Here the database outputs a fully rendered bitmap containing the frame to display. Client script driving everything is just taking that bitmap and pushing it out to the display.

Already feel this way about RAM. Last time I checked a few weeks ago what I spent $2k on three years ago now costs over $24k. At least resale markets are surging which should help for awhile.

Interesting people who make these arguments tend to speak in terms of modalities rather than capabilities. It's a pattern matching token predicting statistical glorified autocomplete not even AI.

All the while completely ignoring demonstrated capabilities or excusing them away under mountains of unfalsifiable rhetoric.

So then WTF is this?

"No person may develop, deploy (either internally or externally), acquire, possess, fund, import, or transfer artificial superintelligence or artificial intelligence systems that display one or more superintelligence precursor characteristics, including any elements sufficient to reconstruct the artificial superintelligence or artificial intelligence systemâ(TM)s capabilities."

"No person may deploy (either internally or externally), release, transfer, or import an artificial intelligence system, including any elements sufficient to reconstruct the systemâ(TM)s capabilities, that may be foreseeably modified to produce artificial superintelligence or superintelligence precursor characteristics."

I'm not sure arrest is on the menu for the current crop of hacking transgressions yet it is being contemplated for other transgressions and civil liability is on the menu. Lawsuits have already been filed related to this shit and more are likely to come.

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Flasst.org%2Fwp-content%2Fu...
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.myfloridalegal.com...

https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.myfloridalegal.com...

There is no solution to the alignment problem and there never will be.

Looks like a job for the consumer. Not sure how it works in the UK... around here FDA pulls the "substantially equivalent" rabbit out of its hat and lets it go bananas.

It should use fancy AI algorithms in the watchdog and have it be powered entirely from a single set of pins.

/w LLMs semantics mostly get abstracted from language. Doing this is more like pissing in the neighborhood swimming pool rather than a jar labeled English.

It would be nice if it worked. Perovskite has been nothing but disappointment for years. I'm not sure I buy lab aging tests because moisture infiltration is what does these things in. When someone sells them with a 20-30 year warranty it will be good enough for me.

Perovskite panels only last a few years... normal mono panels have an order of magnitude longer service life.

Bouncing code off of AI is being baked into software lifecycles and will result in fewer bugs across the board. For open source you will have more people armed with AI finding even more bugs (while annoying the shit out of maintainers). The interesting thing about these models you can run them over and over and over again and sometimes different bugs fall out so the more AI eyeballs the better. Still I expect ultimately the computational cost of finding any remaining bugs to work against attackers.

On the other side of the ledger attackers obviously also have the same AI tools for bug discovery and script kiddies get to use AI to help them automate production of exploit code.

Personally I wouldn't want to be at a three letter agency standing watch as my capabilities are eroded by AI. Zero day factories will probably end up running out of steam in a few months time as the bugs get fixed.

Seems to be a bit of selection bias baked into what AI is being asked to do. Tried AI (GLM-5.3) on new code that has never been executed. Also ran it against code that has been in production use for many years.

The types of bugs tended to be in obscure features, buggy error paths, parsing / protocol pedantry, cut and paste errors especially in various lookup tables, algorithm accuracy, inconsistencies, poor and obscure concurrency bugs. Can't really expect it to have found anything too important as it would have tripped up code and runtime analyzers or angry customers because all of that would have already been discovered and dealt with.

In the new code it found a couple of show stoppers that would be immediately obvious the second anyone tried it in addition to some more obscure things.

While I've not yet seen it discover any magical exploits it did get us to reconsider some questionable security related decisions and make improvements. Unfortunately tends to focus mostly on nuts and bolts rather than higher level machinery.

Been trying to get LLMs to do bug hunting for years and it has never worked. The AI just never had the depth to understand enough of what is going on to say anything useful. They still output quite a bit of crap... some of it isn't the models fault... for example tend to feed it source files one at a time to keep from blowing through too much context. This requires the models to make all kinds of inferences about dependencies it has no real knowledge of... sometimes it doesn't make the right assumptions. Sometimes it says nonsensical things or doesn't seem to "see" its own context perfectly misreading the code and complaining about something that isn't real... still well worth the effort. Amazing this shit works at all.

There is no practical exploit even if the method were applicable which it is not.

I have no idea what you are trying to say here. You seem to be ignoring the point I'm making outright by bringing up enrichment limits that expire under the agreement in 2030 and therefore logically cannot be used as an argument against Iranian acquisition of nukes after that point in time because they would be inapplicable.