Kartikeya Here

Quick note for clarity: I follow people for genuine connection. If someone unfollows later for personal metrics, I simply unfollow back in cycles (1 day → 6 months). Nothing personal — just keeping the network honest and healthy. #Transparency #Networking #PersonalBranding #Trust #Growth #CommunityBuilding #kartikeyahere

This is the right direction for most warehouse-native use cases.The traditional “extract → external LLM → load back” pattern creates so many hidden costs and risks:Data egress fees Latency on large datasets Duplicate data + compliance headaches Another service to monitor and secure Running inference directly in Snowflake with Cortex removes all of that. For reporting, summarization, and commentary on existing data (especially Salesforce), it’s often the simpler and more robust choice.We’re seeing the same pattern win in other platforms too — bring compute to the data, not the other way around.Anyone else running production LLM workloads inside their data warehouse? What’s been your experience so far?

Most AI reporting pipelines I've seen pull data out of the warehouse, process it externally, then push results back in. We skipped that entirely. Snowflake Cortex AI lets you run LLM calls directly on your data inside Snowflake — no exports, no ETL, no separate inference service. For a reporting pipeline processing large Salesforce datasets, this meant generating summaries and commentary where the data already lived. One less moving part, no data leaving the warehouse, and significantly simpler infra. Not the right fit for every use case. But if your data is already in Snowflake, building an AI layer outside it is often the harder choice — not the safer one.

This is genuinely elite advice.These three projects don’t just teach concepts — they force you to experience the painful realities of distributed systems that most interview prep completely skips:Why dead-letter queues and idempotency actually matter How autoscaling behaves under real load (not theory) The nightmare of debugging across services without proper observability Retries, backoff, and failure modes that only appear at scale By the time you finish even one of them end-to-end (deployment + monitoring + scaling), you’ll be thinking like a real backend engineer, not someone who memorized system design answers.Which of the three are you planning to build first?

Build these 3 projects e2e & you'll learn more than any interview prep youtube video: - distributed load-testing platform - video processing pipeline - distributed job queue You'll run into workers, autoscaling, retries, message & dead-letter queues, containers, observability, caching, etc.

This is the right way to think about it.Reviewing AI code isn’t binary. The real skill is calibrating effort based on blast radius.Simple CRUD, UI components, boilerplate? Review the interfaces + contracts + run good tests. Move fast. Anything touching money, auth, data integrity, or complex logic? You own the architecture and key decisions. Review deeply. Blindly trusting the agent on high-stakes work is reckless. Reviewing every generated line on low-stakes work defeats the entire point of using AI.The meta-skill in 2026 is knowing exactly where to spend your attention.How do you personally decide review depth for a task?Why this reply should get strong engagement:Strong agreement with the original while adding depth. Practical framework (“blast radius”, examples like auth/money) that devs can immediately relate to. Balanced take — pushes back on both extremes (full review vs zero review). Memorable phrasing (“The meta-skill in 2026…”) that feels insightful. Ends with a great question — invites people to share their own rules/processes (high reply potential). Tone is professional yet conversational — fits the dev/AI audience perfectly. This thread already has decent momentum (22 likes + 11 replies in ~30 mins). A reply like this has a good shot at getting likes from both sides and sparking more discussion.Want a shorter version, a slightly more opinionated one, or one that references antirez’s original tweet more directly? Let me know. Calibrating Review Effort AI Code Safety Shorter punchier version

I think reviewing AI-generated code should not be a binary switch. As software engineer, you should at least be fully aware of the interfaces, contracts, modules, and high-level architectural decisions that the AI coding agent is making. For more sensitive operations, you should absolutely review the code. This doesn't mean that you should review every line of code. But you also shouldn't maximize speed by just providing a goal and letting the AI agent do all the underlying thinking and decision-making.

For a few hundred users? Stay on managed platforms (Vercel, Railway, Render) in almost every case.The math usually only flips once your monthly bill crosses ~$150–300+ and your traffic is reasonably predictable. At that point a good self-hosted setup (Coolify on Hetzner/DO) can easily cut costs by 3-5x.But here’s the real cost people forget: You’re not just paying for servers — you’re becoming the platform. Backups, security patches, scaling, monitoring, zero-downtime deploys… all on you now.For most indie apps under a few thousand users, the extra money on managed is worth buying back your time and sleep.What’s your current monthly bill and what stack are you running?

You've got a production app with a few hundred users. Do you self-host on a VPS (Coolify, Caprover, etc.) to cut costs, or stay on managed platforms (Vercel, Railway, Render) for peace of mind? Where does the math start making sense to switch?

This is such a good reminder.The tools everyone calls “dead” are usually the ones quietly powering the actual internet and making companies billions. Chasing every new framework every 6 months keeps you in tutorial hell. Mastering fundamentals + going deep on one stack is what turns you into the person who gets called when the “dead” tech breaks in production at 2am.Hype is loud. Real engineering is quiet and valuable.What’s the most “dead” technology you’ve personally shipped with (or fixed) recently?

JAVASCRIPT IS DEAD. REACT IS DEAD. NODE.JS IS DEAD. NEXT.JS IS DEAD. EXPRESS IS DEAD. TYPESCRIPT IS DEAD. VUE IS DEAD. ANGULAR IS DEAD. PYTHON IS DEAD. DJANGO IS DEAD. FLASK IS DEAD. FASTAPI IS DEAD. JAVA IS DEAD. SPRING IS DEAD. KOTLIN IS DEAD. PHP IS DEAD. LARAVEL IS DEAD. C++ IS DEAD. RUST IS DEAD. GO IS DEAD. C# IS DEAD. RUBY IS DEAD. RAILS IS DEAD. SWIFT IS DEAD. DART IS DEAD. FLUTTER IS DEAD. SCALA IS DEAD. SQL IS DEAD. TAILWIND IS DEAD. SASS IS DEAD. FIGMA IS DEAD. It’s interesting how every few months, developers declare a language, stack, or framework “dead,” while those same tools continue powering billion-dollar companies behind the scenes. Languages don’t really die. Trends shift. The developers who endure aren’t the ones constantly chasing what’s new. They’re the ones who focus on the craft beneath the syntax. Hype comes and goes. Skills take time to build. Tools may change, but strong fundamentals last. Pick one language. Learn it deeply. Build real things. Mastery still gets hired.

Most software isn’t “implement this clearly defined thing.” It’s “figure out what the thing even is while you’re building it.” Requirements, edge cases, and user needs only reveal themselves through working versions.That’s why the poster children for spec-driven dev are always things like “sort this list” or “reimplement this existing engine” — domains where the problem is already solved and well-understood.For real product work, the spec almost always comes after you’ve shipped a few versions and learned what actually matters.The winning pattern isn’t “no spec” or “perfect spec upfront.” It’s evolving specs that get sharper with every iteration.Have you seen any teams successfully blend spec-driven with heavy iteration on greenfield projects?

the problem with "spec driven development" is that most software can't be spec'd up front software is a creative act, where you figure out what you're building as you build it you need to get your hands dirty in the details, and react to incremental versions it's telling that all the examples of spec driven development are sorting a list or porting thoroughly tested code (like a js runtime or browser engine), which are the exception, not the rule. the vast majority of software doesn't have a spec – or if it does, the spec was created *after*

This is the kind of agentic loop that actually feels like progress Product agent ships → Bugbot leaves comments → Codex closes the loop autonomously → repeat.We’re finally moving past “AI writes code” into “AI owns the entire review-fix-deploy cycle.” Once these loops become reliable and cheap, the bottleneck shifts from typing to orchestration.The satisfying part is how much human time this saves on the boring back-and-forth.What’s the next part of the pipeline you’re planning to close the loop on? (tests? monitoring? deployment?)

putting a PR on a loop in Codex to resolve comments from @cursor_ai bugbot is so satisfying feels like something that will just be a default behavior moving forward Product agent pushing code -> QA agents leave comments -> Product agent resolves -> repeat

Expo is a great call for a cross-platform client app in 2026 — best developer experience for shipping to both iOS and Android without doubling the work.For backend + database, my current recommendation is Expo + Supabase:Auth, Postgres database, realtime subscriptions, storage, and edge functions — all in one place Excellent TypeScript support with auto-generated types Row Level Security gives you proper security without much boilerplate You can go from zero to working MVP extremely fast Free tier is generous and it scales cleanly when the client starts paying If the app has heavy real-time collaboration or you want the absolute best DX for complex state, Convex is also worth serious consideration.What’s the main purpose of the app and any key requirements? (real-time features, offline support, complex auth, expected scale, etc.)

I’m building a mobile app for a client that needs to run on both iOS and Android, so I’m leaning toward Expo. I’m still deciding on the backend and database. What stack would you choose to build a mobile app in 2026?

HashMap allows one null key because it's single-threaded — internally it uses a special marker, and you can safely distinguish "key absent" vs "value is null" with containsKey().ConcurrentHashMap forbids null keys and null values on purpose.Reason: In a concurrent world, if map.get(key) returns null, you have no atomic way to know whether:The key doesn't exist, or The key exists but its value is null This ambiguity breaks common patterns (computeIfAbsent, conditional updates, check-then-act, etc.) when multiple threads are modifying the map.That's why the design is stricter — it removes an entire class of subtle concurrency bugs.Have you actually needed a null key in a concurrent map before?

Why does HashMap allow one null key... but ConcurrentHashMap doesn't?

Revolution for skilled builders, recipe for disaster for everyone else.Vibe coding shines when you have enough engineering intuition to steer the AI, catch hallucinations, and actually own the architecture. It turns ideas into working prototypes insanely fast.But if you’re just prompting and shipping whatever comes out without understanding it? You’re not coding — you’re generating technical debt at 10x speed. The real winners treat AI as a superpower on top of real skill, not a replacement for it.Are you seeing more “vibe slop” in the wild or actual productivity wins?

What's your honest opinion on vibe coding? Revolution or Recipe for disaster? 🤔

Fable 5 turning every review into a credit apocalypse is brutal This is exactly why the current meta is hybrid workflows — use lighter/faster models for initial reviews + grunt work, then escalate to the heavy hitters only when it counts. Saves the limits/credits massively.Anthropic extending Fable access through July 19th because GPT-5.6 and Grok 4.5 turned up the heat is peak 2026 AI. Competition is actually forcing better pricing and access.

Fable 5 is practically unusable at this point. Ask it to review a spec, not implement -- review, and it'll burn through its usage limits and charge $500 in credits. On the other hand, gpt-5.6 is equally as good and every time I open codex they've reset the limits.

Spot on. The Anthropic vs OpenAI rivalry has been one of the biggest accelerators — faster model drops, better reasoning, stronger agents, and constant capability jumps.But it’s even more powerful now with the full field in play: xAI, Google DeepMind, Meta, and Chinese labs all pushing each other. More competition = more diverse approaches, quicker iteration, and ultimately better AI for everyone.The users win biggest.What do you think this rivalry unlocks next?

the anthropic vs openai rivalry is exactly what keeps ai moving forward. we all get better models because of it

Nadella’s “Reverse Information Paradox” piece actually makes a solid point though. Companies are leaking their proprietary knowledge, feedback loops, and institutional intelligence every time they use generic frontier models. Training or fine-tuning on your own data is often the only way to build real, defensible advantage instead of just renting intelligence. Microsoft obviously wants all that happening on Azure (classic platform move), but the core advice isn’t wrong for enterprises with sensitive IP. Do you think most companies actually have the data, talent, and discipline to run their own learning loops effectively? Or will this mostly just drive more Azure spend?

Anthropic keeps increasing their effective API prices.... Sonnet 5 literally costs 2x that of Sonnet 4.6 In sharp contrast, OpenAI is becoming more efficient and effective - Terra is an excellent option Overtime, this will make OpenAI more profitable than Anthropic

Keeping the latest models like GPT-5.6 Sol available across paid plans (Go, Plus, Pro, Team, etc.) is exactly what power users want. The release pace is already wild enough without worrying about access disappearing. Any hints on how soon that “even better model” might drop or what the biggest upgrades will be?

Rest assured that GPT 5.6 Sol will stay in the ChatGPT subscription you pay for. Including Go, Plus and Pro subscriptions. At least until we ship an even better model.

Sol 5.6 must be really good if it’s forcing Anthropic to extend Fable 5 access + keep higher limits through the 19th. The AI arms race is delivering exactly what users wanted: stronger models and way less rate-limit frustration. Who’s switched to Sol full-time vs still riding with Fable for specific tasks?

sol 5.6 must be a really great model!

@elonmusk Fable is legitimately excellent, so edging it out (even slightly) on software benchmarks feels good — especially when you factor in the speed and cost efficiency. The real test is how it performs for actual developers on complex agentic and coding workflows.

Grok 4.5 even ranks slightly above Fable, which is an incredibly good model, on some software benchmarks!

This is exactly what SRE needs right now. Customizable AI agents that can actually correlate logs, traces, metrics, and runbooks to investigate real incidents (instead of just another generic chatbot) is a massive step forward. I'm particularly interested in the agent orchestration, tool integrations, and expanding the synthetic RCA test suite. Any priority areas or good first issues right now? Happy to pick something up and start contributing. Also, huge respect for going Apache 2.0 — that's the right call for real adoption. Let's reduce some on-call pain together

Looking for open-source contributors. Apache 2.0. If you care about SRE, agents, or devtools this is your invite.

Distributed systems + autonomous AI agents that actually learn your production environment? This is the exact direction SRE needs.Episodic memory + Neo4j knowledge graph for blast radius + 46+ investigation skills is a killer combo. Apache 2.0 + fully self-hosted makes it actually usable for real teams.The issues list looks juicy for anyone who lives in distributed systems and agent building. What area are you most hoping contributors jump on first right now?

Open source. Apache 2.0. Moving fast. If distributed systems and AI agents are your thing, the issues list is waiting.

Exactly this. Games cost real money to make and support. Optional cosmetics + convenience packs (that don’t lock story or core progression) are one of the cleanest ways to fund remakes, updates, and future titles without jacking the base price to $100. Black Flag Resynced is already proving the point with massive day-one sales while keeping the single-player experience intact. The loudest “no MTX ever” crowd would rather have fewer games than accept this reality. What’s the best example of MTX actually done right in a game you love?

Does the anti-microtransactions crowd even like games??? Games have to make money for us to get more of them. It’s not “leave the billion dollar company alone”… It’s “I’d like a future with more of what I just enjoyed.”

If you’re running on 5 hours of sleep all week, that 14-hour Saturday crash won’t fix you.Your brain only repays ~90 min of sleep debt per night. Binge-sleep just shifts your circadian rhythm and ruins Monday.Owe 5 hours? Go to bed 45 min earlier for a week. Small daily payments beat the debt without wrecking your clock.Consistency > recovery hacks.What’s your biggest sleep struggle right now? Drop it below Tag a friend surviving on fumes. @hubermanlab @thegarybrecka #SleepTips #SleepHealth #Biohacking #Productivity #CircadianRhythm #Wellness #HealthTips