Comment Yay! (Score 5, Insightful) 60
How many decades too late?
How many decades too late?
I get what they were going for. But the tech just isn't ready yet, and some things (such as the lack of a screen) will always be iffy, so it obviously should have just been a cell phone accessory. Probably could have worked pretty nice as a cell accessory. But the cell accessory market is hardly as glamorous as the "new device that makes your cell phone obsolete" market.
They also were about a year or two early, and a year or two behind on the frontend devel. Waiting for a cloud response is just not a practical way to do things. With some development work, you could make a hybrid of cloud queries on a high-end model and local queries on a low-end model (like Phi 3). The low-end model could be fine-tuned to act as a "manager", incl. ability to issue commands to find definitive answers (RAG for search results, code gen to run mathematical operations, queries on your your data in the phone, etc), and to command the apps in your phone (there's been neat work on having models learn how to control arbitrary apps). It could also start answering queries on its own while waiting for the cloud, and then seamlessly merge in the results when the cloud stream started to arrive (incl. backtracking if the cloud suggests a better solution). But all this takes time and development work. A rush to market doesn't cut it.
Would also be a lot more useful after some of the anti-hallucination mechanisms mature. RAG helps a lot with that, but it has its own limitations. My money is still on MoEs with cos similarity between hidden states of each expert run fed back into add+norm for the next layer, since a confident answer should get similar hidden states between different experts.
is that someone developing AI for a tech giant will admit it
Elaborate on what you mean by "globalist".
They had a chance to do this right. They could have, say, mandated a series of tests (which they can update the rules for at regular intervals if they prove insufficient) for testing whether models are memorizing and leaking data that they shouldn't. But as always, leave it to the EU to legislate methods rather than outcomes.
And big companies don't want to be fined into oblivion, so yes, it matters. And with rules like:
untargeted scraping of facial images from the internet
You might as well have just entirely banned scraping from the internet. Who wants a diffusion model that's great at everything except has no bloody clue what a face looks like?
At one point, possibly more than a decade ago, Switzerland changed traffic laws and gave pedestrians right of way at crossings.
That's how it works where I am, and pedestrians just walk out in front of oncoming cars without even looking up.
I think the law is backwards, because the person who screws up isn't the one who gets hurt. If cars had the right-of-way, then the person who screws up would be the one that gets hurt, perhaps motivating the pedestrian to look and think.
Why is it considered news when someone says something that is 100% predictable?
How many commercial products do you see on the market?
Nobody is saying that "neural implants" are new. Trying to make them practical, such as not getting rapidly encapsulated and not having a hole in the head allowing infections to spread straight into the brain and having sufficient bandwidth for common tasks - is the point. This is not easy (particularly the first issue), which is why it's taken so long. The old meat cleaver-like Utah Arrays in particular just plow fat pins straight through blood vessels and trigger an immediate immune response.
The only commitments corporations have are to their shareholders.
Which is about the most damning thing you can say about them.
Is an individual's only commitment to acquire the most money possible? Would/should we ever expect them to do anything else?
Doing better than what we've getting should be easy to achieve.
Shh, we're supposed to be bashing tech that utterly transforms the lives of paralyzed people because the guy who owns the company is a jerk.
The GP was talking about "glacier ice that is already in the water", not ice that has newly moved from land to sea.
Anyone who doesn't want to believe something won't have any trouble explaining it away.
And said ice functions as a barrier for the land ice, greatly slowing down its ability to progress into the ocean.
Also, it's an example of Bad Amateur Science that floating icebergs shed by glaciers don't affect sea level. What's true in the case of a glass of water in your kitchen is not true in the ocean. Freshwater ice, melted and diffused in seawater, results in an elevated sea level. It's a small impact (only about 3% that of land-based ice melting), but still meaningful.
Basically: a chunk of ice, floating in water, displaces its own weight in water. At 0C, Freshwater ice is
Seawater at 0CC however is 1,028g/cc, aka 3% more. For a given amount of displacement, there's an extra 3% of freshwater ice volume and mass. This melts to a volume 3% larger than than the displaced volume.
Think of it using a extreme example. Pretend that neutron stars were stable containable liquids and not highly explosive condensed states maintained by gravity, and that you had a bucket containing a thin layer of it at the bottom. You fill the rest of the bucket with a giant chunk of ice. The neutron star "sea level" rise from having the ice atop it is basically immeasurable. That's your starting point. Now let the ice melt. Now the entire bucket is full of liquid. The "sea level" has risen dramatically.
Of course, it's even more complicated than this in reality, because you don't have a separation of saltwater and freshwater, but rather they merge, and the net density isn't exactly a linear weighted average between the two. Close, but not exactly.
"I have just one word for you, my boy...plastics." - from "The Graduate"
