VanillaCafe

Trying to understand it... if by comparison I'm using tmux then switching to something like this adds mouse based window (panel) management?

For anyone wondering why it takes so long to actually switch this stuff out

One counterpoint is do we really NEED to have brightly colored foods? It's a hard problem if you need a food to be bright red. But, that has to boil down to strictly to improving sales, right? Hypothetically, if all the artificial food dyes were banned, then all food companies would be on the same level playing field.

I thought this might be a useful article because I've often had a similar question. But there's a diagram that has text:

More simply put: imagine that you have red, green, and blue light sources. What is the intensity of each one so that the resulting light matches a specific color on the spectrum?

...

The CIE 1931 color space defines these RGB color matching functions. The red, green, and blue lines represent the intensity of each RGB light source:

This seems very oddly phrased to me. I would presume that what that chart is actually showing is the response for each color of cone in the human eye?

In which case it's not a question of "intensity of the light source" but more like "the visual response across different wavelengths of a otherwise uniform intensity light source"?

... fwiw, I'm not trying to be pedantic, just trying to see if I'm missing the point or not.

I don't know why the narrative became "don't call it hallucination".

Context is "don't call it hallicination" picked up meme energy since https://link.springer.com/article/10.1007/s10676-024-09775-5 on the thesis that "Calling their mistakes ‘hallucinations’ isn’t harmless: it lends itself to the confusion that the machines are in some way misperceiving but are nonetheless trying to convey something that they believe or have perceived."

Which is meta-bullshit because it doesn't matter. We want LLMs to behave more factually, whatever the non-factuality is called. And calling that non-factuality something else isn't going to really change how we approach making them behave more factually.

Unless you could somehow make an Alcubierre warp drive.

Even if you can make it, even though it's theoretically possible that the warp bubble could move through space faster than the speed of light, it's a separate and completely open question as to how you might actually get it to move that fast to begin with.

There are a lot of real world caveats that go into those range estimates. I just took my long range Tesla Model Y with an advertised 326 mile range on a multi-day road trip and I was stopping to charge about every 100 miles. I would love to get something with 3x the advertised range.

A pet nit, and the standards probably don't permit this, but for encoding 128-bit numbers, I prefer base-57 in my own implementations. 22 characters for a 128-bit encoding, same as base-64. You can split it into two 11-character 64-bit encodings. You can avoid the two non-alphanumeric characters in base-64 as well as the similar-looking characters like l1 and oO0. And it takes less visible space, so a bit easier for debugging and tabular output with otherwise no loss of generality.

Evidence suggests that none of this is true. Just look at how Google Fiber wasn't able to navigate the myriad regional regulatory roadblocks thrown up by local governments when lobbied by incumbent ISPs. It's reasonable to expect exactly the same scenario to play out if Google tries to make a competitive ride sharing service from scratch.

I think there is true emergence coming from these newest large language models, which is something that can't be captured by reducing the system to merely a mathematical Plinko machine. To understand emergent behaviors, the system really does need to be treated like a black box and experimented on.

And a secondary point, what if a large fraction of our intelligence comes from language, not the other way around. If that's true, then we might be well on our way to building the spark of an AGI.

Love the demo jitters as he picks up the frosted flakes, "Well that's working. That's good."

I can't seem to find the reference, but I seem to recall that Valve had specialized rendering logic for the power lines in Half Life 2?

I presume that's basically the same fundamental problem as rendering line art and star fields...

They successfully show that Niemann was playing many games with a high percentage of "engine perfect" moves, but they do not do enough to show that this is inconsistent with what top players usually do.

I thought the video very much did make that case. A single known cheating game had a 98% correlation (Sebastien Feller Paris 2010), other GMs have generally at most 75% average correlation. The analysis had more than half a dozen games with Niemann at 100% correlation. If that's cherry picking, it seems like there are a lot of cherries to pick.

The real problem isn't the veracity of the information, but the consensus protocol we use to agree on what's true. Before the internet, we were more likely to debate with their neighbors to come to an understanding. Now, with the large bubbles we can find ourselves in, afforded by the internet social media, we can find a community to agree on anything, true or not. It's that lack of challenge that allows false information to flourish and is the real problem we need to solve.

Maybe it would be worth spending some mental cycles thinking about the impacts this will have and how we design these systems.

Regardless if the bot is actually sentient, a portion of the population may _believe_ the bot is sentient.

If in practice the bot reflects back to each user their own views but perhaps more extreme, then this could be a terrible recipe for reinforcing and amplifying negative and socially destructive thinking -- equivalent to social media bubbles on steroids.

This can occur even without explicit bad actors trying to tip the scales toward a specific outcome. This kind of AI bot as-is has the potential to bring out the very worst in at least some percentage of the population.

Hilarious the visceral response here to the apparent and inevitable decline of YouTube content quality with the removal of a dislike count -- when Hacker News itself doesn't show a down vote count nor even a down vote button for a large portion of its population.

Nice. And now the museum can create a piece of art in response called "Repercussions". (Or, maybe it was all one piece of art all along.)

Title should read: China has forbidden under-18s from playing ONLINE VIDEO games for more than three hours/week.

(And even ONLINE may not be quite precise enough.)

The final bit that locked it in for me (and someone please correct me if this is wrong) is that going to the bathroom (which what many people assume is where weight leaves the body) is actually the last stage of "energy input" to the body and doesn't really play a role in "energy output".

If it ever gets to home computing, it will get to data center computing far sooner. What does a world look like where data center computing is roughly 100x cheaper than home computing?

You're not wrong that BMI is okay enough of a metric for most people. But, in terms of a metric that is even more simple (one simple measurement tool instead of two), accounts for athletic builds, and is arguably more predictive of health outcomes (example: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1071369/), it would be great if we could standardize on waist-to-hip ratio instead of BMI.

360 Wh for 18 cubic meters? Is this a joke? That is indeed the volume of a small room. You need nearly 40 of them, the volume of a large house, to give the equivalent energy storage of one 13.5 kWh Tesla Powerwall.

What does the super scheduler that coordinates the two schedulers approximately look like?

For instance, I assume for instance a common workload that would otherwise benefit from the fair scheduler has a fair chance of wanting to do low latency audio. I believe Android has had this issue. As soon as the platform cares about low latency audio it would need to abandon the fair scheduler?

A NOTE ABOUT DEADLINES: While fair scheduling is appropriate for the vast majority of workloads, there are some tasks that require very specific timing and/or do not adapt well to overload conditions. For example, these workloads include low-latency audio / graphics, high-frequency sensors, and high-rate / low-latency networking. These specialized tasks are better served with a deadline scheduler, which is planned for later in the Zircon scheduler development cycle.

Those seem like important workloads. Does this imply that the deadline scheduler runs concurrently with the fair scheduler? Otherwise, what's the point of developing an ideal scheduler for common workloads if it cannot be used for critical workloads. Is it common to run two different schedulers in the same system?

Years later, when we discover these relics and hop on, it’s as if we never stopped biking.

I'd like to challenge this. Barring scientific evidence to this fact, it is at best an anecdotal evidence, and so I will submit my own anecdote:

I didn't ride a bike for about 10 years. When I started riding again, I definitely felt unsteady for the first few days or weeks compared to my previous riding ability. I was unsteady enough that I thought at the time, "People that say you don't forget how to ride a bike are full of shit."

Granted, my ability to ride came back faster than if I was learning from scratch. It might become a discussion determining different shades of "forget" -- but if we get to that point, then we've conceded the crisp assertion that "we don't forget how to ride a bike".

These principles seem to have already been wholesomely applied to Google then Waymo's approach to self-driving cars. This is in stark contrast to the approaches used by competitors such as Uber and Tesla who appear to favor capturing market first and foremost.

It seems a narrow view to take to assume Google's only AI project with mortal human consequence is Maven, then using that narrow view to confirm your own negative bias about profit, perception, disingenuity.

Step 1? Maybe if you don't want your model predicting crime based on race, sex, or other sensitive features, you shouldn't train your model on race, sex, or other sensitive features.

So in terms of training, they went from nothing to "Go Singularity" in about a month? Impressive.

Agreed. For products in Amazon above a certain star threshold (say, 3+), I evaluate given the shape of the review histogram, particularly minimizing the size of the bump down at 1-star and 2-star.

Note: the blog post was written by Eric Brewer who invented the CAP theorem.