alextheparrot

The for-profit is a PBC with the sane mission at the nonprofit [0]

[0] https://openai.com/index/built-to-benefit-everyone/

The paper they published last year goes over some of these transformations: https://arxiv.org/pdf/2510.09263

> Integrating an imperceptible, robust, and content-specific watermark

From the system card someone linked elsewhere in the discussion

No LLMs are calibrated?

I mean, you could also frame this as an issue the electorate could actually prioritize instead of just hoping the courts work it out

The derivative being a grad(ient) student sampling scaffolds against evals + qualitative observations: most prompt-based llm papers

Cancer is a parasitism that kills the host (or the host dies from other causes and it is not self-sufficient). Just because something is defined by uncontrolled self-replication doesn’t mean it is stable to live forever (Which is as much a comment on homeostasis as self-renewal)

It isn’t actually very wrong. Your example is tangential as graders in school have multiple roles — teaching the content and grading. That’s an implementation detail, not a counter to the premise.

I don’t think we should assume answering a test would be easy for a Scantron machine just because it is very good at grading them, either.

> Good evaluations write test sets for the discriminators to show when this is or isn’t true.

If they can’t write an evaluation for the discriminator I agree. All the input data issues you highlight also apply to generators.

I wish the other replies and this would engage with the sentence right after it indicating that you should test this premise empirically.

If this sort of error isn’t acceptable, it should be part of an evaluation set for your discriminator

Fundamentally I’m not disagreeing with the article, but also think most people who care take the above approach because if you do care you read samples, find the issues, and patch them to hill climb better

LLMs evaluating LLM outputs really isn’t that dire…

Discriminating good answers is easier than generating them. Good evaluations write test sets for the discriminators to show when this is or isn’t true. Evaluating the outputs as the user might see them are more representative than having your generator do multiple tasks (e.g. solve a math query and format the output as a multiple choice answer).

Also, human labels are good but have problems of their own, it isn’t like by using a “different intelligence architecture” we elide all the possible errors. Good instructions to the evaluation model often translate directly to better human results, showing a correlation between these two sources of sampling intelligence.

in the app: Settings ~> Data Controls ~> Improve the model for everyone

That’s a premise that would make me consider the wiseness of my actions.

Quippy, but off the cuff: - I don’t go to my present town square(s) socially because it is full of a-social behavior. Same reason to avoid certain bars or clubs, prefer certain parks, or why some are wary of public transit.

- I don’t feel a right to decide the vibe of how a business curates its space. My bakery, coffee shop, local library, etc. all curate a space with an opinion. I don’t feel I have standing to assert that my preferences should dominate their choices.

As an aside, businesses are also an extension of the people, the best ones tend to just not be mode collapsed

Really enjoyed the piece.

A passing thought: the ethe of individuals in the 70s and 80s is important because of the people it informed in subsequent years. While many people still like to hack, code, etc., the relative proportion of people doing this and working in tech continues to diminish as the popularity and importance of the sector grows. I wonder if debt without values / a more cohered zeitgeist is better or worse?

Glibly, I’d also love your definition of the education system writ large.

Bit of a detail, but where are you deriving “with hundreds of terabytes of unified GPU memory” from?

No, but it makes more conceptual sense given the model can consider what was said before it

I love the romance of this piece, but in my experience he’s just describing the difference in expectations of learning biology at a high school versus advanced undergraduate to graduate level.

Romance is for those who care, and most don’t. But it is so, so beautiful once you do.

> “Our best shot at making the quarter is if we get an injection of at least [redacted]% , queries ASAP from Chrome.” (Google Exec)

Isn’t there a whole anti-trust case going on around this?

[0] https://www.nytimes.com/interactive/2023/10/24/business/goog...

You could probably consider learning a sign wave to be a “grammar” related to periodic variations. Grammar in this context feels like “What are the core conceptual heuristics that help guide towards faster understanding”

I'd assume they mean users interacting with the chart vs first load. So the user sees the base chart (Let's say 1MB of data on the server, less depending what gets pushed to the user) and then additional filters, aggregations, etc. are pretty cheap because the server has a local copy to query against

Exactly what I want from DuckDB. Was playing with using it as a quicker cache for a data app backed in Snowflake, wonder how hard it’d be to write the attach for that vs doing it at the client level

The comment was directed at “doesn't this method add another cost and overhead for calling Text-to-Image models”

LLM + Text-to-Image model is exactly how DALL·E 3 is deployed, fwiw

Users here often get the narrative and motivations deeply wrong, I wouldn’t take it too personally (Speaking as a peer)

Not to address the applicability of this, but you can have advantageous fallback behavior and still say “the system isn’t intended to be in the fallback behavior”.

Exactly, litigation has never been applied to content delivered over BitTorrent-style networks