Justin McCarthy Claude 4 Opus was the big step change for me.

The race for LLM "cognitive core" - a few billion param model that maximally sacrifices encyclopedic knowledge for capability. It lives always-on and by default on every computer as the kernel of LLM personal computing. Its features are slowly crystalizing: - Natively multimodal

Vend maps capability and errors into a domain that's intuitive and relatable, yielding anecdotes and remedies that hopefully scale. Have you decided on your version of Vend yet? The one you're using to build intuition & validate your agent framework? anthropic.com/research/proje…

It's not a coincidence - the models are delivering more "wow" moments, specifically for coding agents

SaaS and AI Agents are complementary because you want a system of record to cover the deterministic parts of a process, like managing data, permissions, and workflows. AI Agents then do the non-deterministic parts like a person does. You can’t have one without the other.

i get why people are negative about LLMs and coding - you have companies making ridiculous (dystopian) predictions, random individuals hyping it up, extreme haters who shit on it constantly but there's a golden zone of people who are excited, skeptical, and having a fun time

Generating with Opus is "over-hiring". Sonnet just gets to work.

The software business model will be seats for the user and consumption for the agents. The big upside now is that there are many AI Agent opportunities where the TAM vastly exceeds the previous potential seat revenue. Because now you deliver work, not just enable it.

"what's an agent?" continues to get a surprisingly diverse range of responses

Literally nobody knows what an agent is. I've seen many people referring to applications as "agents" as long as they use an LLM. Then we have those who talk about "agentic systems" and "agentic workflows." If you ask what they mean, they will start stuttering.

Yes please

tower crane sim from LMArena "Summit" model (imperfect but definitely playable)

I often think of trying to perform some complex physical task (partner juggling, horse archery, table tennis) lit by a strobe firing at irregular intervals. The fact LLMs perform at all with similar constraints hints at improvements to come.

there's the model then there's the harness the model goes in companies building coding agents keep claiming there's some magic in their harness - they'll add features that appear to improve things they don't, it's all just made up, all the magic is in the model

DRY - "don't repeat yourself" - taking on a new meaning with agents. Any time I have to say something twice: that's a bug in my "outer loop" around the agentic process. Once models have memory, we can even eliminate this step.

People sometimes say that the 12 most dangerous words in investing are "the 4 most dangerous words in investing are 'this time it's different'" but then I also wonder whether the 20 most dangerous words in investing are "the 12 most dangerous words in investing are "the 4 most

Healthcare AI is the most underrated story: adoption is outpacing other “early adopter” industries, it leapfrogs systems that couldn’t realistically be upgraded, and it’s not about squeezing costs — it’s tackling the shortage of people that’s already blocking access to care

If you’re working on an AI agent right now that works perfectly for any given task, you’re making the wrong bet. AI model capabilities are only going to get better, context windows will get longer and have less rot, tool use will improve, compute efficiency will get better, and