billythemaniam

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.

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

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.

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.

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 gets sore if my back gets too weak. 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.

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.

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.

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

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.

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.

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.

> 99% of apps are basic CRUD apps

Citation needed.

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.

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

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.

Awesome!

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.

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.

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.

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.

Because they are simply not that interested in trying the new thing. I will drink craft beer, but if someone invented the most amazing craft beer ever, I wouldn't go out of my way to try it even if it was free. I just don't care that much about craft beer. Different people are interested in different things...and that is okay.

Storage costs are an excellent reason as someone else said. Another reason is security, less data sitting around to be leaked, either accidentally or purposely, the better. Another reason is privacy, less data sitting around that can be used against you.

Crypto isn't some completely new thing that needs a new set of regulatory rules. The tech behind it is new, but it's really just another currency for the most part. Are there some differences? Yes! But it is not so different that most of the rules either don't apply or it's hard to see how they apply.

In an ideal world, regulatory agencies would also spend significant resources on education. Unfortunately many of them don't have enough resources for rule making and enforcement, their primary mandate. However, there are plenty of educational resources outside the agencies. There are lawyers, accountants, professors, consultants, etc who are experts that you can hire. There are classes available, in some cases entire degrees.

The regulations are designed to protected the general public from companies doing the wrong thing in the name of profit even when the intentions of the people running those companies are good. It's a check on the dark side of capitalism. Are sometimes those rules too complicated or too overbearing? Yes! Are there also people out there just to improve that? Yes, and some of them work in the regulatory agencies!

It's a messy, frustrating process that on the whole seems to be doing a decent job. Most people in US have access to clean water, generally don't have to worry about tainted Advil, or worry their life savings are going to evaporate by just storing it in a bank.

I don't know enough about the specifics of the Coinbase suit yet, but the crypto industry in general seems an awful like the banking industry of the 1800s and early 1900s in the US. During that time there were plenty of companies and people doing the right thing, but there were enough bad actors that something needed to be done. The SEC was born.

This is very likely fallout from FTX, but not because of some nefarious plot against crypto but rather the SEC realized there is a financial sector they need to focus on more closely. A SEC lawsuit is also a starting point for a serious discussion, not a criminal indictment. If the SEC believes there has been criminal behavior, they refer that to the DoJ. The likely outcome if Coinbase has been trying to do the right thing is some sort of settlement (fine) + direction on what to do differently. Then Coinbase will continue on like normal with whatever changes the SEC wanted.

It's not that weird, its there job is to enforce not teach. The FDA, EPA, etc set the rules and enforce those rules. They do not hold a company's hand following those rules.

I really hope just using static types alone doesn't give people confidence to refactor. It of course depends what "refactor" really means. Are we talking about a function, class, module/library, an entire application? Maybe in the function case, static type give more confidence, but in the others the way to de-risk is to have good tests. Static types only help catch one class of errors where refactoring can introduce several error classes depending on the specific case.

In other words, if static types alone are giving confidence to refactor, it is a false confidence except in the most trivial of cases.

In addition to other answers in this thread, those languages came first. Java's main motivation initially was to fix all the memory bugs that come from manually managing memory in C/C++ and give a pure OOP development experience. Finally for the use cases that C/C++ are still good for, you typically don't want memory management or the JVM performance overhead so there isn't any incentive.

While true, GPT-4 kinda just gets a lot of the classic NLP tasks, such as NER, right with zero fine-tuning or minimal prompt engineering (or whatever you want to call it). I haven't done an extensive study, but I do NLP daily as part of my current job. I often reach for GPT-4 now, and so far it does a better job than any other pretrained models or ones I've trained/fine-tuned, at least for data I work on.

Public transit isn't as good, wildfires, and lack of rain means whatever is in the air lingers for longer.

How come? Genuinely curious.