cheald

It still requires willpower, but to use a metaphor, it's much easier to travel down a well-trod path than it is to cut a new path through the jungle. Repetition and consistency establishes the path, so the willpower required to travel down it the next day is reduced over time. Establishing a pre-set time and committing to that time ahead of time removes the "will I or won't I" decision at "go time", when you're most likely to falter.

Depression isn't monocausal so it'd be far too simple for it to have a single solution.

But in general, humans just work better when we're regularly putting our bodies under reasonable physical load.

That's a great way to think about it. "Get moving" is a lot less intimidating and feels more achievable than "exercise" which conjures images of marathons and four-plate squats.

This is a finding that keeps coming up, and I've certainly found it true in my life, but there's a significant chicken-and-egg problem in that depression frequently precludes the motivation to exercise, and if you don't already have a deeply-disciplined routine to overcome the lack of motivation, people won't do it.

Exhortation to develop those good habits in the good times, I suppose.

Rather ironic to complain about phishing attempts with clickbait (which, I largely think of as phishing's kid brother).

I suppose I'm showing my age here, but Stingrays/IMSI-catchers have been around and in use for decades by both federal and local governments, and the problem of mass surveillance is not particularly an ICE problem. In my lifetime, the level of surveillance of the population has increased so dramatically that I'm not sure that younger people actually understand what it was like to live in a world where your every move wasn't monitored, recorded, and archived.

Privacy advocates have been fighting this battle for decades, but they have been utterly defeated because, by and large, people don't and can't be made to care about privacy until they learn the hard way (and when it's too late) why it's so important.

I'm certain it will kill people, but medical error already kills a huge number of people - the exact number is heavily disputed, but in the US the lower bound is in the tens of thousands annually.

The bone density and muscle wasting are a product of a calorie deficit without protein or weight training. You'll see the same effects without GLP-1s if you just diet and don't actually lift.

As confirmed by DEXA scans, I maintained my (very high) bone density and actually put on about 3lb of muscle over a 2-month period on low-dose GLP-1 + a GHRH, while losing about 16lb of my most stubborn fat, including visceral fat that I hadn't been able to crack. I'm in my 40s, and my routine includes weight training, a high-protein diet, and Brazilian jiujitsu.

My approach is minimum effective dose as an amplifier to a good diet and active lifestyle, and it's been pretty damn effective. Primary effects are an effortless relationship with food - the "itch" I used to feel to eat constantly is pretty much gone, and the psychological pain of not eating things is more or less nonexistent. I did a 36 hour fast just to push it a bit, and it was trivially easy.

Side effects include an elevated heart rate (~5-8bpm, in my case taking me into the low 60s at rest) and sweating all the time, both of which are almost certainly due to the glucagon agonist's thermogenesis effect. No problems with constipation, though stools are harder (comparable to those on a keto diet, in my experience). No particular ill effects otherwise. n=1, YMMV.

My experience hasn't been aversion so much as just total apathy. The magic they once held is completely broken, and I'd rather eat real food.

This is, specifically, why oral GLP-1s haven't been a thing until now; peptides are too fragile to survive the digestive system and make it into the bloodstream.

When all one has is a hammer, everything looks like a nail.

Wow, I did not expect to feel the things I did flying through Ironforge.

It's already home to the largest nuclear power plant in the continental US (Palo Verde), and there are essentially no natural disasters to contend with.

There's absolutely nothing wrong with SSE. The problem is that the author is, apparently, stapling SSE directly onto streaming LLM responses, using the LLM inference provider as the "message queue", and then complaining that it doesn't provide seekability and resumption. Things that aren't MQs make pretty bad MQs. The problem isn't the message bus, it's trying to treat an LLM as a message queue.

Use an MQ if you want MQ things. Worse, the conclusion is wrong - pubsub isn't an MQ, though it often uses them. Pubsub is a pattern for message delivery, which by itself, makes no inherent guarantees about message durability or ordering; Rails' ActionCable, for example, utilizes the PubSub pattern (and in fact is built on Redis' pubsub feature!), but due to its implementation, it makes no delivery order guarantees, which is completely inappropriate for LLM inference.

I think the author took the wrong lesson here. I've had doctors misdiagnose me just as readily as I've had LLMs misdiagnose me - but I can sit there and plug at an LLM in separate unrelated contexts for hours if I'd like, and follow up assertions with checks to primary sources. That's not to say that LLMs replace doctors, but that neither is perfect and that at the end of the day you have to have your brain turned on.

The real lesson here is "learn to use an LLM without asking leading questions". The author is correct, they're very good at picking up the subtext of what you are actually asking about and shaping their responses to match. That is, after all, the entire purpose of an LLM. If you can learn to query in such a way that you avoid introducing unintended bias, and you learn to recognize when you've "tainted" a conversation and start a new one, they're marvelous exploratory (and even diagnostic) tools. But you absolutely cannot stop with their outputs - primary sources and expert input remain supreme. This should be particularly obvious to any actual experts who do use these tools on a regular basis - such as developers.

There's also https://github.com/chaiNNer-org/chaiNNer if you want to run arbitrary models. https://openmodeldb.info/ has a great array of them to choose from.

And a $115b burn rate. They're toast if they can't figure out how to stay on top.

It's frequently suggested that you take vitamin K2 with vitamin D, which helps direct calcium to the bones.

"Putting the entire population on vitamin D supplements would be too expensive for the country’s national health service, he told me."

This seems absolutely bonkers. Vitamin D is dirt cheap, and if you can think at all beyond first-order effects, the improvement in immune health alone would likely pay for itself in terms of cost to the healthcare system.

In regards to obesity and fat loss, smooth, consistent weight loss is essentially always preferable to rapid weight loss - both from a whole health standpoint, and from a durability-of-effect standpoint - and sooner is always better than later.

IMO, the ideal way to use these is as an enabler or multiplier for a healthy lifestyle. Using GLP-1-family drugs to enable an unhealthy lifestyle is always going to work out worse for everyone than using them as an adjunct to disciplined diet, regular exercise, and good sleep.

Neat! Is that published anywhere?

It seems conceptually related to ddpm/ancestral sampling, no? Except they're just adding noise to the intermediate latent to simulate a "trajectory jump". How does your method compare?

I do (same username), but I haven't published any of this (and in fact my Github has sadly languished lately); I keep working on it with the intent to publish eventually. The big problem with models like this is that the training dynamics have so many degrees of freedom that every time I get close to something I want to publish I end up chasing down another set of rabbit holes.

https://gist.github.com/cheald/7d9a436b3f23f27b8d543d805b77f... - here's a quick dump of my SVDLora module though. I wrote it for use in OneTrainer though it should be adaptable to other frameworks easily enough. If you want to try it out, I'd love to hear what you find.

Stable Diffusion 1.5 is a great model for hacking on. It's powerful enough that it encodes some really rich semantics, but small and light enough that iterative hacking on it is quick enough that it can be done by hobbyists.

I've got a new potential LoRA implementation that I've been testing locally (using a transformed S matrix with frozen U and V weights from an SVD decomposition of the base matrix) that seems to work really well, and I've been playing with both changes to the forward-noising schedule and the loss functions which seem to yield empirically superior results of the standard way of doing things. Epsilon prediction may be old and busted (and working on it makes me really appreciate flow matching!) but there's some really cool stuff happening in its training dynamics that are a lot of fun to explore.

It's just a lot of fun. Great playground for both learning how these things work and for trying out new ideas.

My wife has chronic migraines which are successfully mitigated by rimegepant (gepants are CGRP antagonists), and didn't have any particular change in headaches from GLP-1.

What has helped, interestingly, is supplemented creatine HCl (she went with hcl because it's absorbed substantially faster than monohydrate). We've learned that depletion of neural ATP levels can result in an energy crisis which results in cortical spreading depression, which stimulates the release of CGRP. (https://www.sciencedirect.com/topics/neuroscience/spreading-...)

She's found that a) daily supplementation of creatine has reduced her headache days, and b) an immediate dose of creatine upon onset of a headache frequently aborts or mitigates it. Her need for the gepants has dropped to a tiny fraction of what it was prior to starting creatine.

She's tried everything under the sun, had all the scans, tried all the meds and procedures, and creatine and gepants are the only things she's found that have worked. She's not a placebo responder, and hasn't responded to about a zillion other therapies, so we're pretty sure it's not just placebo effect.

AI is great. Midwits with AI are dangerous. I've been saying for a long time that the failure mode for AI isn't the AI itself, but the humans using it, and the better the AI gets, the more I think that's borne out.

That's very similar to what I've been experiencing. "This is the best solution, it's what everyone uses" when I know for a fact that it's actually not. Very disappointing when you're trying to solve actual problems.

Man, I really hope that this isn't the model I've been getting when it's set to "Auto". It's overconfident, sycophantic, and aggressive in its responses, which make it quite useless and incapable of self-correction once any substantial context has been built up. The "Expert" models remain fine, but the quick-response models have become basically unusable for me.

I'm afraid it probably is.

Yes, this is broadly true. Just about everything that does not have Linux-disabling anticheat runs wonderfully on Linux these days. You can check https://protondb.com/ to see how any given game runs.

Easily half the other events on the calendar are kid-related. Of course it's going to infer that, absent other direction, the most likely overarching theme of the visible events is "child care".

I don't think it's a personality change, at least insofar as personality is separate from neurochemistry. Ghrelin and dopamine are strongly linked, and dopamine is our central reward-seeking driver; GLP-1s generally reduce ghrelin production, which I suspect helps remove you from a state of being constantly primed for reward-seeking. I noticed this firsthand when I went hard keto for the first time, and could suddently for the first time tell a significant difference between "I want food because I'm hungry" and "I'm not hungry but want food because it's pleasurable".