leon

@awasnikar01 yeah, waking up to see the feature already shipped would hurt

@dotgil Truly hurts

opened my first ever open source PR last night. went to bed feeling like a real developer. woke up to find the maintainer had already shipped the same feature and closed mine lol. lesson learned: for a feature, open an issue before you build it. bug fixes you just send. still counts though. a year ago i couldn't even read a codebase that wasn't mine. no hard feelings, better-shot is genuinely a great tool, that's why i wanted to contribute.

@trevhud i'd still want someone who can write the eval once the usage data exists

Why does everyone think it’s a useful skill to write evals? The best evals are derived from actual usage of your agent/model. It’s a data eng problem.

@jmqcooper yeah, the community keeps adding useful pieces

We keep finding more and more, where the community just makes it that much better

@VishalxKodes dark mode fits the calendar view well

Day 6 Project ✅Login Page ✅Authentication UI ✅Dark Mode DSA ✅Floor & Ceil in Sorted Array ✅First & Last Occurrence of an Element ✅Lower bound 21DaysOfCode

@sreenandhanpp congrats on launching ExpenseFlow on LaunchDock

🚀 First product launched on LaunchDock! A huge shoutout to ExpenseFlow, the first product to be showcased on the platform. An AI-powered expense tracker that helps you manage daily spending, budgets, and even transportation fuel costs all in one place. This is exactly why I built LaunchDock: to give builders a place to showcase what they're creating, whether it's launched or still in progress. Thank you for being the first builder to join the journey. Here's to many more launches! ❤️ check it out : launchdock.space/launch/expense…

@nurnabidesigner would you use it on the phone with that keyboard and mouse

I wish i could install claude code on this.

@kuldeepdotcom i'd search for internships on LinkedIn and contact smaller companies directly.

Resume 👇 Any suggestions on where a fresher should apply for internships or entry-level jobs? I'd really appreciate it. kuldeeprajput.in/resume.pdf

Built projects. Learned new tech. Applied everywhere. Naukri: "Here's another unpaid internship." 😭

@devops_prashant the line by line format makes each YAML field easier to follow

Kubernetes YAML Explained – Line by Line One of the most important skills for anyone working with Kubernetes is understanding YAML manifests. Every Kubernetes resource is defined using YAML, and knowing what each field does helps in deploying and managing applications confidently. This cheat sheet covers: Pod YAML Deployment YAML Service YAML ConfigMap YAML Secret YAML Namespace YAML Complete Application Example (Deployment + Service) Key takeaway: Pod → Runs containers Deployment → Manages Pods and replicas Service → Provides stable access to Pods ConfigMap → Stores non-sensitive configuration Secret → Stores sensitive data Namespace → Organizes cluster resources #Kubernetes #DevOps #CloudComputing #AWS #EKS #Containerization #Docker #PlatformEngineering #CloudNative #K8s #DevOpsEngineer #Learning #TechCommunity #SRE #InfrastructureAsCode

@elshayib_ yeah, the 1000 agents line gets me

"You need to be running 1000 agents building your business and talking to customers and making decisions and doing your tax and making you content while you sleep or else you're falling behind" That's how every Agentic Ai Cringlord sounds like

@_markfenner @DevinAI does Adaptive switch models during a local session or only when starting one

On today's episode of Devinmaxxing: pick your model from your pocket. DevinX now does full model selection for local sessions. Sol, Fable 5, GLM, Kimi, or Adaptive if you'd rather it balance cost for you. Reasoning effort too. @DevinAI shipped its own model picker. 🎥👇

@IamMrfeng try entering a specific feature or use-case page instead of the official homepage

Why is it that when I use the "start with a website" feature in Google Keyword Planner to pull keywords from a SaaS product's official website, I can never find anything suitable? It's actually less effective than researching manually. Am I doing something wrong?

@Ex3NDR i'd stick with the current model until Sol proves itself

I found that if you would build a gym for terminals very easily, purely on vibes using the GPT-5.6 Sol. Then just use it in non-stop loop to improve and build new features. It feels 50x better than GPT-5.5 or Opus, everything is very solid and dont need much of the intervention. But the bottleneck now is the UI, tried a new experiment - rendering UI for humans and for agents. I dont think you really need so much fidelity as pixel, if we can render a screen to a unicode grid it would be sufficient. Or give a model coordinates and way to test them. We dont really need a button specific most of the time, like we rarely care about the font text when whole layout is broken? just build a few components manually or with some dedicated workflow, then put everything together. Most of my problems is composition problems: how many components work together, i bet if we will render it in a simpler form it would work much more reliable.

@maddada yeah, vague prompts explain a lot of the disappointment

The most powerful Agentic IDE keeps getting better! • Native macOS/Android/iOS apps (Win/Linux v soon!) • Native Ghostty terminals • The best Rich Prompt Editor for Agent CLIs • Built in VS Code based IDE • Built in Chromium based browser (agentic) • Kanban board based on beads • Agent Defined Automations (Loops!) • Agentic Artifacts editor: HTML, MDs, and Excalidraw! • Cross-agent and harness orchestration • First-class worktrees support • Remote machines connections • T3 Code integration • Auto persist/rename/resume/sleep sessions • Very customizable settings & hotkeys! Let me know what else you'd like to see, or what would make you try it :)

@Diyaldz i'd stick with the current model until Sol proves itself

GPT 5.6 Sol vs Claude Fable 5 Be honest if you use only one AI for the next year which one you choose?

@0xHumza i'd stick with the current model until Sol proves itself

AI subscriptions are gonna stay heavily subsidized for a while. GPT-5.6 made the math obvious. A model that's maybe 80% of Fable 5, but flat-rate in a subscription, is just better value than a frontier model getting walled off behind API pricing and usage caps. Why it holds: → Most subscription users can't tell thinking mode from the frontend model. Same product to them. → What they can tell is the bill. Price is the only spec they actually feel. → So an 80%-as-good plan at a flat price beats a frontier model gated behind credits and 50% weekly limits. → And the labs are too busy undercutting each other to quietly nerf subscription value. Competition keeps them honest. Cheap, near-frontier intelligence on tap for whoever's paying 200 bucks a month. Honestly, not a bad deal at all.

@viewsfrom02108 a firm date would let people plan around the limit reset

Sick of this gimmick, reset the limits and give us a firm date. You reset it two days before the deadline and act like we didn’t all collectively drain or limits on useless things instead of a thoughtful workflow

@chrisbbh the multi-agent orchestration still needs better coordination between agents

Some thoughts on the GPT-5.6 release & new ChatGPT Codex app. Ultra & Multi agent Overall, multi-agent orchestration is in a good place, but it's clear that it could be many times better. I don't think ultra should be a reasoning level: it's special prompt + max reasoning. I don't think a special keyword or skill is right either (harder to communicate enable/disable to user). It seems more like a proactive subagent orchestration mode, which would have the explicit toggle. I would like to see a lot more work and research in this direction. Ultra & ultracode both seem to oscillate between incredibly useful and incredibly wasteful. Partly because the orchestrators aren't capable / able to select the appropriate models & reasoning levels. And other factors, like context management (inherit full, partial, custom-prompt?), task scoping, and control flow. Claude Code has an interesting model of control flow in Workflows, worth drawing inspiration from. GPT-5.6 Sol I really like Sol. Across reasoning levels, it feels like a considerable step up from 5.5. It's super versatile and scales well to my needs. I'm increasingly finding myself vary the reasoning levels, instead of just defaulting to xhigh. Light reasoning feels great. It's fast and can carry you pretty far! Medium is a great sweet spot for most tasks. I tend to give larger tasks and expect more, so I often use high/xhigh. I'm expecting to use these much, much more than max and ultra. With models like Fable and Sol, it no longer makes sense to default to the absolute highest reasoning levels. They are both very capable (and different!) models, and you won't necessarily get much better results from the higher reasoning levels - except that tasks take longer and cost more. Sol can be incredibly stubborn and chase problems forever. This can be an incredible feature, but in my case, it led to 12k lines of overengineering that I had to reign in. It feels nicer to use. I, like others, find that it is good at knowledge tasks. Even better than Fable on those I tested. Computer use is absolutely incredible with Sol. I feel like OpenAI is leagues ahead of the others here. UX and performance is great. Right now, I've had a /goal running with Sol Ultra for 1d 1.5h. It's eaten through multiple 5h limits - thank god for all the resets! It has used ~100 subagents through the course of the run. It handles compaction very well, keeps on track, and feels like it could keep going forever. This is also a very stubborn run. I gave it permission to adjust course, but that led it to seriously overengineer security. As always, it seems to need to cover all possible edge cases and anything that even looks slightly funny. Of course, this will vary with prompting, but subagents as reviewers seems to steer it in this kind of direction. Luna and Terra Luna on max is great, and so is Terra on high/xhigh! I've used these much less than Sol, but so far am pretty excited to see that they perform well. They don't feel "mini" or "nano". Also, they're super fast. I highly recommend trying them out. The new app It's great to see the ChatGPT app get some love - I was stuck using the web version because it'd regressed compared to that. But also, I don't love how the Chat functionality was pushed into the right corner. For one, generating and viewing images isn't easy in the smaller pane. I had hoped to see greater integration of the Pro series to the Codex experience in this release. I do appreciate the "Add to task" function, though. I've also found myself liking ChatGPT Work on web. Across web, mobile, and the app, the ChatGPT experience now feels more unified and 'smoother' across my use cases. There's still a lot of testing to do - these are just my preliminary thoughts.

@sa_vatsa lmao, Claude opened Fable right after you used it all

@mac_dev9 yep, most frontend bugs make more sense after reading the surrounding code

Hot take: The best frontend developers spend more time reading code than writing it.

@keithinvest_ £1,360 a week from eight chairs gives you a very clear baseline

More Barbershop insight ⤵️ The average chair rent in the UK is around £170 per week. For an 8-chair barbershop, that’s £1,360 in guaranteed weekly income. After looking at this, I chose a different model. With chair rent, many barbers naturally think, “Once I’ve covered my rent, the rest is mine.” That target can unintentionally reduce the drive to keep pushing once it’s been reached. A percentage-based model creates a different mindset. Every haircut increases your earnings, so you’re always motivated to promote yourself, stay busy, and grow your client base. Rather than working just to cover your rent, you’re working to maximise your income—and that benefits both the barber and the business.