sdesol

> Human brains are unpredictable. Look around you.

As it was mentioned by others, we've had thousands of years to better understand how humans can fail. LLMs are black boxes and it never ceases to amaze me how they can fail in such unpredictable ways. Take the following for examples.

Here GPT-4o mini is asked to calculate 2+3+5

https://beta.gitsense.com/?chat=8707acda-e6d4-4f69-9c09-2cff...

It gets the answer correct, but if you ask it to verify its own answer

https://beta.gitsense.com/?chat=6d8af370-1ae6-4a36-961d-2902...

it says the response was wrong, and contradicts itself. Now if you ask it to compare all the responses

https://beta.gitsense.com/?chat=1c162c40-47ea-419d-af7a-a30a...

it correctly identifies that GPT-4o mini was incorrect.

It is this unpredictable nature that makes LLM insanely powerful and scary.

Note: The chat on the beta site doesn't work.

Can you ask for the prompt when chatting?

why not s11?

Yes and no. I've found the order in which you give instructions matters for some models as well. With LLMs, you really need to treat them like black boxes and you cannot assume one prompt will work for all. It is honestly, in my experience, a lot of trial and error.

I haven't found a model at the price point of GPT-4o mini that is as capable. Based on the hype surrounding Llama 3.3 70B, it might be that one though. On Deepinfra, input tokens are more expensive, but the output token is cheaper so I would say they are probably equivalent in price.

Also, best bang for the buck is very subjective, since one person might need it to work for one use case vs somebody else, who needs it for more.

> For example, write a simple pipeline where Claude would scrutinize OpenAI's answers and vice versa.

I'm working on a naive approach to identify errors in LLM responses which I talk about at https://news.ycombinator.com/item?id=42313401#42313990, which can be used to scrutinize responses. It's written in Javascript though, but you will be able to create a new chat by calling a http endpoint.

I'm hoping to have the system in place in a couple of weeks.

For regular users I agree, for businesses, it will have to be a shotgun approach in my opinion.

Edit:

I should add, for businesses, it isn't about better, but more about risk as the better model can still be wrong.

For those not familiar with DuckDB, it's an amazing database, but it is not a replacement for SQLite, if you are looking for a lightweight server side DB. I'm in love with the DuckDB client and use it to query SQLite databases, but due the fact that it only supports one concurrent write connection, it is not suitable as a server side DB.

> We'll have to see if the bump to $22 this year ends up happening.

I can't read the article. Any mention of the API pricing?

> cheapest AI model when asking questions and then switch to the more expensive if it doesn't work.

The thing is, more expensive isn't guaranteed to be better. The more expensive models are better most of the time, but not all the time. I talk about this more in this comment https://news.ycombinator.com/item?id=42313401#42313990

Since LLMs are non-deterministic, there is no guarantee that GPT-4o is better than GPT-4o mini. GPT-4o is most likely going to be better, but sometimes the simplicity of GPT-4o mini makes it better.

Yes, that is the point. The customer service person would be most able to determine if what the LLM said makes sense or not. My point is, we are sold automation instead of a power tool.

I've said before, but I'm not convinced LLM should be public facing. I know some companies have been burned by them and in my opinion, LLM should be about helping customer support people find answers faster.

> Achieving accuracy with RAG and LLMs is a challenging task that requires balancing precision and recall

The challenge is domain knowledge and not tech in my opinion. There are dozens if not hundreds of companies providing RAG and LLM, but the challenge is, like you pointed out, what should you do if you encounter something like "Ankara".

For BestBuy, this might not mean much, unless there is a BestBuy in Turkey. For a government related site, cities and geography is important, so trying to extract additional meaning from Ankara is probably important.

Well if you ask GPT-4o the following:

Give me keywords to search for based on this sentence "international relations about the country with the capital city of Ankara"

You get the following:

- Turkey international relations - Ankara diplomacy - Turkey foreign policy - Turkey global partnerships - Turkey international politics - Turkey geopolitical strategy - Turkey foreign affairs - Turkey global relations - Turkey NATO relations (if relevant to your topic) - Ankara as a diplomatic hub

So it is not unsurprising that the same link was returned

I think anything requiring strong reasoning will probably have issues. However, I think most Enterprises is only interested in knowing that the summary of a document doesn't contain hallucinations, which I think most models will probably get right. If you go by a super majority rule and use 5 models, I think most business will be satisfied that the summary that it was given doesn't contain hallucinations.

However, like you said, we are dealing with a non-deterministic system so the best we can hope for is a statistically likely answer.

Probably in a couple of weeks. It's taken a while to finalize the UX but I know what it should look like now.

I'm working on a rather naive approach that is focused on identifying errors in a LLM response by using LLMs. What I can share right now are screenshots with regards to how it works. The basic idea is you can use other high-quality models to validate and compare against to find irregularities or errors. You can see what it looks like below:

https://app.gitsense.com/--/images/options.png

https://app.gitsense.com/--/images/validate.png

https://app.gitsense.com/--/images/models.png

The basic idea behind my chat system is, every model can be wrong, but it is unlikely that all will be wrong at the same time. This chat system is based on what I've learned when building my spelling and grammar checker. If you look at the following links, you can see that even the best models can get it wrong, but it is unlikely that others will get it wrong at the same time.

https://app.gitsense.com/?doc=6c9bada92&model=GPT-4o&samples...

https://app.gitsense.com/?doc=905f4a9af74c25f&model=Claude+3...

> Different models have different strengths and weaknesses

I would add different errors as well. Here are two examples where GPT-4o and Claude 3.5 Sonnet cannot tell that "GitHub" is spelled like "GitHub".

GPT-4o: https://app.gitsense.com/?doc=6c9bada92&model=GPT-4o&samples...

Claude 3.5 Sonnet: https://app.gitsense.com/?doc=905f4a9af74c25f&model=Claude+3...

I don't think there will be one model that will rule them all, unless there is a breakthrough. If things continue on the same path, I think Amazon, Microsoft and Google will be the last ones standing, since they can provide models from all the major LLM players.

In addition to what you said, faster searches can also provide different search options. For example, if you can execute five similar searches in the same time that it would take to execute one. You now have the option to ask "Can I leverage five similar searches to produce better results" and if the answer is yes, you can now provide better answers and still keep the same user experience.

Where I really think faster searches will come into play is with AI. There is nothing energy efficient about how LLM work and I really think Enterprise will focus on using LLM to generate as many Q and A pairs during off peak energy hours and using a hybrid search that can bridge semantic (vector) and text. I think for Enterprise the risk of hallucinations (even with RAG) will be too great and fall back to traditional search, but with a better user experience.

Based on the README, it looks like vector search is not supported or planned, but it would be interesting to see if SeekStorm can do this more efficiently than Lucene/OpenSearch and others. I only dabbled in the search space, so I don't know how complex this would be, but I think SeekStorm can become a killer search solution if it can support both.

Edit: My bad, it looks like vector search is PoC.

> Who cares if a LLM can spit out an opinion on some political sensitive subject?

Other governments?

> I think the process could be automated

It can definitely be automated in my opinion, if you go with a supermajority workflow. Something that I've noticed with LLMs is it's very unlikely for all high-quality LLM models to be wrong at the same time. So if you go by a supermajority, the changes are almost certainly valid.

Having said all of that, I still believe we are not addressing the root cause of bad searches which is "garbage in, garbage out". I strongly believe the true calling for LLM will be to help us curate and manage data, at scale.

If you look at my post history, you can see an example of how claude and openai can not tell that GitHub is spelled correctly. The end result won't make a difference but it raises questions regarding how else it can misinterpret things.

At this moment I would not trust AI to automatically make changes.

Something that I'm working on is making it easy to fix spelling and grammatical errors in documents that can affect BM25 and embeddings. So in addition to generating keyword/metadata with LLM, you could also ask it to clean the document; however, based on what I've learned so far, fixing spelling and grammatical errors should involve humans in the process, so you really can't automate this.

> Hypothetical answer generation from a query using an LLM, and then using that hypothetical answer to query for embeddings works really well.

This is honestly wear I think LLM really shines. This also gives you a very good idea if your documentation is deficient or not.

Expect to be able to try it by the end of this month if I am lucky. It will be a simple one line docker pull command or if you want to install it, use npm.

Yes I agree if the prompt that was used was less strict with regards to verb tense, prepositions and so forth. I wrote this prompt for use for technical writing, which may not fit the style of the author.

I guess I was a bit direct but I don't fully understand the down vote. I was not implying that the README was bad and it does have corrections that would improve it. My reason for not raising a PR is some repo owners don't care and I really didn't want to go through the effort unless they actually care.

This ultimately why I believe Microsoft and Apple will be the big winners. I suspect a lot of companies will want Microsoft and Apple to sign off on things and Microsoft and Apple are going to make sure they get their cut. We may need a new layer above existing operating systems in the future, to safeguard things.

Once you have created your README, you might want to consider using a tool that I'm currently working on. I'm focused on fixing the "garbage in, garbage out" problem for RAG by focusing on spelling and grammar, among other things. You can see an example of the tool in action at

https://app.gitsense.com/microsoft/TinyTroupe?doc=9bc6270618

This will be open sourced and I hope to have source available before the end of this month.

We are entering the nondeterministic programming era in my opinion. LLM applications will be designed with the idea that we can't be 100% sure and what ever solution can provide the most safe guards, will probably be the winner.