Ricker

@s4yonnara Top tier watch dude, bookmarked

Andrew Ng and Anthropic ran an experiment most people would call reckless. They pasted the entire text of Frankenstein into one prompt. 108,000 tokens, then started asking questions. The bill should have been brutal. Instead every follow-up landed at up to 90% off. What made the difference is prompt caching, and the setup is where everyone falls over: step 1 → everything static goes on top. Tools, then system, then docs. In that order step 2 → mark where static ends. That line is what gets cached step 3 → one stray character above the line invalidates it. Back to full price step 4 → 5 minute expiry, but every hit resets the timer step 5 → cached reads cost a sliver of normal input The part people miss: this isn't just a billing trick. The model stops re-reading what it already knows, so answers come back faster too. Long context stops being something you ration and becomes something you assume. Most builders never touch this setting, then act shocked at their API bill. Watch and bookmark the full 1-hour course below.

Andrej Karpathy just reveales how LLMs actually thinks: "GPT-4 knows it failed. It just won't tell you unless you ask." >80% of GPT-4 errors are recoverable - the model already knows it screwed up. It has 80 transformer layers and spends the SAME compute on every single token as your brain In a 20-minute speach at Microsoft Build, Karpathy reveals the full psychology of LLMs. Worth more than any $500 prompting course you've seen on your timeline.

@0x_fokki 100% dude

@0xRicker It worth my time

@Dipper_pol fact broski

@0xRicker best AI brain on the world

@antisadh watched full bro, thanks

Anthropic engineer: "5 agents run in the background while I sleep. Code migrations that used to take my team 3 weeks I ship before lunch. Every engineer still typing prompts by hand is 6 months from replaced" In 2 minutes he tells you exactly what changed Most engineers are still opening the same PR they filed on monday watch the first 60 seconds. if you don't see your job in it, unfollow me

@insomnia_vip this dude did a great job

A 22 YEAR OLD CHINESE PROGRAMMER IS BUILDING CINEMATIC ADS WITH CLAUDE INSTEAD OF HIRING A VIDEO TEAM One MCP connection turns Claude into a creative workflow that can generate polished commercial videos from nothing more than a detailed text prompt in a single chat Instead of learning complex editing software or paying expensive freelancers the entire production process becomes writing better prompts and letting the tools handle the execution As AI keeps removing production bottlenecks the biggest opportunity is no longer creating content faster but creating more valuable systems around it People learning these workflows early will have the biggest advantage

@phosphenq one of my favorite podcast with him

OpenAI founding member, Andrej Karpathy: "I haven't typed a line of code since December. Code's not even the right verb anymore. I express my will to my agents for 16 hours a day." 1 hour on how he actually works now: agent harnesses, self-improving loops, and zero code. Worth more than any $500 agentic course. Watch it, then read the 12 steps to build an agent that runs for hours without you, below.

@0xClodex top watch bro

Boris Cherny, the engineer behind Claude Code, and Anthropic CEO Dario Amodei, in one interview: Boris: "for me personally, Claude has been writing 100% of my code for at least six months. I feel like I suddenly have superpowers - I have a jetpack, and engineering has never been this fun." he built Claude Code and Claude Cowork after leaving a slow life making miso in rural Japan - one chatbot took his breath away and pulled him back in. then Dario on where this goes: "you automate 90% of the job, people are 10x more productive in the other 10%. but eventually it gets close to 100% - and then you have to find something else for them to do." the man building the tool that writes the code, and the man warning about what happens when it writes all of it. Watch it today ↓

@Dipper_pol i enjoy every video with this chad

Andrej Karpathy builds GPT from scratch in one 2-hour video - before he led pretraining at Anthropic: he starts with an empty file and tiny Shakespeare, and ends generating infinite fake Shakespeare, character by character - the exact same way ChatGPT runs, just token by token. "now we get to the crux of self-attention - this is probably the most important part of this video." then he explains the whole engine: every token emits a query (what am I looking for) and a key (what do I contain), and attention is just them finding each other. by the end he connects it to the real thing: "to train ChatGPT there are roughly two stages: pre-training, then fine-tuning." this one video is the clearest path from "I use ChatGPT" to "I understand how it's built." Watch it today ↓

@gippp69 how this someone even come to this

SOMEONE GAVE CLAUDE FABLE 5 A CONTRACT, A $5 DAILY BUDGET, AND A MANAGER. THIS F**KING DANGEROUS SYSTEM CAN CLEAR A WEEK OF BACKLOG WHILE YOU SLEEP. 00:03 he opens Microsoft’s 4,900 star Agent Governance Toolkit, built for policy enforcement, isolated execution, identity controls, and protection against all 10 OWASP agentic risks. setup starts with 3 files. CONTRACT MD sets the limits, boundaries MD defines what the agent can touch, and signoff SH runs every test before anything ships. four roles split the shift. one model reads the logs, Fable picks the highest value task, another writes the code, and a fresh Fable reviews the final diff. a $5 cap stops runaway sessions, commits above 150 lines need approval, and autonomy unlocks only after 20 runs at a 95% pass rate. one failed run sends the agent back to probation. the repo adds the guardrails, but the contract is what turns Claude from a chatbot into an employee.

@0xMortyx Add quiz is really smart move

Jeffrey Litt's: "make your agent write you a quiz before you ship." Your AI agent just landed a 50,000-line PR. You approved it. You understood maybe 12% of it. You now have massive cognitive debt. 90% of "move fast with AI" advice skips the part where you stop being able to think creatively about your own project. You can't send code for review until you pass it. Watch the full talk. It changes how you think about AI-assisted engineering.

@0xMortyx It really worth every minute of reading

Princeton Ai researchers published a GAMECHANGING 118-page PDF on “SWE-agent: Agent-Computer Engineering” That will teach you more about building agent than a 1 year of paid course Every effective coding agent depends on the same system: Model → Harness → Loop The loop keeps the agent moving: Observe → Act → Execute → Evaluate → Repair → Repeat A strong agent harness provides: → Clear context at every step → Recovery paths when the agent fails → A loop that continues until the task is actually solved The paper shows that interface design can improve agent performance without changing the underlying model. Bookmark & Spend 10 min reading it now, then read the article below.

Ex-Google AI Agent Architect just launched a full course on " Harness, Loop, AI Agent Memory system" 90% AI agents are wasting 60%+ of their tokens wrong 00:00 - Ai Agent Memory 12:06 - Harness, Loop 31:18 - Ai Agent System This 50-minute will replace 10 paid courses and guides on Harness and Looping Watch it today, then read how to build proper agent loops in the article below

@Mnilax 6 min of pour alpha

IBM broke down what actually turns a model into an agent, in 6 min: 00:00 - why a raw prompt guesses, and context fixes it 01:20 - MCP: giving the model safe hands on your real data 02:45 - skills: packaging know-how so it does a task the same way every time 04:15 - MCP vs skills, and when to reach for each that skills part is what separates a chatbot from something that runs real work on its own. the article below is built on it.

@dunik_7 one of the best video about loop and harness

19 free minutes just made the agent-ops course you were about to buy pointless. log every run → trace it → grade it → find the weak step → patch it → run it again that loop is the only reason an agent gets better instead of just failing faster. a loop, a harness, memory, and an eval that can actually fail the work - that's the whole stack the serious people run. one video does what a shelf of $500 courses can't. give it 19 minutes, then keep the full breakdown below.

@qwinsi0x watched a lot videos with this guy, he is real chad

One creator gave GPT-5.6 Sol a single prompt and walked away. When he came back, there was a finished video. Voice, editing, motion graphics, quality checks. He never recorded a single clip and never opened an editor. The model mapped out the entire chain itself: voice through 11Labs, avatar through HeyGen, editing through HyperFrames, and ran the whole thing on its own. The interesting part isn't the result. It's what happened after. Separate agents inspected every frame of the finished video: checked whether the avatar disappeared, whether text spilled outside the frame, whether the facts matched the actual release notes. They found an error. Re-rendered on repeat until it passed. The model never graded its own work. Not once. The price tag: roughly $300 on Ultra mode, 450 million tokens, nine parallel agents. And here's the honest catch: Ultra overdid it here. Too much thinking, too much delegating. The same output on regular "High" would've cost half as much. The most powerful setting isn't always the right one. The article below breaks down the exact "contract" for working with this model, how to set the goal instead of the steps, how to avoid overpaying on an inflated setting ↓

@de1lymoon I also think he has something to teach us. How do you plan to use that?

@0xRicker a guy who worked at google definitely has something to teach. saved it for myself

@raidenfomo this guy is smart enough for tesla

THIS REPAIR GUY TURNED A $240 DEAD RTX 3090 INTO A BOX THAT KILLS $4,200 A YEAR IN AI SUBSCRIPTIONS AND SOLD 21 OF THEM IN JUNE. The card on his bench came in a box marked defective. A gaming cafe outside Shenzhen let it go for $240 because it wouldn't start. He's 27, works out of a spare room, and knows something the seller doesn't: nine dead cards out of ten aren't dead. Seized fans and dried thermal pads. The chip underneath is fine. New pads, two fans, forty minutes with a screwdriver. One card a night, after dinner. A red tin of screws sits at his elbow. Pause at 0:15 on the gold edge connector, that card sold new for $1,499. His buyers aren't gamers. They're people who added up their subscriptions: ChatGPT Pro, Claude, Cursor, Perplexity. $350 a month, $4,200 a year. A 3090 carries 24GB of memory, the same 24GB as the $2,000 RTX 4090. Search RTX 3090 on eBay right now and check. 24GB runs a 32B open model through Ollama, free, twenty minutes to set up. The card pays for itself before the second billing cycle. The models behind the paid tiers stopped being special: open 32B models caught up on most tests. The hardware got thirty times cheaper the day it left the data center. The subscription never dropped a dollar. A subscription is a taxi meter. The car gets cheaper every year. The meter ticks at the old price. Twenty-one cards in June. $680 each, minus the card and $35 in pads and fans. $8,505 clean, out of a spare room. Your card statement has the same five lines his buyers cancelled. Save this before "defective" stops meaning cheap.

@monokern its like from cinema about future

THIS GUY ASKS HIS AI FOR A DAILY BUSINESS RECAP AND IT ALREADY KNOWS EVERYTHING > 4 new clients, $11,500 new MRR > 23 prospects found, 18 cold emails drafted, 1 call booked for tomorrow > best ad identified, weakest one already paused this is what it looks like when your AI actually knows your business this works because the AI has context. everything stored, connected, readable that's what Obsidian does when Claude Code is writing to it after every session > Claude Code runs the pipeline on one command > NotebookLM does the analysis on Google's servers > Obsidian stores everything and Claude reads it all before answering anything vault grows every day. outputs get sharper every week full setup in the article below