crosen99

I got off a quick GPT trying to explore the following theme, and I think the results are promising:

In the ever-evolving landscape of creative expression, a new horizon emerges: AI agents. Imagine authors and publishers releasing these agents as they do books or video games, creating interactive narratives and personalized adventures. This is more than just storytelling; it’s a revolution in how we experience and engage with content. AI agents could soon become our guides, companions, and co-conspirators in worlds crafted by human creativity and AI innovation. Are we ready to embark on this new journey? The next chapter in storytelling is being written, and it’s one where the reader becomes a part of the tale.

Well said, and fully agree. If you horse race two approaches, you can of course arbitrarily arrive at a winner based solely on the version of each approach you choose. You need a deeper look if you want to generalize.

They stumbled into a position where they can make a crap ton of money going up the stack, which can fund the ongoing march toward AGI. (The revenue not only is cash in their pocket, but it’s also driving up their evaluation for future investment.)

we've heard that when people were using chains/agents they often wanted to see what exactly was going on inside, or change it in someway.

I certainly agree, but I'm having trouble seeing how templates help with this. The templates appear to be a consolidation of examples like those that were already emphasized in the current documentation. This is nice to have, but what does it do to elucidate the inner workings?

The biggest challenge I’m trying to track isn’t on the list: online learning. The difficulties with getting LLMs to absorb new knowledge without catastrophic forgetting is a key factor making us so reliant on techniques like retrieval augmented generation. While RAG is very powerful, it’s only as good as the information retrieval step and context size, which quite often aren’t good enough.

You might wonder if these harbingers are actually tuned into an alternate and consistent set of values not shared by the majority, rather than simply irrational consumers. In that same vein, you might wonder if the product managers of these failed products are part of that same cohort.

Same here. I think it's because you typically conquer a language by developing a fairly complete mental model of that language's behavior, but with CSS there are just too many nooks and crannies to get your had around. While it's uncomfortable for me, I find that I do best with CSS when I take a more practical approach and just try to use best practices to accomplish what I want without having the deep understanding that I crave.

I've found that Kevin Powell [1] is a great resource for this approach, and he also helps with the model model aspect as a bonus.

[1] https://www.kevinpowell.co/

I keep going back to LangChain thinking it just hasn't found its legs yet, but every time I do I retreat exasperated. I don't find their abstractions useful or intuitive, and their documentation is woefully scattered and incomplete. Things are moving so quickly with LLMs that theirs is no easy task, but so far they haven't really cracked the nut of making LLM app development easier.

Strong disagree. The way I see it, any successful customer service operation needs to have three things in place:

1. Properly defined and articulated customer service policies and procedures

2. A training program that adequately prepares agents to know, understand and apply those policies and procedures

3. Sufficient inherent capability and disposition of the customer service agents to apply that training

My take is that the (relevant) inherent capabilities of LLMs at even at this early stage are as strong or stronger than what I see in the human customer service agents I typically deal with. And the disposition of an LLM can be arbitrarily good through fine-tuning and prompt engineering.

As such, any company that gets 1. and 2. right has a high chance of providing a better support experience by replacing human agents with LLMs.

AIs, at their current level of development, don’t perceive objects in the way that we do – they understand commonly occurring patterns.

You see this claim everywhere - that AI operates on statistics and patterns and not actual understanding. But human understanding is entirely about statistics and patterns. When a human sees a collection of particles and recognizes it as, say, a car, all they are doing is recognizing the car-like patterns in how the particles are organized that have a strong statistical correlation with prior observations of things classified as a car. Am I missing something?

Not substacks, but I start my mornings reading the TLDR and the TLDR AI newsletters. Lots of overlap with the sort of tech that HN covers. TLDR AI is very useful to keep your finger on the rapidly changing pulse in the world of ML.

This “new technique called implicit code execution” sounds a lot like an early version of the ChatGPT Code Interpreter plug-in.

I wanted to give this a whirl by creating an event triggered workflow, but I got stuck pretty quickly.

I had imagined creating a workflow that would be triggered by the occurrence of some external event, such as a row being added to a Google Sheet. But after struggling with the UX and documentation for a bit, I think I finally worked out that this is not consistent with Fastgen's concept of an event triggered workflow, and that these workflows can actually only be triggered within an API call or workflow that occurs within my Fastgen environment.

Is that correct? If so, I don't believe it's clear or obvious, so perhaps it makes sense to spell this all out more clearly in the UX and documentation. Also, the sort of event trigger I had in mind is an essential aspect of many similar low code tools (i.e. make.com, bubble.io), so it may be worth considering adding this functionality unless you want Fastgen to be all about inbound API calls.

"Radiologists are not performing 2d pattern recognition - they have a 3d world model of the brain and its physical dynamics in their head. The motion and behavior of their brain to various traumas informs their prediction of hemorrhage determination."

Radiologists are certainly performing 2d pattern recognition, as the input they are processing is only 2d even if the details of how that recognition is performed relies on some deeper understanding. Likewise, an AI system performs recognition of 2d patterns based on some deeper "understanding" - in the case of a neural net that "understanding" lies in the complex configuration and weightings of its neural connections built up from vast training datasets.

Even if this dataset is only a subset of the dataset that humans are trained on, we still can't a priori claim that this subset lacks patterns and correlations that escape humans and that allow an AI to make certain determinations better than a human might.

That article links to the below article where Musk lays out his version of the simulation argument. His version is fraught to the brim with confused thinking and non sequiturs, and yet I’ve never seen him called out on it.

https://www.businessinsider.com/the-question-elon-musk-refus...

I'm surprised not to see anything about data-to-parameter ratios for optimal scaling. My superficial understanding per the Chinchilla paper is to target 20 to 1.

I'm also confused about this:

~$1 million: Cost to train a 13 billion parameter model on 1.4 trillion tokens

This is apparently related to the LLaMa paper, but that paper seems to cite 1.0T tokens (rather than 1.4T tokens) for the 13B model. Also, if 20 to 1 is in fact optimal for the data-to-parameter ratio, then using a 100 to 1 ratio doesn't seem like an appropriate way to arrive at a magic number for training costs. The magic number should really be based on an optimal configuration. Or, perhaps, my superficial understanding here leads me to miss some important distinctions.

Could you link to your research and/or describe the models, libraries, data and tests you used for this?

It's amazing how much misinformation and vague information there is on this topic. I tried getting to the bottom of this in the following post in the OpenAI forum:

https://community.openai.com/t/fine-tuning-myths-openai-docu...

Bottom line is that fine-tuning does not seem to be a feasible option for adding new knowledge to a model for question answering.

LLM based autonomous agents remind me a lot of Leonard from the movie Memento. In the movie, Leonard is trying desperately to find the murderer of his wife, but he has one big problem - he can no longer build new memories. So, he needs to develop a system for storing new knowledge and then retrieving the relevant bits every time he formulates a plan for what to do next. Throughput the movie, Leonard makes a number of missteps because his system for recording and retrieving knowledge from this external memory system is imperfect. The movie is very well done.

You'd have to be naïve to deny that AI poses a risk that must be taken seriously. But, to think that a moratorium - which would be ignored by those with the greatest likelihood of causing or allowing AI to do harm - is the right answer, seems plainly silly. It's not surprising that several of the signers have a personal stake in attacking the current efforts.

The letter should instead simply have clarified the risks and outlined a sensible, actionable plan to promote responsible development alongside the inevitable march of progress.

In addition to the choices for how to chunk (i.e. defining chunk size, chunk boundaries, chunk overlap, etc.), there's also the question of what actually gets returned once finding the chunks that match. For example, perhaps I have a document with 100 1-page sections where each section is broken into roughly 5 chunks. I may get optimal performance in my RAG application not by retrieving the top K chunks from the index, but rather by returning the top K sections fom the document, where sections might be scored based on the number and scores of child chunks. It also might be useful to incorporate section summaries, etc., in the retrieval process.

The problem I'd like solved is that when I want to retrieve chunks of data for retrieval augmented generation, it's challenging to optimize the choice of embeddings model, chunking strategy, and overall retrieval algorithm. I'm not sure if that's the sort of problem you're focused on.

This sounds less like Embeddings as a Service and more like Semantic Search (which happens to be using embeddings) as a Service.

I learned long ago that if the focus of my obsessing is on some future state or accomplishment then I will be anxious, miserable and not at all creative or productive. But if I can direct the obsessing to actual activities I enjoy in the here and now, I find my sweet spot.

I forget this constantly, and have to bring myself back to it repeatedly. But whenever I do, I find I’m home again.

Fine-tuning these models reminds me of the good ol' days with tube TVs where the slightest twist of the vertical hold dial meant the difference between a clear picture and useless, dizzying, visual nonsense.

The tensions outlined in this article are very hard to navigate. In some sense it comes down to awareness and handling of "what you don't know you don't know". That skill is arguably a requirement to be a great leader, but it is hard to teach and may largely come down to having a certain inherent capability.

ChatGPT quickly and mightily reshaped my brain. In the past, I would have thought to type keywords into the search box, which is what's expected in this case. Instead, my immediate reaction was to ask natural language questions, which this is app is of course not geared for.

A big thing for me is going simultaneously top-down and bottom-up.

Top-down establishes the high level concepts, underlying motivations, etc. It provides a map of what there is to learn and a scaffolding upon which all new knowledge and concepts can be secured and contextualized.

Bottom-up is where the rubber hits the road. Because disrupting a system and seeing the effects of that disruption is key to learning, bottom-up is a very active and iterative process. Write the “Hello, world!” program, and then see what happens when you change the text, remove the semicolon at the end of the line, etc.

Gradient descent is just how neural networks (including auto-encoders) optimize parameters to minimize the loss function. They do this using derivatives to descend down the slope of the function. Autodiff is one way to compute the derivatives. Maybe we’re saying the same thing.

I don’t believe autodiff is finding the difference in that sense. It’s finding derivatives.