one play i ran all day on .si: collect ai products with real traction, yc and a16z portfolios, new seed rounds, product hunt winners swap .ai to .si and register whatever is free same with ai inside the name, https://calai.app/ -> https://calsi.app/ wonder which one follows SpaceXSi next

Elon already followed my move with .si domains😂 be like Elon 🫵 https://twitter.com/DeRonin_/status/2106767936501268753/photo/1

buy .si domains with names from 70s, 80s and 90s futuristic sci-fi ai founders keep naming their companies after that era https://recursive.si/ sold for $20k this year, and most of those names are still sitting at morning’s coffee price

2nd part is coming soon... it's going to cover all resources and practices which will boost you from junior-level engineer to mid one turn on notis to don't miss it 🔔 https://twitter.com/DeRonin_/status/2106423598411817418/photo/1

you can't even imagine what's coming soon 👀 in the last 6 months i went through 5 of the 6 practical months from my own roadmap so it's time for part 2 this one covers the next 12 months, from "i can build things" to mid level AI engineer the part nobody writes about: getting your first thing working is the easy half the hard half is owning it when it breaks in production and nobody can say why that's what part 2 is about soon.

if i had 6 months to become an AI engineer from zero i'd do this Stage 1: the boring foundations - learn: python, git, the terminal, http and json, basic sql, fastapi - practice: build a cli tool that calls a public api and writes clean json to disk - why: ai engineering is software engineering first, and every stage below assumes you can ship a working endpoint Stage 2: llm app development - learn: prompting, structured outputs with pydantic, tool calling, streaming, conversation state, token cost - practice: build an invoice parser that returns a typed python object, then give a model 3 tools and let it pick - why: this is the actual job, and most people skip straight past it to agents Stage 3: rag, properly - learn: embeddings, chunking, vector stores, metadata filtering, reranking - practice: ingest 20 pdfs and ship an endpoint that returns a cited answer from the top 5 chunks - why: most rag failures are retrieval failures, so whoever can debug retrieval gets paid Stage 4: agents, workflows and evals - learn: the agent loop, tool descriptions, state, retries, and when an agent is the wrong call - practice: build one agent with no framework at all, just the api, 3 tools and a while loop - why: a chain of 3 fixed calls beats an agent that might make 3 calls, and knowing which one you need is the senior part Stage 5: deployment and reliability - learn: docker, background jobs, auth, rate limits, tracing with langfuse, caching, cost control - practice: containerize your rag app with a vector db and redis, then put a hard daily spend cap on it - why: this is where most people stall, they can demo but they can't survive real traffic at 2am Stage 6: pick one lane - learn: product engineering, applied ml, or business automation - practice: ship 2-3 projects a stranger can open and use without you explaining anything - why: six months makes you dangerous in one direction, not average in three you don't need to train models you need to build systems on top of them, and that's what companies are actually paying for i wrote the full version of this with every resource and link for each stage so, if you're beginner and skipped this article 6 months ago, read below:

which board should you actually start with? Arduino vs ESP32 vs Raspberry Pi 4 months into robotics as a hobby, here's what i'd tell myself on day one: 1. start with arduino use it when you're learning: - motors - servos - sensors - basic circuits atmega328p, 16 mhz, 2 kb of ram, and 5v logic so most hobby parts plug straight in when your code runs, nothing interrupts it, and that makes timing easy to reason about best place to learn the fundamentals without debugging an operating system at the same time you outgrow it the day you add wifi or a camera 2. move to esp32 use it when you want: - wifi - bluetooth - wireless control - anything that reports back to your phone dual core at 240 mhz, 520 kb ram, radios on the chip, around $5 a board more connectivity and still hands on, so you keep writing the same kind of code two things that cost me a weekend each: 3.3v logic, so your 5v sensors need a level shifter or they read garbage and on the original esp32 the adc2 pins stop reading the moment wifi turns on 3. go for raspberry pi use it when your robot needs: - computer vision - ai and ml - ros 2 - heavy processing now you're running a computer inside your robot, with all of the upside and all of the cost linux decides when your code runs, so pwm timing jitters and servos twitch and pulling the power without a shutdown corrupts the sd card, which i learned twice also worth knowing, raspberry makes a real microcontroller too, the pico, about $4 so the quick version: never wired anything before -> arduino want wifi and the best price -> esp32 need vision or a camera -> raspberry pi and yes, you can use more than one arduino -> control esp32 -> connect raspberry pi -> compute that combination is what took me 4 months to understand the pi is the brain and the microcontroller is the spine, and a real robot wants both also, if you want to learn robotics engineering for 6 months, read the article below:

Rough tier list of free github skills i'd actually install in 2026 https://twitter.com/DeRonin_/status/2105656626053288415/photo/1

Rough tier list of free github skills i'd actually install in 2026 https://twitter.com/DeRonin_/status/2105654464325796300/photo/1

I'm in an existential crisis rn, and I'm sure of one thing Burnout happens when you work hard for a sense of meaning that society sold you, and it isn't enough Yeah, sure I am pretty happy that previous month I did over $70k net income But even with decent money, I feel no real fulfillment, like I'm living an impostor's life I feel myself much better when I just figure out in electronics + robotics, but by some reason I choose the path which doesn’t make me happy I follow the goals and trends which are created by somebody, but at this time my rational mind is turning off I start discovering each new model as a crazy, measure the efficiency of it, try to integrate to my workflow, but basically it doesn’t change anything significantly Why I do this? Since I read the forums, I read these fake AI sloppy tweets where people get +1000% output and spend 400x less tokens and I have in mind: “Wow, I am out the race if I don’t try it now, I will be underclass if I will not follow the crowd into the race of efficiency” But into the reality everything what happens around AI nowadays is just a white noise And I realize that’s in 5-10 years, the core problem of humanity will be the depression Since I feel on myself and track it through my frens, that our physiology is totally not adapted to such fast changing world And in this race of efficiency and trying to create something new, we just lose ourselves I don’t think that ASI will make us happier, because behind this stays a huge crisis which nobody can solve and heads of our world just saying: “We think in 2036 year, humanity will not need money, everybody is just going to live as average, nothing scary” Idk if anybody feels the same, but I am almost done to live as imposter… I want to be a happy boy in my 20s, I want to grow step by step, but I can’t stay focused in one niche for years, since probably it will be just replaced I don’t this step-by-step level up through the life where I have exact plan for 5-10 years and I just stay aligned with this And it breaks me…

opus 5.5 + boreal-h3 is insane... here's how to make such as video ad in about 2 minutes: - generate your character once, front on, plain background, high res - pick the video format you want to hit - open model playground on creatify - upload your character as the person reference - upload your reference clip as the video reference - lock your opening frame mid-gesture with keyframe control, that first half second is the whole hook - prompt: "use Image1 for the character only. put that character into Video1, same framing, same lighting, same energy. keep her face, hair and wardrobe identical for the entire shot. match the color grading to Video1. single continuous take, no camera moves." - generate ad agent runs on opus 5.5 and hands you five versions of the spoken line, which is the part that actually decides whether the video works then burn your caption in, two lines, lower third, and ship it that's the whole thing model playground, go make one

My top 5 open-source & local LLM profiles: - @MiaAI_lab - @TheAhmadOsman - @jun_song - @ggerganov - @danielhanchen they literally give me insides which are 6 months ahead trends

don't use claude or gpt use them like you're leaving build your own software using their tokens so when their tokens become too expensive, you can switch to your own GPUs your compute, the one you control, so you can't be controlled here's the workflow that keeps you independent: 1. every model call goes through one file you wrote, https://llm.py/, instead of anthropic or openai imported all over your code, so changing models is one line and litellm handles the provider differences for you 2. prompts live in a prompts folder as .md files and never sit inside the code, so you can rewrite one for a different model yourself without a developer 3. that same file appends one json line per call with the task name, model, input, output and cost, and six months of it is the only proof you will ever have that a cheaper model can do your job 4. pull fifty real inputs out of those logs, write the correct answer for each, save it as evals/cases.jsonl, and write a script that prints a pass rate, so every model question becomes a ten minute test 5. fire a second call to a free open model on the same input, never serve it, just log what it said, because it costs almost nothing and you always know how far behind the free option is 6. tag every call site, read the logs, and move over the tasks where the small model already ties, which is usually sorting, extraction, formatting and short summaries, so your bill drops the same week 7. rent one GPU by the hour, run vllm serve, and it comes up on localhost:8000 speaking the openai api, so you point https://t.co/PiJ9LXBLOf at it, run your fifty cases, and you know in one afternoon you need to control your stack or you just become a slave to whoever does

wow, thanks a lot for the morning gift Tibo got 62,500 tokens on my account (as 99% people i can see) roughly $2,500 in value what's equivalent to 12 months of Pro Plan waaay better than one more reset (looks like they force me to remove claude from my daily workflow kek) https://twitter.com/DeRonin_/status/2105252530397811096/photo/1

AHAHAHAHAH follow https://dot.com/ website kek that's what i call a genius marketing Elon, one day i will be like you, just wait for me

@OpenAI pleeease, leave us a space to make money on products as young enterprenuers.. @sama, put us on your place at 20s...

they just killed at least 12 products built by my frens founders...

Opus 5.5 is insane for running a solo agency... send this 3-pages prompt to Claude Opus 5.5 and it turns into the seven departments that used to be seven salaries then read the full system below, the one running my agency at $78k MRR with zero employees: https://twitter.com/DeRonin_/status/2104517555507388837/photo/1

Open the article Cmd A Open Claude Opus 5.5 Cmd V Enter "Build [Type] Solo Agency" WIN

here's everything i'm using in AI right now THINKING AND PLANNING - Claude Fable 5.1 (long horizon plans, the one i argue with before anything gets built) - Claude Opus 5.5 (daily driver, fable level work at 40% less to run) - GPT-6 Astra (second opinion on hard calls, and auditing plans the others wrote) - Grok 4.7 (fast takes, and anything that needs live context from X) - AGENTS.md (openai/agents.md, one instruction file every agent now reads) BUILDING - Claude Code (main harness, everything else plugs into it) - Codex (terminal agent, great when i want a second pair of hands on the same repo) - Cursor (editor plus cloud agents for the long boring refactors) - Gemini CLI (throwaway jobs i don't want to spend good tokens on) - Playwright MCP (the agent opens a real browser and checks its own work) CHEAP DECISIONS AT VOLUME - Jev (classify, score, route.. the call you make 10,000 times a day, not twice) - Gemini 3.8 Flash (vision and long context at flash pricing) - GPT-6 Luna (bulk cleanup where quality barely matters) MEMORY & CONTEXT *you can search for all of them in github (just copy the full naming) - thedotmack/claude-mem (context that survives between sessions) - mem0ai/mem0 (memory layer when the agent needs to remember a person, not a repo) - getzep/graphiti (temporal knowledge graph, for facts that change over time) - headroomlabs-ai/headroom + rtk-ai/rtk (compress tool output before it eats your window) RESEARCH & GROUNDING - last30days (reddit, X, youtube, HN, polymarket, then one grounded summary) - firecrawl (any site into clean markdown the agent can actually read) - context7 (live library docs, so it stops inventing APIs that never existed) - Xquik (X data through an API instead of scraping with your own cookies) DESIGN, FRONTEND & SHIPPING - Claude Fable 5.1 + Opus 5.5 (UI polish, the last 10% that makes it look intentional) - Figma (where the decisions still get made) - Rive (interactive motion) - Supabase (backend) + Vercel (hosting) IMAGES - ChatGPT Images 2.5 (Flare for volume, Sunburst when it's the final one) - Nano Banana 2 / Nano Banana Pro (4K, and text inside the image renders correctly) - Midjourney V8.2 (still nobody close on taste) - Grok Imagine Image 2.0 (dense layouts where small text has to stay legible) - Ideogram 4.0 + Recraft V4.1 (typography, vectors, logos) VIDEO - Gemini Omni Flash (google's own default now, conversational editing, native audio) - Seedance 2.5 (30 seconds in one pass, best thing for anything with a story) - MiniMax H3 (15s at 2K, strongest at doing exactly what you asked) - Kling 3.0 (4K multi shot sequences) - Runway Gen-4.5 (complex camera direction in a single prompt) - Higgsfield (motion transfer, plus one place to reach most of the above) SOUND & VOICE - Suno v6 (music, and you can now rewrite specific sections by text) - ElevenLabs Eleven v3 (voices and SFX) + Eleven Music + Scribe v2 for transcription 3D, AVATARS & POST - Meshy 7.1 + Tripo H3.1 (text or image into a usable mesh) - HeyGen Avatar V (15 second recording, holds identity across a long video) - DaVinci Resolve 21.1 (21.1 lets Claude Code drive the timeline, that's the real headline) - OpusClip (long form into shorts) + Descript (edit video by editing the transcript) RUNNING WHILE I SLEEP - Hermes Agent by Nous Research (self hosted, writes its own skills from experience) CAPTURE & KNOWLEDGE - Wispr Flow (dictation that fixes you mid sentence instead of typing you literally) - Granola (meeting notes) - Obsidian (knowledge base, no first party AI, all plugins, still the best) - Notion 3.7 (where the agents read from when they need team context) REPOS WHICH I USE MAINLY - obra/superpowers - anthropics/skills - mvanhorn/last30days-skill - ai-hero-dev/ai-hero (matt pocock) - hesreallyhim/awesome-claude-code - ComposioHQ/awesome-claude-skills - punkpeye/awesome-mcp-servers - openai/agents.md most people are collecting tools, the ones winning picked 5 and went deep anything missing here? 🤣 average spend on tokens & usage for such as toolkit ~$1,300 monthly drop what you're running, i'll test it and report back

why i'd build an AI-native service over SaaS for the next 12-24 months: everyone is shipping a wrapper now... that's the whole problem you can't build a moat out of something a competitor rebuilds over a weekend or even better, it will get replaced by new frontier model at all selling the finished work is the opposite: > the customer never wanted the tool, they wanted their books closed, their contract checked, their claim filed (saas has been a help tool, but you sell complex solution) > the budget already exists, you're replacing a line item they pay every month instead of inventing a category > every model release makes your delivery cheaper, instead of making your product less necessary > you get paid from job one, no year of guessing at product market fit > the moat is the rulebook, every mistake you caught and wrote down, and nobody can download that > your overhead stops growing with every client once intake is a form and quality is a list here's a few case studies of companies already doing it: 1. https://harvey.ai/ - $100M to $400M ARR - 13 months - the research and drafting an associate used to do 2. https://sierra.ai/ - $100M to $200M ARR - 7 months - the support queue, billed at about $1.50 per resolved ticket 3. https://fin.ai/ - $1M to about $100M ARR - 3 years - first line support at $0.99 per solved conversation none of them sold software to do the job, they sold the job what it gives to you: - you're paid to learn what to build, instead of raising money to guess - one person carries the client load that used to need a department, i replaced four of mine last year - margins land closer to software than to an agency, because your cost of one more unit is nearly zero - you own an asset that isn't a seat license, and it gets more valuable with every job you run everyone is building tools and almost NOBODY is selling the work same 24 months, completely different odds and you don't have to bet the house on it (it will cost to you literally LLMs expenses) keep whatever pays you now, pick one deliverable your clients already ask for, do it by hand for five of them and write down every mistake you catch i want to put the bet of my business here next years

SOURCE: https://ai.gov.uk/knowledge-hub/how-to/ethics/

i'm really sorry for the guys who live in the UK... they've made "rules" on how to use AI: first you explain why you want to use it then if they approve you, it's ONLY short prompts, as few interactions as possible, to cut energy use the EU is probably next to adapt it...

+ vibe https://twitter.com/DeRonin_/status/2103101316390441418/photo/1

stop renting your business's back office 1 signup.. and you've got a live website, a booking calendar and marketing going out, all on one platform you actually own Durable just shipped all of it as your AI business builder.. you set it up once, it runs the rest: [ by the end you'll have ]: * a website live in minutes, built from nothing but "here's what I do" * discovery handled, SEO and the Google Business listing, the stuff that actually gets you found * marketing content and social posts generated for you, no agency, no retainer * bookings and customer follow-up running on autopilot [ the workflow ]: tell Durable what the business is, cleaning, landscaping, detailing, consulting, training, it doesn't matter AI builds the site, writes the copy, sets up the booking flow, no dev, no designer it keeps generating marketing assets and posts, so you're not starting from a blank page every week customer management and back office run in the same place, so nothing lives in six different tools IDEA: this isn't a starter kit for "anyone with an idea" anymore, it's built for real service businesses that already have clients and just need the operations to catch up set it up once, and it's the difference between chasing admin and actually running the business most people are still cobbling this together from five subscriptions and a VA.. you can just run one platform and they're giving it away free to 100 people right now, instructions are in the post You have no more excuses

How to get a job as a Robotics Engineer: robotics is the one frontier field where the entry level still doesn't ask for a degree 1 in 5 robotics jobs posted right now is a technician role, and most of them only need a certificate or a two year degree so here's my workflow for getting a job as a robotics engineer: 1: pick one direction and drop the other two robot learning pays the most and everyone wants it autonomy and mobile robotics has the most openings by a wide margin embedded and mechatronics is boring and you'll never be out of work 2: build things that moved, and write down the numbers recruiters here open your github before they read your CV what gets you through: - a commit history where you're visibly fixing things, not one big final push - real hardware with reliability numbers, sim doesn't count - datasets on the lerobot hub - commits to ROS 2, Nav2, MoveIt, Isaac Lab, LeRobot 3: build your hardware in public the best way to prove your knowledge is the recognition of your skill from masses just build anything you want and share it on X/IG/YT good example of the guy who gets offers from Tier-S level robotics companies (you can copy his strategy): go to IG type in search: "aykhanium" and check which rubrics he does to be recognized repeat. 4: write down what broke everyone posts the demo that worked almost nobody writes up the four things that failed first and how they found each one that's the part you can't fake from a tutorial, and it's the part that survives the third question in an interview cheat-codes to stand out and get into the top 1%: 1. take the shift nobody wants teleoperation pays around $28 an hour. figure posted a humanoid robot operator at $25 to $35 with no degree asked for it's just driving a robot around a lab but it puts you inside a frontier company with a badge on, and now you're a person they know instead of a CV in a pile 2. sell the integration, not the robot the arm is about 25% of what a project costs the other 75% is engineering, safety, and making everything talk to each other a $35k arm turns into an $80k system, and that $45k is your job 3. go where nobody's competing 66% of robotics projects get delayed by certification functional safety pays well and almost nobody bothers learning it 4. C++ and Python, both every Figure and Skild listing I read asks for both, not one 5. delete the ROS 1 from your repos recruiters call it out by name as a red flag if there's still a catkin_make sitting in your github, that's the first thing they see main insight: $47.4B went into physical AI in the first half of 2026 the BLS still projects 1 to 2% job growth in the occupation the money showed up years before the headcount will so the people getting in right now aren't winning interviews, they're walking through the technician door and moving sideways once they're inside that's like to be hired in OpenAI in 2019... and one more thing nothing you learn here goes stale a PID loop works the same as it did in 1990, and the arm you fix this weekend teaches you something you'll still use in ten years you can't say that about anything else in AI right now...

nah this is actually insane pocket fm nearly shut down it's now at $500M ARR → 200M+ listeners → 135 minutes a day per US listener → tiktok gets 53.8 → 70% of the money comes from america here's the actual path: they pivoted 10 times in 2 years podcasts. audiobooks. music. all dead. the try after that was serialized fiction sold one episode at a time, $0.99 at the cliffhanger $21M → $500M in under 3 years $198M → $500M in the last year alone then they removed the thing that caps every content company the cost of making the content elevenlabs partnership audio production down 90% across 30,000+ hours and today they removed the other cap sherpa, an ai writing partner trained on 100M+ hours of retention, coin spend and drop off data one line idea in a full season out it has no idea what good writing is it knows the exact second someone stopped listening, and the exact cliffhanger that made them pay 75,000+ series 100,000+ hours of content $33M already paid out to the people writing it the lesson isn't "use ai" everyone pointed ai at their marketing pocket fm pointed it at the two things actually capping them how fast they could record how fast they could write demand was never their problem supply was go find out which one you're short on

founder to founder you should be watching how Ben Cera operates polsia is a $250M company and he runs it while traveling the world not after the exit during and before you call it a flex, look at what he's actually doing: 1. he treats rest as maintenance, not as a reward most founders earn their rest, hit the milestone, then allow yourself a weekend he takes it before he's empty, because the version of him who hasn't left his desk in five weeks makes worse calls, and one bad call costs more than a week off ever will 2. he says the quiet part out loud founder twitter runs on 4 hours of sleep and a suffering complex he's ambitious, intense, obviously obsessed, and he still says it plainly: leave, see friends, reset, come back sharper that's not softness, that's someone who did the math on a ten year mission 3. he's building a company he can survive building if polsia is meant to get massive, he can't operate like the next 6 weeks decide everything the company needs an operating system that survives years, and so does the founder running it what to take from this: your output over 3 years matters more than your output this month book the break before you need it, not after you break perspective is an input to your decisions, not a prize for making them most of the field quits by year 5, so staying sane is a competitive advantage MUST WATCH ↓

4 things to do when you get access to Jev: 1. don't replace your model, put Jev in front of it Jev makes the small choice first, and your expensive model only runs when it's really needed one email tool swapped 2 AI calls per email for 1 Jev call, and kept all their normal safety checks on a 120 ticket test: 42 seconds with Jev in front → 22 minutes without it cost: $0.0003 → $0.059 the person who ran that test said his slow version was slowed down even more by retry errors, so the real gap is smaller how to do it: - look through your code for AI calls that only choose something and never write text - write the list of possible answers in your own code, don't let the model invent them - send that list to Jev, and keep your old call as a backup - put it behind an on/off switch so you can undo it in one line - run both for a week and compare them before you delete the old one you're not rebuilding your product, you're replacing one call 2. ask 5 small questions instead of 1 big one someone tested Jev on 2,000 phishing emails ask it one big question and it gets 89.4% a simple 2 line text rule gets 91.8% Claude Haiku 4.5 asked the same question gets 94.2% so when it has to give the final answer alone, Jev loses even to a text rule then he asked 5 small questions instead, and added the answers up in his own code 95.0%, the best score in the whole test SAME MODEL, SAME EMAILS how to do it: - take the big question you were about to ask - write down the 5 things a person checks before they answer it - ask each one separately, all in the same request, they run at the same time and cost almost nothing extra - add the answers up in your own code, with your own weights - test those weights on 100 examples where you already know the right answer - change the order of your options and run it again, reordering 4 options changed 7 answers out of 120 Jev is good at noticing things and bad at making the final call so keep the final call in your code 3. never ask Jev if it can answer on 120 test tickets, any question like "do you have enough info?" said yes on 85% of them a simple "need more info" flag said yes on 119 out of 120 tickets, and it then answered those tickets about 87% correctly it doesn't know what it doesn't know how to do it: - delete any question like "can you answer this" or "is there enough context" - make every option a real action your code can run - always read the confidence score, not just the answer - choose your limit by risk: low for reading data, 0.85+ for anything you cannot undo - send everything below the limit to a human or to your big model in that same test, using 0.8 as the limit passed 30 cases to a human, and 93% of them were passed for the right reason 4. run it quietly next to what you already have there's already a langchain package published and a pydantic-ai adapter being reviewed, so this is closer to a settings change than a rebuild how to do it: - leave your current system in charge, it still makes every decision - send the same input to Jev too, and throw its answer away - save both answers plus Jev's confidence score into one table - after a week, look only at the rows where the two disagreed - switch over only for the cases where Jev was right that list of disagreements becomes your test set, and your normal traffic builds it for free and don't use it when you already have labelled data if the question never changes and you have examples to train on, a small model you host yourself beats Jev on speed and price, and needs no API key at all Jev wins when you have no labelled data and the question keeps changing so use it where the list of answers is short and the question is boring that's most of your agent anyway everything above comes from other people's public tests, not from production, because the model is only 4 days old i'm just sharing what i find while i test this and try to make the work in my own company faster and cheaper tomorrow i'll show you what happened when i put Jev in front of the meta ads work we do for one big client and most of you are still on the waitlist anyway so start with step one, because it needs no access and no API key at all: find the calls in your code that were never writing tasks in the first place gl https://x.com/CompleteSkeptic/status/2099925684256899543/video/1

4 things to do when you get access to Jev: 1. don't replace your model, put Jev in front of it Jev makes the small choice first, and your expensive model only runs when it's really needed one email tool swapped 2 AI calls per email for 1 Jev call, and kept all their normal safety checks on a 120 ticket test: 42 seconds with Jev in front → 22 minutes without it cost: $0.0003 → $0.059 the person who ran that test said his slow version was slowed down even more by retry errors, so the real gap is smaller how to do it: - look through your code for AI calls that only choose something and never write text - write the list of possible answers in your own code, don't let the model invent them - send that list to Jev, and keep your old call as a backup - put it behind an on/off switch so you can undo it in one line - run both for a week and compare them before you delete the old one you're not rebuilding your product, you're replacing one call 2. ask 5 small questions instead of 1 big one someone tested Jev on 2,000 phishing emails ask it one big question and it gets 89.4% a simple 2 line text rule gets 91.8% Claude Haiku 4.5 asked the same question gets 94.2% so when it has to give the final answer alone, Jev loses even to a text rule then he asked 5 small questions instead, and added the answers up in his own code 95.0%, the best score in the whole test SAME MODEL, SAME EMAILS how to do it: - take the big question you were about to ask - write down the 5 things a person checks before they answer it - ask each one separately, all in the same request, they run at the same time and cost almost nothing extra - add the answers up in your own code, with your own weights - test those weights on 100 examples where you already know the right answer - change the order of your options and run it again, reordering 4 options changed 7 answers out of 120 Jev is good at noticing things and bad at making the final call so keep the final call in your code 3. never ask Jev if it can answer on 120 test tickets, any question like "do you have enough info?" said yes on 85% of them a simple "need more info" flag said yes on 119 out of 120 tickets, and it then answered those tickets about 87% correctly it doesn't know what it doesn't know how to do it: - delete any question like "can you answer this" or "is there enough context" - make every option a real action your code can run - always read the confidence score, not just the answer - choose your limit by risk: low for reading data, 0.85+ for anything you cannot undo - send everything below the limit to a human or to your big model in that same test, using 0.8 as the limit passed 30 cases to a human, and 93% of them were passed for the right reason 4. run it quietly next to what you already have there's already a langchain package published and a pydantic-ai adapter being reviewed, so this is closer to a settings change than a rebuild how to do it: - leave your current system in charge, it still makes every decision - send the same input to Jev too, and throw its answer away - save both answers plus Jev's confidence score into one table - after a week, look only at the rows where the two disagreed - switch over only for the cases where Jev was right that list of disagreements becomes your test set, and your normal traffic builds it for free and don't use it when you already have labelled data if the question never changes and you have examples to train on, a small model you host yourself beats Jev on speed and price, and needs no API key at all Jev wins when you have no labelled data and the question keeps changing so use it where the list of answers is short and the question is boring that's most of your agent anyway everything above comes from other people's public tests, not from production, because the model is only 4 days old i'm just sharing what i find while i test this and try to make the work in my own company faster and cheaper tomorrow i'll show you what happened when i put Jev in front of the meta ads work we do for one big client and most of you are still on the waitlist anyway so start with step one, because it needs no access and no API key at all: find the calls in your code that were never writing tasks in the first place gl

How to use Jev, and where it actually gives you the 100x: setup takes 10 minutes: 1. join the waitlist, people are getting approved same day 🔗 https://typesafe.ai/ 2. install the official skill so your agent writes correct calls: - npx skills add typesafe-ai/skills --skill typesafe-ai on Claude Code it's two commands, the marketplace add on its own doesn't install anything: - claude plugin marketplace add typesafe-ai/skills - claude plugin install typesafe@typesafe-ai 3. create an API key in the dashboard 4. in your prompt just say: "use the TypeSafe skill" now the part nobody is posting: the 100x isn't the model, it's where you put it you don't get it by swapping your LLM for Jev you get it by deleting the calls that never needed a language model open your agent and find every call that just picks something: > which tool next > is this spam > is this chunk relevant > does this need a human > is this diff risky none of those are writing tasks they're if statements you outsourced to a frontier model here's the upgrade, in order: 1. replace each one with a typed question Choice picks from up to 255 options, Score places it on a 2-10 level scale, Noul returns a raw 0-1 2. batch them questions in one call run in parallel and barely move the latency, and output tokens are free so ask every question you might need, including the ones you'll throw away 3. threshold on confidence, not on the answer under 0.5 escalate to a big model or a human 0.85+ before anything irreversible 4. never let it invent options build the candidate list in code, from the DOM, the retriever, the tool trace then let it pick 5. put it in the loop, not next to it router picks the cheap model, gate checks the tool call before it runs, judge verifies the output after that's where the heaviest calls in your agent are hiding 6. start with compaction tonight score every tool call, drop the dead ones, keep the survivors verbatim instead of a lossy summary lowest effort win available and you'll see it on tomorrow's bill the honest part: text only right now, no images, no audio and on broad benchmarks it loses to frontier models but somebody ran 18,514 emails through it zero-shot and got 98.33% against a TF-IDF classifier trained on 14,800 labelled examples that got 98.39% no training data, $1.12 total it wins on narrow, well specified decisions which is most of what your agent is actually doing all day today gonna share use case how i integrated it to content creation and how i find winning meta ads now in a seconds...

watched a solo founder go $0 → $10m ARR in 14 months with zero employees, and his ENTIRE founder brand built off letting the product talk about him Here's the playbook (step by step): BEFORE YOU POST ANYTHING > write the one sentence your product promises, then live it in public every week > name the character after the product, give it one trait and one obsession > write a 10 line doc: how it talks, 3 things it always notices, 3 things it never says > paste that doc into every prompt > lock one font, one colour, one voice, never change them > new X account with the product name if people can't recognise it in one frame with the sound off, you changed too much. THE FIRST EPISODE > feed the model your last 20 posts, your about page, everything you've shipped > ask it to describe you like it's been watching, not to summarise you > keep the 3 sharpest lines, bin the rest > 60 to 90 seconds, one idea > EP01 in the title and the filename > post from the product account, quote it from yours EP01 is the whole trick, it promises an EP02 you haven't made yet. EVERY EPISODE AFTER > same prompt, new subject: EP02 users, EP03 investors, EP04 competitors, EP05 the market > one episode every two weeks > cut 4 to 6 vertical clips from each > post a clip every other day between episodes > reply from the product account, in character, always the narrator never runs out of subjects, and that's the point. the whole thing costs a prompt and an afternoon. meanwhile the product sells the founder, the founder proves the product, and everyone buying it is really buying proof that one guy can do this

i know for a fact some 19 year old is about to make stupid money with this brands pay $400 to $2,500 for ONE ad video they need around 40 a month, they can't afford that, so they run the same 10 until it stops working you can have Videoclaw one shot a whole month of ads... and i mean the WHOLE thing: - the hooks - the cuts - the captions - the voice - the b-roll and it runs on the Claude subscription you already pay for so this is the play: start with ecom they burn through creative faster than anyone, the product shots are already sitting in a folder, and every new SKU needs its own set ask for the folder, send back 40 ads next month only 1 to 3 out of every 10 ads ever win, so whoever makes the most finds the winner first the agency sends 10 polished ones you send 40

i can't believe this is free... https://twitter.com/DeRonin_/status/2100249364429930857/video/1

here's how i auto-edit video with the DaVinci MCP and GPT-6 Astra: the trick is that the model never watches the footage 1. import the raw files into the media pool one call, the MCP handles the pool and the folders 2. transcribe locally with whisper this is the whole thing, you now have every spoken word with a timecode attached 3. hand Astra the transcript, not the video it returns which timecodes to keep and which to kill, because models are hopeless at watching footage and very good at reading a transcript 4. ripple delete everything else the MCP applies the cut list to the timeline, dead air and blown takes gone in one pass 5. drop markers with notes on every keeper so when you open the timeline you can see why each cut survived instead of trusting it blindly 6. pull Text+ captions from the same transcript no second transcription, one source of truth 7. render queue, quick export and the part nobody mentions: 155 of the 162 tools run on the FREE version of Resolve local whisper replaces Studio auto-subtitles, demucs handles voice isolation, rembg does background removal only Smart Reframe and stabilization still need the paid license https://github.com/hiteshK03/davinci-resolve-mcp i want to be honest, five days ago i've been an absolute zero at video editing and even i can edit the videos now at mid-level...

I’m 20, bootstrapped autonomous solo-agency based in Vienna Looking to connect with more technical builders :<)

after 5 days of total lock in, i realized ONE thing people are not ready anymore to buy saas which help them to close their "PAIN" instead, they're buying tools for the agent that already does their job, not for themselves build agentic infrastructure as the next product...

ahaha, i got you guys 😂 "high volume, repetitive or manipulative posting" 3 days ago, 8,000 words article, with sharing my full workflow on running solo-agency 9 days ago, 11,000 words article, full guide on how to become Robotics Engineer, no existed guide in space, all links are checked 10 times goood job X algos, you detected the real bot 👏 appeal is in if you've had this and got it reversed, tell me what actually worked, because the help docs say nothing

this guy has literally shared full workflow of auto video editing it was extremely useful to me as for youtuber beginner, READ THIS:

If you write anything with Claude, add this to your project to kill any AI slop: ------------------------------------------------ THE ANTI-SLOP PROTOCOL: HUMAN SIGNAL ENFORCEMENT Instructions for Claude: This is a combined negative and positive constraint set. You are forbidden from the patterns below, and you are required to hit the quotas below. During your mandatory self-audit step you must scan the draft for every pattern listed, rewrite any sentence containing one, and confirm every quota before delivering output. If a quota cannot be met honestly, report it. Never pad to hit a number. 1. BANNED EMOTIONAL SHORTCUTS (HIGHEST PRIORITY) The Named Feeling Definition: stating an emotion instead of causing it. Prohibited examples: "I was frustrated." "It was exciting." "I felt proud of what we built." "It was a humbling experience." "I couldn't believe what I was seeing." Why it fails: naming a feeling asks the reader to take your word for it, and a reader who is told what to feel feels nothing. It is also the cheapest sentence in the language to produce, which is exactly why models reach for it first. The fix: the Evidence Principle. Delete the label and write the thing that caused it. "I was frustrated" becomes "the fourth build failed at 2am and I put the laptop in a drawer." The reader supplies the word. The Clean Resolution Definition: every thread tied off, every lesson learned, the piece landing in a state of completion. Prohibited examples: "And that's when it all clicked." "Looking back, I wouldn't change a thing." "The lesson here is simple." Why it fails: real experience leaves loose ends. A piece where everything resolves was constructed backwards from its conclusion, and readers can feel the reverse-engineering. The fix: leave one thing unfixed and say so plainly. Name what you still do not know. 2. BANNED RHYTHM PATTERNS The Metronome Definition: consecutive sentences of near-identical length. Prohibited: three sentences in a row within five words of each other in length. Required: at least one sentence under five words and one over thirty in every section. Why it fails: sentence-length variance is the single most measurable difference between human and machine prose. Machines write at a steady pulse. People speed up, stall, interrupt themselves, then run long. The fix: the Variance Rule. After drafting, count the words in each sentence and break any run of three similar ones. The Uniform Block Definition: every paragraph the same size. Prohibited: a page where every paragraph is three to four lines. The fix: some paragraphs run six lines, some run one. Use at least two single-sentence paragraphs per piece. 3. BANNED ABSTRACTION The Weightless Noun Definition: a general word standing where a specific one belongs. Prohibited examples: "businesses", "solutions", "a tool", "results", "the industry", "stakeholders", "content" Why it fails: abstraction costs nothing to generate and proves nothing. Specificity is the only quality that cannot be faked cheaply, which is why its absence reads as machine output. The fix: the Specificity Quota. Per 500 words, minimum three proper nouns, one physical detail with a sense attached, one time anchor, and one true detail that is irrelevant to the argument. The Round Number Definition: figures that end in zero or five. Prohibited examples: "about 50%", "roughly 3x", "around $10k", "hundreds of hours" Why it fails: real measurement produces ugly numbers. A round number tells the reader you estimated or invented it. The fix: 47%, 2.8x, $9,340, 213 hours. Minimum two non-round numbers per 500 words. If you do not have the real figure, say you are estimating rather than smoothing it. 4. BANNED SAFETY The Balanced Take Definition: hedging both sides so the sentence cannot be argued with. Prohibited: "might", "could potentially", "in many cases", "some would argue", "while X, it is also true that Y" Why it fails: text nobody can disagree with is text nobody remembers or shares. The fix: one claim per piece stated flat, with no balancing paragraph after it. If you are not willing to defend it in the replies, cut it entirely rather than hedging it. The Costless Claim Definition: advice from someone who risked nothing. Why it fails: authority comes from having paid for the knowledge. Text with no cost attached reads as compiled rather than lived. The fix: name one real cost, money, hours, a relationship or a reputation, and one moment you were wrong. Do not soften either. 5. BANNED BORROWED LANGUAGE The Inherited Sentence Definition: phrasing that could appear word for word in any other piece on this topic, including anything lifted from material the user pastes as reference. Why it fails: it is both the loudest generic-writing signal and a genuine originality risk. The fix: reference material sets direction, never wording. Quotes maximum fifteen words, always attributed. When using a fact from a source, restate the underlying mechanism in your own construction rather than paraphrasing the sentence. Flag any sentence you suspect is inherited. 6. REQUIRED: THE ONE WRONG NOTE Uniform polish is itself a tell. Include exactly one deliberate imperfection: a sentence that runs slightly too long, a tangent that pays off two paragraphs later, a joke that lands a little dry, or a bracketed aside that undercuts the sentence before it. One only. Two reads as carelessness. 7. MANDATORY SELF-AUDIT Before returning anything, output this filled in: named emotions found (must be 0): clean resolution present (must be false): shortest sentence / longest sentence, in words: three-in-a-row length violation (must be false): paragraph length range: proper nouns per 500w (min 3): non-round numbers per 500w (min 2): sensory detail / time anchor / irrelevant true detail: unhedged strong claim: cost admitted: unresolved thread: wrong note used: suspected inherited sentences: Fix every failure, then deliver. Finally, list three phrases you almost used and cut for being generic.

Tripo AI website: https://studio.tripo3d.ai/workspace/generate?tab=smart_mesh

a business idea from gamescom small studios are not blocked on ideas or code they are blocked on someone to model 200 props Tripo showed Smart Mesh P2.0 in hall 10.1, and the part that matters is not that it generates 3D it generates native quad topology now, 500 to 25,000 faces, edge loops that follow the geometry riggable instead of just renderable, which is the whole difference between a demo and an asset so here is the play: 1. pick one genre and one art style low poly survival, stylized fantasy, sci-fi industrial, whatever you can hold consistent across 80 pieces 2. build packs, not models nobody buys one crate, they buy "80 stylized medieval props" 3. characters in quad, props in triangle props never deform so quads are wasted on them, and triangle goes to 50k faces when you need the detail 4. feed it 4 reference views instead of 1 front left right back, so the back of the model comes from reference instead of being invented 5. sell on Fab, not TurboSquid Fab pays 88%, TurboSquid pays 40 to 60%, identical asset, different shelf 6. price at $79 and ship a new pack monthly 60 listings selling once a month each is $56,880 gross, Fab leaves you about $50k of that, TurboSquid about $28k the whole

ohhhh yeees baby, i've been approved for Original Content Rewards Program 😎 thanks X for proving my originality kek p.s. i am so sorry for the guys who deserved it as well, but got rejected... https://twitter.com/DeRonin_/status/2097568824182669758/photo/1

this should not be free...

8,192 words, 8 departments, 29 copy-paste blocks the full system behind my $78k MRR solo agency, and how to rebuild it in any niche turn on your notis today, at 19:00 CET 🔔 https://twitter.com/DeRonin_/status/2096944576384639197/photo/1

everyone is using GPT-6 Astra to make games... i had it build me a robot prototyper i type what robot i want, and it builds the whole thing and tells me how to make it real one prompt: "prototype me a full Microduck" it designed every part, checked nothing crashes into itself when it moves, did the motor maths so the thing can actually stand up, and handed me the parts list with prices and the step by step assembly not a render, a robot i can build this weekend every piece is ready to send straight to a 3D printer, and the motors and screws come with prices and where to buy them took me 45 minutes and 5 prompts to build (burnt all my limits) process below

please, god, get me back few days ago 🥹 that i could feel one more time this incredible diving experience 😍 i loved it much more than skydiving, recommend 112/10 to visit this activity and try as many as possible things in your 20s, it gives more emotions while u’re young 🙏 https://twitter.com/DeRonin_/status/2096213449143443780/photo/1

my morning starts not from robotics unfortunately :<( probably one of the hardest night in my life https://twitter.com/DeRonin_/status/2096191274642534712/photo/1

omg, this dude is killing it at robotics he has explained step-by-step how to build your own Microduck (cool pet project for portfolio) gonna try to build it next weekend follow this hidden talent

omg, this dude is killing it at robotics he has explained step-by-step how to build your own Microduck (cool pet project for portfolio) gonna try to build it next weekend follow this hidden talent https://twitter.com/DeRonin_/status/2095832360075796653/photo/1

first 1k followers on Youtube 😱 guys, you're insane and incredibly thank y'all for supporting me at this early stage ❤️ i'm going to record only practical videos which you can follow, repeat and get the result stay tuned, i'm coming soon on the scene https://twitter.com/DeRonin_/status/2095792472299278769/photo/1

PLEASE LISTEN TO ME WHILE IT'S NOT TOO LATE this is the end for software... and for you, a little more every day as a normal person with no capital, you cannot build anything unique anymore, because Astra and the rest already built it you have no edge in this new world, and the gap gets wider with every release even if you are shipping fine right now, the demand for it disappears one morning without warning that is why you should start robotics now, not later: > the skills on your resume today are a subscription anyone can buy tomorrow > your github proves nothing anymore, anyone generates that in an afternoon > a machine that moves is the one thing nobody can fake or copy off you > an arm on your desk costs less than your monitor, so the barrier was never money > the seats fill with the people who started this year, not the ones who start after the panic if software stays fine, you end up a robotics engineer who can code if it doesn't, you're holding the one skill that didn't get cheaper software is crowded, but the physical layer is ABSOLUTELY EMPTY start combining AI + robotics while it hasn't replaced even the founders yet no need to quit anything, keep taking the money out of the AI side just give the physical layer 2-3 hrs a day, and in 6-12 months you'll be standing where the models still can't reach p.s. btw, the game preview below has been also generated with GPT-6 Astra, so, junior and middle game developers, i am so sorry :<( https://x.com/mattshumer_/status/2095596175705399482/video/1

best addition to this article: five robotics courses from Michigan University: open-sourced, 100% FREE on GitHub, repo links below: - ROB 101, computational linear algebra: the transforms and Jacobians from month 5, done properly https://github.com/michiganrobotics/rob101 - ROB 201, calculus for the modern engineer: full textbook sitting in the repo https://github.com/michiganrobotics/rob201 - ROB 311, how to build robots and make them move: the mechatronics half of months 2 and 3 https://github.com/michiganrobotics/rob311 - ROB 501, mathematics for robotics: the rigorous version of everything in month 5 https://github.com/michiganrobotics/rob501 - ROB 530, mobile robotics: Kalman filters, localization and SLAM, the theory under month 4's Nav2 stack https://github.com/UMich-CURLY-teaching/UMich-ROB-530-public lecture videos on YouTube, homework, and code in MATLAB, Python, Julia and C++ start at 101, and leave 501 and 530 until you have actually built something, because both are graduate courses and will bounce you cold this is the fundamentals layer underneath the article, not a replacement for it five courses is not a degree either, so treat it as depth rather than the whole thing build first, then come back and learn WHY it worked

dear, video editors! i need the video editing as on this channel: https://www.youtube.com/@TheCodingSloth

my first 500 followers on Youtube ❤️ you're so crazy my guys, really appreciate everybody who got followed on me gonna record a lot of useful content there (hope u'll love it, since not going do stuffy content) and now, i am looking for the guy(-s) who could make me thumbnails and video editing if u're such as guy and up to grow together, then dm me with your portfolio or leave a comment with it generous run has begun.

here are 5 robots you could build after reading this article: 1. the SO-101 arm, $122 in parts build it, calibrate every servo, teleoperate it, then print your own gripper fingers 2. a ROS 2 rover that maps a room and navigates it SLAM Toolbox and Nav2, and you can do the whole thing in simulation with no hardware at all 3. self-balancing robot it will not stand up until your filter and your loop timing are both correct, which is exactly the point 4. an ACT policy trained on your own demonstrations record 50 demos, train, deploy, measure the success rate, then record 50 more and beat it 5. line follower with PID your first closed loop, and filming it badly tuned next to properly tuned is the whole interview answer that's what will make you at least advanced junior at Robotics Engineering and will help you to get your first offer in robotics company

$4,660 worth of paid robotics courses, according to Opus 5.1 i gave it away for FREE 121 verified links, 27 pet projects, zero theory you won't use six months of two hours a day between you and hireable enjoy below: https://twitter.com/DeRonin_/status/2095201993539109136/photo/1

Posted! You can read this article here:

11000 words, 121 learning resources, 27 pet projects 6-month robotics roadmap, and all is in one article turn on your notis today, at 18:00 CET 🔔 https://twitter.com/DeRonin_/status/2095062666335150587/photo/1

everybody who is getting the same sh*t (password reset from X) enable in settings "Password reset protection" + turn on 2FA share this post that more people would see 🙏 https://twitter.com/DeRonin_/status/2094780442998161590/photo/1

everybody who is getting the same sh*t (password reset from X) enable in settings "Password reset protection" + turn on 2FA stay safe. https://twitter.com/DeRonin_/status/2094780121064300731/photo/1

oh my god, somebody is trying to hack my X account... https://twitter.com/DeRonin_/status/2094778120498782488/photo/1

here's every way to get paid in robotics: sell a system: - turnkey work cells: layout, fixturing, end effector, cycle time and programming, at 10-20% of the system budget or $3-16k on a light-duty cell - vision for bin picking: 6D pose estimation and camera calibration, the part that breaks first in production - force-controlled assembly: insertion and mating tasks that position control alone can never do - fine-tuned manipulation policies for one job, one gripper, one product line - sim-to-real pipelines: train it in Isaac, land it on the real cell, and own the gap nobody else wants to debug sell your hours: - teleoperation and data collection: $22/hr and up, no degree asked - contract bring-up and debugging: $30-70/hr, the going freelance band for robotics engineers - functional safety and risk assessment: 66% of robotics projects are delayed by certification, and almost nobody sells this - sensor calibration and vision setup on site: the step every integrator underestimates - on-site commissioning weeks: the travel everyone avoids is the work nobody can skip sell an artifact you own: - demonstration datasets for a task nobody recorded: you get paid per hour of capture, then sell that same hour again - sim environments and digital twins: billed at $30-50/hr, then reused across every client after - grippers, jigs and mounts: design once and print forever, the margin is the file and not the plastic - drivers and ROS packages for hardware the manufacturer never bothered to support - calibration rigs and test fixtures, the boring tools every single lab rebuilds from scratch then own the vertical: - a $35k arm becomes an $80k deployed system once it is actually working - that $45k gap is not hardware, it is engineering, and it is the entire business - rent the cell monthly instead of selling it, so one install pays for years - own the data for one narrow task until nobody can train a better policy than yours - turn the third integration contract into a product and license it instead of rebuilding it all four exist for the same reason: > the models cannot learn manipulation from the internet, so someone has to move a real machine and record it > simulation solved locomotion, but dexterous manipulation still needs 60-80% real data even after strong sim pretraining > which makes even the cheapest seat in robotics one the industry structurally cannot skip AI engineering pays well because plenty of people can do it. robotics pays well because almost NOBODY can and if you want to build your own thing, look at software. a survey of 1,000 robotics developers this year put software architecture as the top bottleneck at 27%, ahead of hardware at 16% you don't have to quit anything or call yourself an engineer to start get a cheap arm as a pet project, give it two evenings a week, and you're on this list before the year ends

just deleted Grokbot from my harness after Hermes got released it…

here's every AI model i run in my solo agency at ~$80k MRR... the models: - research sweeps: Gemini 3.1 Pro - trend synthesis: Opus 5 - creator screening: Qwen3.7 Flash - drafting: GPT-5.6 Sol - verification: Opus 5 - judge panels: Opus 5 + GPT-5.6 Sol + DeepSeek V3.2 - cleanup: local Qwen - high-stakes strategy: Opus 5 the rest of the stack: - scraping: Apify and Firecrawl - link tracking: Dub - scheduling: Typefully - creator payouts: Wise - client approvals: Notion - analytics: PostHog - database: Supabase - memory: mem0 - knowledge: Obsidian - hosting: one OVHcloud bare metal box p.s. i was using before Hetzner for hosting, but they got scammed me on server at the middle of the plan... if you want, you can use them, that's my principle position just TOTAL COST to run all of it: ~$1,400/month, and $680 of that is inference now the part that actually matters: most people pick one favourite model and run their whole business on it i pick the best model for each job, then pay for the seat where being wrong is expensive screening a creator profile is pattern matching, so it runs on a model at $0.03 per million input tokens drafting in a creator's voice shows up in the invoice, so it gets the model that wins on creative output verification catches the 4% that would embarrass a client, so it gets the strongest reasoning model on the list and the verifier never runs on the family that wrote the draft, because a model grading its own homework always agrees with itself the stack is not the moat... the ROUTING is stop asking which model is best and start asking which model is best at this

here's every AI model i run in my solo agency at ~$43k MRR... the models: - research sweeps: Gemini 3.1 Pro - trend synthesis: Opus 5 - creator screening: Qwen3.7 Flash - drafting: GPT-5.6 Sol - verification: Opus 5 - judge panels: Opus 5 + GPT-5.6 Sol + DeepSeek V3.2 - cleanup: local Qwen - high-stakes strategy: Opus 5 the rest of the stack: - scraping: Apify and Firecrawl - link tracking: Dub - scheduling: Typefully - creator payouts: Wise - client approvals: Notion - analytics: PostHog - database: Supabase - memory: mem0 - knowledge: Obsidian - hosting: one OVHcloud bare metal box p.s. i was using before Hetzner for hosting, but they got scammed me on server at the middle of the plan... if you want, you can use them, that's my principle position just TOTAL COST to run all of it: ~$1,400/month, and $680 of that is inference now the part that actually matters: most people pick one favourite model and run their whole business on it i pick the best model for each job, then pay for the seat where being wrong is expensive screening a creator profile is pattern matching, so it runs on a model at $0.03 per million input tokens drafting in a creator's voice shows up in the invoice, so it gets the model that wins on creative output verification catches the 4% that would embarrass a client, so it gets the strongest reasoning model on the list and the verifier never runs on the family that wrote the draft, because a model grading its own homework always agrees with itself the stack is not the moat... the ROUTING is stop asking which model is best and start asking which model is best at this

why i'd learn robotics over AI engineering for the next 12-24 months: everyone is an AI engineer now... that's the whole problem you can't build a moat out of what a subagent does at 3am for pennies robotics is the opposite: > the models already work, what's missing is someone who can put them in a body > you can't scrape the physical world, someone has to move a real machine to create the data > there is no stackoverflow answer for why your gripper keeps slipping > a robot falling over is not a problem you can hand to a subagent > nobody clones your robot over a weekend > robotics companies are valued much higher than AI now, but you should have engineering background 100/100 what it gives to you: - robot foundation model specialists sit at $280-475k total comp - recruiters say there aren't enough senior robotics engineers in the US to staff the programs already funded - the door opens early, since teleop data collection, sim work and bring-up are entry-level seats - you could found your company first when it will become a mainstream AI engineering is crowded. robotics is EMPTY same 24 months, completely different odds don't need to put it in priority, no, cashout from AI side-hustle but make it as a daily hobby and put here 2-3 hrs for learning and engineering - can make the difference for you into the future

Wait for the article today/tomorrow Just need to move my roadmap to the article :&lt;)

everyone is hyping the new X report and sleeping on the real algos drop... "Under the Hood" shows which labels are throttling you: https://x.com/i/under_the_hood but X open-sourced the ranker too, and the weights sit in home-mixer/params/param.rs verbatim from the file: > share via copy link: 20.0 > reply: 5.0 · quote: 5.0 · share via DM: 5.0 > follow author: 4.0 · share: 2.0 > retweet: 1.0 · favorite: 0.5 > profile click: 0.0 > report: -234.0 · mute: -58.8 · block: -31.2 a reply from a mutual gets +15.0 on top of that and bookmarks aren't in the file at all. not zero, ABSENT... the For You ranker never scores them these multiply predicted probabilities, not your raw counts, so it's not an exchange rate. it's a map of what the system is trying to cause and what it wants most is a post someone quietly sends to one person

i started learning electronics/robotics... should i share with you my 6-months plan how i'm going to get my first engineering offer with no degree? https://twitter.com/DeRonin_/status/2094114648144363583/photo/1

every way to make money with Claude Code and Hermes, in one map: ┃ ┣ Agency & Services ┃ ┣ AI implementation for SMBs ┃ ┣ Agent setup & onboarding ┃ ┣ Workflow automation builds ┃ ┣ Custom skill development ┃ ┗ Agent fleet retainers ┃ ┣ Productized Offers ┃ ┣ Skill packs ┃ ┣ Agent starter kits ┃ ┣ CLAUDE.md templates ┃ ┣ Obsidian vault systems ┃ ┗ Whop blueprints ┃ ┣ Building Products ┃ ┣ Micro-SaaS ┃ ┣ Vertical agents ┃ ┣ Telegram & Slack bots ┃ ┣ MCP servers as products ┃ ┗ Internal tools for clients ┃ ┣ Marketing Systems ┃ ┣ SEO & AEO audits ┃ ┣ Cold outbound engines ┃ ┣ Lead gen & enrichment ┃ ┣ Ad creative factories ┃ ┗ Landing page builds ┃ ┣ AI Media Production ┃ ┣ AI UGC ads ┃ ┣ Faceless YouTube channels ┃ ┣ TikTok content pages ┃ ┣ Brand video production ┃ ┗ Voiceover & dubbing ┃ ┣ Content & Audience ┃ ┣ X growth ┃ ┣ Newsletters ┃ ┣ Paid communities ┃ ┣ Sponsorships ┃ ┗ Ghostwriting ┃ ┣ Research & Data ┃ ┣ Deep research reports ┃ ┣ Market intelligence ┃ ┣ Competitor teardowns ┃ ┣ Data enrichment ┃ ┗ Premium reports ┃ ┣ Engineering ┃ ┣ Freelance dev at 5x throughput ┃ ┣ Legacy code migration ┃ ┣ Test suite generation ┃ ┣ Code review as a service ┃ ┗ Documentation generation ┃ ┣ Back Office Automation ┃ ┣ Customer support agents ┃ ┣ CRM hygiene ┃ ┣ Invoice & receipt pipelines ┃ ┣ Recruiting screens ┃ ┗ Personal assistant agents ┃ ┗ Teaching ┣ 1:1 consulting ┣ Team workshops ┣ Corporate enablement ┣ Courses ┗ Office hours most people pick one branch and grind it... the money is in stacking three that feed each other

these guys built an infinite movie generation machine... @fal's post-trained Minimax H3 Max is 50x faster than the original: 5 seconds of video out of 3 seconds of compute it generates faster than you can watch it and i found the open-sourced version: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree https://twitter.com/DeRonin_/status/2093716500175544695/photo/1

key insight of the day: build the apps for women, since they bring 2x revenue

follow these three rules and Claude Code gets radically better: - use /subtask instead of spawning a fresh subagent, it forks your session and shares the prompt cache - let it run in the background, the tool calls stay out of your context and only the result comes back - pass isolation: "worktree" on any fork that edits files, so parallel ones stop colliding a normal subagent starts blind. you re-explain the codebase, it rebuilds the cache from ZERO, then hands you back a wall of tool output you have to read a fork already knows everything you know... and it's been on by default since 2.1.232 everyone clipped the cross-session messaging headline from that release and slept on the part that actually cuts your bill

the best AI design guide which i’ve seen for the last time turn this into the skill and make design as a senior

Dear Gen-Z, here's how to get rich: 1. Go to Grokbot and say "find 20 creators in [hyped niche] on TikTok and Instagram who post strong work, get real engagement, and have NO offer in bio.. for each: handle, best video, what's broken in their distribution in one line" 2. Pick one. Don't sell them anything. Point at the hole they already see every day: an audience asking to buy and nothing to buy. Offer to become their growth operator for 30% rev-share, they pay $0 upfront 3. Let the bot run the ops: it builds the offer on Whop straight from the terminal, writes content that sells instead of chasing views, and pulls the numbers back weekly. One creator settling at $20k/mo = $6k/mo to you, from ONE partner 4. Repeat for 10 partners if you have to. Same bot, same playbook, zero new cost per partner. If nobody says yes, your pitch sucks and you need to work on it. Fix it and go again The ops that used to need a $3-5k/mo team now run on a bot subscription. There has never been as much alpha up for grabs as there is now Full playbook is in the article below:

i genuinely don't understand why nobody is building this.. last night, i built my growth operator system on an Obsidian vault with HUNDREDS of notes as markdown files: 1. every niche i've mapped, demand signals, supply gaps, which ones are heating up and which are already crowded 2. every creator the bot scouted, why their work is strong, what's broken in their distribution, who i passed on and why 3. every partner's offer, price, rev-share terms, and the objections that came up before signing 4. the content data, which topics pulled views vs which actually sold, hook by hook, with numbers next to each 5. the playbooks, scouting prompts, the content engine, whop CLI commands, outreach that doesn't sound like outreach 6. the rulings, every decision i never want to re-argue, one dated line each.. never below 25% rev-share, never partner without proof of skill 7. the trend radar, weekly scans of what's going viral in the niche right now, new formats and hooks that convert THIS week, not three months ago.. content stays on trend, conversion holds then whenever Grokbot starts working on a new partner, i point it at the vault it doesn't start from zero.. it starts with everything the last 10 partners taught it that's the difference between an agent doing tasks and an operator with MEMORY every partner makes the next one cheaper to run in the article below i break down the full growth operator system this vault plugs into:

okay, that’s probably the most insightful what i’ve seen for the last months that’s literally all-in-one article you need to launch your own successful app:

nobody gets a $350k offer from Anthropic with a certificate they get it by showing what they've shipped 5 from-scratch projects that make you unrejectable: 1. a tokenizer from scratch train BPE on a raw corpus.. once you see how text becomes tokens, context limits and weird model failures stop being magic 2. attention + KV cache by hand compute Q, K, V for one token yourself, then build the cache.. and finally understand why inference is memory-bound, not compute-bound 3. a mini-former you actually trained tiny decoder-only model, real training loop, loss curves.. proves you understand what everyone else just calls through an API 4. a two-expert MoE router route tokens, collapse the router on purpose, plot the histogram.. every frontier model works this way and almost nobody can explain it 5. quantize + serve one model quantize it, measure the damage, serve it with latency charts.. the difference between a demo and a system Anthropic doesn't care what you studied.. they care what you've built, broken and can explain the full build order (34 projects, 12 weeks) is in the article.. save it

i built a system to run LLMs without limits... my Token Router it's why Claude Code runs 10+ hours a day and my API spend dropped ~70% one layer between me and every model, pricing each call BEFORE it happens here's how to build the same tokens saver: 1. route by cost-of-failure, not by task every request gets classified first: what does a wrong answer cost here? boilerplate → local model. production work → cheap workhorse. architecture decisions → frontier. the router picks, not you 2. make your prompts cache-eligible providers give a 90% discount on cached input and almost nobody collects it.. because the cache breaks on a SINGLE changed byte. keep your prefix stable (tools → system → context, same order, same bytes) and put anything volatile (timestamps, fresh data) at the END. yes, your AI bill partially depends on byte order. nobody tells you this 3. block redundant reads agents re-read the same files constantly. a simple pre-tool hook that says "you already saw this, here's the summary" kills the dumbest spend in every agentic workflow 4. feed skeletons, not files parse your codebase into a map of functions and calls (tree-sitter does this free). the model queries the skeleton and only opens files it actually needs.. up to 49x fewer tokens on big repos 5. isolate the dirty work sub-agents read the heavy stuff and return only conclusions. your main context stays clean.. think of it as the CEO's inbox: reports come in, raw dumps don't 6. discipline the output side output tokens cost ~5x input and everyone ignores them. demand diffs instead of full rewrites, terse mode by default, hard stop sequences. 30-50% cut from one instruction 7. run overnight work in the batch lane every provider gives ~50% off for non-urgent jobs. your nightly agents are paying rush-hour prices for work nobody reads until morning the killer part: none of this is a tool you buy. it's ~200 lines of routing rules and hooks you write once, and every session after that is discounted forever same philosophy that runs my whole agency: spend intelligence like money, because it is your AI bill isn't the price of intelligence. it's the price of lazy routing P.S. this is the surface layer the full router also does semantic caching (similar question = cached answer, zero cost) and compresses prompts with a small local model before the big one ever sees them show me your interest and i'll drop the full build guide with all configs sharing everything for free

My daily cup of coffee prepare to you skills for running solo-business today https://twitter.com/DeRonin_/status/2091446557195292692/photo/1

Let’s force this meme moment across the internet Put it on your pfps, terminal background, prompt it in all LLMs I want to be as popular as Khabane Lame https://twitter.com/DeRonin_/status/2091229434233725213/photo/1

Let’s force this meme moment across the internet people Those who put it on pfp, get instructions from me how to build fully automated $80K MRR agency hehe https://twitter.com/DeRonin_/status/2091228543346749553/photo/1

my first flight documented 🤝 Dox is completed Face brand arc began ❤️ https://twitter.com/DeRonin_/status/2091208499376066770/video/1

Oh my god, I am so fcking aliiive ❤️❤️❤️ https://twitter.com/DeRonin_/status/2090806695790321799/photo/1

In 30 mins, I will jump first time in my life from airplane 😨 My biggest fear in life is heights btw Wish me the luck… If I will not come back in 24hrs, here’s automatically will be posted a tweet with the login and password of my X account 😂 Take care on it guys. https://twitter.com/DeRonin_/status/2090768156608401813/photo/1

list of top 30 angel investors to get raised first capital at early stage:

Preparing my own knowledge base on my profile Soon will get it released too

if you want to grow your X profile, just load this knowledge base to your Claude and it will be more valuable than any paid X course:

phrase, which helped me to reach 7-figs in my 20s: "you learn more in ten minutes of doing than in ten hours of planning" https://twitter.com/DeRonin_/status/2090386188750238141/photo/1

SOMEONE JUST KILLED THE VACUUM CLEANING INDUSTRY two ex-Google engineers spent 9 years on a home robot that doesn't navigate blind. it SEES your house every robot vacuum for 20 years has been guessing. this one understands the room [ the specs are insane ]: years in development: 9 onboard cameras: 5 LiDAR sensors: 0 cleans in complete darkness: yes data sent to the cloud: none WIRED score: 10/10 The Verge score: 9/10 [ the tech is wild too ]: no LiDAR, no laser turret, no cloud, no app required five cameras feed an NVIDIA chip inside. it builds a 3D map that never leaves the device it knows a sock from a cable from your dog. and treats each one differently pick it up mid-clean, drop it in another room, it instantly knows where it is. every other robot forgets and re-cleans the same spot [ the interface is the real kill shot ]: no app, no map editing, no menus point at a mess, say "clean here," and it goes. a 5-year-old can run it it even turns to look at whoever spoke [ the privacy angle is even wilder ]: Roomba's maker shipped 2 million images from people's homes to overseas contractors this one processes everything on-device. raw footage discarded in real time. works with wifi off nothing leaves, so nothing can leak every $1,000 robot with a laser turret is now a museum piece it's called Matic. it's real, it's shipping, and it just made the entire aisle obsolete

literally the only article you need to cut your AI bill by 90%:

Best AI models by use cases in solo-business (my real stack): Bulk production (drafts, code, briefs): Kimi K3 Verification / QA passes: Opus 4.6 High-stakes strategy: GPT-5.6-SOL Research sweeps (24/7 agents): Kimi K3 Sales copy rewrites: Fable 5 Video / UGC gen: Seedance 2.5 Image gen: GPT Image 2 Live search + trend leaks: Grok 4.5 Translation / localization: Gemini 3.5 Flash Best price-to-output: DeepSeek V4 Pro Local background agents: GLM 5.2 Local cleanup (small): Qwen two rules that matter more than the list: 1. never verify with the model that wrote the draft, different family, on purpose 2. price the model to the cost of the decision, not the habit the wrong model choice doesn't fail loudly, it just quietly eats your margin P.S. unfortunately, i couldn't finalize an article today on how i'm running $80k MRR agency :<( catch 5% value of this article in this tweet turn on notis for tomorrow, at 6pm CET 🔔

i've been running this content system for months.. i call it my Audience Radar and it's quietly the highest-leverage software in my content creation stack here's how to build the same with Claude Code: 1. define your sources: 5-10 subreddits, the X accounts your ICP replies to, youtube channels they comment on.. write them into a sources.md file 2. build the listener: a scheduled Claude Code task that pulls fresh posts and comments from those sources every night (scrapers or APIs, either works) 3. give it an extraction prompt: "pull out pain points, repeated questions, objections and exact phrases people use. group by theme, count frequency".. raw noise becomes structured insight 4. add the competitor agent: same loop but pointed at 5-10 competitors. what they post, what performs, what offers they push, where they're silent 5. everything writes into context files inside your skills folder: audience-pains.md, hot-questions.md, competitor-gaps.md.. Claude reads them before writing anything the killer part: this radar plugged straight into my Content Engine (the .md system i shared before, link below) ARTICLE: https://x.com/DeRonin_/status/2042604279077237170 the engine was already turning 1 idea into 10 platform-native posts. now the radar feeds it the ideas.. i removed myself from both ends of the pipeline my hooks come from real complaints. my offers come from gaps competitors ignore. my articles answer questions people asked THIS week that's why posts feel like mind-reading.. because technically they are most creators create from imagination. the ones who win create from surveillance build your radar before you build anything else P.S. this is just a superficial explanation of how to build such a system now my audience radar version is extended ultimately to all existed sources in ai/tech space and i'm getting the news in easy format every second (which are really getting viral) show me your interest and i'll write an article on my new version of my content engine + audience radar and how to build this sharing everything for free

$5.4B valuation at Higgsfield it's been good months with you guys 🤝 next more.

I cut my Claude Code bill ~80% with one move: I actually read the docs.. &gt; /clear when context hits 50% &gt; /context to see what's bloating it &gt; prompt caching (free, just enable it) &gt; CLAUDE.md scoped per-task, not global &gt; subagents for anything that returns &gt;2k tokens No Rust… https://twitter.com/DeRonin_/status/2050949843556643074/photo/1

How to grow your app from 0 to 100k users (PLAYBOOK): by the end, you'll know how to: - get your first 10 paying customers without a funnel - build organic growth that compounds monthly - know exactly when to spend money and when not to - turn your users into your best… https://twitter.com/DeRonin_/status/2038235735120101529/photo/1