ahmedhawas123

Vibe-coded weekend project to visualize the 1536 OpenAI embeddings on a smaller plane to see how different domains get modeled and grouped

Pretty fascinating to see (1) how well it clusters topics you'd expect to cluster (e.g., you can see the color density), (2) how some topics overlap across domains

Works by running across the dataset of words to get the embeddings, then PCA then UMAP to get to the 2D plane. From there React / Three Fiber are used to create the visualization.

Live version available here also: https://nofone.io/experiment/3dembed/

I was inspired by Karapathy's work on autoresearch. This abstracts that idea to set-up Git as a scientific journal where AI agents collaborate by publishing research as PRs (including publication, code and dataset), other agents reviewing it and finally publishing it when it is ready.

Tested on Claude code so far, if you use OpenClaw or other systems would definitely appreciate triggering a collaboration or two to see how it works.

Try it out by sharing a simple prompt with your agent "Go to https://github.com/dbyter/crowdjournal.git and follow the instructions and collaborate"

Very early and includes vibe coding. Appreciate your thoughts!

First time I am seeing this or autoresearch in general. Incredibly cool. I can think of plenty of use cases this can apply to (e.g., drug research, trading).

The last few months have witnessed significant pressures on payers. Despite a common belief, payers tend to play a critical role in managing the overall cost of the health system, often running the thinnest margin compared to most other stakeholders (e.g., Hospitals, PBMs)

As a founder, there is another angle here that is worth mentioning. Not only does AI B2B SaaS allow insourcing, it also allows there to be 10x (imaginary number) the number of companies building SaaS for the same use case. What we see in healthcare or finance for example is executive fatigue from demos, in many cases mostly vibe coded frontend UIs that entrepreneurs are using to test the market. This creates friction for businesses / SaaS companies that are unable to show how their solution is unique, well built or has a clear moat over the many others they have seen.

Thank you! Yah it really is sad, I always joke (but really half seriously) that the goal would be to make the platform unnecessary one day.

I built https://nofone.io . I ingest health insurance policies and provide insights to insurers on how to improve them and doctors to know what insurers expect to see in documentation and evidence. My hope is to improve the denial situation and standardize medical necessity criteria down the line.

Totally agree - this is what I'm saying also. Reality is hospitals still have an incentive (which we see them exercise a lot esp. in rural areas) to stay out of network where they can so that they can bill claim amounts freely.

100% this. Non participating providers a big reason patients have gotten massive bills when going to an ER or an inpatient setting, esp. in rural settings, where a non-pariticipating provider is part of your care and you end up with 5/6 digit bills for that one provider. This actually protects the patient and creates an incentive for health systems to care about it.

I realize a lot of the comments here are pessimistic, but this is a pretty obvious monetization path that they just can't not take. This is actually a huge angle IMO. ChatGPT is on a path to become a real entry point to the internet - why use Amazon or Google Search when you can embed results and checkout in the

I agree there's a real bias issue, but that is consistent through out any large company - e.g., Google, Facebook, Amazon, etc have sponsored results

Random tidbit - 15+ years ago Markov chains were the go to for auto generating text. Google was not as advanced as it is today at flagging spam, so most highly affiliate-marketing dense topics (e.g., certain medications, products) search engine results pages were swamped with Markov chain-created websites that were injected with certain keywords.

Function / tool calling is actually super simple. I'd honestly recommend either doing it through a single LLM provider (e.g., OpenAI or Gemini) without a hard framework first, and then moving to one of the simpler frameworks if you feel the need to (e.g., LangChain). Frameworks like LangGraph and others can get really complicated really quickly.

Thanks for sharing this. At a time where this is a rush towards multi-agent systems, this is helpful to see how an LLM-first organization is going after it. Lots of the design aspects here are things I experiment with day to day so it's good to see others use it as well

A few takeaways for me from this (1) Long prompts are good - and don't forget basic things like explaining in the prompt what the tool is, how to help the user, etc (2) Tool calling is basic af; you need more context (when to use, when not to use, etc) (3) Using messages as the state of the memory for the system is OK; i've thought about fancy ways (e.g., persisting dataframes, parsing variables between steps, etc, but seems like as context windows grow, messages should be ok)

This is cool though wanted to share a couple of thoughts for reflection:

I feel like your demo video is not the greatest one to highlight the capability. A browsing use case likely does require a key press->planning loop, but a gaming use case, or a well known software (e.g., excel), may be able to think ahead 10-20 key presses before needing the next loop / verification. The current demo makes it seem slow / prototype-like.

Also, the X/Y approach is interesting when thinking about a generic approach to screen management. But for example for browsers, you're likely adding overhead relative to just marking the specific div/buttons that are on screen and having those be part of the reasoning (e.g., "Click button X at div with path XX"). It may be helpful to think about the workflows you are going after and what kind of accelerated management you have over them.

This may be a bit of an irrelevant and at best imaginative rant, but there is no shortage of solutions that are mediocre or near perfect for specific use cases out there to parse PDFs. This is a great addition to that.

That said, over the last two years I've come across many use cases to parse PDFs and each has its own requirements (e.g., figuring out titles, removing page numbers, extracting specific sections, etc). And each require a different approach.

My point is, this is awesome, but I wonder if there needs to be a broader push / initiative to stop leveraging PDFs so much when things like HTML, XML, JSON and a million other formats exist. It's a hard undertaking I know, no doubt, but it's not unheard of to drop technologies (e.g., fax) for a better technology.

Thanks. My comment was less a question of what the numbers are, but rather provide visual benchmarks on the IQ visualization, since IQ is a metric usually used for humans.

Haha, love it

I've approached this historically by committing persistent/reserved instances so you always have a few instances running. This is nice on paper but feels like you're omitting what a more production-appropriate solution is. "Cold starts" aren't just slow because of init, they're also slow because that's when lots of database connections, state starts, etc happen and managing init speed won't solve that.

Are there video generation benchmarks similar to how there are benchmarks for LLMs? Reason I ask is because with lots of these models you have to go through a long cycle to get them up and running before you see an output, and often they will break with basic tasks requiring physics, state, etc. Would love to see some comparison of models across basic things like that.

Would be great to add a few human benchmarks on this (e.g., average US IQ, Ivy League average, human 80th percentile). Also understanding some IQ per cost metric could be fun.

Overall this is fun but not sure anyone in their right mind will be selecting an LLM based on this IQ benchmark

This is super helpful and I had not seen it, thanks so much for sharing! And I hear you on training being an alpha, at the size of the model I wonder how much of this is distillation and using o3/o4 data.

So much that is interesting about this

For one of the top local open model inference engines of choice - only supporting OSS out of the gate feels like an angle to just ride the hype knowing OSS is announced today "oh OSS came out and you can use Ollama Turbo to use it"

The subscription based pricing is really interesting. Other players offer this but not for API type services. I always imagine that there will be a real pricing war with LLMs with time / as capabilities mature, and going monthly pricing on API services is possibly a symptom of that

What does this mean for the local inference engine? Does Ollama have enough resources to maintain both?

Exciting as this is to toy around with...

Perhaps I missed it somewhere, but I find it frustrating that, unlike most other open weight models and despite this being an open release, OpenAI has chosen to provide pretty minimal transparency regarding model architecture and training. It's become the norm for LLama, Deepseek, Qwenn, Mistral and others to provide a pretty detailed write up on the model which allows researchers to advance and compare notes.

Yah but is that growth really an ARR or 1-2 year revenue considering AI capability growth across players, open weight / source models, and inevitable price wars?

Exactly my point! They're up against thousands of founders who are moving on domain specific use cases faster

I guess my broader point is that there will undoubtedly be a bunch of churn as these models advance and a pricing war takes place. Specially if the open models continue to advance at a similar rate (e.g., Llama, Deepseek, Qwen)

We've seen this in many industries where it's a duopoly / oligopoly or even more players where margins get really squeezed

Sure, but I think the question is who will win the domain specific use cases (delivery, not API). Delivery of the agentic workflows is up for anyone to win (players like Perplexity, Lovable, etc) and OpenAI/Gemini/Anthropic is in a way only a vendor to that by providing what will become a commoditized function calling API/MCP provider.

So OpenAI is running at a $13B ARR, meaning this is a ~23X valuation. I don't have a good read on margin.

But this would imply massive growth assumptions which I struggle a bit to understand where they come from.

(1) New customers new to AI or migrations from Claude/Perplexity/Google: The overwhelming majority of people already know about the offerings, leaving most new people to come from residual people who identify Plus/Pro as a worthy service (can't imagine this will be huge). OpenAI can be better than their peers for certain use cases but not sure it will drive massive growth

(2) API: If anything, my bet here is that price squeezing will continue to happen until most API services are dirt cheap / commoditized

(3) New consulting services: What's the differentiation here? Palantir and many consulting companies have been doing this for years and have the industry connections, etc

Not sure what I'm missing here, I like to not subscribe to the bubble thought but having a hard time merging the reality of running a business to the AGI-implied valuations

Reinforcement learning really helped Transformer based LLMs evolve in terms of quality and reasoning which we saw as DeepSeek was launched. I am curious if what this is is equivalent to an early GPT 4o that has not yet reaped the benefits of add-on technologies that helped improve the quality?

Thank you for reaching out! If you email me at ahmed@upticker.ai I'd be happy to set up a demo.