craighay1

Really the only potential upside to this, and it's a niche one, is for time or security sensitive compute tasks where the raw data originates in orbit. Every happens over inter satellite link and there's no downlink to Earth until the end of the process (downlink is still a problem as Earth is mostly ocean and wilderness)

E.g. one satellite's wide area sensor payload is processed and "potential wildfire detected". The result is passed to another satellite with finer grained sensing capabilities which is due to pass over in the next X minutes which then tees up a capture.

The difference between those past successes and this is that with none of the previous doomsday weapons has it been possible to cobble something together from COTS parts and open software. Nor have they had a large degree of crossover with your common garage tinkerer hobbies like drones and robotics. Nuclear material and chemicals are likely easier to put controls around or get names put on a list if/when acquiring them.

Applying controls to software and embedded hardware is a bit more difficult. Then you have things like precision GPS systems that might just be too useful for civilian uses to not allow free access to. It feels like pandoras box might already have been opened on this one without some backpedalling in certain areas.

Is the majority of current LLM research reaching for the mythical grail of AGI? Or is there enough research effort being poured into how to retrain networks efficiently and how to finetune accurately on small amounts of data?

My personal knowledge base and personal correspondence is miniscule in comparison to the great swaths of the internet these things are trained on. Vague, inaccurate or non-comittal answers based on my personal data are perfectly well served by my magic 8 ball at the moment.

Is this because application developers are simply going where the people are? Or is there better data harvesting to be had on your average iOS/Android app with blindly accepted permissions than there is in the average desktop browser.

It was unclear from the article whether this applies to only credit card transactions or also Mastercard debit cards. At least it is possible to choose not to use credit cards, a bit harder with debit.

I'd love to see how this data is "anonymised".

On a related note, in the UK at least, all of the big supermarkets have switched their loyalty card (data harvesting schemes) from offering "points" per unit spend to offering direct large discounts on certain items in store. Makes it much more difficult to justify opting out of giving away the details of your personal shopping habits to these corporations.

I sometimes wonder if modal editing (like vim uses) might be a good approach for navigating/editing longform text on touch devices.

It does seem like a missed opportunity to have taken the keyboard/mouse approach and then transferred it to touch devices. Even the keyboard layout has no real advantage for two thumb typing on a screen.

Approaches that adapt the interface whilst leaning heavily on letter based inference could be interesting for one handed / single digit entry of letters. Something like dasher: https://www.inference.org.uk/dasher/dashersummary.html

https://en.m.wikipedia.org/wiki/Dasher_(software)

But the author works at Google, so can surely pass the message on up the chain that this design decision is stifling innovation in touch-based text editing?

In cases like these what it to stop the author (the prompter) from lying and claiming that they in fact created an original work using MS Paint? I don't think it's currently neccessary to show evidence of process (e.g. how you made something) in any copyright claim.

Is all of this AI generation in creative fields simply devaluing digital creation and will it encourage a move back to traditional media? This particular image is in the style of an oil painting, but good luck getting an AI model to generate an _actual_ oil painting.