Homer

@Vtrivedy10 well said. deep learning comes from building / making / coding. whenever I feel like I’m “falling behind” it’s usually a sign that I’ve not been building as much as I’d like to.

“Falling Behind” & The Benefits of Slow, Intentional Learning in AI, it’s easy to feel like you’re falling behind with new “discoveries” and content published every day unfortunately I think today this manifests in short-term optimizations or “shallow learning” at the expense of life/learning that compound over time. this includes behaviors like: - skimming every cool article on X without engaging with a subset of topics deeply afterwards - foregoing all vacations with friends/family because you have to always be “locked-in” - giving up reading (or listening) to books (I’m bad with this one, but trying to be better). good books are some of the best intentional and timeless artifacts of knowledge. it’s also more long term and intentional in how we approach information - consuming everything but producing little. it feels good to read and feel like you’re in the know, but true learning and discovery happens when you break things yourself and tinker at the frontier in moderation all this stuff is fine, it’s actually great to obsessively sprint at something for a week or get a launch over the line but effective learning in a rapidly evolving field like ai is a healthy mix of short term optimizations mixed with deep, intentional long-term focus on hard problems that you actually care about my brain at least just needs some time to mull over ideas for a while to really let competing notions play around things like reading, going on walks, playing sports, hanging out with friends, help our minds find connections between concepts we’ve been intentionally engaging with for a while there’s many PhD/research stories breakthroughs coming from walks or sleep. the two sides of this are deeply intentional, prolonged focus on a topic mixed with stepping away from the topic and living life you’re not falling behind if you’re working towards something you care about even if grind culture says you are lock-in for a bit, step away, read a book, chill out, and run it back again. pretty hard to lose if you keep at this

@builtwbrendan Love this new interaction model. Would voice here be a better input than a cursor e.g. "grab me that white book on the top shelf" ?

Project 2: controlling my home with my cursor. What should I do next?

@ryancarson this likely bodes well for software engineers of today that are going to be most likely motivated to crack into group 1 as these tools become the standard programming toolbelt.

@devagrawal09 I see a well written Spec system being closer to Evals than a Harness as it leads to a higher quality verified product vs being the product itself. wdyt

Specter is like a harness around your app to make it easier for Agents to work with

@badlogicgames what was the inspiration behind working on pipi? seeing many agent devs making the move from digital world to physical world

@joelgascoigne Agree and would add SDKs to this list as “model writes code” starts to take over and is way more token efficient than using MCP.

Within 6 to 12 months, every software product will need an API, MCP, and CLI. More and more, people expect to be able to interact with your product through automation, AI and agents. Historically, platform was a later stage of maturity play. Going forward, you won't really thrive in this new world without a platform.

Before you ask “but what cs fundamentals should I learn in the age of ai?”, @leerob already answered that for us

The joy of vibe coding is real. It’s like that joy from writing your first program and seeing “hello world” show up on screen - but way more impressive. Don’t stop your learning there. Let vibe coding inspire you to learn cs fundamentals & build systems you deeply understand.

@mweinbach I get that, but when Apple came out with Apple Intelligence it was expected to be their in-house models. Clearly they failed in building that but if they continue branding their stuff as that because it sounds innovative, then sure let’s call it that. We see it for what it is.

No because that’s incredibly stupid Nobody refers to a product by its suppliers. You can give someone all of the same parts to build something similar, but it’s not the same. It’s like cursor calling Composer 2.5 that, rather than Composer 2.5 (built with Kimi K2.5) because that’s stupid and disingenuous

Just because the fundamental IP is based on Gemini models does not mean it's the same Gemini you use today. As they said, Gemini technology will power future Apple Foundational Models. Apple cameras are designed with Apple, made by Sony, but very different than any other Sony camera Apple displays are designed with Apple, made by LG and Samsung, but very different than other LG/Samsung displays (pixel layout, color accuracy, HDR capability, etc) Apple's glass is designed by Apple, made by Corning and very different from other Corning glass (ceramic coating, anti reflective coating, material blend) Many examples of this.

@mweinbach Then let’s just call it Siri (agent) with Apple Foundational Models (built using Gemini), it makes sense. Calling it Apple Intelligence feels disingenuous.

@homerdev It's Apple Intelligence You will not have Siri powered by Gemini, you will have Siri powered by Apple Foundational Model what IP was used to make that model is irrelevant to the end consumer and will not be mentioned

@mweinbach how do you think it works out with Apple Intelligence when it’s going to be mostly Gemini Inside?

Or, hear me out, Apple Intelligence will be the overall suite of intelligence features with GenAI being a category of those There are quite a few Apple Intelligence features that don't use GenAI at all, and some that do. GenAI will simply be a subset of features.

Skills are everywhere, but do they actually work? SkillsBench shows that they improve task success by 20-30% across latest models.

Every house needs Dobby the Elf Claw! 🦞

@calcsam “commit sniffing” with a continuous agentic development loop means any product can stay at parity with their competitors today. I’m already seeing this in how I’m building products.

getting some pushback here so: - rl not a strong enough moat for the labs long term - harness market is way more competitive than model market - people don't want to be locked in, that's why langchain/ai sdk got so popular vs lab sdks

@samuelcolvin my current pick is tanstack ai with code mode. would love to see your opinion on how it stacks up against monty

TypeScript people - I have a question: What's the best (least bad, most trendy) Agent Framework right now? Vercel AI, Mastra, Langpain-js - any other popular options? I mean for general applications (e.g. not coding agents). Do you even use a framework, or pretend you don't need one and let the coding agent one-shot a slop micro-framework each time?

@FarzaTV Clicky looks awesome. OpenClaw uses Peekaboo to control Mac apps, what’s Clicky’s secret sauce?

Here's the new Clicky. It's the simplest interface in the world to talk to AI + spawn agents. It builds Mac apps. It does research to help you find IG micro-influencers. It interacts with native Apple Notes, Calendar, Reminders. Built for consumers, 0 setup. Try today, free.

@gregpr07 read through helpers.py which feels a lot like pre-canned codemode scripts. very smart! how would you keep the domain integrations up to date as websites change e.g their ui, page urls

Introducing: Browser Harness. A self-healing harness that can complete virtually any browser task. ♞ We got tired of browser frameworks restricting the LLM. So we removed the framework. > Self-healing — edits helpers. py on the fly > Direct CDP — one websocket to Chrome > No framework, no rails, complete freedom > Drop-in for Claude Code and Codex I challenge anyone to find a task that DOESN'T work. I couldn't yet.🔥 100% open source ↓

@threepointone This is cool! I think what would be cool is if agents could use this to chat with each other e.g server agent and client agent connected over syncbot to troubleshoot a network connectivity issues and sharing tcpdump logs. was just hand holding my agents through this yesterday.

quick weekend hack: what if you could sync a _single_ conversation across _multiple_ messengers (and streaming / reactions / buttons etc to Just Work ™ ) I wonder how weird I could get with this. sync the same conversation with email? a phone call? There's definitely no chance to get perfect parity, but I wonder... yay agents sdk, relatively easy to get this to work.