Comcast and Verizon were granted special tax breaks in order to reach 98% diffusion of 'broadband' in PA. By 1998. They still haven't met the 1998 definition of broadband, but Comcast has exclusive licensing to the region I live in. I have only starlink, dialup, or god freaking awful geostationary satellite as an alternative, and it is not legal to start another cable provider, even if I wanted to here.

Those cell companies throttle down to LTE all the time. I tried the T-mobile thing once for about half an hour. It flip flopped connection protocols in the middle of conference calls and dropped in the middle of downloads in order to hop between connection types. It was damned near unusable.

One person can download the whole archive and make it available. Silly talking head man knows that, but doesn't want anyone to panic. I've never actually seen anyone make a security by obscurity argument unironically in defense of a breach.

yes and it maximizes engagement. period. it doesn't differentiate between positive and negative engagement. They have the tech to do that, but it would be expensive to run it at that scale for every comment on every post all day long.

IF they don't pair engagement signals with sentiment analysis on those same ones and treat them like a whole signal together (subtracting points for negative sentiment and adding it to positive) then they will all do this. The algorithm is public. We already know it doesn't use sentiment.

When you think about it, you're allowed to walk in anywhere. Look at whatever you like. Listen to whatever conversation you want to. This is well and truly already established. Now that someone can remember everything they see and hear, suddenly we're talking about privacy issues. Lol.

We still should be thinning. That jackass in the article saying thinning couldn't be done is an idiot. You don't fix it in a year, you start the practice of thinning so that a decade from now it can be under control. Cutting emissions is the same kind of a solution, except the timeline is EVEN LONGER.

Now, we will watch the scramble from all the telephone companies seeking to rebrand their handsets as 'wearable tech'

They can be called in the same way, but not all of them are as good at working in a particular harness as each other. Anthropics models have been trained to call tools in a certain way, the chinese models may or may not work in a way that is compatible. Tool calling ability (not just whether or not you can do it, but how and when it decides which tools to call and why) differs WILDLY between open models.

National security isn't part of the picture unless these things are running in China, or phoning home. Something that is easy to prevent (for now)

No. RSI doesn't mean a model modifies its own running code/weights. It trains another copy of them as a baseline for the next round of the experiment.

They are. It's just not happening on the versions we have access to. Look at Sakana. RSI just involves defining an experiment and parameters, and varying the values of the parameters, trying out a training run, and record the results. If it works well, use the results as part of the base model for the next iteration. All of that can be automated. I have no idea why people are saying they can't be. If people iterate on them and change one variable at a time, and use that as a means to improvement, an AI can be tasked to do the same. And they can also be trained to design the experiments themselves.

Small teams running the show is the new normal. AI coding assistants mean that software engineers must also be product engineers, well enough to explain how to build features from soup to nuts, to LLMs. I know. this is what I do for a living now.

You can turn off auto-updates, but you'll be sorry in other ways.