Claude Code is now available to $20 per month customers

[kewms]

@jonmoore

And you’ll notice that there have been no deep fake hysterics in the media ref Veo 3, as Google’s DeepMind are the nanny state of the major Ai vendors.

Sure there have: Google’s Veo 3 Can Make Deepfakes of Conflict, Riots, More | TIME

[kewms]

@uimike

I mentioned medicine, a close friend is deeply involved in using it, in DNA / RNA / Protein analysis, and is very enthusiastic about LLMs.

Is your friend in fact using an LLM, or some other kind of machine learning model? The conflation of different kinds of models – which I think is deliberate on the part of LLM advocates – makes it very difficult to discuss which tools are (and aren’t) capable of which tasks.

[jonmoore]

I stand corrected.

Weirdly, even after disabling my ad blockers, none of the videos play. But it’s a Time piece, so I believe the words.

I would have expected DeepMind to have put stricter guardrails in place, particularly now that the videos have convincing audio.

DeepMind have however come up with an invisible watermark, for generative content, which I hope gets integrated into all their generative output. The watermark alone isn’t enough, internet users need to have the tools available (possibly browser plugins) to identify deep fake content. Until recently, most deep fakes weren’t convincing, but that’s not the case with Veo 3.

That Time piece put me in mind of Mountainhead, which I enjoyed much more than I thought I would, but then again, The Guardian’s Peter Bradshaw gave it a 4 out of 5, and he’s a real barometer as to my own tastes, so I should have ignored the Rotten Tomatoes reviews.

[rkaplan]

@cgrunenberg

That’s just a simple script.

I am certainly an enormous fan of 99% of the AI features which you have added.

Correct me if I am wrong or if this has been revised since I last tried it but what I found when I used Scripting Assistant previously were two notable issues -

1 - If the script does not compile then it gives up

2 - I cannot iteratively ask it to add features to the script or change some aspects of how it functions

For these reasons, I concluded that it is more effective to use Claude directly (either in Claude Desktop/Claude Web or in the DT4 Chat) rather than to use Script Assistant.

Did something change with Script Assistant or did I misunderstand how to use it when I tried early on?

Or maybe the advanced Claude models are now much better at creating scripts with Script Asssistant? I have not tried it with 3.7 or 4.0 yet.

[rkaplan]

@kewms

Is your friend in fact using an LLM, or some other kind of machine learning model?

Agreed it is likely a different type of AI. Similarly tremendous advances have been made using Ai for imaging/radiology in medicine but those are not LLMs.

[jonmoore]

@rkaplan

Agreed it is likely a different type of AI. Similarly tremendous advances have been made using Ai for imaging/radiology in medicine but those are not LLMs.

AlphaFold is what springs to mind.

This is the technology that won the team behind it (Demis Hassabis & John Jumper) a Nobel Prize for chemistry in 2024.

Much as it isn’t an LLM, the underlying Transformer architecture is the exact same deep learning architecture that makes LLM’s possible. And it was in fact only the fact that DeepMind shared their breakthrough with Transformers with the outside world, that allowed OpenAI to use that same underlying architecture for the creation of ChatGPT.

[rkaplan]

@jonmoore

Much as it isn’t an LLM, the underlying Transformer architecture is the exact same deep learning architecture that makes LLM’s possible. And it was in fact only the fact that DeepMind shared their breakthrough with Transformers with the outside world, that allowed OpenAI to use that same underlying architecture for the creation of ChatGPT.

True - but the issue of hallucinations can be managed more effectively than with LLMs.
Nothing is perfect of course- even humans.

[cgrunenberg]

@rkaplan

Did something change with Script Assistant or did I misunderstand how to use it when I tried early on?

This template didn’t change recently (and probably won’t as we have some other ideas in the pipeline).

Of course like every other template it’s customizable (but it’s recommended to customize copies). In this case by changing the extension to .scptd and editing it with Script Editor or Script Debugger.

[jonmoore]

@rkaplan

True - but the issue of hallucinations can be managed more effectively than with LLMs.
Nothing is perfect of course- even humans.

My main point was that Gemini, ChatGPT and Claude were all made possible because of the work of DeepMind.

Much as LLMs still appear to be benefiting from Ai “scaling laws”. I believe that their capabilities will hit a soft ceiling in the near future and other Ai techniques will need to be devised (and not necessarily based on current machine learning techniques), to go beyond up-scaling frontier model pre-training and post training via e.g. reinforcement learning (both human and algorithmic feedback). A modern take on symbolic approaches to Ai, and a simulation of basic human “common sense” are possible additions to the knowledge stack, that may help guide the LLM layer to hallucinate less.

Who knows the specifics, but simply scaling the pre-trained LLM models in isolation doesn’t seem the logical way forward to me. Particularly when you consider the sheer amount of Ai generated slop that already exists on the public internet.