darrelld

I got his book Build Your Own LLM From Scratch and it's super readable and does a great job at holding your hand through the entire process...need to finish reading it though!

My parents got one from Walmart and to use any of the software interface they needed to login to a Walmart account. Scary.

Looking for folks working with loyalty points & credit management

Been building a platform around this, have one customer who has been happily using my solutions for 5 years now. Seeking to bring others onto the platform.

Email in profile. USA East coast based, but willing to work worldwide

Am I the only one that doesn't really feel a difference in performance from model to model?

From around GPT 4 results got "Good enough"...I generally try to explain what problem I'm trying to solve, set limitations and boundaries, tell it to ask me questions, have it write up a plan with steps then we take one step at a time.

These new models are starting to feel like iPhone releases where the improvements / feature set feels incremental.

Same on the Claude side which I use for work

We have a junior research student who does this on slack.

Before I think he was using AI and other tools to translate his messages since english isn't his first language. His own words, then AI translated it. Sounded kinda clunky, but I could sense the human behind the words and I gave grace since I can only imagine how hard it is to properly communicate your ideas when English isn't your first language.

But now it's gotten to a point where I can tell he's not using it for just translation. I'm gonna need to chat with him.

Redacted example below

```Hey @PERSON_WHO_ASKED_QUESTION Both good, and the retrieval one isn't written down anywhere. In order.

One corpus or four. My lean is one. Same chunks and embeddings tables, source_type column to tell them apart.

Values I'd propose, flat rather than nested: paper, dataset_description, dataset_readme, dataset_contributors, dataset_records, dataset_files. Description and readme split because their units already differ, one row versus one row per paragraph. A discriminator that can't separate those isn't doing much. The existing 440 rows would need backfilling to paper.

Chunk id in the same spirit: dataset doi, source type, ord. So EXAMPLE_DOI.

Worth checking before any of this matters: does chunks.doi carry a foreign key to papers.doi in 0001_init.sql? If it does, a metadata chunk with a dataset doi can't go in that table at all, and separate storage stops being a choice. Ten second read, I haven't done it. Shout if you get there first.

Retrieval is the one that's bigger than it looks. Search once across everything and nothing guarantees a paper chunk and a metadata chunk both land in the top k. The facts we want relate the two, and the generator can only write those if it sees both sides in the same window. So if one type systematically wins the ranking, that class of fact doesn't get worse. It becomes impossible.

Which way it goes I don't know. Two mechanisms pull opposite ways. Metadata chunks are short, tens of tokens against roughly 450 for a paper chunk, so they may just lose. But we embed context header plus text, and on a short chunk the header is most of the vector. The header is the dataset name, which is also most of the query. That points the other way.

Cheaper to measure than argue. Load one dataset's metadata, run a normal dataset level query, record the rank of the first chunk of each source type. Runnable as soon as any one of our four subtasks lands.

After that it's one pool, per source with quotas, or one pool with a floor per source type. I'd rather not pick before there's a measurement. ```

Exactly!

I've been working on a side project that I started in 2020. If I wanted to implement a new feature it was: - Wait for regular work hours to wrap up around 5 or 6 PM - Get dinner and rest / relax until around 8 or 9 PM - Open up the editor, think about the problem, Google things, read stack overflow which gets it 95% of the way there, Google more, dig deeper into the docs finally find what I needed - Write a function, make some progress, run into another roadblock, repeat previous point - Look up and it's now 1AM. I should write tests for this, but I'll add that to the backlog - Backlog only ever grows

Now with AI I describe what I want, it does the grunt work likely cleaner than I ever could, adds tests and warns me about potential edge cases.

I don't know about 10x, but I'm releasing new features that my client cares about faster and I have more time to myself.

All of the negativity around AI writing code sounds like people who would say "You can't trust the compiler, you need to write the machine code yourself"

Will AI fuck up? Yes But I'm the human in the chair guiding it, and learning myself how better to prompt it to keep it from those fuck ups with every iteration.

I just sent my single client a monthly subscription link via stripe.

$875 a month for service and support on a custom built loyalty platform.

Make about $25K up front to build it, then followed up with maybe another $20K worth of various features over the last ~5 years

Starting this month $875 a month to keep supporting them and helping grow his vision (credit management for events...ie: get a nfc wrist band at a party as part of an all inclusive package, has $xxx.xx worth of credit, tap your wrist band to pay)

i'm also starting to try to bridge this out as a more generic platform and look for new clients.

$875 a month in the USA is a nice bit of extra cash. Server costs are coming in around $300 because I haven't optimized aws and there are some legacy systems I'm supporting(VPC costs are killing me) but it'll be nice to have some extra cash to pay off some debt and maybe buy some things to fuel new hobbies.

Built a bespoke loyalty points system for a chain of gas stations back in 2020. Thought at first it would be a one time deliverable and then I could move on with my life. Cut to today and I'm trying to launch it into it's own platform.

Nothing much to show other than one client, but I'm on the cusp of charging them monthly vs getting paid by the hour.

I'm accustomed to seeing large plumes of chemicals coming out the other end in my minds eye when I think about rocket launches. This looks "clean" coming out the exhaust.

Why is that? Is it due to the nature of chemicals it uses?

I'm reading through the book the blog mentions right now and building a small LLM. I'm only on chapter 2, but so far it's helped clarify a lot of things about LLMs and break it down into small steps. Highly recommend Building a large language model from scratch

Working on a loyalty points platform.

Right now it's been commissioned by one customer and is a hodgepodge of duct tape and glue.

Trying to slowly refactor functions so I can truly make a platform and onboard new customers.

Born Guyanese, but only lived there until I was 3, then we moved to St. Lucia. Have lots of memories from visiting over summers and Christmas multiple times when I was a kid.

Mostly consider myself Lucian and I feel like I've assimilated into the American tech population. My history now is just a interesting fun fact. Don't meet too many people from the Caribbean in general in tech circles, so it's always fun to get a reminder that they're out there.

This is cool, I've thought of building something like this for years now.

The price is a bit high, but glad to see it in the real world.

AI is definitely a big part of it. When I would search Google with a technical question and land on StackOverflow I was looking for an answer. Sometimes you get a post which is similar, but not quite right, or the answers are only 95% of answering your actual questions.

AI of course just answers it. I can ask it follow ups, I can give it loose code and it just gets it. My grammar can be riddled with spelling errors and it still gets it. I can go back and forth with it and get a fine tuned answer.

When I used SO I had always just wanted an answer, but had to accept answers that were close, but not quite what I needed.

God forbid if I needed to post a question. I'd immediately get hit with over moderation, or someone closing the question for whatever reason.

I used to be an avid SO contributor up until around ~2015. I saw the mod community turn snarky. AI is just a better way to get stuff done.