LevinX | AI Tools, Agents & Automation
10.1K posts
LevinX | AI Tools, Agents & Automation
@levinxhq
Daily AI breakdowns - new tools, agentic AI, AI video, and using agents to actually run businesses. I cut the hype. Follow for the signal, not the noise.
Katılım Temmuz 2012
3K Takip Edilen993 Takipçiler
20 years ago today, Twitter was first officially launched to the public
Today, as 𝕏, this app continues to connect the world through free speech and the global conversation 🎉
English
@Steady_Finance Geopolitical oil drama isn’t mainly about gas prices at the pump, it can indirectly squeeze stock markets through higher inflation and interest rates. That’s the real risk to watch.
English
Everyone's watching the ships.
Wrong screen.
Hormuz doesn't take your money. The chain it starts does.
Oil into inflation. Inflation into a Fed that can't cut. And a 10-year near 4.5% that every stock you own now has to beat.
I'm watching one number, and it isn't #oil.
Which one, and why, in the new video:
youtube.com/watch?v=9kDLE3…
YouTube
English
Japan just did the thing everyone keeps saying the U.S. is about to do.
It passed a bill that opens the door to crypto ETFs and cuts the tax on crypto profits from 55% down to 20%.
That 55% number is the real story. For years Japanese investors got taxed like crypto gains were top-bracket income. Now it drops toward how stocks are treated.
Meanwhile the U.S. is still stuck trying to get its own crypto rules through Congress. The CLARITY Act keeps stalling.
So the map is shifting. Asia is writing clear rules while Washington argues about what the rules should even be.
Money tends to flow to wherever the rules are actually written down. x.com/cryptorover/st…
Crypto Rover@cryptorover
💥 BREAKING: 🇯🇵 Japan just outpaced the U.S. on crypto regulation. Japan OFFICIALLY passed a bill paving the way for crypto ETFs and cutting crypto taxes from 55% to 20%. While the U.S. struggles to pass the CLARITY Act, Asia is going ALL-IN on crypto.
English
An ad agency used to spend 8 to 15 days and $10K just to show a client what the ad would look like. inVideo just did it in half a day with one agent.
A boardomatic is the rough version of the ad you show before you shoot anything. Storyboard frames, a scratch voiceover, temp music, timed to the real edit. It is the thing that gets you signed off.
The old cost was not the shoot. It was the illustrator, the editor, the voice artist, and the three rounds of notes in between. Weeks of people time to produce something everyone agrees is disposable.
An agent collapses that into one pass: script to frames to voice to a cut on a timeline. The output is not the final film and it is not supposed to be. It is a decision aid, and decision aids only need to be clear.
Here is the copy-today version if you sell any creative work. Stop describing the deliverable in a deck. Build the ugly moving version of it in an afternoon and put that in the pitch instead.
The client is not buying your taste. They are buying certainty about what they will get. Whoever removes the guesswork first wins the job.
The pitch used to be a promise. Now it is a preview, and a solo operator can build one before lunch. x.com/invideoOfficia…
Invideo@invideoOfficial
If you're an ad filmmaker or an AD, you know a boardomatic is what gets you signed off before production - and that it usually eats 8–15 days and $5–10K. We got it down to half a day with one agent. Something to put in front of your team, your DOP, or your client before a single day of shooting.
English
Andrej Karpathy says he has not typed a line of code since December.
Not slower. Not assisted. Fully delegated.
His three shifts:
He went from writing 80% of his code to writing basically none of it, handing whole features off to agents.
The new skill is removing yourself as the bottleneck, arranging things so agents run completely autonomously without you in the loop.
The bigger bet: the customer is not the human anymore, it is agents acting on behalf of humans, and most of the apps we use get refactored away.
Credit: @karpathy · via @saranormous / @NoPriorsPod
English
@NaoyaCreates the studio part is easy now. keeping one character consistent across shots is still the grind.
English
Everyone thinks AI filmmaking is one text-to-video prompt. The real unlock is wiring the models together so one person runs the whole studio.
fal's new walkthrough connects Claude Code to their MCP, which just means Claude can call fal's image and video models directly instead of you clicking around a dashboard.
You describe a shot, Claude picks the model, sends the job to fal, and pulls the render back. No copy-paste between five tabs.
Claude Artifacts become your review room. Each generated frame or clip shows up in a live preview panel you can approve or reject before spending on the next render.
Claude Skills lock in your style. You write the look once (color, framing, character rules) and Claude reuses it on every shot, so scene 12 matches scene 1.
Copy-today version: one person storyboards, generates, reviews, and re-renders a full sequence in an afternoon, with the model consistency handled for you.
The pipeline is the product now, not the single clip. x.com/fal/status/207…
fal@fal
learn how to COMBINE Claude Code and the fal MCP to design your own custom "AI Filmmaking Pipeline!" our latest YouTube video covers: - how to set up Claude Code and the fal MCP - how to create Claude Artifacts for creative review - how to create Claude Skills for consistent creative prompts watch the full tutorial here: youtu.be/IQ7PGe5mPMU?si…
English
The AI token price war is the most underrated line in this whole list. It is the reason a one-person shop can now run agents that cost a team a salary last year.
Here is what a price war on tokens actually means. The labs are cutting the cost of a unit of AI work, not the price of a subscription. Same output, cheaper input, month after month.
So the math flips. A job you would not automate at 10 dollars a run becomes obvious at 40 cents a run, and you start pointing agents at boring work you used to eat yourself.
Copy this today: take the one task you do every week that you priced out as too expensive to automate six months ago. Re-price it at today's token cost. Most of the time the answer changed and nobody told you.
The bitcoin part of that conversation rhymes with it. He is not calling a price. He is watching the Fed, stablecoin flows and tokenization, which is plumbing, not a chart.
Two different stories, same shape. The cost of doing the work is falling, and the rails the money sits on are getting rebuilt underneath everyone.
Cheap tokens are a business model change disguised as a pricing update. x.com/APompliano/sta…
Anthony Pompliano 🌪@APompliano
I spoke with @jvisserlabs this week about: - is the AI mid-cycle slowdown over? - AI token price war - why he is turning bullish on bitcoin - tokenization & stablecoins - Siri vs Grok voice mode - Samsung and bottleneck stocks - Fed's July rate decision - gold and silver - regional banks YouTube: youtu.be/TP9AEulCw9g?si… Apple: podcasts.apple.com/us/podcast/has… Spotify: open.spotify.com/episode/5uCek2… TIMESTAMPS: 0:00 - Intro 0:42 - Grok vs. Meta's AI price war why cheaper AI won't cut spend 10:03 - Apple's Siri stumble & price hikes 14:50 - Using Grok in a Tesla 17:28 - Short sellers, Samsung's earnings & the mid-cycle slowdown 23:34 - Why Jordi is turning bullish on bitcoin 29:55 - Tokenization, stablecoins & the AI-crypto nexus 33:33 - Michael Saylor's Bitcoin sale — does it matter? 38:02 - Where else Jordi is deploying capital 42:04 - Does the Iran war actually matter for markets? 45:00 - The new robotic hand demo that changed everything 50:18 - Jordi’s upcoming video
English
A $770K gain gone in one trade because holding felt like doing nothing. The chart rewards patience, not reaction. x.com/coinbureau/sta…
Coin Bureau@coinbureau
🚨WHALE WIPES OUT ENTIRE $770K SK HYNIX GAIN, GOES LONG ON NVIDIA A whale locked in $770,000 in profits on a $30 million bet on South Korea’s SK Hynix, only to immediately hand back the entire gain after piling into U.S.-listed Hynix shares. Undeterred, the trader has made a fresh bet on AI by opening a $20 MILLION long on Nvidia.
English
The 80s raiders stripped companies for parts. The next ones buy boring businesses and rewire the guts with AI. x.com/TheIcahnist/st…
The Icahnist@TheIcahnist
The Venture Raider Era The New '80s Barbarians are the Venture Raiders. You're going to see corporate raiders pairing with VCs to carve out traditional businesses and reposition them for AI bottlenecks. Venture Raider Model: 1) Carve out a high-quality company from a sinking conglomerate. 2) Partner with a VC that brings AI portfolio companies and an innovation, moonshot mindset. 3) Deploy AI to rebuild operations from the ground up. 4) Use buy and build to create a one-stop-shop ecosystem.
English
Everyone reads this as free alpha. It is actually a free org chart.
Under the hood these repos are not one big model picking stocks. They are a room of small agents with job titles: a value agent, a sentiment agent, a fundamentals agent, a technicals agent.
Each one writes up its own view with reasoning attached. Then a risk manager caps the exposure and a portfolio manager makes the final call and sizes the position.
So the real output is not a buy signal. It is a paper trail you can argue with, which is the part chart watching never gave you.
One honest note. Most of these open source funds ship as research and backtesting tools with a loud disclaimer, so plugging your brokerage into it on day one is how you donate money.
The copy-today move has nothing to do with stocks. Take the shape: three specialist agents that each write their reasoning, one manager agent that weighs them and decides. Point it at your keywords, your pricing, your inbox.
That is why it feels illegal that it is free. The value was never the model. It was the committee, and someone just published the committee.
The edge is not the trade. It is the reasoning trail underneath it. x.com/EvanLuthra/sta…
Evan Luthra@EvanLuthra
An AI hedge fund just went open source. And honestly, it feels illegal that it's free. You pick the companies. AI digs into the research, quant models turn it into trades, and every single decision shows you the why behind it. Chart watching is officially optional.
English
Everyone will argue about which model won. Look at what stayed the same: both clips came out of Seedance 2.0.
The renderer is the constant here. GPT 5.6 Sol and Fable 5 are not making the video. They are writing the instructions for it.
So this is not a video model test. It is a test of which language model writes a better shot description, with the same engine executing both.
That is the whole game in AI media right now. A good prompt model turns "guy jumps off a roof" into a real shot list: camera move, lens, lighting, the beat where the action lands, how long it holds. The render tool just does what it is handed.
Copy this today. Take one idea, ask two different models to rewrite it as a 40 word shot description with camera, lighting and action spelled out. Feed both into the same video tool. Watch how far apart the results land.
Then keep the version that won and save it as a template. That template is your asset. The video models will get swapped out every few months, but a tested prompt spec keeps paying.
The render is the cheap part. The instructions are the product. x.com/higgsfield_ai/…
Higgsfield AI 🧩@higgsfield_ai
Action scene comparison round. GPT 5.6 Sol vs. Fable 5, generated with Seedance 2.0.
English
Everyone is watching the samurai clip. The actual product in that post is the eval set.
Higgsfield is running the same scene through different engines and lining the results up next to each other. Same brief, different models, one comparison.
The models are the part that will be replaced. A new one lands every few weeks and swapping it takes an afternoon.
What survives is the eval. Your own private set of prompts that represent the work you actually get paid for, run against every new model that ships.
Copy this today. Write 10 prompts that look like your real client jobs, save them in one folder, and run all 10 the day a new model drops. Keep the outputs with the date and the model name on them.
Do that through three releases and you stop reading benchmark threads. You already know which engine handles your kind of work and which one quietly got worse.
One person with a good eval set beats a team guessing from the leaderboard. x.com/higgsfield_ai/…
Higgsfield AI 🧩@higgsfield_ai
Eval set, samurai sample: GPT 5.6 Sol vs. Fable 5. Generated with Seedance 2.0.
English
Everyone will call this a video tool. It is really the death of the "we need a video guy" line for solo builders.
HyperFrames turns your Figma mock straight into a launch video. No handoff, no editor, no re-drawing the design in some other app.
How it works: copy your Figma link, hand it to your AI agent, and invoke /figma. It reads every hex, every font, every frame and renders the video to match your design exactly.
Setup is one line: npx hyperframes@latest skills. Then your agent does the frame-by-frame work you used to pay someone for.
The real shift is where the design lives. Your mock is now the source file for the ad, not just a picture someone eyeballs and rebuilds.
Copy-today play: draft your product screen in Figma, point your agent at it, and ship the promo clip the same afternoon you finish the design.
The design was always the video. The tools just finally caught up. x.com/HeyGen/status/…
HeyGen@HeyGen
Figma -> HyperFrames Designers can now ship launch videos straight from mocks Just copy the link, hand it to your agent, then invoke /figma Every hex, every font, every frame is exactly your design $ npx hyperframes@latest skills
English
Nobody checks your AI certificate. They open your repo and read what you actually shipped. x.com/doublenickk/st…
Shadow Nick@doublenickk
99% OF PEOPLE FAIL THEIR AI CERTIFICATION EXAMS BECAUSE THEY STUCK TO DRY TEXTBOOKS INSTEAD OF BUILDING IN PUBLIC Passing the CCAF or any top-tier AI exam isn't about memorizing learning modules, it’s about proving concepts against raw reality by coding live on screen. The most heavily tested concept across all core competencies is the fundamental mechanics of a complete AI agent loop, and most product managers still get the exam traps wrong. The ultimate strategy to master and monetize AI agent architecture: The Agent Loop Blueprint: Understand the exact transition when a model pauses for tool_use, executes external code to fetch live data, and feeds the results back to close the cycle until the stop reason changes to end. The Build in Public Advantage: Ditch the silent studying that leaves your bank account empty, document your errors and architecture fixes live, and immediately differentiate yourself from certified amateurs who can't deploy a real script. The Exam Trap Defense: Stop falling for tricky exam scenarios by manually building an agent loop from scratch so you intuitively see how core networks and gNodeB bridges actually interact under pressure.
English
A Bitcoin treasury just sold $87M of BTC to fund AI data centers. When the hardest asset gets traded for compute, compute is the scarcer one. x.com/Alaouicapital/…
Alaoui Capital@Alaouicapital
Empery Digital sold 1,400 $BTC for $87.1M. Not because Bitcoin failed. Because AI infrastructure became the higher priority. Empery Digital is cashing out Bitcoin to invest in AI data centers and strengthen its balance sheet. Fascinating shift in priorities. Yesterday, companies were buying $BTC. Today, they’re selling it to build AI. The capital rotation is getting interesting.
English
HeyGen avatars now come back with a transparent background straight from the API. Sounds like a file format detail. It actually removes the last manual step in AI video.
Until now you got a green screen. Someone had to open an editor, key out the green, fix the fuzzy edges around hair and shoulders, then export. That is a human in the loop on every single video.
Set output_format: "webm" on /v3/videos and the file itself carries the transparency. No green to remove. Nothing to clean up. The avatar just sits on top of whatever is behind it.
The part people will skip: it is on their MCP too. So an agent can request the clip, get a ready-to-composite file back, and drop it onto a screen recording without a person ever touching a timeline.
Copy this today. Record one screen walkthrough of your product. Write 20 short scripts for 20 different customer problems, loop the API, and stack each transparent avatar in the corner of the same recording.
That is 20 personalized demo videos from one recording and one loop. A year ago that was an editor on retainer.
Every time a format gets an alpha channel, a job gets an API call. x.com/HeyGen/status/…
HeyGen@HeyGen
Transparent WebM output on the HeyGen API, CLI, and MCP Set output_format: "webm" on /v3/videos, and your avatar comes back with a transparent background, a real alpha channel, and no green screen. Take a look at the before and after videos below ↓
English
Full session with @sampriti0 of @cognition on @OpenAI Build Hours (Agent RFT), reposted by @0xCodez: x.com/0xCodez/status…
Codez@0xCodez
OpenAI just launched a 3-hour course on building agentic systems from scratch: 00:00 - Building workspace agents 37:42 - Deep dive into OpenAI AgentKit 1:24:01 - Agent memory (loops, context, evals) 2:21:44 - Agentic training (OpenAI playbook) This 3-hour watch will replace you 10 paid agentic courses. Watch it today, then learn how to build a team of AI agents from scratch in the article below.
English
Cognition fine-tuned GPT-5 to plan, and Devin stopped overthinking.
Their research engineer broke down exactly how, on OpenAI Build Hours:
1. Train ONE narrow step. Given the user's prompt, pick which files to edit. Read only tools, nothing else.
2. Reward = F1 score on the files a human actually edited. Precision and recall balanced, so the agent neither guesses wildly nor plays it safe.
3. The result: getting through planning mode went from eight to 10 back and forths with the model down to four.
4. Every rollout trains in its own VM, because the agent has a real shell. In his words: "in case the model goes crazy and runs rm -rf."
The lesson is not that you need a bigger model. It is that an already frontier model, pointed at one narrow task with one clear reward, beats that same model guessing.
Credit: @sampriti0 of @cognition · via @OpenAI Build Hours
English