Happy Autumnal Equinox! Today is the middle of the autumn season, and when the days rapidly shorten.
––– 凄い –––
Memorizing π to 100 digits. I memorized π to 50 digits, and still remember it to this day. I just can't utilize the memory palace technique which is what a lot of people use. As this ronin gets older, the ability to recall gets more challenging. The memories are still there, which is why you know you've recalled something correctly. But the ability to access them seems to be impaired somehow. I've started taking Cocoa Via supplements since a few months ago, and I have definitely noticed an improvement in quickness of recall. This is one supplement that I can say has made a noticeable difference. It helps with cardiovascular health, too, which is a plus.
––– 凄い –––
Here's a doctor who has come to realize that using LLMs for medical answers has limitations. He goes through some techniques that he thinks will improve accuracy. Maybe they will to some degress. But that's the thing about using LLMs – they silently fail. You'd never know if the confident answer was wrong in some way, and we're trained to accept and learn from authoritative sources, which may be the wrong thing to do in this case. We badly need medical benchmarks on these models, but it's difficult because these models soon learn to train on the benchmarks, and they become less useful.
––– 良くない–––
Trump is officially renaming artificial intelligence to "super intelligence". It's just a label, and it won't change anything. It's having a larger working memory and faster processing, which is great to have at your disposal if you are a human who knows what to do with it.
––– 凄い –––
Jev is the latest craze in ML circles now. It's a classification model, where instead of letting the LLM issue the most probable token at the end of inference, you constrain it to defined categories. In this way, you force the model to commit to a type of answer and remove some of the ambiguity. This guy tried it with racial and gender categories on various traits, and gets what you might expect. He didn't provide information about the base LLM model that made the inference. Whether or not one accepts the results, it is what the training data reflect.
––– 凄い –––
Compositor is a free Photoshop alternative for the Mac. I use Apple's Pixelmator Pro, but I might give this a try.
––– 凄い –––
Did OpenAI solve the wrong Navier-Stokes problem? OpenAI may have provided a solution to one scenario, but the broader N-S problem is still unsolved.
––– 良くない–––
Apple's MacStudio M5 Ultra beats DGX Spark. And they only tested this with 256GB of unified memory. Imagine 512GB, which will be available end of October. I gotta get me one of these.
––– 凄い –––
One minute park is a typical GenZ idea. You supply a one-minute video of a park you like, and when someone else visits the website, they'll see a park video. Mental health break, I guess. This is what we old farts did with webcams back in the day. That was the only exposure to the outdoors some of us had during the day.
––– 凄い –––
No one wants to move to Oregon. Oregon and Colorado now have every other home for sale on the market undergoing price cuts. Oregon takes the lead in real estate price cuts. Washington is not far behind. Maybe people just set their initial prices too high. Maybe Californians are more realistic.
––– 良くない–––
