rooftopzen

Sam Altman has been a joke for awhile now, heard only his investors defend him for their next round increase - is that who you are?

Dove deep into this - 25+ security issues; no thx

I agree w post, and relevancy from its original date in 2020, but curious on what original intention was to repost from such a long time ago, see link belwo

https://news.ycombinator.com/from?site=alexw.substack.com

You naively replaced deterministic process w probabilistic process - following a trend that is uneducated.

I am taking screenshots of blogposts like this for a museum exhibit opening next year - lmk if you’re willing.

Caveats: 1. It's 101 pages (do # of pages correlate with aggressive effort to be authoritative, e.g. 'state of the world' pdfs. 2. This appears to come from the AGI existential threat doomer camp - does it even have any validity? At first glance, it appears both absurd in terms of presumptive risks, and also AI generated (this is biased to the pages I'd read) 3. MLCommons has a more scholastic approach to ranking models on potential harms, curious on reception to all approaches (pros vs cons of each)

Lol

Yh I get your point - post is not necessarily designed to prove AI use (it's already highly probable, and not necessarily bad by itself in theory) it's the implications of it that are more interesting than deterministic evidence of it, but by showing evidence of it being likely - updated the post to reflect a better baseline.

Not following exactly, so apologies if I'm misinterpreting, but I'm the author and updated this post (transparently) with nuance I'd recently learned about that explains this (somewhat) - the larger bills contain entire pages with only headings that contain emdashes - removed the headings from analysis so that the emdashes per page are only from the legislative text itself. For the conservatively / minimal difference, we're still looking at a 30% increase from a decent baseline.

I'm the author and updated this post - after looking into this, the larger bills contain entire pages with only headings that contain emdashes - removed the headings from analysis so that the emdashes per page are only from the legislative text itself. For the baseline, over 50% of bills found on congress.gov are 1-2 pgs, after reading a few I decided some rationale could exist to remove them from the baseline - even after all these adjustments, we're still looking at a 30% increase from a decent baseline of similar bill size. It's evident when reading the text below headings (as a human!)

If I’m following correctly, this drama is like a Netflix series (and I agree, crazy stuff we couldn’t make up). If it’s only the Trump admin policies here, yeah it’s crazy bold (and which I state the ethical implications of).

Science fiction plot twist could involve anything completely crazy in the bill no one notices by having to use an LLM to read and that is open to interpretation enough to be only decided by the courts later.. I didn’t look for anything hidden and vague; but how would one really know lol.

Share IT is from 2024, but the 2017 tax cut bill is interesting (lots of emdashes there that deviate from the avg) - you’re correct on the additional need for text analysis in this case. Bills I’d found from earlier in 2024 that are publicly available do not have emdashes outside of the table of contents, which is built into the average - curious how/why they are used so much in this bill from 2017, now wondering how they got into any potential templates (or not), and adds the confound of how much this is AI or template (or requirements, or something else) Thx!

Exactly the point of the post (900 pages falls in line with “flooding the zone” mentality), but we don’t have data on the intent.

Lol can you elaborate

First, see below for Toll et al 2020 and I used autocorrect for grammar. Sorry you were dismissive before looking it up, is more a reflection of your bias.

https://liu.diva-portal.org/smash/get/diva2:1591409/FULLTEXT...

Second, I noted all caveats with an LLM counting that - I actually presumed I undercounted, but it had been noted that a simple ctrl-f found 3.8 per page rather than 9.8 per page (counting only single emdashes not double). The actual number doesn’t matter so much, since low bound is absurd difference from baseline bills I checked from earlier this year and 2024, where they do not exist outside of the table of contents.

4.x emdashes per page (low bound) is absurd, and the implication of this is the point you (respectfully) missed.

Compares emdashes per page in the bill vs the same from the average bill sent to Congress

Important topic but is expected behavior (questionable research if implying this is something that happened randomly):

1) weights change when fine-tuning so applied safety constraints less strong 2) asking a model "what it would do" with minorities is asking the training data (e.g. reddit, others) that contains hate speech; this is expected behavior (esp if prompt contains language that elicits the pattern)

lol no comment - the post states:

>> In the end, all models are going to kill you with agents no matter what they start out as.

Also historically happens during wartimes: https://niemanreports.org/press-access-to-satellite-images-i...

Bad idea. Will be textbook case of boomer toy executives misunderstanding nuances of AI:

https://www.psychologytoday.com/us/blog/parenting-matters/20...

The guys comments on AGI do not follow scientific method - my hypothesis is Meta wants to have enough AGI influencers so that they can dictate definition as goalposts change, not quite based on skill set obv.

Seems AI generated, if not, nothing new here. Post regurgitates info known for long time and misses largest issues of nuance of “LLM-as-a-judge” as if written in 2023 and audience is living under rock (why?):

>> This is where LLM-evaluators (also called “LLM-as-Judge”) can help

>"When it comes to vector search, it's not just about matching words. Understanding the meaning behind them is equally important."

This statement ^ is clearly incorrect on its premise -semantic meaning is already vectorized, and the problems with that are old news and have little to do w indexing.

I went through the article though, and realized the company is probably on its last legs - an effort that was interesting 2 years ago for about a week, but funded by non-developers without any gauge of reality.

If we need a fraud detection algo what tech bro demographics would be stat sig?

Pypi downloads used to be more granular (can't recall when this changed), but still a publicly avail dataset in GCP

AI development no longer exponential; it appears growth is now linear - is this a surprise?

Assuming prompt injection issues will never be prevented 100%, is there any solution or will AI tools be stuck here?

Here's a write up the differences of GraphRAG - TLDR is it's expensive but can improve accuracy (possibly pertaining to LLM definition of the nodes), so if it's worth it to your app. https://medium.com/@jonathan.bennion/graphrag-analysis-part-...

Did you read it? That paper mentions "Comprehensiveness" and "Diversity" lift to a degree of "substantial improvements over naive RAG baseline" - could they be any more vague?

AI safety != protection from AGI (lots more to unpack under AI safety umbrella)