mehulashah

For those of you from the AI world, this is the equivalent of the bitter lesson and DeWitts argument about database machines from the early 80s. That is, if you wait a bit with the exponential pace of Moores law (or modern equivalents), improvements in “general purpose” hardware will obviate DB specific improvements. The problem is that back in 2012, we had customers that wanted to query terabytes of logs for observability, or analyze adtech streams, etc. So, I feel like this is a pointless argument. If your data fit on an old MacBook Pro, sure you should’ve built for that.

I am excited to see Databricks turn into the next Oracle. This type of acquisition was inevitable. The king is dead! Long live the king!

And yes, congratulations to the Neon team! (Nikita is, after all, YC)

Love it, though if it’s sufficiently instructive, I bet it will go beyond education.

There’s something beautiful about not being riddled with previous artifacts and starting clean with how you imagine you want to build your system. If the system is large enough, you can’t do it that often.

First — 0.69% of total exports is blood still is surprisingly high! Remember, you can get blood from humans in your own country, and prepare it there. So, why do you need it from the US?

Second — it’s amazing the detail that you can achieve from public data.

Third — I’m left wondering if a true “Deep Research” like tool would be able to provide the same analysis. I find that Deep Research is fine for secondary sources, but not for Deep Analysis of primary source data.

I agree. As a sponsor, we don’t also do it any longer. The incentives are fake. You want people to be excited about using your tool, not using it in some contrived way to win a few bucks. When the use is real, the feedback is real. Otherwise, it’s the blind leading the blind.

The hackathons I remember as a college student were different. They were much more free form. You were asked to create something around a broad topic: eg graphics, operating systems, modeling a neuron. You were allowed and encouraged to go outside the box.

Getting the prize was fun, but a working artifact was much more satisfying. That’s why we did it.

So, yes, it’s too commercial these days.

There’s a larger problem here beyond careerization that is going unsaid. People are using these metrics as an assessment of their self worth. I’ve had top researchers point out that such and such publication was accepted in the industrial track vs research track, and how important it was to keep that flag in the bibliographical data, even well after publication in a top conference. Most papers research or industrial stay in obscurity. The truly novel ones will catch like wildfire. The metricization of research and academic standing is the underlying culprit.

itch?

I find that people often conflate search with analytics when discussing Deep Research. Deep Research is iterated search and tool use. And, no doubt, it’s remarkably good. Deep Analytics is like Deep Research in that it uses generative AI models to generate a plan, but LLMs operations and structured (tool use) are interleaved in database style query pipelines. This allows for the more precise counting and exhaustive search type use cases.

I’m sorry that this happened to him. There are innumerable ways to let go of people with dignity. This isn’t the first time I’ve heard of this treatment from Google. It will not be the last. Google is a great company with great people. Google is incompetent at handling layoffs. They could learn from banks on Wall Street. The banks do layoffs routinely and have a process that’s more empathetic than Google, and that’s not a high bar.

We should celebrate this for what it is: another brick in the wall that we’re building to achieve practical quantum computing systems.

I'm so confused!

I'm sympathetic to the sentiment - founding engineers take on much of the same effort and ups and downs as founders, the relative difference in stock is not commensurate with the work.

While I don't like VC math and we (I'm CEO and co-founder) at Aryn AI give more than 1% for founding engineers, numbers like 5% to 25% are really a reflection of the risk and not the work. Prior to an equity round, you can give that away and you're treating the founding engineers like minor founders. Post an equity round, there isn't enough left to give everyone that and still build a decent team.

After an equity round with decent funding for 2-3 years, the co-founders have convinced someone to part with lots of money with no guarantee of an exit. There are innumerable YC startups and well-known VC-backed companies that have been sold for parts. After an equity round, you are joining a startup with some runway that you would not have gotten on your own, and hopefully, working on an opportunity that is outsized. Otherwise, you shouldn't take it.

Outsized means that even if the founders are gazillionaires, you may be 1/20 of a gazillionaire, which is still a gazillionaire. Otherwise, it would have been better for everyone to have worked for some big tech firm anyways. It really is a zero to one game, and you don't get to do it often. So do it if its compelling and fun and the outsized reward is and has always been a lottery ticket.

(CEO of Aryn here: https://aryn.ai)

Nice post and response to the previous one.

It’s important to remember that the use cases for VLMs and document parsers are often different. VLMs definitely take a different approach than layout detection and OCR. They’re not mutually exclusive. VLMs are adaptable with prompting, eg please pull out the entries related to CapEx and summarize the contributions. Layout parsers and OCR are often used for indexing and document automation. Each will have their own place in an enterprise stack.

(Disclosure, CEO of Aryn (https://aryn.ai/) here)

Good post. VLM models are improving and Gemini 2.0 definitely changes the doc prep and ingestion pipeline across the board.

What we're finding as we work with enterprise customers:

1. Attribution is super important, and VLMs are there yet. Combining them with layout analysis makes for a winning combo.

2. VLMs are great at prompt-based extraction, but if you have document automation and you don't know where in tables you'll be searching or need to reproduce faithfully -- then precise table extraction is important.

3. VLMs will continue to get better, but the price points are a result of economies of scale that document parsing vendors don't get. On the flip side, document parsing vendors have deployment models that Gemini can't reach.

It seems that Beej built this guide to help his students. More of them are probably using Git and GitHub in anger than the other tools.

Building a new kitchen consumer device is nearly impossible. This is amazing. Congratulations!

I don't agree at all with this post. YC's mission is to help founders start companies and help them build sustainable businesses. Towards that they've made it super simple to get started and funded, and provide guiding principles and support in the form of advice and a strong network to help the startups survive. They were never about prestige -- that came as a byproduct.

Scaling this model is hard, however, and I think they may be running into scaling limits. There just may not be that many fast growing businesses (startups) every year. We won't know until years from now.

This advice is much more general than academic research. It applies to startups as well.

Put it on the resume. I'd hire you.

So I contradicted myself above. Aggregation is probably not monotonic -- the output of the aggregation is not a subset if the input is a subset.

Examples of monotonic computation that fit are traditional database operators: filters, projections, joins. Interestingly enough, there is some coordination needed for joins, i.e. the items that match need to be in the same place at some point. But, that's not the type of coordination that the paper is pushing back against. It's still possible to distribute the computation without worrying about ordering.

Examples on non-monotonic computation (i.e. that don't fit) include aggregations, in particular those that include time series with windows. The answer depends on what is in the window. So, if you take an average and only include a subset of the data in a window, the output will change when you add more data in the window. At least, that's how I interpret the monotonic requirement.

This is a cool result from a couple of famous database experts. Though, it seems that the only set of practical programs that satisfy the CALM property are set-based operations found in database systems. These are expressive, but not expressive enough to cover all possible computations. So, is it the case that all computations can be decomposed into a partial order of CALM programs each of which don’t need coordination except at the points of order. If so, is this approach analogous to Lamport’s CStructs in Generalized Consensus?

Yes. Based on the definition.

This restore my faith in positive human creativity. It makes me wonder what I can do to encourage the constructive collaboration among the younger generation. I agree that todays doom scrolling has destroyed so much of todays potential.

Yes, and I will fill in definitions of obscure functions. If its named well, ChatGPT often can figure out what its doing. (No idea how this works!) I'm sure there are more streamlined ways, but this works well enough for me.

I do believe that ChatGPT as a programmer productivity tool is its most valuable value proposition to date. In addition to generating code, it makes it much easier for me to chase down obscure error messages and potential root causes and workarounds. It's definitely not perfect. But, as part of the Interact->Generate->Verify cycle that modern AI has transformed other disciplines (eg. materials science, protein folding, mathematics, and more), it serves as a valuable component of each of these.

I wish we didn’t think of understanding math as being smart or dumb. Like, it was a singular ability. Math is a subject deep, wide, and rich of ideas as well as results. For me, learning and understanding new concepts and results is like discovering new mountains and trying to ascend them. I won’t be able to ascend most, but I can certainly sit back and appreciate and enjoy their magnificence.

> “The point is, all of this made me feel very small and powerless after I realized the sheer size of the problems I was staring at. Nowadays, to me it's about creating good jobs for my friends, helping our customers, and taking care of the community. Importantly, I realized that this is still making a bigger positive impact than what I could have done alone just as an individual hacker or engineer.”

I couldn’t have said it better. This is why I’m an entrepreneur and still take the pains to do it.

Classification is just a start. Wondering if it's worth doing something more -- like turning all of the text into Markdown or HTML? Would anyone find that interesting?