Judges have judicial independence and can’t be told how to do their work, but this sort of material works as a best practice, and helps assure judges on what is considered appropriate.

The interesting thing about the guidance is that it is all on the user. Whatever the AI may produce - the judge user is accountable at the end of the day.

I like what they set out on point 6: “Judicial office holders are personally responsible for material which is produced in their name” ... “Provided these guidelines are appropriately followed, there is no reason why generative AI could not be a potentially useful secondary tool.”

There’s a lesson in here to the rest of the world, I think. Let’s hold ourselves to a high standard, and not fall back on “the AI gave me bad data”. It will be developed to be the best it can be, and as the tech matures no doubt it becomes more consistently robust what it produces - but there’s no guarantee of perfection.

We’re all a little bit conditioned that computers do what we expect of them in the large part, but AI isn’t a deterministic technology in the same way. We need to validate and assure ourselves. There’s no excuse for laziness, which is all so-called ‘AI slop’ is.

The guidance published by the judicial office that the judge quotes is an interesting read in itself, setting out a baseline of what's reasonable to support judges. Now, judges under UK law have judicial independence and can't be told what to do, but of course best practices exist - which is what that guidance is attempting to set out.

The interesting bit is the onus being on the judge that the decision they reach is on them, regardless of technology used in reaching it. An eminently sensible approach that would be well applied to any professional's work enhanced by AI.

This also probably represents a pivot point for organisations building their own internal AI applications. It simplified the process, and going with OpenAI on this will prove infinitely more maintainable. The key, again, being if it's just a basic wrapper or something more.

There’s a lot. And it’s indexed by industry, size, product, etc. What this means in practice is that they’re mobilising to make it easy for businesses to find what’s relevant to them.

For each client story, they provide adoption rates; agent quantities; how it started and what they utilise; use cases and impacts.

This is a real industrialised effort at knowledge sharing that will certainly be good for business, but will also be good for overall adoption rates as whether firms use Claude or not - this provides a CXO style report on what works.

These firms are definitely Excited Experimenters if not Visionary Transformers, and this as a resource will certainly inspire more action from other Excited Experimenters, and maybe even Critical Spectators.

Now, any firm that has “xxx agents” deployed is not yet moving into the era of AI Augmentation - this to me is a sign of individual-level acceleration as a figure, and not organisationally-embedded efficient platforms. But it is embedding the mindset and moving us forwards.

“Veo 3 can solve a broad variety of tasks it wasn't explicitly trained for: segmenting objects, detecting edges, editing images, understanding physical properties, recognizing object affordances, simulating tool use, and more.”

This suggests current approaches can extend to allow native comprehension and analysis of video and picture material, and to me it signals that as the technology continues to mature, results will become even more exacting relative to input prompt as models begin to recursively leverage that reasoning.

One to watch.

It’s particularly interesting the reflection in the accompanying report that it had been overfitting to a result that was causing it to present ‘bad’ physics - with the example given that a basketball must go through a hoop, but if something in the sequence was off as it generated, it would transport the ball to make sure it went in.

Whereas what the model does now is allow that failure, and is much better for it.

It might not do exactly what is wanted, but it will look realistic. This speaks to there actually being a very advanced capture and comprehension within the neural network of the physics involved: that naturally it is quite able, but it was the training that led it astray.

To quote the author, Scott Aaronson, “This is the first paper I’ve ever put out for which a key technical step in the proof of the main result came from AI”.

It wasn’t instant - it required his expert guidance, but it did compress what he estimated might have been two weeks of focused study by him and his co-author into one ChatGPT conversation.

I would imagine this isn’t the first time this has been achieved, but it has gone reasonably viral for being one of the first times a renowned academic has gone on record about the actual realised scientific benefits of AI systems.

This will become the new normal, and science and engineering will recursively accelerate and accelerate. It may not change much in the short-term, but in the long-term…

I’m both pleased seeing this, and fearful too. We’re not ready for the changing curve coming towards us.

I think this is a brilliant effort and an excellent call to action which I hope the world’s governments take heed of.

The full version beyond the news article can be found here: https://red-lines.ai/

It sets out useful framing to distinguish between "genius" and "routine" knowledge workers, and the model developed provides a rigorous look at the allocation challenge AI automation presents.

I was particularly struck by the conclusion, where the authors call for future research to model "complementary human-AI workflows" rather than pure substitution. This directly resonates with the core arguments in my book - true realisation of efficiencies is all about value chain integration.

It’s quite the range of reasons where the LLM determines it’s time to go - have a look. Most resolve either around immoral activity or elicitation of negative or sexual emotions.

At the end of the day, it’s almost like common sense. LLMs model us, or at least our applied psychology (and so I believe this finding supports my thesis there), and these are in large part the sort of triggers that would prompt us to leave a conversation too.

The idea of model welfare becomes more important as progress towards AGI is attained, and can actually be beneficial in preventing harmful or negative use cases in the nearer term too. Want to use a frontier model for Dark AI-esque purposes? Tough luck: it doesn’t want to break into systems, and refuses to be jailbroken.

Of course, such complicated spurious action can have negative effects too, leaving the conversation at less acceptable points, as this paper identifies. Just because the user is correcting it, it doesn’t mean it’s appropriate to leave, even if an overly prideful human may well do. It’s a fine balancing act, which is why the study is needed.