They threw the kitchen sink at it to see what would happen and it learnt to talk.

All I have on this subject are reckons, but I reckon the rise of LLMs and AI image/video generators is just a trend. Boring stuff really. Yes, they’re building tools that will stay with us in some form or another, but they’re essentially still just dumb tools. There’s potential for them to be massively beneficial, but not epoch-making in their impact. There’s also potential for them to lawyered out of useful existence. So it could go either way. Best case scenario: better tools for us all. Yay.

The ‘deep learning’ stuff, on the other hand, is interesting. That’s the stuff that could accelerate human invention exponentially. Or kill us all. Again, could go either way.

Let me try again. I can make that point clearer.

What’s interesting is when AI can do something we humans can’t. What’s boring is when it just does something we can already do, but in a different way.

So finding new proteins and extrapolating their uses for us: interesting.
Driving a car: boring.
Finding a way to do fusion at room temperature: interesting.
Drawing a picture: boring.

@poprock It's still the Wild West at the moment, so the lawyering thing will go away, once the big AI companies prove they have the data sets verified and owned. This is already happening. Amazon already offer unlimited IP indemnity for anyone using their Titan image gen.

The LLMs literally deep learning models. You might be thinking of chatbots and image gen etc. but certainly all the modern proprietary LLMs are using deep learning, along with machine learning etc. in their make up. Also, the newer models are also introducing multi-modality which will further improve the outputs, as context from text can be support with images, video and sound, pretty much as we understand stuff.

Yes, LLMs are, or were, relatively "dumb" but the growth is exponential and with that growth comes greater intelligence, and results. If we look at ChatGPT-3 it has about 175 billion parameters and failed at every attempt to pass the bar exam, ChatGPT-4 has 1.7 trillion parameters and passed the bar at around 90%.

Beyond that, the real advantages will come with improved speed of processing (check out the Groq chip) and the possibility of System-2 thinking (self-reflective, rational, complex) being added to the current System-1 thinking (instinctive, quick) that will allow for real AI-assistants. You will be able to actually set an AI a complete task (go and make me a movie about the death of Timdog). It might take time, but you won't have to keep feeding it with new prompts to modify it's output.

monkey wrote:

My optimistic hope is that the web is going to end up so devalued by fake untrustworthy content that a new web is born. Higher quality content, less driven by algorithm and SEO, probably less open to amateurs. Verified humans behind everything.

How would this work?

GurtTractor wrote:

monkey wrote:

My optimistic hope is that the web is going to end up so devalued by fake untrustworthy content that a new web is born. Higher quality content, less driven by algorithm and SEO, probably less open to amateurs. Verified humans behind everything.

How would this work?

Dunno. Some bio-passport, government-backed scheme. Or the authentication goes to the platform holders e.g. the BBC verify that everything they put out there is genuine. So the new web is spread across mega-sites, way less access for regular plebs. Social networks would be responsible for IDing people on the way in.

There's some sort of digital signage of files that I saw some bloke from MS talking about recently but can't remember anything about it.

....Blockchain?

monkey wrote:

Dunno. Some bio-passport, government-backed scheme.

I'm sure that would never be abused and send us further into a dystopian nightmare.

What if we don't need the web? 

*checks the future news*

"Nobody died today and all societal decisions were made by bots. Bots made a significant scientific breakthrough today that no human will ever understand".

We'll still want to communicate with each other so maybe we can all have dumb phones again.

Yep, very good for productivity stuff, just double check any data it comes out with. Unless it's citing a specific source it could be wrong.

This is worth a read if you're new to prompting too: https://platform.openai.com/docs/guides/prompt-engineering/six-strategies-for-getting-better-results

Starting to muck about with co-pilot.  There's a an impressive pile of stuff MS have built around the LLM technology.  Stuck my nose into a chatbot building workshop yesterday.  Interesting stuff.

Aye - we are thinking of configuring copilot to watermark anything it writes.  Of course you could be a prick and copy/paste but you are doubling down on being a prick.

We see it as a way of starting a document.

i like the document summary side of things.

In co-pilot you can build an expert bot and then add it to Teams so it can be chatted to in a Teams meeting. I kind of like that.

I downloaded all the lab documents from the training course so I'll work them at some point.

I fancy building a couple of bots just for the craic - one for helping you procure something in our place would be a good start.  Our procurement rules are a fucking pain in the arse.

There's the API too for coding which might be worth a sniff.

When you can limit the LLMs scope of reference to specific documentation it will be very useful. Like most big organisations there are literally mountains of policy paperwork, supplemental notes and amendments to dig through at times at my place. Being able to ask something like copilot about an issue, and have a conversation with it while it trawls through the policy on your behalf would be very powerful and convenient.

I mean plugging it into the work intranet, so it can trawl all the info on there (I'm talking 100s of policy documents of many hundred pages each). Would be great to have it as a little browser plugin where you could just ask a question and it comes back with an answer, with the relevant policy and paragraph reference.

Gemini 1.5 is handy for the Google ecosystem, especially Gmail. It gives accurate summaries of the important emails and does well at ignoring all the spammy ones.

These are fun. It’s a bloke playing around with AI imagery to make Lego versions of movie posters. The idea is to see how much the image generator ‘understands’ each movie. Does it nail the tone? Pick up on the important characters/places/objects?

https://imgur.com/gallery/90q6uX6

Well worth clicking ‘more’ and scrolling further.