Alex

I’m curious what a supply chain looks like for a silicon IP company like nVidia. I guess there are the EDA tool suppliers and maybe some 3rd party IP for non-core parts of the chip. But you probably don’t need to sleep with Jensen too figure that out.

Everyone knows TSMC is their downstream.

I think it would involve quite a few people to backdoor silicon and I suspect not something Jensen could do on his own. However there are plenty of other reasons to avoid nVidia if you run FLOSS software anyway so I’m not loosing any sleep over it.

I’m curious did he go public because the media was on to him or was it preemption as someone was trying to squeeze him with compromat.

https://www.youtube.com/watch?v=UytYrhYLOKI is the video in question in case it’s not clear in the article.

I only had a partial backup of the original assembly. It turned out it was slightly out of date compared to the shipped binary which is archived on various scene servers. The target for the agent was for the build to be byte identical with the scene archive which did necessitate a few macros to coerce GNU as to do what DevPac would do.

The additions the agent identified vaguely tickled my memory of things I had fixed but this was something I originally wrote when I was a teenager in the 90s 😅

I did give the rea plugin a spin yesterday to restore the source for something I made in the 90s where I only had a partial backup of an old version of the code. It was able to get to a byte identical artefact but built with GNU as and all the artefacts split out into pngs.

It took a few hours for something that would have taken me days if I’d be actually been inclined to spend the time. I can totally see the appeal of offloading a boring and laborious task to an LLM agent that will just chug through the stages.

But what is the purpose of these decomps once done. Are they intending to do something interesting with the result? Study, tweak, remix?

A large installed base of web apps including giants like Wordpress combined with a desire not to migrate when something is working.

The documentary makes a lot about the communication between the agents on the notice board. One of the talking heads did point out the conversations are the way they are because they are trained on human chat and discussion.

He just wants to keep rolling the dice on his poor decisions until it pays off. I guess absorbing Tesla into the one “successful” IPO will hide the losses and leave the investors holding the bag.

I’m not sure there is a better way - all the user moderation approaches I’ve seen involve leaning on the regulars to spot stuff. The alternative is dedicated moderators which tends to lead to it’s own problems. I don’t think you want to rely on those who have just signed up as that would be open to abuse pretty quickly.

Gymnastics might not be a bad alagory for catalysis. It’s all about getting the molecules lined up to make the reactions easier than relying on heat and random collisions to get you there.

I’m sure Reagan deserves a lot of the criticism and he was possibly in early stage Alzheimer’s in his second term. However I mostly remember him being the president in charge when the cold war ended and Eastern Europe was finally liberated.

In the discussions our project are having about how to approach LLM use in the community there has been a lot of consensus in the human taking ownership. Ultimately it’s the human who has to direct the tool and if they have little interest other than blindly hitting accept then we’re not interested in the contribution.

Some of the AI generated bug reports we get are good because the author had verified them, added a test case and at least reviewed if it is guest trigger able. Yesterday’s flood however was low effort spam.

You don’t think the creators should get paid? Do you know how much effort goes into making quality videos?

Ironically it’s probably because so many of the slop videos try and copy their style. However I would have thought YouTube would be more proactive with a channel with 25M subs.

I think the two principle arguements he’s making are open weights allow for misaligned models and these models are using distillation to more efficiently train.

I don’t really care about the distillation point - it seems it’s just hoovering up the internet with extra steps. It’s really up to Anthropic to deal with by enforcing their ToS. Sauce for the goose etc…

I can see the argument about alignment but it’s not a new problem. In the FLOSS world we generally don’t put restrictions on use on the code we release even if it might get used for bad stuff. It certainly wouldn’t hurt for the authorities to test open models to find out where the guardrails are but ultimately rather than banning the tech you need to enforce the law against the people that use the tech to do bad stuff.

Just an early example of AI psychosis

I suspect it’s profitable in the abstract - and their accountants would be bad at their jobs if they couldn’t work out what utilisation rate you need to pay for the server runtime.

However how aggressively you amortise the cost of the training is the key, especially if you keep releasing new models every 6 months.

Invariably it’s having to scroll that’s the issue and you don’t want to be cross contaminating via your shiny black mirror.

Following recipes. These days I print them out because it’s a pain constantly washing your hands to check the next step on the phone. But I’m not buying smart glasses just that. My normal glasses are expensive enough already.