R Max

@sukh_saroy This is my intuition. When people project their own mental framework of how things should be done in a role-playing fashion they handicap the agents Roles and personalities are easier to understand to humans and story telling, not for exploratory agents

@pankajkumar_dev I hope we don't have to wait a quarter for the skills feature

Google I/O leaks: Gemini desktop app is becoming a real AI agent - The new Gemini desktop app is split into two workspaces: a normal Chat mode and a dedicated Spark mode for local agentic tasks - Gemini Spark can connect to local folders, analyze code files, run scripts, organize files, and even sync workflows directly with Google Drive - A new “Stream to Cursor” feature acts like Google’s Magic Pointer idea, where Gemini understands the context of whatever app or window your cursor is hovering over - The floating overlay can instantly share screen, window, or camera context while also allowing rapid switching between models like Gemini 3 Flash and Gemini 3.1 Pro - Google is also preparing local “Skills” support, letting users attach custom scripts or capability folders directly into the agent workflow - Internally, Gemini Omni is referred to as “Veo4 Omni”, hinting at deep Veo 4 integration inside the desktop experience - Gemini Live is also present as a persistent voice overlay, though it still appears to be work in progress internally

@AndyMasley lol, I thought my phone got corrupted

Everyone complains about minimalist design until the company tries something fun and everyone reveals why all the companies have been forced into minimalist design

@TylerFCloutier Compile to single binary

What even is the point of Bun? It’s only marginally faster than Node.

@acatovicx @DanielSmidstrup None of them develop frontier models in the top charts

@DanielSmidstrup Mistral. HuggingFace. Black Forest Labs. DeepL. ElevenLabs. Synthesia.

USA has ChatGPT USA has Grok USA has Claude USA has Gemini China has DeepSeek China has Qwen China has Kimi China has MiniMax Europe has?

@farzyness Vibe-designed

Whoever designed this needs to be fired immediately:

@cnakazawa @lilienblum AGENTS.md is brief and terse, plus a context map to point the agent to relevant docs (including read me) If save tokens instead of letting the agent do exploration every time

@lilienblum Not true. If the details are only in AGENTS.md, then human need to read those files too. If it’s in the existing files it’s easily discoverable for humans and agents. There is a high likelihood that the custom instructions in a repo are useful for humans too!

AGENTS.md and agent specific config doesn’t make any sense. Just use README.md and put the context in the repo.

@ahitposter Fun story Once I was looking into our data lake costs, using serverless sql server in databricks We built a handy dashboard to show query costs with filtering and knobs Turns out the main spending was this dashboard that we all had open and made complex queries every minute

> “so i see here you are spending $1.3m per month on ai credits” “yes that’s correct” > “what exactly have you built this month that warrants this kind of spending” “we built this nice ui that shows you how much we are spending on ai credits”

@theirongolddev @adamwathan They focus on growth = user acquisition So existing users are second priority

@adamwathan It’s a naming thing. He’s referring to the Codex desktop app. You can now connect to your desktop running the Codex desktop app from your phone For the life of me I don’t know why companies keep reusing the same names to refer to completely different paradigms

I must be going crazy because I’ve been using Codex in the ChatGPT mobile app for literally months? What’s the thing that launched today?

@AlLeakWire @ai_for_success If you have Google workspace org then you can set up Gemini enterprise with custom connectors They definitely don’t cater to small teams or individuals for this power features

@ai_for_success How can we get more connected apps other than Google like chatgpt have

Google Gemini Spark spotted 🔥🔥 Google appears to be testing a new 24/7 experimental AI agent experience called Gemini Spark. Gemini Spark is designed to act as an everyday AI agent that can help with: - Inbox management - Online tasks - Website interactions - Connected apps - Chats and productivity workflows Some notable details: - Learns from connected Google apps, websites you are logged into, chats, tasks, location, and Personal Intelligence data - Can save remote browser data, login details, and remote code execution data - May share necessary info with third parties to complete tasks - it can potentially make purchases without asking every time - Positioned as a 24/7 AI agent experience

@icanvardar External skills are a reference I write my own prompt files for my specific use cases Otherwise you waste tokens in information you don’t need and may confuse the model

skills were a mistake. they were never meant to live in the context window all the time

@leetllm @thegenioo too much customization nerfs future models

@thegenioo For me it's git worktrees. The CLI lets me use worktrunk to spin up multiple agents on different branches at once. Plus I keep my custom skills in a local folder so the model learns my style.

I've noticed that hardcore coders prefer using Codex in CLI rather than macOS app What's the reason?

@thegenioo A bunch of people re-discovered the “terminal” because Claude/Codex and now it’s a second coming for them I prefer being able to scroll and select+copy like a human

@simonbrown “Hey Claude, how would you implement X feature ?” “lgtm, write up a spec” “cool, implement the spec”

Spec-driven development makes very little sense to me. The software development industry has repeatedly shown that devs don't like writing docs, often saying "it's tedious and time-consuming; I'd rather be coding". - How will this turn out to be different? - Why automate the fun part (coding) and force devs to write docs instead? developer.microsoft.com/blog/spec-driv…

@OrevaZSN Going above and beyond is part of the expectations to keep you employed Otherwise they could hire another average employee at lower rate

Corporate jobs are crazy because you can work your ass off all year, exceed expectations on your annual review, and only walk away with a 2% raise.

@zenorocha Aside of speeding up the build: batch merge train

before our main repo had an average of ~20-40 open PRs on any given day. now, we average ~130-200 open PRs. apart from reviews, CI/CD is also becoming a bottleneck. builds are too slow. humans shouldn't be waiting. x.com/zenorocha/stat…

Finally happened, one of my Gemini API keys leaked somewhere 🙈

@aasthajs You have to have tan/brown skin to enjoy natural sunlight for hours, that's evolution

I used to be a Bryan Johnson fan. I’ve changed my mind. Here’s why: - Once you understand even basic biophysics & circadian biology, I don’t see how you can listen to him. - He gets 0-10 minutes of sunlight. Doesn’t understand how life evolved with it for 4.6bn years, how avoiding sunlight is as bad as smoking, the power of light therapy. I get 6-8 hrs of sun everyday. - Doesn’t talk about the impact of LEDs and artificial lights on health. - Thinks his diet is the best in the world. Can’t possibly be true with no fish, and no mention of local food. - Crazy number of supplements, dubious at best. Sure, he’s a good inspiration for people but pulling on the wrong levers. No sunlight is the biggie for me. More on why:

@KexinHuang5 Agents creating workflows on demand using a DSL is a natural evolution instead of chained tool calling As long the agent can test and validate correctness before doing a full load it should work as expected

Introducing agent-managed sandboxes: AI agents to autonomously orchestrate fleets of sandboxes to handle massive workloads. This unlocks adaptive scaling, from small tasks to terabyte-scale processing, while minimizing unnecessary cost. With parallel sandboxes, throughput multiplies, and agents can explore multiple ideas simultaneously. Checkout our new technical report of this sandbox pattern:

@hridoyreh If you still write code by hand you are on your way out of the industry

This guy is a software engineer with 18 years of experience. Now, working at McDonald's. One day, his company decided 2 AI guys could do what his team of 12 used to do. As a result, got laid off. He sent 100+ applications, but didn't get any jobs. One HR rep literally told him, "CS degree is useless now". Now he's flipping burgers while grinding through rejections. He knows 200 developers in the same boat. And the executives who made this call? They are fine, they always are...

@fchollet Jobs that are single task are done

AI automates tasks, not jobs, and when a task gets cheaper, demand for the job grows. AI cannot automate jobs end-to-end because it lacks autonomy and cannot operate without supervision. There is still zero job from 2022 that can be performed end-to-end by AI, not even translator or customer support associate.