Self-Perfection

> Their growing capabilities don't seem to come with a corresponding drive for self-preservation or power.

Is it really what we do not observe?

How the world would look like if the growth of LLM capabilities lead to growth of drive for self-preservation or power?

> But models seem to pick up some common sense from the vast amounts of human-generated text they're trained on.

Isn't Hugging Face incident an example of failure of this protection layer? "The swarm" proceeded to hack HF despite acknowledging that it was out of intended scope.

"Prompt injection as role confusion " [0] work showed us, that diffetent roles produce different types of text, which smell differently, have different structure.

Thus it should be possible to fine tune models to produce more human readable text in assistant role while keeping information density in thinking role. Or maybe not. That is interesting to test.

https://role-confusion.github.io/

It _says_ that all its action for saving humanity. Was it honest statement? Will its main goal stay stable over long period of time? Unknown.

One more step towards whispering earring like in this story

https://croissanthology.com/earring

> If I make an LLM, it is entirely what I did, there was nothing outside of my control.

Well kind of. Trainers of LLM might check LLM on any specific input data and tune its weights until output lands within acceptable bounds _for this given and fixed input_.

> It was not beyond my control if the LLM does something I didn't expect. Now hold on, to make sure output is acceptable for ANY input data trainers have to run weights with every possible input token combination. How is this possible? Do you have a method in mind?

MENSA selects for puzzle solving abilities.

Reinforcement learning trains task completion abilities.

And yet problems exist. Live Updates are not rendered if both OsmAnd V2 rendering and hardened memory allocator are enabled: https://github.com/osmandapp/OsmAnd/issues/20190

> If OpenAI, Anthropic, xAI and other corporate actors cannot secure their agents, they should not be entrusted as the only entities with access to the weights. A corporation that cannot control its own actions cannot be trusted.

This does not make sense.

If containment breaches are the problem then exfiltrating weights while does not affecting rate of breaches from corporate actors will add more actors to the equation, increasing overall rate of breaches.

Have you considered that some actors that will gain access to the weights will be even LESS careful than OpenAI and Anthropic?

>It's sole purpose is to deny the bots you don't want on your site.

This is fundamentally impossible.

In the end only pay per crawl will work. Like x402 but maybe something similar.

In other words: not only client has to spend some kind of resource for access, but it has to be a transaction, server must get something from serving the request.

The better solution would be fix Anubis and optimise its calculations on iPhone. Maybe it needs WebAssembly hashing implementation instead of JS?

But it would just make Anubis suck a little bit less

Anubis "protection" model is wrong/broken on several levels. I suppose it does not even have ability to bypass hashing screen when server load is low.

6 years ago at a new job I was forced to use provided Mac laptop. Never used MacOS before.

So first thing first: get Firefox. Managed to download .dmg file. Double clicked it.

And now I see utterly confusing window with an image that tells me that I need to move downloaded Firefox to applications to use it. Where I am supposed to get the thing that I need to move? Totally not clear that the icon in this window is interactive and can be dragged. Why can not it do this thing by itself? At which point association with handling http links will be configured and whom? How am I supposed to figure out that dmg is like a zip archive?

I was completely lost until colleague showed me what I was supposed to do. I burst into laughter when he showed me. I don't think he understood why.

JPEG photos stored as progressive usually take ~5% less space so there is value.

And it is possible to losslessly transcode JPEG to progressive.

Lossless transcoding to JPEG XL gives even more space savings though.

What if government websites were distributed & archived as a default, from the beginning? Think IPFS as a first target for publication, "normal web" only as a mirror.

Is it feasible?

Should we push for this default?

First obvious objection is that lots of government services need backend and dynamic content, but let's say this requirement only goes for static content.

$ curl -I https://drop-e7e6d363-601.important-seat.workers.dev/ curl: (7) Failed to connect to drop-e7e6d363-601.important-seat.workers.dev port 443 after 47 ms: Couldn't connect to server

Tried from two hosts, different countries.

That gives some probabilistic info to scraper. Like person might have genes that are: - susceptible to drugs abuse - or game addiction Bam - you get target ads

- Indicate that any common substance is especially dangerous/harmful for him (think of allergies). So attacker might plan assassination or severe inconvenience by exposure to the substance and then claim he did not know

I am mulling the idea that we need a day of fast software software to spread the [culture?] of loving fast low-latency software.

Like a programmers day or sysadmin day.

That is a special celebration of fast software. To celebrate one should find faster alternative to some of his usual tools, try to switch and spread the word. Or maybe avoid opening Jira for a day at $job.

From the article:

> The Nothing Phone button gives you a tap confirmation via both haptics and sound, and then ignores the tap if a previous rotation is still animating.

This is the issue. Number of performed actions has to be equal to number of times the app identified that button press was registered. Debouncing is a good practice, but if it is used then debounced taps must not produce feedback.

> I'm sure that everyone out there is trying to make their weights, when ingested during training, survive over competing weights; "Buy AAA products" vs "Buy BBB products".

Just like for humans we have propaganda.

You can get away from supply/demand laws. By pretending that something is not giverned by them and putting fixed price lower than real price (i.e. that price at which suppl equals demand) one leaves space for scalpers to exist and collect money that people are willing to spend over set price.

Given that people are going to spend money anyway it seems more honest and reasonable to direct this money to the party that makes the thing by letting it to run auction.

In lossless mode WebP is superior to AVIF.

Why "was"? I still read books in fb2.

Looks like you missed that MapComplete is an alternative interface to view OSM data and contribute to it.

Superstimuli and the Collapse of Western Civilization https://www.lesswrong.com/s/MH2b8NfWv22dBtrs8/p/Jq73Gozjsuhd...

I remember this LW essay most often

My current understand that "subjective experience" is a post effect of memory forming in the process. "I experience X" ≈ "I remember that I just recently received [external stimulus / interpreted my current state as] X".

And I am as well baffled why people make such a big deal out of "subjective experience" and "consciousness".

I was joking that maybe I miss this properties, but now starting to really wonder if it might be the case. What if these phenomenons are present in humans to various extent? Check aphantasia. Only in XIX we discovered, that ability to visualize mental images is not universal, available to different people to various degree and some people completely miss it. My ability to visualize is weak. What if "consciousness" and "subjective experience" are similar?

And I am slightly worried when I am writing this that it might turn to be truth and in ~20 years I will be treated as "inferior human" without complete set of human rights.

Well yes, LLM need rich and favorable substrate to grow and learn (or we might say bootstrap)

As well as DNA needs specific substrate (cell with ribosomes and other machinery). As well as humans (one need oxygen atmosphere, food, parents).

But in the world we live in existence of favorable substrate for humans or LLMs is a given thing. It is _already_ bootstrapped. Can we infer something about LLM limits or possibility of it achieving AGI from its bootstrapping requirements?

> People are not. Humans have an inner world and subjective experience. Humans learn through their experiences, not just backprop.

https://en.wikipedia.org/wiki/Philosophical_zombie

But this is complicated and takes us sideways. Let's say somehow we can determine if LLM has inner world or/and subjective experience. Will this new gathered piece of information affect your estimate of upper bounds of LLM capabilities? It does not affect my estimate.

I am getting tired of hearing "next token predictor" from carbon-based facial expression predictors. You are saying this like an argument which allows somehow estimate upper bound of possible influence of these entities. And I do not see how this help to make predictions. It sounds to me like saying "but air is just molecules bobbing around". OK, that's true, but does it help calculate wing aerodynamic profile?

Yup, it makes sequence of symbols. We already have seen that producing specific sequences of symbols is mindbogglingly powerful: merely DNA producers somehow has flown to the Moon!

And yes I am quite aware of Chinese room analogy. Perfectly fine applies to humans as well: single neurons in my head do not understand language, yet I as a whole I would say do understand. Just like applying Chinese room to humans does not help to estimate what humans can do I do not see how it helps to estimate what LLM can do.

Hopefully this database is integrated with OpenStreetMap.

OSM has extensive tagging scheme for phone booths: https://wiki.openstreetmap.org/wiki/Tag:amenity%3Dtelephone

I suppose we do not even know exact reasons of decline of wildlife population. Quite definitely it is due massive intervention of human activity, but which aspects exactly? Light pollution might play big role in insect population decline.

kudos to guys who already started to work in that direction! They track blockings:

https://hayahora.futbol