibiotics at a pharmacy four blocks away. The price of the medicine was free with insurance and $15 dollars without.

Strep tests cost $35 dollars retail. The price to visit the clinic was $200 dollars ($165 sans the price of test). The per hour cost of medical care was $990 dollars.

OpenAI has a capped-profit model so I really don't think it will ever go public.

Well OpenAI's explicit goal is AGI. While this is farfetched, it would clearly make them the most valuable company in the history of the world and the ability to monetize something like AGI would essentially be unlimited.

OpenAi isn't really interested in making money from selling products. So instead they have worked out a deal with MS where they give them the best new tools they can come up with in exchange for tons of cash and compute. They are basically a research lab of which MS owns a significant, but not controlling, share. ChatGPT itself is probably not going to make any money. It is a very cool tech demo that functions as an ad for MS products. The money making part comes in when MS starts to bundle OpenAI models into MS products which then get sold to businesses. And if there is one thing MS is very good at, it is selling software to businesses.

Something that I think is often overlooked when talking about this is a large chunk of tech labor are millennials. These people are now in their thirties or early 40s. Many now have families and want different things from work than they did in their 20s. This is totally natural but it does create new tensions.

Deepmind has been doing some interesting work around Retrieval Enhanced Transformers (RETRO) models [1] that might be relevant in this context.

[1] https://www.deepmind.com/blog/improving-language-models-by-r...

a) No idea. I read slowly, cannot spell, get letters confused, and frequently don't realize small words are missing from sentences. It is in very large part why I went into math.

b) Not if everything I write is going to now be run through these detection algorithms and flagged.

I guess that's fair. I just personally don't think the additional gain is worth taking away your child's privacy.