Saw this via @sundeep. It truly is double edged.

There are a number of people I know who started to take notes by hand again

The more I build with AI, the more I realize that I need to spend more time thinking about the agentic systems design optimized for that workflow and within my parameters (cost, perf, local etc) My previous approach was much more of a massive 2nd brain but I find this to be too expensive, slow and brittle in many ways as a global solver of every workflow.

Shared this analogy in a recent convo RE AI coding: You can easily spend all your time upgrading the bike instead of riding the bike

TIL a new word, hyper-sycophancy: a way to describe how LLMs try to please their humans (and so often are very agreeable)

It's amazing, summer starts and literally overnight, the kids go from waking-up at 7AM to regularly after 1PM :) I wish I could sleep that much.

Am curious of long-running agent tasks that people are finding useful. I'm able to get what I need to get done w/ Claude/Codex via a head-ed/less browser. The only use cases where I could maybe leverage a long-running agent is w/ web scraping or some deep research type task with a local model. Asking b/c I'm debating btwn investing in an Apple Mini vs an inference-optimized machine running Spark OS. And no, I'm not running /goal to recreate large SaaS to just burn tokens.

Looking to upgrade my support agent and have been looking at agentic RAG. Was just pointed at this visualization for chunking strategy: chunkviz.up.railway.app

Saw this and realize that this is true of many things, particularly in a GenAI era.

Yah, I just used the word token-flexing b/c that’s where we’re at

I have to commend @linear. They might be the single-most used bot in my Slack usage history by a mile as most Slack-bots didn't make it past the initial week.

The Anthropic double-dip is real. Practically every builder I know has an enterprise Claude sub (day job) while also paying for a personal Claude Max sub. When in software have we seen this kind of PMF?!

Having my middle schooler join me at the gym this summer. No matter how hard he lifts, the next day, he’s fine. Me on the other hand, I’m like crawling out of bed and going straight to the foam roller :)

Spent a few hours playing with Fable last night: 1. Yes, it uses more of the wkly quota but not nearly as much as peeps describe on social. Note, I am selective and/or predictive about when to let AI run or stop and I do lean on my tech background and own intuition. 2. It certainly seems more comprehensive. Things like me prompting to make sure we discuss something first or add a test for this or consider performance or other, it seems to take more of that into account. In that way, it's much closer to a more mid-level engineer vs an intern. 3. I'm not able to tell if it's actually producing a better result from a systems design or code quality perspective BUT b/c it's more comprehensive, it makes me think that it's doing a better job and that might be enough. Would I pay for this out of pocket? Not yet but that's mostly b/c I enjoy staying involved in each step of the process and having the opportunity to apply my own taste, judgement and reviews. That said, I could see how if I was building something more disposable or one-shot and had a different goal in mind, I'd just let it cook.

Would bet money that the #1 AI coding workflow for most teams has been audit my existing code. Consumers are the silent winners here. Every app or website you use is going to load faster than ever before!

Some really polarizing feedback from a mix of early stage VC friends over the past wks. Some more enthused than ever w lesser capital and time required to execute. Others, almost bearish and completely unsure what to invest in as models swallow workflows.

A real challenge w GenAI is product bldg often needs some dev process stability and thus why post-mortems are often done at the end of a cycle vs real-time. With the pace of innovation, the new normal is to reset the stack every x wks and that has some real downsides.

I don't follow the latest/greatest with open-source models. What is currently suggested for running local that can be distributed via an install or other for the purposes of text chat classification. And yes, this is for my kid's homework project.

Lots of peeps mistakenly conflate opensource models w running locally. The problem w that is no way you’re loading a 1T parameter model into your 64GB MacBook. Long Micron.

Lots of talk of recursive agents but anyone who has ever debugged recursion knows what a nightmare it can be. I’m prob missing something but am curious to see how adding and maintaining evals at each recursive subagent is going to yield a net time save.