threeseed

You are trying to control a system that is inherently chaotic.

You can probably get some where by indeed running a task 1000 times and looking for outliers in the execution time or token count. But that is of minimal use and anything more advanced than that is akin to water divining.

> an mcp server that lets the agent raise its hand instead of hallucinating

a) It doesn't know when it's hallucinating.

b) It can't provide you with any accurate confidence score for any answer.

c) Your library is still useful but any claim that you can make solutions more robust is a lie. Probably good enough to get into YC / raise VC though.

> Currently, we do not offer protections against man-in-the-middle attacks. As a result, if someone—a malicious insider or X itself as a result of a compulsory legal process—were to compromise an encrypted conversation

I assume this means that the "encryption" is about as strong as base64.

But it’s not just about coding quickly but also correctly.

Coding LLMs do not solve the problem of it hallucinating, using antiquated libraries and technologies and screwing up large code bases because of the limited context size.

Given a well architected component library and set of modules I would bet that on average I could build a correct website faster.

a) Big data is more than just the size of the data. It's about how you treat that data i.e. instead of doing expensive and brittle up-front RDBMS modelling you instead dump it all into a data lake and figure out how to handle the data at run-time. And it is still the standard pattern in almost all companies today.

b) Nobody was choosing MongoDB solely for performance. If it was you would choose some in-memory K/V store. It was about it being the only well supported document store that was also fast and scalable.

c) Stripe’s DocDB is an extension of MongoDB Community: https://stripe.com/blog/how-stripes-document-databases-suppo...

MongoDB as a company is growing 20% y/y and 2B in revenue.

So very far from being a legacy product.

I still use it for new projects because (a) Atlas is genuinely a solid offering with a great price point for startups and (b) schema-less datastores have become more of a necessity as our control of data has decreased e.g. SaaS companies dictate their schema and we need to accomodate.

People used to tell me all the amazing things no-code and low-code was able to do as well.

And at the end of the day they went nowhere. Because (a) they will never be perfect for every use and (b) they abstract you from understanding the problem and solution. So often it will be easier to just write the code from scratch.

> The speed at which LLM stuff is progressing is insane

You clearly haven't been following the space or maybe following too much.

Because the progress has been pretty slow over the last years.

Yes modals are cheaper and faster but they aren't substantially better.

They never said it was useless. You just invented that straw man in your head.

3D printing is largely used for prototyping where its lossy output is fine. But using it for production use cases requires fine tuning it can be 99.9% reliable. Unfortunately we can't do that for LLMs hence why it's still only suitable for prototyping.

People said the same about VB style coding then low-code and now AI.

They have been wrong every time and will continue to be wrong.

I have worked for a dozen companies all earnt more than $20b a year in revenue. That includes two banks and a hedge fund. All use the cloud.

You must be living under a rock if you think the cloud isn't secure enough for the enterprise.

> Apple

I can't imagine Apple being interested.

Their priority is figuring out how to optimise Apple Silicon for LLM inference so it can be used in laptops, phones and data centres.

Cost is irrelevant when compared to the salaries of the people using them so they will do basic cost controls but nothing too onerous. And cost is never a reason to prevent solutions being built and deployed.

And most enterprises aren't even doing anything advanced with AI. Just doing POCs with chat bots (again) which will likely fail (again). Or trying to do enterprise search engines which are pointless because most content is isolated per team. Or a few OCR projects which is pretty boring and underwhelming.

Only an insignificant minority of companies are running their own AI LLM models.

Everyone else is perfectly fine using whatever Azure, GCP etc provide. Enterprise companies don't need to be the fastest or have the best user experience. They need to be secure, trusted and reliable. And you get that by using cloud offerings by default and only going third party when there is a serious need.

Yes. And now they have a massive vested interest in drawing startups into the space.

> YC doesn't push founders to start a specific type of business

Of course it does.

Founders look at YC batches, see that it is 99% AI companies and are then forced to also go in that direction if they want the benefits of the accelerated YC path.

And YC deliberately chooses founders with AI companies because they have an investment thesis that is different from "request for startups". Garry Tan has been a massive e/acc fanboy since the beginning and genuinely believes that AI in every use case will advance humanity. And the partners all align with this.

This is all inarguable because amongst the tens of thousands of applications there are surely many amazing non-AI companies. Is this implication that they are all worse than what was selected in the batch ?

Internet only became a fad once it was already large and had tens of millions of users.

I remember the pre-Web days of Usenet and BBS and no one thought those were trendy.

AI is far more akin to crypto.

Those large number of outdated dependencies are in the LLM "index" which can't be rapidly refreshed because of the training costs.

MCP would allow it to instead get this information at run-time from language servers, dependency repositories etc. But it hasn't proven to be effective.

Yes. The top performers in each batch have offers before demo day and some won't even present.

There is a secret funnel from YC to a select group of top tier VCs.

a) The next two years are the reckoning for a lot of these AI startups. In the enterprise space everyone was being pushed to trial AI products to see whether they can deliver the ROI that was being marketed. Newsflash: it hasn't. And many of these startups will see serious churn.

b) Everyone needs to stop perpetuating the YC lie that they invest in the best founders and they just happen to want to do AI. It's rubbish and insulting because it implies that only young, male, SF-based founders can be the best. Instead it's clear that YC has been aggressively pushing AI which makes sense given they are a significant investor in OpenAI.

> I don’t use RAG, and have no doubt the infrastructure for integrating AI into a large codebase has improved

It really hasn't.

The problem is that a GenAI system needs to not only understand the large codebase but also the latest stable version of every transitive dependency it depends on. Which is typically in the order of hundreds or thousands.

Having it build a component with 10 year old, deprecated, CVE-riddled libraries is of limited use especially when libraries tend to be upgraded in interconnected waves. And so that component will likely not even work anyway.

I was assured that MCP was going to solve all of this but nope.

How are we in 0.1 of GenAI ? It's been developed for nearly a decade now.

And each successive model that has been released has done nothing to fundamentally change the use cases that the technology can be applied to i.e. those which are tolerant of a large percentage of incoherent mistakes. Which isn't all that many.

So you can keep your 10x better and 100x cheaper models because they are of limited usefulness let alone being a turning point for anything.

Because lack of transparency in pricing affects competition.

Which is bad for consumers and the broader economy.

a) At least with Github Actions it is trivial to support polyrepos. At my company we have thousands of repositories which we can easily handle because we can sync templated CI/CD workflows from a shared repository to any number of downstream ones.

b) When you are browsing through repositories you see a description, tags, technologies used, contributors, number of commits, releases etc. Massive difference in discovery versus a directory.

> For both vibe-coded and enterprise codebases

What in god’s name does this even mean ?

> if they're going to go into consumer hardware that's going to need to be an all-consuming strategic pivot

You have to build a phone.

There is no other way to get the data you need to make XR glasses, AI pebble, Rabbit etc work the way people expect without it. Because Apple and Google are well within the rights to deny the siphoning of your private data to a company like OpenAI who only exists because of large scale trademark abuse.

OpenAI is becoming like Tesla.

Their first mover advantage is gone, they are unable to innovate on their core product and everything that is said or done is hyped to the extreme.

And now trying to move into adjacent categories where there is no clear problem to be solved and is putting them up against the biggest players in the industry.

Because the platform doesn't need to expose the details of buyer/seller.

Most tickets these days are digital.

There's a UK show called Hunted where contestants go on the run and are chased by law enforcement.

It is amazing to see just how quickly they are found because of the all the CCTV cameras.

Except that in my experience half the time it will modify the implementation in order to make the tests pass.

And it will do this no matter how many prompts you try or you forcefully you ask it.