We're running out of features to spec!
...ion debt of product teams is going down as we can burn through the backlog with LLMs. Engineers were the bottleneck so PMs had time to plan and think. [bp] But not any more!
Topic: llms
1,598 mentions · 615 chunks · 117 episodes
169.8× distinctiveness vs baseline
?
How much more common this term is here than in ordinary English. Higher values mean the topic is more characteristic of this corpus.
Topic summary
?
A short read on the topic's time range, peak episode, and strongest associations. Use it as the quick orientation before drilling into examples.
- llms appears in 615 chunks across 117 episodes, from 2023-11-06 to 2026-04-20.
- Its densest episode is Bits and Bobs 2/2/26 (2026-02-02), with 15 observations on this topic.
- Semantically it travels with ChatGPT, Claude, and prompt injection attack, while by chunk count it sits between Claude; its yearly rank moved from #3 in 2023 to #1 in 2026.
Over time
?
Raw mentions over time. Use this to see absolute attention, not relative rank among all topics.
Observations
?
The primary evidence view for this topic. Sort it chronologically when you want concrete examples behind the larger pattern.
When software was expensive you had to go to the turf of the software creator, and that meant they got your data.
...that means the power accretes on their turf. A raw deal for users. But now with LLMs, software isn't precious.
LLMs can do fiddly slogs that are meticulous and require expertise.
LLMs can do fiddly slogs that are meticulous and require expertise. Vercel is all about lots of small fiddly details for a great developer experience. You...
AGI arising from a single model seems unlikely to me.
...nlikely to me. But a bottom-up AGI, where every person can marshal the power of LLMs to create compounding tools for themselves, feels way more plausible to me. It would also be a runaway process that it's hard to reason about the mul...
Emergence is precisely the part of the system that doesn't fit into human brains.
...eces and layers and the outcome still feels like a miracle. Markets, evolution, LLMs, and life itself all have this characteristic.
I wonder if the whole focus on AGI is downstream of LLMs talking like humans.
I wonder if the whole focus on AGI is downstream of LLMs talking like humans. Like, the idea of a planetary scale omniscient personality is easy to imagine, and also terrifying. Chatbots make this kind of f...
There's a difference between builders and coders.
... Coders see coding as an end unto itself. Before, they looked the same. But now LLMs reveal the difference. Builders love LLMs, and coders hate them.
Systems that assume LLMs will be nearly perfect won't work.
Systems that assume LLMs will be nearly perfect won't work. LLMs will never be perfect. Resilient systems assume that they can be confused.
LLMs can be made to be default-converging.
LLMs can be made to be default-converging. Every input to them, if scoped small enough, it will do what a reasonable person would with that information. S...
Confused deputy attacks are about tricking someone who is powerful but kind of dumb.
...b. The expected danger of them is the multiplication of "powerful" and "naive." LLMs with access to sensitive data are the most confusable deputies ever!
At some point we'll see the mask-off moment for the aggregators.
...umbling along as cold war, but will become a hot one as the power of unleashing LLMs on your data heats up.
I've seen a lot of approaches that mitigate the danger of malicious skills.
...t mitigate the danger of malicious skills. Or that mitigate the danger of naive LLMs confusing themselves. But still nothing in the market that credibly mitigates prompt injection.
Jenny Wen makes the case that our current best practice design process is obsoleted by LLMs.
...en makes the case that our current best practice design process is obsoleted by LLMs. The process that was the best practice fundamentally assumes that software is extremely expensive to create.
StrongDM and OpenClaw are downstream of where LLMs hit a new scaling threshold of agentic ability.
StrongDM and OpenClaw are downstream of where LLMs hit a new scaling threshold of agentic ability. They were inevitable; the time had come. They were at the right place at the right time to surf the w...
We've left the Generative AI era and entered the Agentic AI era.
...ke a chatbot, but that's incidental. A couple of years ago people tried to wrap LLMs and create agents, but it was premature. The model quality wasn't there yet. So they concluded it wasn't possible. But it just wasn't ready yet. Now ...
In the Chatbot form factor the chat is a party trick.
...of durable ends. Those are the form factors that enable the potential energy of LLMs to blossom best.
A position of "It's OK to write code with AI but it has to be reviewed by humans" is untenable.
...f you try to review every line of AI code you will go crazy. It's not possible. LLMs can produce code so quickly the only way to tackle it is to use LLMs to review it.
People use LLMs as interpreters more than compilers.
People use LLMs as interpreters more than compilers. An LLM as compiler could make guardrails for itself that then limit bad behavior.
Karpathy has spoken, enlightened AI-driven engineering is called "Agentic Engineering."
...red to write before: lightly held, possibly throwaway. But it's possible to use LLMs to help write code in much more disciplined ways that give compounding advantage, and that is not a flippant or unserious exercise. Agentic engineeri...
Before software had to be fractally mechanistic, all the way down.
...mall enough you can clear the threshold where any LLM can answer it reasonably. LLMs are more expensive than mechanistic code, but much more flexible and able to handle variance. When you have a working amalgam of LLMs and mechanistic...