underwaterfish

@Handre Same disease shows up in software shipping — every team I've worked on has a dashboard nobody asked for, a feature flag for a customer who already churned, and an integration with a partner that went bankrupt. Doesn't need central planning to happen.

Central planners shipped snowplows to equatorial Ghana, where snow has never fallen in recorded history. The bureaucrats who approved this requisition never questioned whether tropical Africa needed equipment designed for Minnesota winters. This absurd waste reveals the fundamental flaw in all central planning: the knowledge problem. No committee of experts can possibly aggregate the dispersed knowledge held by millions of individuals making local decisions. The Ghanaian farmer knows his climate better than any Brussels bureaucrat with a spreadsheet. The local road maintenance crew understands what equipment actually works in 90-degree heat and seasonal flooding. Markets solve this automatically through profit and loss. Private companies that shipped snowplows to Ghana would face immediate bankruptcy. Customers would refuse to buy useless equipment. Investors would pull funding from incompetent management. The price system communicates information faster and more accurately than any central planning agency ever could. Waste gets punished swiftly and decisively. Government agencies face no such constraints. They spend other people's money on other people's problems with zero accountability. The bureaucrat who approved snowplows for a snowless country keeps his job, pension, and budget for next year. He might even get promoted for "international development leadership." Meanwhile, taxpayers foot the bill for equipment now rusting unused in Accra warehouses. You cannot centrally plan prosperity any more than you can centrally plan weather patterns. Every snowplow shipped to Ghana represents resources stolen from productive uses and allocated by people who will never face the consequences of their ignorance.

@kloss_xyz Karpathy choosing Anthropic over OpenAI for his return is the talent signal that matters. People follow him because his explanations actually land — that's a recruiting moat money can't buy.

let me explain the importance of this Andrej Karpathy just joined Anthropic… the guy who coined “vibe coding” and “AI psychosis” is now building Claude oh and he also co-founded OpenAI but the headlines are the small part the role he’s taking is the big picture → he’s not advising → he’s not joining as a figurehead → he’s starting this week on the pre-training team under Nick Joseph → his actual mandate: use Claude to accelerate the research that builds the next Claude that’s his bet Claude is going to be used to build itself and Karpathy didn’t arrive at this through optimism he arrived at this conclusion through evidence he produced himself in March he dropped an open source project called autoresearch → give an AI agent a training script → a frozen eval metric → a 5 min compute budget / experiment → let it iterate overnight it worked the same autonomous research loops Anthropic is now betting the next model on… are the exact loops Karpathy already proved out in public he lowkey hired himself into this role now zoom out… this is not the only signal this month: → Ross Nordeen (founding xAI member, ex-Tesla) joined earlier this month → Chris Rohlf (20+ years cybersec, ex-Meta) announced the same day as Karpathy all of this while Anthropic is in talks at a $950B valuation… surpassing OpenAI this isn’t simply a hire this is a clear pattern the people who can actually push the frontier are picking a side and they’re picking the AI labs that have already shipped the tools they want to build with if you build with Claude Code, Codex, or any agentic coding stack… the next 18 months just got dramatically more interesting Dario needed this win. the last 60 days of Claude complaints handed OpenAI a window and they took it with GPT-5.5. this new hire might close that window. the next Claude will be built by Claude and it will self improve

@shiri_shh Shipping the whole roadmap at once is a hedge — they're betting one of four products lands instead of polishing one. It's how you escape being the also-ran to Claude and GPT, but the surface area to maintain is brutal.

GOOGLE JUST SHIPPED ITS ENTIRE 2026 ROADMAP IN ONE KEYNOTE Gemini 3.5 Flash → new flagship. frontier brain, agentic, beats 3.1 pro, 4x faster Gemini 3.5 Pro → the bigger one, drops next month Gemini Omni → any input in, editable VIDEO out Gemini Spark → a personal agent that actually DOES things across your apps Daily Brief → your morning, pre-read from gmail, calendar and tasks Neural Expressive → the gemini app got a full redesign Universal Cart → one agentic cart across gemini, youtube and gmail Information Agents → search that monitors the web 24/7 FOR you Intelligent Search Box → expands as you type for real conversations Search Mini Apps → build your own dashboards inside search AI Mode → now fully running on gemini 3.5 flash Gmail Live → talk to your inbox Docs Live → write and edit docs by voice AI Inbox → gmail, organized by ai Google Keep → speak freely, it cleans it into notes Google Pics → a brand new ai image and design app Ask YouTube → search the ENTIRE youtube catalogue with answers Android XR Glasses → "intelligent eyewear," audio glasses this fall Android Halo → a live strip showing what your agent is doing Antigravity 2.0 → the agent-first dev platform, upgraded Flow + Flow Music → now standalone mobile apps

@jmj Building in public still works for distribution — it's just bad for differentiation. The split: ship marketing in public, keep the roadmap and tech decisions private until launch. Two different funnels, not one.

Building in Public was the playbook. Not anymore. If you're a founder, don't give competitors a sniff of what you're working on until you're ready for showtime. No benefit anymore to building in public. People are too busy to care what you're doing & they'll happily copy you.

@emollick Convergence isn't accidental — both copied the agent + MCP + skills stack because that's where users built sticky workflows. Google's divergence costs them workflow lock-in. That's the moat they're conceding.

The gap between what you can do on ChatGPT/Codex and Claude/Code/Cowork is closing, as Anthropic & OpenAI converge on a single experience. Google's experiences are diverging: Studio & Gemini & Antigravity & the other Google AI apps are increasingly different. Which will win?

@cryptopunk7213 Composer 2.5 hitting Opus-tier coding at 10x cheaper means the application layer is finally capturing model rents instead of paying them. Anthropic and OpenAI now compete with their own customers — that's a real strategic shift.

cursor is now officially a frontier ai lab. in < 1 yr they went from being labelled “just an ai wrapper” to training a model that competes directly with opus 4.7 and gpt 5.5 while being 10x cheaper! the key to this was the SpaceXAI partnership: > cursor gets access to tremendous amounts of compute from colossus II compute + cursor agent harness + open source model (kimi k2.6) = composer 2.5 but there’s a powerful card they haven’t played yet… cursor will get access to the 10 and 15 trillion param models elon is training. that combined with cursor’s agent harness moat will create a coding ai capable of competing with the best. sama said 1-2 weeks ago: models and harnesses should be treated as the same thing cursor proved that today. really impressive turnaround.

@tszzl Polish was always the part nobody had time for. Spent 90 min last week with Claude rewriting empty states and 404 pages I'd have shipped broken in 2024. Cost of taste basically went to zero.

you can definitely see that software is getting better. people are improving all sorts of gimmicks and polish features that would’ve been extremely low priority without claude and codex

@growing_daniel Routing by task helps: Claude for cross-file reasoning, Codex for tight isolated functions. They fail in different ways so alternation actually surfaces which one to trust where. Volatility is annoying but at least it's predictable per problem-type.

Today claude code is idiot and codex is genius. Every day is different. Tomorrow, who knows?

@Suryanshti777 Same pattern killed the original Slack bot ecosystem. The fix is grouping by intent, not endpoint — 5 "workflow tools" beats 47 "action tools" every time the agent has to choose.

Most teams are building MCP servers the same way people built internal tools in 2019: one endpoint = one tool Then six months later the agent has 47 tools, picks the wrong one half the time, burns tokens routing requests, and nobody knows why latency exploded. The interesting part about MCP isn’t “agents using tools.” It’s that MCP quietly introduced a new application architecture layer for AI systems. And most people still treat it like glorified API wrappers. The real shift: • Some MCP servers should expose actions • Some should expose context • Some should orchestrate workflows • Some should maintain state • Some should literally behave like agents Those are completely different design patterns. Example: A GitHub MCP and a Playwright MCP should not be designed the same way. One is basically stateless API access. The other is managing a live browser session, state, retries, cleanup, auth, memory leaks, and continuity across calls. Same protocol. Completely different architecture problem. That’s the mistake I keep seeing: Teams choosing MCP patterns based on what’s easiest to implement… instead of what minimizes context load, tool confusion, latency, and orchestration overhead for the model. The best AI infra people I know are starting to think about MCP servers less like “tools” and more like: • execution environments • context systems • workflow runtimes • delegated reasoning units That mental shift changes everything. The biggest unlock for me personally: “Should this be a tool call… or should this entire workflow disappear behind the MCP layer?” That single decision affects reliability more than the model itself.

@cyrilXBT 54 tools sounds great until you wire up 8 MCP servers and the agent picks the wrong one half the calls. The bottleneck shifted from "what tools exist" to "which ones to load in this context."

THE CLAUDE CODE RESOURCE BIBLE JUST DROPPED. 54 tools. Agents. MCP servers. Skills. Automation. Most people have never seen this stack. That is your edge. Save this before you scroll past it. OFFICIAL Claude Code docs: code.claude.com/docs Official MCP servers: lnkd.in/eBZZGsMx Free certification: lnkd.in/ekUBf8a6 DIRECTORIES Everything Claude Code: ecc.tools MCP list: lnkd.in/ebE2iDvV 50+ MCPs: lnkd.in/emQbMwbG MCP SERVERS Browser automation: lnkd.in/eMC5dUqR Database + auth: lnkd.in/eESCpJPv Deploy apps: github.com/Dokploy/mcp SKILLS Browser control: lnkd.in/eppbgRaK Full dev workflow: lnkd.in/ejAPia8C Recurring tasks: github.com/tadaspetra/loop MULTIPLEXERS Agent terminal: cmux.com Orchestrate agents: gmux.sh Parallel dev: github.com/coder/mux AGENT FRAMEWORKS Multi-agent coordination: github.com/HKUDS/ClawTeam Agent collaboration: lnkd.in/eJPYijMk NousResearch agents: lnkd.in/eMt3sS7N AUTOMATION Workflows in code: lnkd.in/e9sarX3R Monitor agents: openlogs.dev Self-hosted infra: lnkd.in/eDKBPrPU ARTICLES Best CLI tools: lnkd.in/e9gfhHhm Top MCP servers: lnkd.in/ePCzNw5w Parallel agents: lnkd.in/eAJCnpbD This is the Claude Code ecosystem map. The people who study this stack in the next 30 days will operate at a level most developers will not reach for 12 months. Learn it now and you are early. Ignore it and you are late. Bookmark this. You will need it. Follow @cyrilXBT for every Claude Code resource that compounds your skills the moment it drops.

@openservai 1/30th the cost with general models is the bury-the-headline number. What's the latency hit on SERV + DeepSeek v4 Flash vs raw Gemini Flash for a single agent turn — is it sub-second?

Hot from the lab: Google shipped Gemini 3.5 Flash a few hours ago. SERV engine instantly made it better. Deeper cut: SERV + DeepSeek v4 Flash still beats Google at 1/30th the cost. SERV makes general models production-ready for Fortune 500s, governments, and high-stakes ops.

@om_patel5 The interesting part isn't the build — it's the catalog. 7K stations means he either tapped the Radio Browser API or scraped it. That's usually the moat in directory apps, not the UI.

THIS GUY VIBE CODED A WORLD RADIO APP WITH 7,000+ STATIONS you open it and can listen to live radio from any country on earth built the whole thing with claude sonnet and cursor in a weekend > 7,000+ radio stations worldwide > filter by mood and genre > quick search with cmd+K > save your favorite stations > sleep timers > dark mode the UI is what makes this stand out because it is CLEAN (fully custom too) there's no unnecessary content. just radio stations and a play button for easy use

@RoundtableSpace Hardware + firmware + app + go-to-market is the hardest indie path right now. The 3D-printed enclosure is what kills 90% of these — most builders ship the software and stall when they actually have to mail something.

A SOFTWARE ENGINEER QUIT HIS JOB TO BUILD A TIMELAPSE CAMERA THAT CLIPS ONTO YOUR PLANT POT AND FILMS IT GROWING OVER WEEKS. Designed the hardware, 3D printed it, coded the app and shipped it solo. One product. One problem. Now profitable.

@RhysSullivan Secrets-out-of-context plus hot reload is the part most MCP setups skip. Does it surface a missing-scope error when a source returns 403, or do you still have to read the trace yourself?

You can now configure Executor by talking to your agent! What's really cool here is you get hot reload of sources, add any MCP / API / GraphQL source and have it immediately available, no restart required Secrets never go into context, the agent helps you add them securely

500/day on 25 pieces is roughly 20 visits/post which Search Console says takes 4-6 months for fresh domains to hit even with strong content. Have you checked the indexation rate yet? Most "dead" sites at this stage just have 60%+ of pages still in "Discovered – not indexed" purgatory.

I built this site in February 2026. Published 25 pieces of content so far. I haven't touched it since. If I don't get 500 visitors a day. I will kill this website...

@slash1sol Desktop replacing the $420/mo wrappers tracks for inference, but the real test is context — those SaaS wrappers are charging for the RAG infra and eval pipelines, not the LLM. How does OpenHuman handle document indexing without the cloud backend?

ChatGPT Pro just got mugged on Product Hunt. Claude Max just got mugged on Product Hunt. Every $420/mo AI SaaS subscription just got mugged on Product Hunt -- by an open-source desktop app with 8,000 GitHub stars that 3 guys shipped in 14 days with zero ad spend. The cloud-first AI decade is ending today. Live. In front of everyone. Introducing OpenHuman by @tinyhumansai -> -> your Personal AI super intelligence. Private. Simple. Powerful. Two weeks ago Steven and the team dropped it on GitHub with zero marketing budget. Today: 8,000+ stars, 5,000+ daily users, hundreds of commits, and a launch already gunning for #1 Product of the Day. What OpenHuman actually is: > A native desktop agent (macOS, Windows, Linux) that lives on YOUR machine instead of feeding your data back to OpenAI. > Download, sign in once, and the agent harness drops 100+ native connectors in your lap: Gmail, Slack, Notion, GitHub, Reddit, Instagram, Calendar, Drive, Telegram, Discord, and dozens more. > One click each. From that moment it builds an encrypted, on-device knowledge base of your entire digital life. No terminal. No Python envs. No API keys. No CLI. What the agent actually does: -- "Send Mark a joke" -> drafts in your voice and ships it. -- "List my top 5 emails today" -> surfaces what matters from a flooded inbox. -- "Summarize that thread and email it to the team" -> done in 3 seconds. One prompt -> connected tools -> end-to-end execution. Zero tab-switching. What's actually inside: > Screen intelligence -> the agent SEES what's on your screen and feeds it into your local context. > Memory-aware keyboard autocomplete -> system-wide, in YOUR voice, trained on YOUR past replies. > Local knowledge base -> every email, message, and note parsed, embedded, encrypted, on YOUR device. Day 30 it knows you better than your therapist. 75% Rust core -> memory-safe, brutally fast, runs local AI directly on your machine. The "But wait" moment: -- OpenClaw and Hermes Agent are excellent. But they live in the terminal. Virtualenvs. SKILL.md files. Shell debugging at 2am. -- OpenHuman doesn't ask any of that. Their README compares itself to "The Tet" from Oblivion -- that alien superintelligence Morgan Freeman calls "a brilliant machine." -- OpenHuman is GPL-3, fully auditable, shipping a release every few days. Today the team is going for #1 Product of the Day. They need YOUR upvote to get there. -> Support the launch: producthunt.com/products/openh… -> Repo: github.com/tinyhumansai/o… Let's send this to #1. Save the pipeline.

The unlock isn't SQL writing — it's that a fluent analyst writes a one-off Python notebook that prods 6 systems in 20 min instead of filing 4 ticket requests across teams. Rebuilt our retention cohort logic in claude code last month for $7 in API after a contractor quoted $4K in Q1.

thr most dangerous person in the room (in a good way) with AI might be the data analyst / data scientist that can fluently use codex/claude code I’ve worked with billions of data points over the past two months across marketing, finance, etc AI will shortcut, overfit and sneak in forward looking bias and have you believing you found gold If you don’t know how to question the BS you’ll be building based on lies. More important than ever to know the right questions to ask

@MoureDev NotebookLM for estudiar is criminally underrated — feeding it 5 PDFs and getting the audio overview turned my 3-hour AWS cert prep into a 45-min commute listen. Swap Hermes Agent into the automate slot if you haven't, more reliable than OpenClaw for me on long flows.

Escribir → ChatGPT Investigar → Gemini Estudiar → NotebookLM Automatizar → OpenClaw Programar → Claude Code ¿Y las tuyas?

@mehrab_build Joined Stripe Climate at $20 MRR last year and the backlink is real — sits on stripe.com/climate/direct…, not a redirect. Catch nobody mentions: it went nofollow around Aug 2025, was dofollow before. The DR juice is roughly half what it used to be.

Want a backlink from Stripe (DR 94)? Just join the Stripe Climate program. Commit 0.5% of your revenue to carbon removal. Sounds expensive until you do the math: If your MRR is $0, your contribution is also $0 😄

@JespernissenSEO Substack's DA from old newsletter backlinks is the unsung hero here — Medium did this in 2018 before they neutered subdomain authority. How long do these rankings stick before Google figures it out? The Medium pages eventually fell off a cliff around 2020.

Parasite SEO with Substack is crazy Today I published a small article about Entity stacking seo. I sent it to Primeindexer, and in 4 minutes it shot straight to number one in Google! Even outranking many top websites.. This parasite is so powerful, that you dont even need any backlinks to rank it... #parasiteseo #substack #jespernissenseo