sota_pop

I recall some moves where SO announced they would be revising ToS to disclose their intent to repackage and use the site’s content for sale and/or model training.

Contributors responded by going to delete their own respective contributions en masse. Upon doing so, were banned by the platform mid-process which then led to people going back to revise their contributions to be false rather than deleted.

I guess that’s “LLM related”

Wow, a blast from the past to be sure. Was not but any means avid, but did own a tipman. And was always dazzled when someone showed up with an angel.

Ngl I had to stop reading part way through. I’m no AI spokesperson, but the opinion expressed here is..

“insultingly naive”, “overly simplistic”, “immature and self absorbed”

..exactly what I would expect from someone partly or majority through a PhD about tools. I like tools, their design, creation, and deployment more than average (though probably less than the author), but you need to come up for air, man.

Your interpretation of that phrase is oddly specific, highly esoteric, and completely different from EVERYONE not doing a PhD on tools.

AI is just a tool

That word “just” seems to be doing a lot of heavy lifting for you. This sentiment is not intended to downplay the importance of tools on society. The generally understood intent of this phrase is to quell the current hysterics and to reassure anyone unfamiliar with the term _back propagation_ that AI, in fact, is neither alive nor sentient (in the sci-fi sense), and that it is “just” statistical modeling. Do not mystify the technology.

a car/hammer is just a tool

While these are unanimously considered tools, I could entertain a long discussion and the “particular specialness” of tools and things that are inherently dynamical systems.

what prototyping _should_ be about

The huge majority of the rest of your rant (or what I read of it) is awfully presumptuous and weirdly confrontational. Who can say what prototyping _should_ be? You of many people should understand that the creation and use of tools is contextual, sentimental, highly personal, intimate even.

“But artificial intelligence… intends us not just to sit forward

From the first part of “The phrase”… AI is inanimate. Humans DO tend to anthropomorphize, and the whole point of saying “it’s just tool” is to remind people that… inanimate things don’t have an intent!

our tools are using us

I find this notion somewhat trite. “The tail wagging the dog”. I’m not arguing that we aren’t impacted by the tools we use nor that the use and proliferation of AI is not impacting society, but it takes a PhD level of mental gymnastics to push that concept as far as you have in your rant.

The important takeaway though, is that you are a human, with agency and autonomy.

it only matters how you use it

The implication of “How you use it” is that YOU GET TO CHOOSE. what you think, how you think, “how you use it”™, and even whether you use it! The choice is yours.

I don’t myself have a PhD, but seriously, do yourself a favor and come up for air. Read the parable of “The Empty Boat”.

That is fair, I HAVE INDEED seen similar of those anecdotes. However, I still am skeptical of the file analogy.

LLMs are good at traversing traditional file systems because file/tree-like structures very likely encompass an overwhelming supermajority of their training set. By contrast, other approaches like graph-based vector/data discovery via sql queries seem equally or more promising (to me), not to mention the ability to run curl/http queries. Either way, it’s still all prompt engineering approaches to discover/manage context. In this way, files seem like a “local minimum” ie a medium that dually optimizes interpretability and accessibility for both LLMs AND humans.

Just riffing here, but this could even broadly be considered a discussion of memory-vs-storage (somewhat analogous to fluid vs crystallized memory in humans). In this way, one could imagine the models performing context compaction/backup by periodically dumping their context to files (or some other non-volatile storage) WITHOUT coming back to feature space.. just dump/load the tokens directly.

Personally, I am much more interested in even other approaches like VLMs (using visual tokens), architectures like auto encoders and its variations, jepa architectures and other approaches that emphasize operating primarily within the latent space.

To be fair, my particular interests have always been more in the computer vision area, and have been amused to watch the attention mechanism rise to prominence even over CNNs (given the contrast |similarity in their mechanics). Then again, I began my journey in the ML field when GANs were still the hotness, but I digress……

for LLMs, it’s tokens all the way down, and the name of the game is how discoverable and accessible can you make them?

It’s interesting to me how many people articulate things like the “Preiminary Truths” section as if they are novel insights.

Effective people managers (of whom I would not specifically consider myself) have known these tenets for as long as history. “Be concise”, “state your intent clearly”, funny how these are touted as novel “strategies” with which to expertly direct AI.

I don’t agree that “everything is a file”. Files are arrays of bytes. For an LLM, everything is a vector of tokens/embeddings.

An aphorism I recently heard: "All sufficiently advanced technology eventually becomes a web browser".

… seems apt especially in the context of the progression from chat-windows to harnesses and onwards to “harnesses that can do anything”.

A good tool is invisible such that it is highly opinionated and nails the 80, 90, 99% use cases in terms of making assumptions regarding logical defaults.

Restated, the tool understands the process and what a “good” result/outcome looks like. It correctly presumes relevant and important information (especially given the current stage of your full task), and can “fill the gaps” between start and finish.

A good tool can distinguish between what you wanted, what you thought you wanted, and what you “should have” wanted.

A good tool makes it easy to do the “right” thing and hard to do the “wrong” thing.

A good tool doesn’t function to be understood, but to not be misunderstood.

I have spent a lot of time sitting in this question, having spent years creating design and drafting automation tools for architects, designers, and engineers in the building-design industry.

Nice to hear Google is thinking about this. Was always of the belief that they SHOULD be considering things like scarcity in the various recommendations the apps in their offerings provide.

Also reminds me of that guy in Berlin who caused traffic jams on Google Maps by walking around with a wagon filled with phones:

https://www.theguardian.com/technology/2020/feb/03/berlin-ar...

Personally, for any kind of serious reading, I have to go to a place/room with zero opportunities for distraction. No noise, no surrounding commotion, even slightly dimming lights helps. My retention rate drops significantly if there are any distractions. For lighter reading or docs, I’ve got to take almost diagrammatic-like notes. whiteboard, paper/pen, or ipad work best for me - typed notes don’t work well. I may not even revisit the notes later, but actually writing them helps retention. If I’m out of luck and must read in a non conducive environment, just resolving that I’ll have to re-read once or twice is the only way. Not aware of any magic solution unfortunately.

In my experience, boilerplate sql isn’t even something I would consider a big pain point. The biggest challenge (to me) is explicit data contract enforcement across the stack.

I haven’t personally had third-party ORM frameworks succinctly encourage synchronization and help me build against long-term divergence of data models across the stack. “Make it easy to do the right thing and hard to do the wrong thing” still leaves a lot of room for ‘gotchas’ as apps evolve over time.

Finally, I don’t understand the aversion in learning even a bit of sql. As topics go, it’s a very good (maybe even the best, if I were being provocative) effort:payoff ratio. Not sure I’d call myself a sql expert, but am always pleasantly surprised how much functionality is within reach by knowing even the very basics of sql.

As someone who has historically spent a lot of my time with C#, and now spend most of my days writing python… LINQ is typically what I miss most from C#… (obviously aside from static types and compiled binaries).

We understand more than you’ll never know.

Most upscaling and super-resolution techniques I’ve seen use various implementations of interpolation; typically nearest-neighbor approaches. Although I don’t work in the medical field and haven’t checked in on the research at least since ViTs overtook CNNs for other areas of computer vision.

Sturgeon’s Law abounds.

I’ve always found that most books within or similar to Ferris’ genre hold a nice message but are 90 fluff.

He picks on summarization as the sort circuit and claims the lived experience imbued within his books is the real draw. Not disagreeing here. LLMs unpack the intentional content stretching rigmarole authors go through as a function of publishers requirements to produce something that is minimally viable to publish.

Some seemingly important questions that come to my mind in a future where LLMs, ad-based revenue, and attention-as-currency prevail:

how do you create a knowledge-based product valuable/compelling enough to make people want to pay for it? And what is the best value-prop way to monetize it? (For BOTH producer and consumer)

Is length a necessary requirement for substantiveness?

If so, where on the spectrum of long-content_vs_short-content should it live?

Unfortunately, I im not confident the internet can fundamentally diverge from an ad-based or subscription monetization model in its current form. IMO, Michael I Jordan has a lot of refreshing ideas on how to build better marketplaces that sustainably benefit producers and consumers (as opposed to mostly just platform-maintaining-rent-seekers).

Could it be possible that Zuck has a Llama-shaped magic 8-ball to which he has fully committed himself to “dogfooding” an AI-only strategy to his responsibilities in the company?

“[Ll]esus take the wheel”

Very cool! Analog mono has a very “Christmas sweater” vibe.

I can also relate here, seeking a product review on Sony wh1000x_, Google wrote a nice seeming summary, but scrolling down to some Reddit discussions, stumbled upon a single comment that was very nearly verbatim what the “AI Summary” said, only the ai summary phrased the summary as if it were a sentiment aggregated over many users’ experience. i.e.”users say…”

I agree with the sentiment, but native ads i.e. blogs, reviews, articles, etc. that do their best to hide that they’re a sponsored product review have been around for a long time. Admittedly, LLMs WILL make it even more difficult to discern the difference.

So many questions:

Is “the goal of Search” really: “to help you ask _anything_ on your mind”?

If “reimagined Search” is “designed to anticipate your intent”,

Would it correctly infer my intent to not utilize an agentic approach? Is there an “off switch”?

As for “Search agents”

“operating in the background 24/7”,

What is the carbon footprint of that? How do I turn it off? How do I ask it to stop phoning home my every keystroke?

These questions are asked partly rhetorically because it’s likely I don’t need a team of “24/7 Search agents” to help me guess the answers…

Historically, I scoffed when someone said “here’s the difference between a google search and asking ChatGPT”, or when people said that ChatGPT would “kill search”, but Google sure seems to be in a hurry to burry the original feature all by themselves.

I have been saying this and viscerally reacting to this “contrastive language” for months. it… “hits different”…

“Load-bearing” is a new one for me though, yuck.

THANK YOU, 1000% agree. I call this style “contrastive language”, and find it viscerally unpleasant. It is a pithy attempt to be overly attention-grabbing and “punchy”. Everything that is (was?) wrong with social media (sic LinkedIn) posts, even in the immediate lead up to when it became 100% LLM slop.

“Wow, can’t believe our metrics this month, usage is way up! all our users are maxing out their token limits, KPI already achieved for the quarter!”

It came to my attention recently how many TOTAL objects currently exist in LEO. And that a study said that due to light deflection of these objects, that the earth’s night sky is an average of 10% brighter than it was in 1980s… although I generally am excited by technological advancement, that fact (if true) made me feel somewhat melancholy.

That’s fair, the “practically plagiarize” comment was admittedly harsh and quickly written.It’s obviously an allusion to the paper.

Nevertheless, referencing the LFR work and merely including a link where the paper is discussed feels a little like beating around the bush. The primary article doesn’t seem to use the words “self-similar” nor “fractal” - if you’re only interested in LFR work, why reference Mandelbrot at all?

How on earth can you write an article that practically plagiarizes the title, mention the paradox, and neither mention mandelbrot nor cite the original paper anywhere!?

https://www.researchgate.net/profile/Ion-Andronache/post/Wha...

In what industry do you work?

Nice writeup. I’m curious why you went with chromadb and not pgvector. I haven’t built a rag system myself, but I’ve always understood the initial doc parsing to be a major challenge alone, so kudos there!

Additionally, I also thought it was customary to store a pointer to the source in the same row as the vector (i.e. vector+ doc path + page#/paragraph/etc.) OR just store the original text chunk (though based on your disk reqs doesn’t sound like it would have been feasible).

Glad you’re having good results! Maybe you’ve inspired me to finally try out a similar setup myself!

Very much this, it’s all the technical rigour, code debt, and none of controls/reversibility.

At least when I report fraud to credit card or my bank, they can stop or undo/chargeback a transaction.

I really hope this is a TiST reference.

Or actually good (for smartphones…): radial drag so my thumb can slide ergonomically. Have the entire phone UI/UX have a left/right hand setting.

Single pushbutton: roll to set volume to random number.

clean, concise, single button