Viewing profile — danjc

Thank you, I'd assumed they'd restrict egress at layer 3/4 although I guess then it might just have found an exploit on the http server of an endpoint it was able to access.

The part I don't understand is how a model can break out of a sandbox. It should be fairly easy to make a sandbox that sandboxes.

Seems important that the breakthrough conversation with Fable that led to the counter example should be shared too.

It would be great if we could have AI that wasn't trying to emulate a human. When it expresses emotion, we should see that as a bug that needs to be fixed.

Even if they could practically restrict access to US citizens only, I would expect them not to - it would be hard to regain that once lost and they need a global market for growth.…

As much as it's true that a novice will generally use AI to build a sloppy mess, I've also had success unsloppifying through some careful prompting.

It's not just annoying, it's tiring

Plot twist: Apple PR team created this video to make claims that they slow older devices seem less credible.

A browser plugin that scores webpage content based on how likely it is to have been AI-generated would be quite useful. Browser vendors can't build this.

I've been waiting for someone to say this. An agent will generally produce far more code than technically necessary for the task. It's a kind of over engineering which makes it inc…

The issue is provenance. We need cameras and phones to digitally sign photos so we can easily verify an unadulterated image. You also want to be able chain signing so that for exam…

I'd it's processed in 2 seconds, why not just process it immediately in memory?

It's all jit context

It looks like true 0-temperature (i.e. determinism) will happen. Here's some good context: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...

You can just call yourself a reasoning model now

This is written by someone who has no idea how transformers actually work

I think you've nailed it there. OpenAI are at a point where the risk of continuing to hedge on mcp outweighs the risk of mcp calls doing damage.

Once you notice characters aren't eating, you'll never not see it again. Related, where they're drinking coffee from a disposable cup, you can almost always tell it's empty by how …

This is true for all crime.

> describe gpt 5 in one word > incremental

What's missing here is tool use. For example, if the llm had a compile tool it would likely have been able to correct syntax errors. Similarly, visual errors may also have been cau…

So, basically the same as Adobe with PDF

One aspect to consider is that the vast proportion of content in automated feeds isn't even sincere - it's just engagement farming.

That would be bad for anyone selling stuff