jcattle

What I can not wrap my head around: How are these models trained?

What training mechanism or model architecture provides the glue to go from human text to images?

Don't you need to have millions of really descriptively labelled images?

Ah I thought you would link https://xkcd.com/605/

I was wondering why I don't get that animation. Turns out I turned on reduced motion in Windows 11, to get snappier UI and they actually set some accessibility tag correctly for the animation!

Same as for banks. Downloading some verification app with a confidence inspiring 1.2 rating on the app store, getting on a call with some random gig worker looking like they are taking the call in their living room and wiggling your ID around while giving a thumbs up is not the way I would like to prove my identity to a bank.

But there's no alternative. The EU digital identity wallet would be the alternative. You control and exactly see, what kind of information the bank is getting from you and you can be sure that it doesn't flow through some sketchy third-party identification service.

The thing is that LLMs can only ever go forward. It might go into one direction, once that direction is in the context it will realize that it doesn't work, but can't delete. So it has to backtrack (or double down, see the whole seahorse emoji breakdown).

Something something bit flips, something something stateful

Yep. I don't see 500 million being even close to enough to develop broad-spectrum preventatives that would be easy enough and safe enough to administer, to get over 67% of the population to take them.

Yes all good points showing issues that academia has at the moment.

However I often see this going from "there's issues" to discounting academia altogether and positioning private labs as a good or only alternative.

After all, most people in the open science collaboration which published the seminal paper kicking off the replication crisis were from academia.

There's this crowd on HN which is very vocal against academia. From what I've seen, the main points are that academia isn't efficient, most of the science coming out of academia is useless and that the whole system is just a waste of taxpayers money. Instead, what is often argued, all good research is done in private labs. Then pointing to SpaceX, Moderna, OpenAI, Google, etc.

And while it is very true that often the research coming out of Academia is useless, what is always neglected are the roots of the research done in private labs.

When Jürgen Schmidhuber and team published their work on Neural Nets back in 1991 it was also useless. Unless you had a supercomputer and very, very deep pockets you were not going to do anything with what came out of their lab.

But still, 30 years later here we are, standing on top of the shoulders of this useless research.

It's tech from the 80s. Look up the Nishika N8000 and Nimslo 3D.

Basically it's multiple lenses next to each other, each capturing a small slice on the 35mm film. Every lens has it's own shutter, which is triggered at exactly the same time.

This wasn't too involved and quite cheap to implement with analog tech in the 80s/90s, but if you want to do the same thing with digital there's quite a bit more to consider. Here's a cool video of someone building a digital stereo camera: https://www.youtube.com/watch?v=_aofxbH0elo

The hard part with digital boils down to: Cheap camera modules are hard to calibrate to the same parameters and sometimes impossible to set focus, so pictures look the same. And taking pictures takes quite a bit of processing power, so if you want to take 4 pictures at once it gets a bit tricky with just a cheap raspberry or similar.

there's also https://www.myvocab.info/en

From what I can tell they actually have a bit more robust science behind their algorithm (and a lot less questions to answer)

Then I have a backup yubikey at home for services which allow to register two keys. For other's there's still good old password+some second factor.

Pretty happy with having a yubikey on my keychain. Log in someplace new? plonk in your yubikey and off you go!

Yes, you could do a lot more with the data. The only other thing I did was a wordle style guessing game: https://guess.landshade.com

But the data is open source, so if you want to dig around my badly documented code and raw satellite data yourself, be my guest: https://github.com/jonasViehweger/LandShade/tree/main/data

That's pretty cool. In a similar but very different vein: A few years ago I took twenty years of daily satellite imagery and computed the mean color for countries and the world https://www.landshade.com/

But in doing that you really do notice how everything concerning colors is just a bit arbitrary. You get raw reflectances from a scientific sensor on a satellite with specific spectral bands and sensitivity within those bands. And then you try and map this scientific sensor to the sensor that is your eyes, to try and emulate what we would actually see if shot up into space.

There's some really cool science around that if you're a color nerd: https://www.sciencedirect.com/science/article/pii/S003442571...

A coding agent should make short work of that. However I'm a bit doubtful if the results would actually be meaningful.

I'm thinking that some of the LLM-isms are a bit more complex than just repeated phrases. It's often more that short, punchy writing style with quick setups and punchlines. But would be interesting nonetheless. I really think that some things (like "solving real problems" or "it's not"/"this isn't") would show up.

This blurb gave me the idea to try and quantize this. Scrape the top HN blogs over the last few years and see how occurences of common phrases change.

I'd expect to see a huge increase in "solving real problems" over the last months.

Thanks I'll check out that video

Very nice visualizations, thanks for that!

One thing I still struggle with in my head is how these vision embeddings can then be used to give LLMs eyes.

Because you somehow need a giant training set which describes images in natural language, no? Is that actually how it works, or is there some smart trick so you don't need to pay labellers a bunch of money to look at pictures and describe them.

If you ask him 100 times in a row I would bet that by the second, maybe third, time he will not answer with his song but with: what the fuck is wrong with you?

Nicely circling back to LLMs not being able to learn and form memories.

The author mentions that in the blog. The majority is the delay of the updated file actually being written to disk by the device.

The author did not find a solution to trigger file write earlier/more frequently.

Good to know, thanks! This article was such a breath of fresh air compared to the usual "LLM-assisted" writing you get.

Just sentences like this:

This isn’t just a problem for far-off countries of which we know little, like the EU and the US and China. Here in the UK [...]

So good! I feel like I'm becoming an old cynic but if it's the tenth time on the day that I read an overdramatized "It's not X, it is Y" in an article, actually good writing just hits different.

I'm not suprised that in the swiss economy no one bats an eye at 1000 CHF bank notes. After all the swiss are historically known for being the classy alternative to launder and store your ill gotten gains from, for example, your stint as the dictator of an African country.

But there has been some changes in recent years so I don't know how it is today.

Yea that part also stood out to me.

Everything else is also chock-full of plausible sounding but baseless claims and generalizations devoid of any nuance.

For the people inside the Foundation: this is not a moment to manage. It is a moment to decide.

What does that even mean? Moment to manage what? Decide on what?

I guess it's about habit. In the end you are communicating. If I get into the habit of being rude while communicating with a machine, I would be afraid of this habit spilling over to my communication with other humans.

I can't with these AI generalizations for big effect.

This is the standard tech playbook. Fire the engineers who know how the system works, fire the ones organizing labor, hope nothing catastrophic breaks before you can ship something splashy. Twitter did it. Meta did it. Salesforce did it. Google did it. We have all seen this movie.

Just fluff without any substance.

Is that the standard tech playbook? What did Twitter, Meta etc do? "Ah you know, didn't you hear? They did that thing. With that splashy release."

I think a more higher level "internal representation" was meant. The internal representation of knowledge.

Sure we know that a model stores weights. We also know that a neuron transmits electric pulses. That doesn't mean that we know how knowledge is represented in our brain.

But even if RCP 8.5 did happen, it was not a humanity ending scenario.

Can you do it like the IPCC report and assign a confidence to that claim?

Mine would be: RCP 8.5 ending humanity (very low likelihood, low confidence), based on absolutely nothing.

I would wager that investment in battery tech, solar, wind and electric vehicles was largely driven by the trust of companies in institutions that carbon pricing will continue in the future.

I'm not a macro economist, I also haven't looked for any sources on this, but this is my guess.

Same for LEDs. I would guess that adoption, investment and improvement of LED tech was driven in large parts by a clear roadmap abolishing incandescent lighting.

There's a huge number of people in developing countries (think Indonesia, Brazil, Nigeria, Kenya, Mexico) where the country just lagged on internet build out. Now in these countries the price of a subscription will be a lot lower, than the >100$ you pay in the US, but since (simplified) the only additional cost per customer is the cost of the end terminal, it's still worth it for the ARR. The break-even point per customer will just be further in the future.

Also I wouldn't underestimate the amount of people living in rural areas of the US, Canada, Australia or Germany.