

If someone buys a synthesizer from a store, no one is writing an article on how neat it is that they bought a synthesizer.


If someone buys a synthesizer from a store, no one is writing an article on how neat it is that they bought a synthesizer.


I didn’t say folding, it was literally bragging that they could get it to just get clothes off a bed and into a basket


They can barely get a robot to slowly pick up clothes in a controlled, small environment.
They have been having a very hard time with physical world stuff. They’ve mostly been making them do stunts like dancing and running with the occasional tele operated demo and supremely basic things like the picking up clothes thing.
They don’t have the volume or quality of training data to pull the same stuff they have pulled off in pure digital terms. Quite critically they cannot capture tactile part of human work, and that causes a great deal of limitation.


Yeah, the FAA incident may or may not be related, but it certainly seems just like an averted disaster at my workplace where someone proposed a Claude translated network config to go from one vendor to another and I had to manually interject and insist on a review and found that claude completely screwed up something that would have done… well pretty much exactly what happened to the FAA, had it been actually applied.
The banking issue they flat out mentioned it was AI related, but they only explicitly admitted it was trying to remedy an AI detected security issue. Question is whether they took the AI suggested remediation and got screwed by it, or if a human boffed the remediation, and it’s tough to say without the bank coming out and being explicit about it.
Microsoft came out kind of vague and said some change was made to Windows OS in July that caused the catastrophe that they had to unwind. Again light on detail, but given their excessive public bullishness on full on agentic AI and the sort of change described, it’s almost certainly a vibed change in pursuit of what would likely be considered a premature optimization. Huge risk, no reward sort of change for the sake of someone’s AI KPIs, probably.
A bit grumpy because in my work, I try to be as hard on the AI skepticism as my company will take, and onboarded a pretty all-in AI coder and thought I had at least gotten them to actually make sure their change at least specifically did as they intended, and not merely passed the automated test suites. I sat with him interactively on some pull requests to make sure he actually made a plan to try out his changes and validate the behavior with his own eyes, and thought I instilled a good habit. So shocker as the first ‘big test’ after I trusted that he actually vetted some of his changes fell over. His pull request even explicitly listed the scenarios and claimed that each one was tried. He then had to admit that Claude just made up that it tested the scenarios tested but he figured it was a safe bet if Claude ‘audited’ it that it was ok if it lied about actually trying the scenario…


Yeah, the things is the processes at work are all driven by optimism, “won’t fix” is not an available option for a bug, and “not bothering” isn’t a valid answer to a PR. And both the owner and reporter must agree on any closure.
So we end up with perma-issues that can’t be closed.
I can effectively block the PR, but so long as the submitter wants it open… Doesn’t matter how senior I am.


In TOS, they had the ‘Euginics Wars’ as a plausible WW3, but did not explicitly name it, but that would have been the 90s.
But other than that, the year asserted in First Contact is… oh no…


Just so long as your ultimate pull request is plausibly human viable.
Have been fighting someone because they submitted a 15,000 line pull request for ‘one’ feature and no one is going to review that mess. I gave a very quick ‘for example’ showing in the first 30 lines of changes that it’s screwing around with stuff that has nothing to do with what the sloperator said they pointed out.
Now they are complaining that I won’t take code even though Claude audited it and said it is all good…
(This is work, so I don’t have the luxury of just permanently ignoring them).


It’s bad enough in commercial software land where I regularly hit issues because multiple vendors did the same thing in incompatible ways.
Now slop encourages people to never try to get on the same page.


Even if pure generated content is exempt from copyright, asserting some human treatment applied to the slop brings copyright back into the picture. Hard to prove what is and is not treated by a human if someone publishes material with such a claim.


Pure slop is not copyrightable is likely the case (the couple of cases where they ruled that works with zero human involvement do not grant any human copyright). However, the moment a human tweaks it, that resultant work can qualify for copyright. Consider that Steamboat Willie Mickey is public domain, but despite being derived from that design, Fantasia Mickey is still copyrighted. In the cases where courts ruled the human didn’t have copyright, the human had loudly and proudly proclaimed that there was no human in the creative process. If they claimed they applied some transformative work to it, then that would be copyright eligible. Further, a person is not compelled to divulge the ‘raw’ materials, so even if the slop they tweaked would be public domain, they are under no obligation to show the ‘pre-human’ form of the data.
On the matter of LLM works being derived by virtue of training, unfortunately it is looking like the courts are broadly siding with the AI companies. Not only is it not inherently infringing to train models using material without express permission, it is also the case that if the output is infringing, the user, not the provider, is on the hook for the infringement.


First have to get courts to recognize that the training is subject to copyright for a license to apply in the first place, which isn’t looking good.
I fully anticipate that there absolutely will be copyright consequences for AI generated knock-offs of big software, but conveniently the well resourced AI companies won’t be on the hook, just the end users will be left holding the bag. By that point, the license goes back to working as intended already.


I guess the key is that ‘open source’ generally means ‘free use’ versus ‘source available’ which at least in the old days used to be a supremely common arrangement for commercial software.
I will also agree that every other IP pretty much is effectively ‘source available’, you see a book, there’s no hidden secret form you need to construct that work. Nothing inherently stops you from reproducing a modified copy. Sure there are more convenient assets (having distinct tracks that are used to compose a song versus only the post-mix recording), but generally other IP has to decide how to deal with remix, sampling, etc, without some fundamental barrier like ‘it got compiled into machine code which is hard to modify’.


A fire? At a sea parks?!


It certainly seems supremely pointless…
There’s no accumulated karma, no internet points to be gained. The posts are so milquetoast as to not even be useful for manipulating anyone.
Guessing it’s someone just researching or enamored with LLM shenigans. The most obnoxious sort of behavior that helps make people hate AI.


A buzzer, a bell, a klaxon… Audio saying “caution: chlorine levels too high”


You are correct. Further, AI kind of fell into their lap, though they did have the core competency to sustain supporting that.
However, even if Huang did not drive the success directly, another leader could easily tank things. Just like how when a good president can’t magic the economy into being great, a very bad president can screw it up.
I would say the sort of potential CEO to secure leadership of the second largest market cap in the world is the most likely sort of terrible leader to do that. The sharks that circle for that sort of prize are the ones most eager to burn it all to the ground for the sake of short term self enrichment.


And I would say people are also WAY too confident in their take on how far “clean room” applies to LLM, or how much it covers copyright in practice.


It’s still consistent with my experience, but in certain contexts, can be corrected if the output can be processed in some context that can auto detect wrong output. For example if you have a unit test to pass, then it can be wrong several times but doesn’t matter because it auto runs and gets directed to try again.
In contexts without the possibility to objectively measure correctness, it is still pretty bad.
Ok, but recent stock and financing give some reason to believe that the expansion is going to slow down at least