This is right, but only the first step in the right direction.
If agents are handling discovery and the purchasing path we may also see a lift in quality and value of our purchases across the board. Agents will want to find the best option, not the best marketed or trending option.
I say “may” because this assumes the agents won’t get manipulated like search/social was. This is surely part of the end game for the big personal assistants like Instinct etc. - they aggregate demand and then control the flow of commerce, just like paid ads have done.
for a long time it felt like the world was fully optimized against you
you go to buy something, you get a bunch of ads and upsells and micro optimizations to get you to spend more
you go to consume some specific content and the algo puts in things that are hard to ignore and designed to get you to spend more time
agents are the first time i've felt like maybe things can go in a different direction. send your robot into that mess and it can ignore it all and do your task
i'm not naive, this is going to be an insane battle with entrenched interests doing their best to fight it
but there's a small chance that in the end we can finally be free of this (until the next new problem)
This is an important thing to understand about the models. They are not just predicting the next word. This is what we built them to do, but it's not all they do. They have a plan beyond the next word and context of the universe. This is how humans think. By designing models to predict the next word we actually built thinking machines.
Don't be fooled by the "just predicting the next word" narrative out there.
"To make them is not to understand them."
Emmanuel Ameisen (@mlpowered) has spent the last 3 years doing interpretability research at @AnthropicAI, looking inside LLMs like Claude to figure out what they're actually doing. As he puts it, models are "grown," not built.
We discuss:
🔸 His discovery that models plan many tokens ahead
🔸 Why anthropomorphizing AI can actually be the right approach
🔸 What the "J-space" reveals about how models think
🔸 How interpretability connects to AI alignment and safety
Listen now 👇
https://t.co/qEYdKfpkab
In the age of agents, the whole company lives in the repo.
Marketing site. Docs. Changelog. Wiki. Decisions. All markdown. All in the same diff.
One agent can read, build, and compound across the whole thing.
https://t.co/A97K4OhOtO
Everyone yelling about the DB not in the same data center is technically correct, but completely wrong from a founder simplicity perspective. This is not the problem to solve today.
Also, I’ve seen this play out in the real world with read heavy apps with 100+ requests per second. One time we accidentally provisioned a DB in the wrong region, a neighboring region. We saw DB queries to that instance up ~250ms. No good.
When they are in the same region but not same data center we are seeing sub 20ms increases. Just not worth the squeeze in many cases. Id much rather have the DX of @PlanetScale or @neondatabase.
@Shpigford I noticed you are using Rails and Inertia in your skills. How are you liking this stack?
I have a large codebase that is Rails, Hotwire, ViewComponent. It has a minimal frontend right now, as it's mainly an API.
We are considering expanding the frontend, but the current stack doesn't feel well suited for our AI friends.
@inertiajs@rails@evilmartians
@levie The companies that figure out how to productize pieces of this process will mint.
We need layers of abstraction and products to own pieces of it. It’s still way too much for most companies to implement, let alone maintain, with or without outside help.
I met with a mid-market manufacturing company yesterday. They make hinges and casters for automotive and industrial.
The CEO explained the RFQ process and it was clear the whole thing should be taken over by AI. Currently it’s 5 CSRs and a team of engineers manually responding to each request. They only win 15% of them this way, and each one can take a day to multiple weeks to respond to.
I’m urging them to just start building. It could be the single most impactful thing they do this year.
I met with a mid-market manufacturing company yesterday. They make hinges and casters for automotive and industrial.
The CEO explained the RFQ process and it was clear the whole thing should be taken over by AI. Currently it’s 5 CSRs and a team of engineers manually responding to each request. They only win 15% of them this way, and each one can take a day to multiple weeks to respond to.
I’m urging them to just start building. It could be the single most impactful thing they do this year.
@Shpigford This looks cool. I just started using Obsidian for my personal knowledge base. I don't touch it, but Claude organizes it for me.
Why should I switch to Clearly from Obsidian?
I wonder how much of this is capacity related (actual downtime in this case) vs. internal errors being raised somewhere down the stack (could also be capacity related, but not on the initial request). The internal errors aren't getting picked up by an uptime status monitoring tool.