jerrygenser

Even if tesseract accuracy is low, if the tesseract result in addition to the image is then passed to the LLM, it can result in a much more accurate OCR.

For example, GPT4 with some vision capability would be able to fill in the incorrect OCR with the additional word co-occurrence understanding.

I've tested this approach with purely text LLM to correct OCR mistakes and it works quite well.

Also note that in some newer OCR pipelines that don't involve LLMs, there is a vision component and then a text correcting model that is in some ways similar to some forms of spell check, which can further improve results.

> but that is really a gross generalization as stuff like this shows.

I think this research actually still reinforces that we still have very little understanding of the internals. The blog post also reiterates that this is early work with many limitations.

AI used to be autonomously controlled systems. Now AI is re-brand of machine learning by marketing departments, etc.

This seems more like AI than the very recent rebrand of ML. Machine learning can be a sub-system of AI systems but in itself was not thought of as AI for a long time.

Yep, and the only acknowledgement to llama.cpp is that llama.cpp is a "supported backend" (see last page of their GitHub repo readme)

And the acknowledgement was only added after community outcry. If not for that they would have kept on as if llama.cpp didn't exist.

Implying ollama has other back ends which it does not.

Found this fun stack overflow post:

https://stackoverflow.com/questions/49299000/what-are-the-se...

Robots that are powered by AI might be able to.

> So it's not like this model is alien to Our Kind Hosts at YC. They understand that this is a crap shoot with a slightly tilted table, but they're optimising for staying out of the zones where everybody else is betting and not that much more.

Do you think this still applies, given recent waves of following the crowd in the last few years like crypto, and now AI? It seems that YC is actually in exactly the same hype zones as everyone else these days.

I spent a little time googling around but even wayback machine is only showing results in 2024 for the domain.

Was there a different project that was being worked on previously? Or is this the only product being launched since YC S19?

Does this assume retail price? Maybe they are getting a discount for buying in bulk

He co-founded angellist which seems to be at least a moderate success

I wonder if Ollama will or plans to have other "Supported backends" than llama.cpp. It's listed on the very last line of their readme as if the llama.cpp dependency is just incidental and a very minor detail rather than Ollama as a deployment mechanism for llama.cpp and gguf based models.

> they must release training data and source code under an OSI approved license

The source code is also apache 2.0

Anyone have benchmarks on how the llama 3 8b model performs when quantized to varying degrees? I reckon many people will be running these with llama.cpp or similar.

I can believe there are no plans, right now. But having raised over $100mm, the VCs will want to liquidate their holdings eventually. They have to answer to their LPs after all, and be able to raise their next funds.

The primary options I can think of that are not full acquisition are: - company buys back stock - VC sells on secondary market - IPO

The much more common and more likely option for these VCs to make the multiple or home run on their return is going to be to 10x+ their money by having a first or second tier cloud provider buy you.

I think there's a 10% chance that a deal with Google is done in the future, so their portfolio has Firebase for NoSQL and Firebase for SQL.

PG tweeted today in support, kind of: https://twitter.com/paulg/status/1780897522778554841

YC backed company: https://www.ycombinator.com/companies/bloom-institute-of-tec...

They are embedded into a particular semantic vector space that is learned based on a model. Another feature vector could be hand rolled based on feature engineering, tidf ngrams etc. Embedding is typically distinct from feature engineering that is manual.

Numpy might not be able to handle a full o(n^2) comparison of vectors but you can use a lib with hsnw and it can have great performance on medium (and large) datasets.

The acknowledgement is the absolute last line of their readme as one of the "Supported backends[sic]"

They ended up (attempting to) address this issue by including it on the last line of their readme as one of the "Supported backends[sic]".

https://github.com/ollama/ollama/issues/3697 https://github.com/ollama/ollama/commit/9755cf9173152047030b...

llama.cpp only support BERT architectures for embedding but not with classification heads - although there is a feature request to add that

Yeah the naive approach would to join on key. Otherwise can join on a transaction id and key. However it's definitely complicated to get right.

You can fine-tune a 60mm parameter (e.g. distilBERT) discriminative (not generative) language model and it's one or two order of magnitude more efficient for classification tasks like sentiment analysis, and probably similar if not more accurate.

Is there another managed service of Citus that is NOT offered by MSFT? I understand it's legally possible, but is anyone else actually hosting a managed Citus?

openssh does not use xz but in some distros xz is used when it interfaces with systemd

Citus was bought by Microsoft so now it's only offered as Azure managed service.

I have a probably janky habit of creating slack reminders and then rolling them over by some mix of 3 hours, next day, or next Monday.

Should add (2021)

At this scale the negotiated price with AWS could be a lot cheaper even than the sticker price on 3 year reserved instances.

React can be added by rendering into a dom node that's not necessarily the root of your app.

Yes, it can be used to render the entire UI app with rich state management across all pages, routing, etc.

However it can just as simply be used to render 1 complex part of an app that has an isolated but possibly complex state. In this scenario I think sprinkling react is reasonable.