billythemaniam

vor 2 Jahrendiscuss

Yes, federated queries (external tables) are supported but that is a lot slower than ingesting the data into Snowflake's storage and querying it. Since Snowflake's pricing model is based on computation time, querying external tables are usually more costly because of worse performance.

billythemaniam

vor 3 Jahrendiscuss

I suspect ignoring them is more effective than essentially complaining back.

billythemaniam

vor 3 Jahrendiscuss

Trial and error. I recommend looking up non-weighted exercise, stretches, and activities that target the body parts you are interested in, and start trying them.

billythemaniam

vor 3 Jahrendiscuss

I'm not surprised. LLMs are not good at those problems. There is a lot of hype, it not good at every problem, but it is quite good at some of them. If you have a classical NLP task, then it is good, particularly GPT4. If you have a generative problem where you don't care about mistakes, such as grocery list, marketing copy rough draft, etc, then it is good.

LLMs are not good at search, math, encyclopedias, logic engines. Maybe some day they will be, but not yet.

billythemaniam

vor 3 Jahrendiscuss

Similar experience for me. I found that many of my aches and pains as I get older are resolved by exercise and stretching. For example, I hurt my lower back in my 20s. Not extremely bad, but bad enough that it still gets sore if my back gets too weak after 20 years. I stretch and strengthen it without using weights, and it eventually gets better. I stopped running for a decade then decided to pick it back up. My knees were sore or hurt mildly initially, but eventually they got strong enough that they got better. Running is also great for strengthening the lower back.

billythemaniam

vor 3 Jahrendiscuss

CockroachDB and Yugabyte are Postgres compatible. GCP Spanner also has a Postgres compatibility layer. CockroachDB and Spanner support foreign keys across shards which is an interesting feature.

billythemaniam

vor 3 Jahrendiscuss

Google engineers invented most of the primary innovations in NLP in the last decade, such as transformers and word2vec which started it all, that led to the current stack that OpenAI is built on. Their best public models certainly lag OpenAI, but it certainly isn't vaporware.

billythemaniam

vor 3 Jahrendiscuss

Without hard evidence, that's a great example of a conspiracy theory.

billythemaniam

vor 3 Jahrendiscuss

Drug trials are already double-blind and highly regulated, particularly stage 3, so further controlling for the corruption you are suggesting doesn't seem necessary. If there is evidence (hard evidence not hearsay or conspiracy theories) of that sort of corruption, then I think what you suggest is a good improvement.

billythemaniam

vor 3 Jahrendiscuss

It is extremely common among research universities where they try to own all your ideas. I remember a professor had a side business cleaning pools which had nothing to do with his research area. The university found out and wanted their cut. He gave it to them because it was better than being fired and sued.

These types of contracts are illegal in CA, but legal in many other states.

billythemaniam

vor 3 Jahrendiscuss

I agree LLMs alone aren't good at search but their embeddings replace the need for stemming, manual synonym lists, etc in most cases. LLMs can also be used for query understanding which can improve the keywords submitted to the engine and extracting the best snippet for a highlight. LLMs + search are better than either alone. However LLMs still have an inference performance/cost issue which may make them unsuitable for some search use cases.

billythemaniam

vor 3 Jahrendiscuss

> 99% of apps are basic CRUD apps

Citation needed.

billythemaniam

vor 3 Jahrendiscuss

There are definitely use cases where you want a render loop in web, pioneered by React, instead of traditional approaches for a web application. I think the pendulum swung too far and too many sites are being built as a SPA, but there are absolutely some "very good reasons" to build a SPA with React. As always, it depends on the use case.

billythemaniam

vor 3 Jahrendiscuss

That is machine learning or more colloquially: AI. A pretty cool result regardless.

billythemaniam

vor 3 Jahrendiscuss

I found that stemming the text before generating vectors helps increase recall and the vectors still capture context, etc. However it does hurt precision because some information is lost by stemming. The more recent vector training algorithms are better able to capture semantic, syntactic, and contextual similarity without a lot of preprocessing. So I have found that vectors can replace all the nonsense that used to be needed to increase recall: stemming, manual synonym lists, etc.

However vector similarity search only helps with the literal text search not ranking. Tf/idf, bm25, page rank, learn to rank ML, etc are still needed to rank documents. Whenever I find a new vector search engine, I always look to see what ranking features it has beyond vector similarity.

billythemaniam

vor 3 Jahrendiscuss

Awesome!

billythemaniam

vor 3 Jahrendiscuss

Imagine if every big box store parking lot was covered with a solar panel roof? Customers walk to the store in shade and out of rain, cars aren't extremely hot in summer when customer returns, electricity for EVs right there, excess can go to grid or batteries, not disruptive to anyone or ecosystem, large sizes. Obviously someone has to pay for it which is always the tricky bit.

billythemaniam

vor 3 Jahrendiscuss

No it's designed to generate text: summarize some text, grocery list for a steak dinner, name ideas, short stories, etc. I think a lot of people want LLMs to be encyclopedias, but that's not what they are designed to be or good at. The fact that they can do encyclopedia tasks at all is amazing.

billythemaniam

vor 3 Jahrendiscuss

Yeah pretty much. The data is pretty clear that life for most of the global population is getting objectively better overall such as declining poverty rates. Climate change is certainly looming and starting to have a real effect, but it is only starting. But a lot of people feel "the world is getting worse" because most news is negative and most social media is overly positive, meaning it makes people feel like they don't measure up to the influencers. I think the word "propaganda" implies a coordinated effort which I don't think is the case, but the current state of media seems to hurt mental health and biases perceived reality to the negative.

billythemaniam

vor 3 Jahrendiscuss

Classic NLP tasks (eg classification, summarization, translation, etc) just work with GPT-4 mostly. It is probably still possible to beat GPT-4 with a fine-tuned model, but it isn't easy. The open source LLMs are pretty good too at the classical NLP tasks, but still need to be fine-tuned in many cases. However I bet eventually open source LLMs will get close to GPT-4. What this means is that at a minimum, LLMs will be used to replace "legacy" algorithms for classical NLP tasks to boost accuracy. Also more people who have a problem that can be solved/improved with ML, but currently is cost/time/expertise prohibitive will use LLMs.

comments