merry christmas

the only things in the new testament about jesus’ life after age 12, before starting his ministry at age 30, are the verses mark 6:3 and matthew 13:55

in most english translations, they say jesus was (the son of) a carpenter, but this isn’t quite right in the original greek, the word is τέκτων, tekton, a general term for artisans, craftsmen, and builders - indeed, this word has the same root as τεχνολογία - technologia, the study and discourse of craft

jesus spent his early adult life as a builder, making useful things

good morning from san francisco.

now that the dust has settled on the first round of the openai debacle, it’s time to start asking some questions

two of the board members who voted altman out, helen toner and tasha mccauley are deeply enmeshed in ‘effective altruism’ ‘effective altruism’ as a movement has a lot wrong with its ideas, some of which i’ve documented before, but this isn’t about that - it’s about basic competence

https://twitter.com/atroyn/status/1725936159665377333

regardless of how exactly this shakes out - and it’s looking increasingly likely that this was a poorly executed power grab by mental defectives - it’s important to remember that ‘effective altruists’ have been consistently wrong about literally everything they completely failed to predict the dominant paradigm for ai (llms), instead building castles out of sand on the idea of recursively self-improving agents and ideas like “instrumental convergence” and “orthogonality”, none of which apply because an llm has no agency

i spent the weekend working to try to get more data about the question of whether or not gpt has world models

my results seem to suggest it doesn't, but i found some interesting things along the way. writeup below. the entire question really kicked off for me when people were using gpt's performance on chess under certain conditions as strong evidence for the idea that it has world models

at the same time, it couldn't play well from random positions

https://twitter.com/atroyn/status/1708532375968530603

making retrieval practical for ai applications takes a lot more than repackaging a vector database built for semantic search, or adding vector search to an existing database. in the first place, the core vector search and storage architecture for ai is fundamentally different from that made for semantic search

in semantic search / recommender systems, all the data in a single very large index is accessible to every user

theodore ’ted’ kaczynski regains self-awareness in the basement of ‘the annex’ of the department of social relations, harvard, cambridge.

the heroic dose of lsd administered personally by henry murray has had unanticipated effects on his mind.

it is spring 1961.

ted is 19. his inherent shyness has melted away. something has reformed itself in his personality.

furthermore, murray’s constant psychological assault no longer touches him. it’s a feeling as if these people and their cruel experiments don’t matter.

he leaves the study.

here is an argument against the possibility of recursive self improvement of any 'intelligent' computer program, based on kolmogorov complexity. intelligence is the ability to make correct predictions about the state of the world given available information.

each program which makes predictions about the world has a kolmogorov complexity corresponding to the length of the shortest string which can express that program

have been discussing what exactly this means and felt it might be useful to clarify what i'm talking about when i say 'do engineering' in this context

i'm referring to the empirical approach to engineering which marks the evolution of any technology from experimental to useful

https://twitter.com/atroyn/status/1645504550890258432

the process goes like:
- identify an interesting phenomenon in the function of the system.
- empirically characterize the appearance and process of the phenomenon
- develop a predictive model of the phenomenon
- develop a theory which explains the model
- repeat

last weekend we ran the first ever chatgpt plugins hackathon, with support from @OpenAI, @Replit and @retool

by all accounts, it was an enormous success.

here are the winners:

https://twitter.com/atroyn/status/1644788509541277696

@OpenAI @Replit @retool category: civic duty

find relevant legislation in progress about a political issue you care about, find out what stage its in, then find the contact info of the relevant people and get your voice heard

https://twitter.com/mpopv/status/1645248186838290432

TREMBLE, ENEMIES OF TECHNOLOGICAL PROGRESS

CHROMA HAS SECURED FUNDING TO SHAKE THE VERY FOUNDATIONS OF THE POSSIBLE

A NEW WORLD IS BEING BORN

https://twitter.com/jeffreyhuber/status/1643987813346742274

i am, obviously, incredibly excited about this next stage for @trychroma.

a.i is the kind of once-in-a-generation shift you hope to see as an entrepreneur, and it feels like we're uniquely positioned to ride it

i have for a long time been prepared to be swayed by the arguments of @ESYudkowsky and of the alignment/safety community in general

i cannot get past this part in his piece in 'time'. this reads like 'and then the ai invents magic' @ESYudkowsky this is the part i struggle with most in trying to understand safety arguments, and believe me, i really have tried - sometimes i feel like the person on earth who has read the most safety / alignment literature while staying unconvinced

it's 2043.

agi has been with us for anywhere between ten and twenty years, depending on who you ask.

it's a sunny friday afternoon in san francisco. i'm walking to meet an old friend for coffee at flywheel.

it's still there, right at the edge of golden gate park. i've got my airphones set to minimal distraction - apple's ar thing worked out by the way, these are pretty good - i want to think 'organically' for a while, take the world in as it is instead of how it could be.

tomorrow there will be time to look ahead again.

chroma engineer @HammadTime has shipped a brilliant demo showing how you can use chroma's javascript API, together with vercel and the VoxCeleb dataset, to find out which celebrity you sound like. try it out!

https://twitter.com/HammadTime/status/1633222219219472384

@HammadTime building this kind of application is super simple with chroma - check out the write-up here, give it a shot for yourself!

trychroma.notion.site/chroma-blog-f6…

i don’t startup hustle post much, but one of the more gratifying things about building chroma has been the trust and confidence other successful founders of really great companies have in us if you like what we’re doing, we welcome your contributions via pull request or even github star github.com/chroma-core/ch…

why do most approaches to controlling model behavior rely on acting only in the model's input domain?

prompt engineering, fine-tuning etc. all feel like trying to pick a lock with a wet noodle. mechanistic interpretability (valuable work) follows the execution path and attempts to map it onto a human-understandable computation, but that doesn't quite feel right either.

from some simple premises it's fairly straightforward to empirically estimate a lower bound on the physical complexity of a 'superintelligence' only two premises are necessary:

- a superinteligence is a configuration of ordinary matter

- a superintelligence will quickly consume all resources in its lightcone

announcing the chroma embedding database. it's the easiest and best way to work with embeddings in your a.i. app.
trychroma.com we designed chroma to be as easy as possible to get started with.

just `pip install chromadb` and you're good to go. no configuration, no additional installation necessary.

docs.trychroma.com/getting-started

buddy, you failed to predict or even notice the possibility that one of your most significant donors with whom you corresponded regularly was an outright scammer, i don’t think it’s ‘the public’ who has problems understanding risk

https://twitter.com/anderssandberg/status/1625528237559279618

effective altruism trying its level best to divest itself of longtermism as if people don’t also have long term memories forum.effectivealtruism.org/posts/TykvLXN7…

stable attribution has been out for about 24 hours now, it's certainly created a response.
as promised i'll go into a high level technical overview in this tread.

https://twitter.com/atroyn/status/1622355473193381888

first, i want to repeat, we cannot detect whether images people upload to stable attribution are generated by stable diffusion or not.

our algorithm works on all input images. the results are only meaningful for images generated with the vanilla open source stable diffusion.

announcing stable attribution - a tool which lets anyone find the human creators behind a.i generated images.

stableattribution.com given any image generated by stable diffusion 1.x or 2.x, stable attribution finds the images in the model's training set which most contributed to the generated image

how the fuck are longtermism and effective altruism compatible?

how can you simultaneously believe that you ought to measure the impact of charity in order to allocate to it effectively, while believing that the best allocation is to some immeasurable distant future? this is a rhetorical question, the reason these beliefs are compatible in many people is because they really truly believe they’re that good at predicting the future