midas vision

Somebody recorded Karpathy's workshop for university faculty. Thank god they uploaded it. Andrej Karpathy, who led AI at Tesla and co-founded OpenAI, said: "For the last 15 years I wanted to work on AGI. Now that it's here, no matter what I choose, I feel like I'm making the wrong choice." If someone who helped build modern AI feels lost, your confusion isn't a skill gap. The field is just moving faster than anyone can track. no clear path → no wrong move → just start → clarity comes after This one's worth more than any career advice you paid for ↓

A Google DeepMind engineer showed me how he cut token usage by 95%. Fable 5 made that trick a lot more valuable. The article below includes the exact config he uses and why it works.

Rockstar spent years building GTA 6. Someone spent two hours building their own version with Claude Fable 5. We're entering a very strange era.

@DoeOnChain this changed how i think about it

Anthropic already explained how to prompt Fable 5. Almost nobody is using it this way. They wrote down six habits their own engineers use. It's public, and it's buried. One of them stops Fable from silently downgrading you to a weaker model mid-task, and most people never notice it's happening. Ten minutes later I had rewritten how I use Fable. Spend ten minutes on this while Fable is still included. It'll save you hundreds later. The full breakdown is below.

instead of watching netflix tonight spend one hour with this it shows how to run claude as a team of models instead of one chatbot > fable 5 plans and delegates > opus handles the hard reasoning > sonnet does the grunt work all inside claude code then he builds an interactive solar system and a starcraft-style game from basically one prompt most people will scroll past this the few who watch it will get how the fast movers are actually shipping watch it, bookmark it now

Claude Fable 5 is getting ridiculous. Someone built an entire survival game from one prompt. From a single idea → game world → crafting → enemies → a playable game. We're getting closer to prompting software into existence.

Fable 5 is back... but this isn't the same model everyone was waiting for. Anthropic had to tighten the safety classifiers before bringing it back. That means more routine coding requests can now get silently routed to Opus 4.8 instead. You only get Fable for 50% of your weekly limits until July 7. After that, every message comes out of usage credits. The funny part? After watching someone test it for six minutes, I don't think most people should be using Fable for everything anyway. The workflow below explains when Fable is actually worth using.

Anthropic brought Fable back today. I completely forgot how weird the last few weeks were. Everyone wanted Fable back. Very few people know what to build with it first. The workflow below is where I'd start.

I thought Google was winning because of better prompts. Then I spent 45 minutes listening to Addy Osmani. I never got the secret prompt I was looking for. I found something much more useful instead. The article below is the simplest version of it.

a guy says claude code and youtube made him $19,000 last month his whole channel is built on one thing that is happening right now while everyone watches the world cup, he turns the same moments into faceless videos before the search wave dies the full setup is broken down in the article below

@Raytar instant bookmark

My friend launched a big as f*ck design agency last month that already runs completely on AI. In his first month he made $37k. All because of this video. Ad creatives that used to take 2 hours now take just 5-7 minutes. Now his entire team runs everything on Claude and loops. Claude creates the designs while loops manage everything with zero human input. Watch the video first. Then see how to build loops yourself.

Someone quietly re-uploaded an Anthropic workshop that should have stayed inside the conference. I don't think they realized what they gave away. It's the closest thing I've seen to Anthropic's internal playbook for building AI agents. Halfway through I stopped watching and rebuilt my own workflow. The article below is the simplest version of everything I learned.

@cyrilXBT super helpful.

IF YOU HAVE BEEN HOLDING OFF ON HERMES AGENT BECAUSE OF API COSTS, READ THIS FIRST. Four real ways to run it for free. Local, through Ollama, complete privacy and no rate limits. Cloud, through free tiers from Groq, OpenRouter, or NVIDIA NIM, no card required. A subscription you already have, GitHub Copilot connects natively. A fallback configuration so a rate limit on one provider does not stop your whole session. Hermes Agent is free and open source. Now there is no excuse on the model side either. Bookmark this. Follow @cyrilXBT

@noisyb0y1 people are sleeping on this

THIS DEVELOPER BUILT A FLEET OF AI EMPLOYEES WORKING 24/7 - AND THE GOVERNMENT WANTS TO TAKE IT AWAY FROM HIM 16:00 he says it straight - "I have a fleet of employees working 24/7, never eat, never sleep, never ask for time off, never need health insurance - this is only possible with local models" the US government decided who can use the best models - and we are not on that list - we have officially entered the age of hand selected winners every 20 minutes his AI scans Reddit and Twitter looking for business opportunities - with cloud models this would cost tens of thousands per month Mac Mini is already enough to start - $499 once, zero subscriptions, zero limits, zero data leaving your machine the window is closing - those who buy hardware today will depend on no government in 2 years

@cyrilXBT saving this

WHAT HAPPENED WHEN I HAD CLAUDE ANALYZE MY TRADING JOURNAL EVERY WEEK FOR A MONTH. It found my real win rate by setup type, not the one I assumed in my head. It found the specific hour of the day I make my worst decisions. It found that my best trades almost always have one thing in common that has nothing to do with the chart. None of this came from a new strategy. It all came from data I already had sitting in Obsidian, just nobody had actually looked at it properly until now. Bookmark this. Follow @cyrilXBT

who is the steam machine actually for it is $1,349 and runs linux $899 buys a windows rtx 5060 pc that plays more games and runs them faster valve built the expensive way to do less

@Volknasty SteamOS 3.8 runs on custom builds now. look it up.

@callsofmidas Lol this build won't work with Steam OS. Retards all around

valve wants $1,349 for the steam machine this guy built a far better one at home for $906 ryzen 7, rtx 5060 ti, nvme ssd, quiet high-pressure fans, the whole thing stronger parts, lower price, and you own every screw

@RoseCityWil fair. but "console price" shouldn't mean laptop money.

@callsofmidas This isn’t made for someone who can build their own PC. It’s made for someone who wants to get into PC gaming but wants a console like experience. Valve could’ve taken a hit on it considering most people will just use their Steam Library but I respect that they kept it open.

@mikenevermiss respect for sharing.

“within 12 months, AI will be writing essentially all of the code.” that’s what anthropic CEO Dario Amodei believes. save this post before you forget. right now, there are 2 types of people using AI. the first asks a question, copies the answer, and moves on. the second builds systems that let ai do hours of work in minutes. the gap between them is growing fast. if you want to stay ahead, start using these 3 features: 1. projects stop starting every chat from scratch. upload your project files, coding standards, and documentation once. now AI understands your project before writing a single line of code. 2. artifacts see your app or website update live. test layouts, click buttons, and make changes before exporting the final code. 3. claude code instead of chatting in a browser, claude works directly inside your terminal. it can search your codebase, fix bugs, build features, and automate work that normally takes hours. the people who learn to work with AI today will have the biggest advantage tomorrow. bookmark this post so you can come back to it later. 👇