Thanks

Looks like this GPU shortage is the real deal. We would be very close to multi-modal by now Sam Altman said. 1 million tokens!! insane. Seeing some research on multi-prompting to mimick reasoning too.

Apparently, this article got taken down at OpenAI request. I'm seeing a lot of complaints on Twitter about quality and speed. I think they are obviously having serious trouble handling the growth of users and growing user load on their systems.

https://web.archive.org/web/20230531...g/openai-plans

There's a Chat AI like ChatGPT for Android. Perplexity. I downloaded it from the Google Play store. Looks like it does the same stuff. Not good at one search i did. Did good on other stuff.

I asked ChatGPT to compare its knowledge model as of its September 2021 cutoff to the research the Web Requests plugin enables it to do — and the results did not disappoint ��

https://twitter.com/JD_2020/status/1...CKBd_0YHw&s=19

This channel seems pretty good for getting an overview of new developments. It's mostly about LLMs and primarily ChatGPT:

https://www.youtube.com/@ai-explained-

I've been playing around with this, it's pretty impressive:

Caught some of this deep conversation on possible future AI systems with Lex Friedman and Matthew Mcconauhey. Starts at 1:40:00 i like the positive scenerios. Much like I mentioned with AI being a perfect friend and assistant.
https://youtu.be/8qgIFD0PaWo

Always more stuff coming out. Now every couple days it seems. Gonna have to to checkout the new Google search beta. Think this is it. Looks like they are gonna fill it full of ads already. I hope they lose the AI search wars to Bing or ChatGPT.

https://blog.google/products/search/...announcement-/

The pace of AI

The last 3 days:

New AMD AI Chip
Salesforce AI Cloud
ChatGPT Workspaces
UK Gov't: Data Access
AI "Last Beatles Record"
92% Programmers Use AI
Adobe Generative Recolor
Google AI Search 2x Faster

Here's what you need to know:

https://twitter.com/aakashg0/status/...lfPH2_TUA&s=19

It's getting to where i can't tell fake from real now

https://twitter.com/_akhaliq/status/...hFkbssgvw&s=19

Seems to be getting better at math. The magic seems to be in the correct prompting.

This was an impressive paper with some big implications

GPT-4 was given 4,550 novel questions representing the entire “MIT Mathematics and EECS undergraduate curriculum, including problem sets, midterms, and final exams”

With good prompts, it scored 100%

https://twitter.com/emollick/status/...W26UJxHZA&s=19

Off topic, but speaking of google, I clicked on this link last night and it frightened me.

Levis and google collobarated to put a camera in a slieve. Its five years old, though.

https://atap.google.com/jacquard/products/levi-trucker/

Ray-Ban put a camera in one of their wayfarer lenses. There was instant kill it with fire responses so probably won't take off either.

Back on topic, our privacy days are numbered.

I'm conscious of posting private info sparingly anywhere online.

I envisioned (before chatgpt: which caught me off guard) that there'd exist AIs that will be able to piece everything I ever posted online, together, find a pic of my face, house address and output it one of the days. Which is now very evident.

Am I paranoid. do I have stuff to hide. Maybe....

On the flip side, as you say, it will be hard to tell what's real and what isn't which may act as some defense, maybe.

An idea I've been playing around with:

Me: Go back even further than existing mental frameworks and think more like their operating system. So maybe start with the 3 laws of thought that people come intuitively equipped with and start build from there

Me: Right but the important thing is we don't need to learn them; we know them intuitively.

Bard is multi-modal now. This is crazy.
https://twitter.com/alexandr_wang/st...22PJKvYfA&s=19

You can upoad a pic and it tells you all about it. Not sure if ChatGPT is yet. Bard might have just gotten ahead in some ways. I tested it with uploading a pic myself and the response is amazing. It seems the pace of advancements are still progressing along at a fast rate.

Dude i follow is already pointing out some LLMs are good at training themselves
https://twitter.com/abacaj/status/16...9NGdcUVtw&s=19

Https://Heypi.com is faster at response time than the others right now.

I'm really surprised because Sam Altman was saying their multi-modal was far off bc of lack of chip availability. It was supposed to be next year or way later until multi-modal. I dont have GPT-4 i just use Bing chat in creative so i dont know if GPT-4 is multi-modal yet. Once the LLM, neural net can see the world like humans AGI is def getting closer.

Does this mean we will have to complete more annoying methods of proving we are humans than clicking fire hydrants soon?

I'm seeing anecdotal stuff of Bard converting mobile screenshots to code.

I don't know anything about IT security. I know some of these devs in Twitter are predicting it to get "weird" for awhile. Mostly because of voice impersonator software. I mean no system is ever perfectly secure even pre 2010 or whenever. I think people are just hoping something changes. Or auto-GPT or bots are still not good enough for mass problems. Auto-gpt is pretty wack. I dont know enough about systems to have a good idea on whats gonna happen. Im not really worried.

Protonmail has a pretty cool CAPTCHA system. It's a picture with a little jigsaw puzzle piece cut out and you have to drag and drop the puzzle piece in the right spot. I think monkeys could solve it though. But monkeys aren't computers and we are kind of smart monkeys, so ...

I think ChatGPT is kind of a moron. I asked it questions about Escape To Witch Mountain and it said a line from that movie was used in School of Rock and I got into a big argument with it. The relevant quote is "Tia, we need my harmonica." I asked it what movie that was from and it said School of Rock, and I don't think that's correct. I then had a long argument about Tony and his harmonica and his use of it to focus his telekinetic powers and ultimately ChatGPT just ended up agreeing with and mirroring me, and I think that's ultimately what it does, cause I've had arguments with it about other movies too. It just eventually starts agreeing with you and mirroring you. It's pretty much a perfect life partner. They need to find a way to have it control and interact with a sex toy.

Oh yeah it was totally wrong on major plot points of Silver Bullet. It said Jane killed the werewolf by shooting it in the eye with a bottle rocket. Marty shot the damn thing with the bottle rocket but that didn't kill it, just blinded it, then he shot it with the pistol and killed it with the silver bullet. The AI was wrong on every major plot point from that movie.

WSJ story:

WASHINGTON—The Federal Trade Commission is investigating whether OpenAI’s ChatGPT has harmed people by publishing false information about them, posing a potential legal threat to the popular app that can generate eerily humanlike content using artificial intelligence.

In a civil subpoena to the company made public Thursday, the FTC says its investigation of ChatGPT focuses on whether OpenAI has “engaged in unfair or deceptive practices relating to risks of harm to consumers, including reputational harm.”

https://www.wsj.com/articles/chatgpt...y-ftc-21e4b3ef


PairTheBoard

If anybody has openai ChatGPT-4 payed version I'm curious about feedback on this web request plugin. Maybe it's on GPT3 too. I use other equal LLM's for web search like PerplexityAI, Bing chatbot. But this looks super cool if it runs AutoGPT by default. Josh developed the web request plugin. It's his.

https://twitter.com/JD_2020/status/1...mZPLTbnRQ&s=19

More and more models coming out. First Llama then this one. It's on huggingface. I can't find a link. Not sure if it's an app yet. Hearing it is better than GPT-4 and Llama on many tests. It's smaller too.

Alibaba bets on AI democratization �� The Chinese e-commerce giant expands accessibility in AI development by opening up its large language model, Tongyi Qianwen, with a seven-billion-parameter model, Qwen-7B, for global use.

https://www.cnbc.com/2023/08/03/alib...e-to-meta.html

So cool. It's open-source too!

My hunch is all these breakthroughs or near breakthroughs like with the room temp superconducter, oral cancer killer and the PCNA isomer cancer killer are from proprietary LLM's that are well trained.

Heard there are a lot of studies on databases like pubmed on drugs that never got funded because of the man-hours needed to search 1,000's or 10,000s of scientific papers. Now you can scrape a database and fine-tune a search with LLMs. Research seems to be speeding up due to the algo's and LLM doing the "thinking". Best example is the 5,000 yr old cuneiform translations

https://www.thedailybeast.com/this-a...s-into-english

I have a lot of time spent with the nuances of chatgpt (thru langchain also).

Its evolving fast.

Its interesting at this time it is VERY good at refactoring code for me, I understanding coding concepts well enough but I'll never have enough experience with the language to be strong and efficient with syntax. But the nature of the 'ai chat program' and 'stackoverflow' the reddit of programming I think make it the number one skill it has.

I thought it might be interesting to mention because also in contrast, especially up until the last upgrade, and still sometimes now, it is terrible with recursive functions (ie functions that call themselves from inside themselves). Its just a little bit of a rabbit hole for it atm, kind of a funny turing trip we will no doubt surpass.

An example that is relevant to me. Let's say you are coding a project and some aspect causes you to feel that most of the graphical aspect of the user interface should be controlled via css styling sheets as opposed to javascript. Just a trend in the design evolving. Later you suspect the opposite and wish you evolved the project and code the other way.

You can quite easily and feasibly feed the project into chatgtp and ask it to refactor the design to rely on javascript based styling vs stylesheets.

This would be more like a mathematical translation than an Advanced AI capability for the technology. So it comes out quite accurate and using commonly used/tested design templates (because thats what it knows best!).

So the design and building process has an aspect that was impossible before, infeasible. Especially for the hobbyist.

This lets me also explore design principles which I couldn't otherwise hope to get to as a hobbyist and where I suspect it would take more years than I have in the field to start to really explore.