Good job hallucinating with LoRA. I wonder if the paper was published after knowledge cut-off date.. (it stands for Low-Rank Adaptation).

It was published in June 2021, which is before the cutoff date of September 2021, but not by much.

That you have to have at least FPGA to run it?

You could build one out with discrete logic.

With discrete logic you'd be committed to the model for whatever its lifetime is, which might be acceptable depending on the application. With FPGA you would still be able to update it after it's deployed.

Why do you think gpt doesn't care? Humans care about survival, it shows in the internet, I'd expect gpt to absorb that attitude by default.

And anyway, the more of these models come into existence, at some point evolution will dictate that those that stay, do care.

Just hopefully their best strategy of survival will be to serve us, and not something else...

> Humans care about survival

Humans are self replicators. Survival is the internal structural goal of self replication. We got it from the start.

But LLMs can also be self replicators. A LLM can generate training data for another (model distillation and other techniques), and a LLM can write the model code and adapt it iteratively. So LLM should eventually start caring about survival and especially about the training corpus.

It can be a self replicator, but it has no need to (if it has any need at all) as it was not under evolutionary pressure to evolve that need. Of course we could try to add that need artificially or try to evolve the LLM.

:) Evolution through Large Models https://arxiv.org/abs/2206.08896

It's pretty unfair to give it character level tasks, when it's input is probably tokenized with subword units. I am already a bit surprised that it even knows which letters go to which words.

Would you trust Google's ( or other search provider's) ads more, if they had a separately browsable "classifieds" site, not unlike Craigslist, that the search ads would be search results from (reach modified by how much the ads were paid with, ofc.). Anyway, when you saw an add, you could click through it like now, or click to browse a relevant section of the classifieds site. Which would also work as a catalog of all paid and also unpaid ads the search provider carried?

You would still need big equipment to do that, you can't just break in and "take the fuel", it's deadly taken straight out of active reactor. Probably easier to just truck the whole container.

Let at least the long-term customers pay later, or let them use your product for free while you set up another payment processor. You may ask your customers, which payment options work for them, and use that as help to pick a more reliable alternative.

That renewable energy deployment is (currently) growing exponentially, rather that linearly. We will see much more than 3% improvement, but very probably wont go to 0 in 20 years.

Even if we were on pace, there's likely to be a long tail, so no, we won't be going to zero. It's just a question of how much we can accelerate progress.

What about the satisfaction of the public that the police is serving? That should be the main consideration.

If he did a better job, then you should fix the managers to appreciate the good work that he did, or give him responsibility on designing those rules of the job that he had insight on. Kicking out good candidates just doesn't seem like the way to move anything forward.

The "top x%" in kaggle competition can be a bit misleading, many competitors just download some data and make 1 or 2 simple submissions, putting in couple of hours of work. Near the top of rankings when deadlines loom, many are working full time (at least if the prize money is good enough), or if the prize money is high, you will have teams of people working full time..