ralusek

What's interesting is that the way in which they're childish is actually extremely human. In fact, one of the ways that you're often taught to draw more realistically is to stop thinking of the concepts as icons you're drawing the outlines of, and instead sort of blur your eyes and see things as they are: hues and values. In other words, become a camera or a printer that has no idea what it's capturing or printing other than a grid of values. That is how you achieve realism.

The fact that it has clearly iconified these concepts in its mind and is tracing the outlines of the things it thinks/expects to go where is very human.

One benefit of an open source one is that you can, as a large corporation, run it "locally" within your own data center. Even fine tune it.

It's funny that they triggered the infamous "Code Red" moment in OpenAI when the 3-3.1 models came out. I switched to using them for a lot of single-shot LLM calls because they were fast and cheap. Their only area that they were lacking in was agentic/tool calling.

Needless to say 3.5 was a disappointment. Curious to see 3.6.

If you want an LLM to have knowledge about something, the knowledge has to either exist in its weights or be provided to it in its context. Because context is expensive and limited, and models tend to get dumber the more their context is filled, there is usually more that you'd need to put into context than can reasonably fit in it in order for the model to answer questions about your data. So your options are basically

1.) stuff it into context

2.) figure out a way to determine what to put into context based off of what is being asked of the model (RAG)

3.) change the weights of the model to have knowledge of your data baked into it (fine tuning)

This is like a single prompt AI. The term labels are illegible against the white background. The connections don't make any sense. "Training vs Inference" connects to Large Language Models and GPU. It doesn't connect to RLHF. Machine Learning connects to fewer things than Large Language Models. The whole thing is insane.

MoE is literally exactly what they're describing. The classifier being described is baked into the model and is the thing that makes it MoE.

Imagine I had [a model that was good at math], [model that was good at code], [model that was good at writing], [model that was good at general knowledge]. If I then had [a model that was good at determining whether the user query would be best served by one of those models and sent it to it, leaving the rest of the models inactive], that is the platonic version of what MoE is. In practice, it works a bit differently. It instead basically restricts the number of pathways that can be utilized in solving problems during training, which allows for "expert neuron groupings" to form and "classifier layers" to form earlier on in the structure, but the effect is the same (better, even, since it allows some overlap between structures of experts). It also allows "routing to an expert" to happen token-by-token rather than at the prompt level.

I didn't know people could just have GPT running on their own hardware. How does one...do that? Do you have a special relationship with OpenAI and they lock down your servers or something?

This is kind of insane reasoning. It's basically asking "what is their pay going to be justified by once their pay isn't justified?"

Guess if people who write articles like LLMs

Claude performs 50 microappreciations per second

1.) China is not communist, even remotely so. China is fascist in every sense of the word.

2.) Authoritarianism can move faster than anything. They can just say "wipe out that village, build the coal plant there, data center here, fab here.

3.) If it's red tape and regulation holding the US back, then that's clearly not "capitalism."

I don't use Gemini cli because they're so bad at agentic work/tool calling. I use their chatbot all the time, though.

I use both constantly for different things. You don't need to be a one-model Andy

I have built a few voice based integrations into my applications that use these live agents (gpt and gemini), but they are always too expensive to be viable. I have to end up hacking up context and turning on and off in ways that are very fragile. It'll end up being $2-5 for about the 30ish minute sessions I typically end up with, and it throws the price of the product I'm making completely out of whack.

I have the same response every time I hear this: like 95% of application CRUD plumbing is much better served by an ORM. It gives your application typed versions of your data, lets you work with objects rather than rows, which are almost always more useful, is much easier to read, etc. Then for the 5% of critical/complicated queries: just use SQL there. In fact your ORM almost certainly has an escape hatch for you to do that.

European statecraft has led to a massive bureaucratic blob whose best and brightest just go to America.

To me in second grade, Red Alert was literally just building Tesla Coils.

Given my memory these days, I can't keep that either.

This implies that corners were cut. They were not. They went through the full regulatory procedures.

That's absolutely not true. The standard for new vaccines, iirc, required a period of something on the order of 7 years. Time, in this case, is not a function of procedure that can be expedited in an emergency, but is actually an important element in and of itself. Many issues do not manifest immediately and actually need follow up over time.

The crazy thing about these vaccines was that both mRNA vaccines and the viral vector vaccines were completely new platforms, never deployed at scale. They work entirely differently than all other vaccines. Up until this point, vaccines all delivered the antigen in one of 3 ways: you get a weakened virus, you get a dead virus, or you get the antigen itself (subunit protein like Novavax). Both the mRNA (Pfizer & Moderna) and the viral vector (J&J) vaccines worked by getting either mRNA or DNA (viral vector) into your cells, and then having your own cells produce and express the antigen themselves. Basically the difference between server generated code or shipping the JS for you to run the SPA on your own client.

One of the crazy things about this was that it wasn't obvious what the implications would be of having our own cells expressing the antigens (and thus flagging themselves for destruction by our immune system). This was particularly concerning because the cells that were shown to be doing this, despite the complete lie that kept being repeated of the vaccine staying localized at the injection site, were found all over the body. In the case of the viral vector vaccines, at least they were being delivered by a vessel (living adenovirus) that our bodies have had billions of years of evolution to determine where they might end up. In the case of the mRNA vaccines, though, the vessel was a lipid nanoparticle with an exceptional ability to deliver payloads basically anywhere in the body. Note: the attention these lipid nanoparticles had received prior to their use in mRNA was their ability to deliver payloads to places that are notoriously difficult to reach, notably their ability to cross the brain blood barrier. So you have delivery mechanisms delivering a payload that makes our cells into antigen factories, shown to be producing them all over the body, and targeting themselves for destruction by the immune system/causing an increased immune response in these areas.

And then, for the icing on the cake, there was mounting evidence that the antigen itself was actually likely destructive/problematic.

I could go off forever on this topic. The amount of obfuscation and gaslighting was insurmountable for anybody that was even remotely interested in figuring out what was happening. From a personal perspective, my trust in many institutions was permanently shaken.

My favorite part about this/what blows my mind is that his system has him editing singular tiles at any given time. He seemingly only gets to see what it actually looks like at intervals like 15 years apart. There are probably entire epochs of his system that he'll never actually see laid out because they've since been overridden.

My understanding was it was an OPTIONS preflight request that is made.

I understand CORS and I don't.

TL;DR: It's a restriction your browser gives itself. If it's on Domain A and it sees a request going out to Domain B, unless Domain B responds saying that it's expecting traffic from Domain A, the browser prevents itself from making the call.

I think the part about it that is off/silly to most people is that it's not a normal security threat model, because a malicious client could simply just...not impose that restriction on itself. You're perfectly capable of going and curling that same request to that backend, or calling it from an app, or any number of other things. So it's not really protecting your protected resource, the backend, from malicious clients.

All of that is where I feel like I understand clearly. The part I fail to retain is the exact scenarios it does protect against, which IIRC, are basically about attempting to protect your users from being misguided on other clients that are acting as your client, something like that (but again, this literally only applies to browsers). It's just kind of a weird niche problem that I often find myself thinking "I mean why is the user on another client and have allowed themselves to authenticate on that client with my server...this sounds like the user's fault."

ORiON Nero 6.6 was my shit. Please give it a spin.

There were definitely many keygens I would open just to have on in the background.

When people say the AI bubble is about to bust, I don't think anybody means that "the use of AI is going to go away." AI is absurdly useful. I think what people mean is "the valuations of these companies will have to snap to a reality that is actually attached to their market value."

Those prices, what a disappointment.

We used to not know, but now because open source models are being hosted and served by people whose only incentive is making profit on directly running inference, we have a ballpark idea.

If my grandmother were to find out that housekeepers occasionally do actually take things, it would set us back decades.

It can also do that for any picture of a man.

The human mind is capable of the same thing, you know? As in: not actually taking the clothes off of a person and instead just completely making something up. I hereby give permission to all AI, and human minds, to completely make up what I look like naked.

What's puzzling or schizophrenic about that? Those seem like two very natural factors that would be in tension with one another and have to be balanced.

PaymentMethods = a specific credit card, debit card, etc. Payment Method is basically a term of art so ubiquitous that it's user-facing in UIs and has nothing to do with Stripe.

PaymentIntents is definitely a Stripe abstraction, however, but that's one that I like. It's been a while since I used it, but I remember liking that it allowed me to bundle up everything related to the payment, i.e. the amount, the payment method, etc, and pass it around between server, client, and different views in the client, such that you could really build the exact payment flow you want without touching PCI data.

The Stripe abstractions I have always felt are much clunkier are the distinctions between Products/Prices/Subscriptions/SubscriptionSchedules, etc. A lot of "what lives where?" with those; very clunky to work with.