> when you put these models in a loop, you are able to do more and more insane and cool things...
Please share some of these insane things that you speak of..
Please share some of these insane things that you speak of..
use your imagination to solve problems people face and pay $$$ for today that is error prone and hard.
i’ve got agentic workflows for the particular industry im building for, one of which that replaces the need to hire $500+/hr services.
in this particular workflow (don’t want to reveal too much, sorry this is my competitive advantage but you can figure it out for your own workflows) a $4/1M model ingests a file that is currently used in a extremely complicated program that few people understand how to use.
it parses the data, loads it into a database, and then spawns a bunch of other agents that check the data against work in flight. there’s checks for bad data. in that case, more agents are spawned that reach out to the involved people or parties for clarification. if it cannot figure something out it reaches out to the right contacts for more information. while this is happening, more agents begin doing work that involves continuous reconciliation against 100s or 1000s or more things in flight.
as files are uploaded, or updates from people come in, agents do work to ensure things remain on track.
people are able to work across languages and cultures, and my agents ensure that while people can make mistakes, it will catch them in real time and ensure continuously monitor the situation.
it’s pretty nuts how much inefficiency agents today can solve. it takes patience to run tests and tweak shit until it works.
*** the really cool thing is that more capable agents can continuously monitor how things are going and improve the workflow itself… so all i need to do is maintain the actual tests. ****
i loved writing tests back in the day to ensure i built good software. today we write tests to ensure the business can run.
I could not think of a worse technology to use for an ETL pipeline than throwing LLMs at it and asking it to vibe out the correctness of the data every time it runs.
98% of people think AI means what gemini tells them when they do a google search.
of that 2% who go beyond... maybe 20% of those are using AI to code.
most SWE still think "using AI" to code means the copilot pane they open on the side in VSC. they get sloppy code output and think "ai sucks!"
so of that 20%, maybe 5% have actually explored what an "agentic" workflow even means. they might use skills, set up their code base so AI almost NEVER makes a mistake. maybe 1-5% of that 5% actually went deeper, and those are the people who built tools like Cursor or Harvey or whatever other "agentic" companies.
you are still thinking about the old world where you obsess and define data models and bike shed over data integrity, all the while you have a "temporary" table with 3 attributes that gets 2 million queries per second that's now holding up a bunch of other shit that's also glued together.
> I could not think of a worse technology to use for an ETL pipeline than throwing LLMs at it and asking it to vibe out the correctness of the data every time it runs.
like i said, if you are still having quality issues in 2026, that's a skill gap.
also, agents now continually improve the process.
for a business, the only thing that matters is transaction log. for 99% of businesses, swe are a cost.
again this is like baby steps on the journey. we are still only 3 years into this technology being opened up to masses.
i'm sure people thought computers were dumb, or that cars are stupid because the first cars were moving slow af. "we have horses, why do we need to build out roads to get anywhere"
but you're free to feel smart doing 20/20 hindsight on things 100+ years from the future
> of that 2% who go beyond... maybe 20% of those are using AI to code.
> so of that 20%, maybe 5% have... maybe 1-5% of that 5% actually went deeper
How can you write any of this when you said you just started liking LLMs with Astra, a model that released two weeks ago. You're whipping out a bunch of made up stats trying to describe large swaths of programmers while admittedly being unfamiliar with the capabilities until two weeks ago? Why are you so antagonistic towards people's skepticism if you used them for years and found no value in them? This sounds more like a comment trying to generate massive amounts of FOMO. "If you REALLY put a lot of your money into them, that's when they shine! Don't believe the haters!"
> like i said, if you are still having quality issues in 2026, that's a skill gap.
> also, agents now continually improve the process.
> for a business, the only thing that matters is transaction log. for 99% of businesses, swe are a cost.
Maybe you should try building something for more than two weeks before making such big declarations, because this really does read like someone who is excited about a novel they wrote last night at 3 am
been building stuff for years, so i'm not just some w2/1099 type guy who bike sheds for a living lol
this stuff is new... obviously it is a fast moving space.
already been through one IPO where i was an early employee so i’ve never really had to work with morons.
So what ever shit the "AI" gets wrong, it is the user's fault, right?
If there is no such guarantee, why should I spend time writing an elaborate skill file? There is no telling when the model chose to ignore stuff in it.
I don't understand how people can work with something like that!
when you have chains of agents working seamlessly, where agents are managed by themselves (like spawning a new chain to do some scope of work), and it just sort of works on a infinite game loop, your system continually keeps improving as it works. at least, that is the goal with the systems i like to build.
you have to do your due diligence, obviously, as you would with any TOOL you use. in that case it means having specific goals, methodologies, etc. that each model has to follow. (if one is a "hey grok explain this" type of person when it comes to using ai, ngmi)
we have one agentic workflow where there are 4 different models that can be spawned (to handle tasks of various complexity), that work against an API (the source of truth). their goal is to work any time a specific file is uploaded, and handle it. it is a complicated file, with 1000s of line items, with varying amounts of uncertainty involved.
it happens in the real world today, where it costs $xxx,xxx per year to do. because it involves many people and companies... it is done totally manually today..
how it works in order to get to that resolution is different based on the complexity of the task... sometimes there's bad data in the mix because people make mistakes when they create stuff (we're not special). in this case it requires finding out whether it is indeed bad data or actually correct. that requires setting up a scheduled job, and handling it when there is a resolution.
i don't code this part. how this gets accomplished... the agents are able to work autonomously to handle any edge case, in order to complete the goal. the goal in this case is to ultimately record transactions (a goal of a business is to make money, believe it or not).
does this not make sense? jeez hacker news used to be an imaginative place.
the model that we pay $10-20 per 1M tokens will be $1 per 1M token next year.
And the model next year that we can pay $10-20 per 1M token will be even better than Astra and Terra/Sol/Luna. It is an exciting time.
The non-sense is the part where you think putting agents in a loop is going to yield better results indefinitely...
life of a w2