Dungeon Crawler Crawl would be absolutely next level with an interactive experience that showed Carls stats, a 3D map of the floor, ability to explore his inventory and that of his team, and his current quest and mission status. And boss stats. Driven completely by the story as you read/listen

I find Opus 5.0 so unbelievable cringe to work with. "The answer is not what you expect, and the why *matters*" I can't imagine the folks at Anthropic actually enjoy using this model. My Claude .md actively tries to get it to not use this kind of phrasing, but it just can't help itself.

Graph Engineering is just a just building workflows that run workflows.

Quick update on @SubspaceHQ I've decided to switch to maintenance mode going forward. With the latest release, Subspace is now Free, and I've refunded our existing annual and monthly subs. I'm going to take a pause and start exploring what's next for me and the direction of my career.

Blink 182’s California album with John Feldman and Matt Skiba (and not Tom DeLonge) was their best album since Enema of the State IMO. I love Feldman’s work. Goldfinger and his stuff with The Used was just genius. open.spotify.com/track/12YGK...

Bored to Death

blink-182 · California (Deluxe Edition) · Song · 2016

open.spotify.com

I used GLM 5.2 all day (~$100 via openrouter) - very impressed. ++ great root causing some hard perf issues + clear comms, straight to the point (vs Opus verbosity) + good UX/UI < Opus, > Codex +/- very thorough (sometimes too? burned tokens trying to fix a flaky test)

I listen to/watch all of @clairevo’s How I AI podcasts but this one on using GLM 5.2 is a must watch if you are using Claude Code/Codex/Cursor and are battling with rate limits and costs (🙋‍♂️)

We've been through enough of these cycles to know now when people start getting excited about an open source coding model (deepseek, kimi, now glm 5.2) you know we're about to get some major upgrades on the frontier model side.

Hit Codex weekly rate limit yesterday AM. Man am I feeling it. IMO, Opus 4.8 is just not even in the same ballpark. Hoping Anthropic's rumored Sonnet 5 release is really strong.

Any AI obsessed PMs or any of my friends/former colleagues down for a nerdy chat about AI/Agents and how PM is evolving? Looking to chat with a bunch of folks this week - for fun, to reconnect, to vent, to get feedback on something you’re tinkering with - anything! DMs/email open!

Super interesting episode on “Free Range AI Consultant”. As I mentioned in my recent posts seems like the gap between what the most AI pilled know and can do and what normal humans know and can do with AI is massive. I imagine that will be a big problem worth building a biz around for 5-10 yrs

Life is just a lot more interesting if you look at any seemingly dumb thing in the world and ask yourself "if the person who made this decision was actually very smart, why would they have made this decision?".

I honestly didn't believe hardware + local AI was going to be THAT big of a deal but now I get it. You can only get that much better for most tasks from Fable, so local model on your own hardware -- will be massive. Very bullish now.

My other key takeaway from Fable. If I (as AI pilled as one can get) am getting as much value from the previous gen models to the point that using the upgrade for most normal work doesn’t make that much sense ... Hard to imagine what even better models will be used for by normal humans.

I feel like there is a really big gap in the market for a product that can tell you if any of the cool skills that keep coming out actually improve the speed + quality + token-efficiency of your results. github.com/DietrichGeb...

GitHub - DietrichGebert/ponytail: Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.

Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote. - DietrichGebert/ponytail

github.com

It really is funny how much better Claude Code is at design than Codex. I've tried enough times to know it is definitely not a fluke thing. Add it to the pile of reasons you should really be using both these agents not silo'ing yourself in one.

I’m really glad Anthropic seems to be revisiting their stance on SDK and Claude -p usage. Cautiously optimistic they realized they should be encouraging such such use.

Being in my bubble it's pretty wild chatting with friends/colleagues/neighbors about how slow AI adoption is even at companies you'd think would be packed with AI pilled engineers.

This is what has been really hard about Subspace lately. Retention & conversion are good/great. Early paid growth experiments were promising, but not worth it yet.

One of the most annoying Opus 4.8 things Me: "we need to do X, let's play on using Y following the same approach as Z"

Custom Autocomplete UI for files, skills, mcps, shell commands, etc. was one of the first features I built into Subspace from the earliest days. I improve it constantly. It has gotten really really good, and fast. Here's one very obvious example (video) of how much better it is

Idk maybe just me but between the World Cup and the Knicks and the beautiful PNW weather lately just feels like everyone wants to be out, be seen, be in community lately. Feels like we may finally be healing post pandemic.

Loved this episode with Mark Pincus founder of Zynga. Really resonated with the challenge that “better” for founders/product folks is often actually “new”, which requires way more thought and experimentation. Have a feeling I will absolutely devour this book. open.spotify.com/episode/27A...

The hidden pattern behind successful products | Mark Pincus (Founder of Zynga)

Lenny's Podcast: Product | Career | Growth · Episode

open.spotify.com

Proactive Agents are how products start to build themselves. The mistake is thinking it’s one agent. It’s not - it’s dozens, each with narrow responsibilities, each with a system designed to enforce quality and human-in-the-loop in the right spots.

My takeaway from the craziness with Anthropic and the US Govt right now. We need better Open Source models in the US. Meta and increasingly xAI/Cursor generate significant revenue from adjacent businesses, would love to see them take on this important role.

One of those 'living in the future moments' - Had a question about Subspace I couldn't answer - Had an agent add the necessary instrumentation - Scheduled that same agent to review the data starting 7 days from then - 7 days later got an email with the complete analysis

Fable is a pretty amazing model. Its expensive, chews through usage, and soon will be at API prices only. I think as models become smarter and the old models become "good enough" for the vast majority of daily tasks, it will finally become much much more obvious when to use certain agents/models

The biggest story of the new Siri is the impact it will have on ChatGPT and Claude consumer subs. Why would anyone pay $240 per year min, or $0 w/ads for something not integrated with their phone/apps when they get all of that and more for free with an iPhone. No wonder OpenAI is all in on Codex

Shipped a few quick QoL updates to SPECTRE - light mode plan so you can use /spectre-plan for any task - /spectre:plan now 2 gates (high level design + final) instead of 4. Less friction, better results. 🚀 repo 🔗 below

You give an agent a goal. It pursues the goal proactively, brings you the decisions that you and the agent decide are worth your judgment, and competes for your attention. You own the gates, it owns the motion. Agent pursues, Human approves.