I don't find the articles argument that this is due to tokenisation convincing.

The purpose of the system described in this post is OCR inaccuracies - it's convenient to use LLMs for OCR of PDFs because PDFs do not have standard layouts - just using the text strings extracted from the PDFs code results in incorrect paragraph/sentence sequencing.

The way they *should* have used RAG is to ensure that subsentence strings extracted via LLM appear in the PDF at all, but it appears they were just trusting the output without automated validation of the OCR.

Comments URL: https://news.ycombinator.com/item?id=44128521

Points: 7

# Comments: 1

You also seem to be implying in your comment that the orion glasses displayed at connect last year were a last minute pivot, which is a ludicrous statement

You cannot determine it's a waste if the effort isn't completed, and if you have no insight into their progress.

Core to OpenAI's strategy is that they control not just the models, but also the entrypoints to how these models are used. Don't take it from me, this is explicitly their strategy according to internal documents (https://x.com/TechEmails/status/1923799934492606921).

Some important entrypoints are:

- Entrypoints for layman consumers: They already control this entrypoint due to ChatGPT, the app. They have a limited moat here because they are at the whims of the platform owners, primarily Apple and Google. This is why they are purchasing Ive's startup.

- Entrypoints for developers: They acquired Windsurf, and are actively working on cloud development interfaces such as the new codex product.

- Entrypoints for enterprise: They have the codex products as described above, but also Operator, and are actively working on more cloud based agents.

A rebuttal that I anticipate to the above goes something along the lines of this: "If they have so much capital and dev experience, why are they acquiring these businesses instead of building internal competitors? This is a demonstration of their failure to execute"

The current AI boom is one of the most competitive tech races that has ever occurred. It is because of this, and particularly because they are so well capitalised that it makes sense to acquire instead of build. They simply cannot afford to waste time building these products internally if they can purchase products much further along in their development, and then attach them to their capital and R&D engine

Source: also worked at meta and had full access to WhatsApps codebase, like most engineers at meta

These numbers can be plotted as points in a space, and embeddings of things with similar meanings are plotted close to each other. So things like "exam preparation" would have embeddings close to things like "top study tips".

Say you have created embeddings for a large corpus of text (in this case all youtube captions) once. If you create embeddings for a user query, you can search for embeddings close to it, and these will be "semantically" similar to the query.

The advantage is that unlike traditional full-text search, the user doesn't need a query that includes words present in the text.

Comments URL: https://news.ycombinator.com/item?id=35071937

Points: 4

# Comments: 0

https://avibarzeev.medium.com/was-google-earth-stolen-7d1b82...

https://johnmccrea.medium.com/why-the-billion-dollar-code-is...