imranq

I mean if someone likes the subject, then they should invent a project that forces them to build something or write something up. Otherwise its easy to "fake" learning by doing the motions on a bunch of tutorials and quizzes.

I would not start with any course unless you had a project in mind that would take the knowledge from the course to get started. Otherwise you risk wasting a lot of time for knowledge that won't help you in any way and will get outdated in a few months anyway

I had a student who was basically performing at a 6th grade math level despite being in 11th grade. I told that student to go through Khan Academy starting from Pre-K all the way to 12th grade and don't continue unless they achieved mastery. That student went from the C math student, to ultimately scoring a perfect score on the SAT II Math. So I highly recommend everyone start for the absolute basics, even if it feels like a blow to the ego

Wouldnt this be possible with a solution like Guidance where you have a pre structured JSON format ready to go and all you need is text: https://github.com/microsoft/guidance

This is pretty cool, do you find the server farm of older computers valuable for your own work?

There must be. Since there is a finite amount of real world reward signal captured by these models, there can only be a finite amount of grounded knowledge that can come out. Usually without reasoning, the output knowledge "yield" will be far lower than input.

The interesting piece is if we can get LLMs to reason well, that finite reward capture can take LLMs much farther than possibly humans could do with the same signal.

This is an impressive demonstration of GPT-4's logic and reasoning capabilities, but the DM result uses a real-world reward signal (e.g. the list is sorted correctly) to validate its results whereas we can never be sure that GPT-4 outputs are hallucinated or not. Also this experiment would never be done if not for the DM paper in the first place.

Brilliant idea - I think I'll try this in my neighborhood and at work where people seem to be walking alone all the time

Maybe this is the answer...when humans are no longer required for a field to move forward, prestige becomes meaningless and everything becomes a means toward self-fulfillment not competition

Nice! There is probably a whole space of applications that help users effectively prompt engineer a workflow with a UI

There are probably ways to embed cryptographic hashes within images. Any device that creates images from the real world could have secret keys that can be used to validate any image created by said device.

We will still need a centralized party that holds the secret keys for validation through

I can definitely see this being useful for work and school, but not sure if I would use a foldable for everyday tasks.

Casing these phones seems like a challenge

I suspect that there's a sweet spot that combines a collection of several "neurons" and a human-readable explanation given a certain kind of prompt. However, this "three-body problem" will probably need some serious analytical capability to understand at scale

A notebook and pen seems to do the trick. In the beginning of the day, write an entry of what I'll aim to accomplish and when. At the end of the day, judge myself on my productivity and write a plan for the next day. That alone has been a force multiplier of productivity for me

Also remembering to have fun and enjoy life always (surprisingly!) seems to help

Eventually I hope the compute required to run these systems become local, but yes I do find Bard's "personality" to be a little more controlled and cleaner than chatGPTs. However, chatGPT 3.5 usually gives better responses (e.g. specific book recommendations I ask it).

We have a bimodal distribution. Elite universities whose prestige is already established have an "embarrassment of riches" in terms of applicant pool, whereas middling and low ranked universities without a moat are facing enrollment cliffs.

Applicants are trying to go big or go home (and learn online instead)

I remember Musk even posting a picture of him and Brin partying after this story came out.

Seems like another WSJ hit piece to discredit Musk for no good reason (other than to appease short sellers)

This is on GPT-J which has 6B parameters. I wonder if this scales well to much larger models like Llama 65B or GPT3

I thought they were on contract? hopefully they don't go on strike again at the same time the power went out like last time

I wonder what will happen to isomorphic labs which Demis is also leading

There are plenty of LLM benchmarks that are used to test performance, some of them are: * Winogrande

* BoolQ

* PIQA

* SIQA

* HellaSwag

etc...

I like a lot of the LLM use cases mentioned here. A couple more are:

- conducting literature reviews (stay sane while researching LLMs!)

- Talking to textbooks / AI teaching assistants

- language learning with a companion tailored to your level and interested

LLMs are so hyped and written about these days that it would be hilarious if the next version of GPT trained on todays internet would be biased towards praising itself

It's interesting how all it takes is the right bag of numbers (the weights of gpt) and the entire economy can be upended

"Make it work, then make it right, then make it fast" is brilliant. This should be a universal principle for almost everything

For all the hype, the benchmarks they report don't seem super compelling. Financial questions need to have 100% accurate responses (not 60% or so they report) otherwise they are worse than useless and can cost billions of dollars in losses

I suspect they must be selecting different models based on the question asked. I've noticed this myself asking simple questions vs difficult questions, chatGPT is much slower on the difficult questions even though each token has the same computation cost if given to the same model

Check out the web demo for the 14B parameter model: https://huggingface.co/spaces/BlinkDL/ChatRWKV-gradio

I tried a few queries. It works okay for basic prompts for coding questions but can give pretty good results with more context / prompt engineering

The stats shown are pretty compelling that this transformer + RNN approach works. Looking back at the Llama paper (https://arxiv.org/pdf/2302.13971.pdf), many of the results are comparable between llama and RWKV. E.g. Llama 13bn scores 80.1 on PIQA and RWKV scores 77.5

I wonder if RLHF would boost performance for this architecture in the same way

How does a 3rd party know the traffic figures for Google / Bing? I see they are using a service called Similarweb based in Tel Aviv, but its still a mystery to me how you can get this information without access to internal analytics tools

Maybe they define social good as the collapse of democracy, and the commodification of attention?