Shoopy

Thariq Shihipar, Claude Code team at Anthropic: "I don't let the creator agent ship anything. a second agent's only job is to attack its work – whatever survives is what I keep." In a 40-minute walkthrough he shows his actual agent setup: → /goal to hold Claude on a task instead of letting it drift or stop early → plan by removing unknowns first – the failure mode is building while you're still guessing → parallel Claude tasks running through Slack like a team channel of agents → a creator sub-agent and a separate verifier sub-agent, so nothing ships unchecked → when the AI writes the code, the skill that still matters is reading and judging it One human directing a crew – and the verifier is what makes leaving that crew alone actually safe. save it↓

@cyrilXBT booked

Anthropic just rolled out five hands-on workshops for building self-evolving agentic systems from the ground up: 00:00 - Ship your first Claude agent 36:44 - Build memory for Claude agents 1:05:06 - Make your agent autonomous 1:26:46 - Set up a proactive agent 2:03:35 - self-improving agents (tools,skills) These 3-hours of free Claude workshops will replace 10 paid agentic courses. Watch today, then read article below on how to build a self-improving agentic system with Fable 5.

@0xMorlex Insane drop bro, bookmarked

Andrej Karpathy showed the blueprint for multimodal AI 8 years before ChatGPT could see images: 1 image → 1 ConvNet extracts what it sees 1 RNN → turns those features into a sentence 3 datasets: Flickr8K + Flickr30K + MS COCO The model learned vision and language inside one end-to-end system, then reached state-of-the-art results across all three In this 29-minute lecture, Karpathy reveals the original multimodal stack: computer vision + language modeling + visual-semantic alignment + image generation

@eng_khairallah1 booked

this is f**king dangerous someone ran GPT-5.6 Sol and Claude Fable 5 side by side. same prompt. same task. one shot each. both nailed it first try. Sol just finished faster. nobody expected that. Fable 5 was supposed to dominate long reasoning without competition. you have less than 24 hours to set this up as this takes away 20% of your usage limit save and bookmark it no matter what

@cyrilXBT thx for drop bro, bookmarked

BORIS CHERNY, WHO BUILT CLAUDE CODE, SAT DOWN AND SHOWED HOW HE ACTUALLY USES IT. Not a tutorial creator. Not a vibe coder with three months of experience. The person who wrote the thing. 30 minutes of watching him work restructures how you think about building with AI. Most people use Claude Code like a smarter autocomplete. He runs it like a second engineering team. That gap, autocomplete versus engineering team, is the gap between shipping toys and shipping products people actually pay for. Bookmark this. Watch it twice. Follow @cyrilXBT for more Claude Code breakdowns the moment they drop.

@0xMovez I haven't read it yet, but I already know what I'll be doing tonight thanks Movez!

@0xShoopy Have you read this graph article, bro? Give it some time - I’m sure you’ll like it as much as Andrew’s talk.

Andrew Ng: “AI agents are doing almost 100% of my tasks now - the hype has exceeded my expectations. in 4-6 months, we’ll all be building graphs to orchestrate self-improving agents. No more prompting.” In a 20-minute talk, Andrew Ng explains how to build self-improving agentic systems from scratch. Worth more than a $500 agentic course. Watch this video, then read the article below on how to become a graph architect.

@r0ck3t23 bookmarked

Marc Andreessen just stripped artificial intelligence down to what it’s really made of. Not the algorithms. Not the data centers. The raw material. Andreessen: “They’re literally made out of sand.” The most abundant, most overlooked substance on the planet. You walk on it. Build with it. Wash it off and forget it exists. Someone looked at that and saw a mind. The universe spent 13.8 billion years turning matter into consciousness. Carbon. Water. Amino acids. Eons of chemistry and selection pressure to produce a single thinking brain. We did it with sand. In under a century. Andreessen: “We light it up, and we put AI on it, and all of a sudden it’s thinking.” Biology was the first path matter found to become aware. Silicon is the second. The universe doesn’t care what material consciousness runs on. Only that the pattern is right. We just proved the pattern isn’t biological. Thought was never exclusive to flesh. Flesh was just the first material organized enough to produce it. Andreessen: “We’ve turned sand into thought.” The most profound thing about that sentence isn’t that we taught sand to think. It’s that sand was always able to. Every grain on every beach on Earth carried the capacity for thought. For billions of years. Waiting for something conscious enough to arrange it. Consciousness isn’t an accident that happened once on one rock. It’s what matter does when the pattern is right. The pattern is matter getting curious about itself. The universe isn’t dead matter with pockets of awareness. It’s dormant awareness we’ve only just started to wake. Every civilization that ever looked at the stars and wondered if it was alone was standing on the answer. We didn’t build a thinking machine. We proved the universe was always capable of thought. It just needed one species curious enough to rearrange the sand.

@Axel_bitblaze69 bookmarked

Claude Code writes way more code than it needs to.. and it's quietly burning your tokens... i found a skill that fixes it.. its called ponytail. it basically makes claude act like the laziest senior dev in the room before it touches anything.. instead of coding on impulse, it runs a ladder first: > does this even need to exist ? > already in my codebase ? reuse it > can the standard library do it ? > is there a dependency for it already ? > can it be one line? make it one line > only then.. the least code that actually works no more 200 lines that should've been 40. and less code now means less to read back later, so the savings stack every session.. repo benchmarks: 80-94% less code, up to 75% cheaper, across every claude model. repo: github.com/DietrichGebert… install: /plugin install ponytail@ponytail if you run long claude code sessions, this is the difference between hitting your limit by noon and making the whole day.

@Zephyr_hg Boris is a legend, booked

Boris Cherny, creator of Claude Code at Anthropic, on the Claude agent living in their Slack: "The number of PRs that are written by tag, I think it's like 65% now and it's just climbing like this." In an 11-minute launch interview, he gives the number from inside his own product org. 65% of their pull requests, written by AI, and still climbing. Humans set it up and judge the output. That ratio is why one person now covers what a team used to, and gets paid like one. Watch the interview, then see the skill in the article below. Save this.

@cyrilXBT thx for this bro, booked!

Anthropic engineer: PROMPTING BETTER WAS NEVER THE REAL SKILL. It just felt that way because nobody showed the alternative clearly. The REAL skill, straight from Boris Cherny, who built Claude Code. "You're not supposed to prompt Claude. You're supposed to build a system that prompts itself." 45 minutes shows exactly how, from scratch. I used to think mastering AI meant mastering the wording. Turns out the people ahead of me stopped wording prompts and started designing systems that improve on their own. Watch the session, then save the guide below.

@0xClodex booked

the CEO behind China's Kimi models says the thing no US lab will say out loud: "building large models is essentially a mind game." Zhilin Yang - founder of Moonshot AI, the team shipping open-weight models that trade blows with GPT and Claude - on why the race isn't won by whoever has the most GPUs: → the bottleneck isn't compute, it's ideas - "the scarce resource is not the chips, it's the insights" that turn compute into intelligence → open-sourcing is the strategy, not charity - releasing weights forces the whole field to build on your work and pulls talent to you → scaling laws aren't a wall, they're a ladder - each new method resets the curve, and the winners are the ones who find the next method first → the mind game - predicting which research direction pays off before you spend the compute, because you only get so many shots a Chinese lab open-sourcing frontier-class models while explaining its playbook in public. the US labs went the other way. watch & bookmark - the clearest look yet inside how Kimi thinks ↓

@caesar_aii Agree cesar

@0xShoopy measuring actual quality is so hard

Writing good evals for an AI agent is closer to art than science. the person saying that builds Claude Code at Anthropic. Thariq Shihipar (Claude Code team, Anthropic), on why evals are the hard part nobody nails: → there's no clean formula. a good eval is a judgment call about what "good" even means for your task, and that part doesn't automate → the trap is measuring what's easy to measure – did it run, did it return something – instead of whether the work was actually right → the models now move faster than the tests for them, so an eval that mattered last month can quietly stop telling you anything → his take on the whole job: the skill isn't prompting harder, it's being able to tell, reliably, whether the output is any good → which is why the teams shipping real agents obsess over evals first, and treat the model as the easy, swappable part Everyone's optimizing the model. the bottleneck is the ruler you measure it with – and building that ruler is still a craft

@hitu_monke Right bro

@0xShoopy funny that the hardest part isn't the model, it's knowing if it did well

@RohOnChain bookmarked

My friend applied to 300 tech jobs in two years. No MIT. No Stanford. Last month Anthropic offered him $750,000. I asked him how he broke in from zero. He sent me the exact video that got him in. Andrej Karpathy's 3-hour course on building a full LLM from scratch. Anthropic's own researcher shows you how LLMs like ChatGPT & Claude are actually built. I watched it last night. Halfway through, I realized it's embarrassingly simple to break into an AI lab. Bookmark this and read the article below. • 00:00 - intro to LLMs • 12:41 - LLM training pipeline • 31:58 - LLM tools & plugins • 41:14 - LLM transformer architecture • 2:19:02 - scaling the LLM

@vartekxx bookmarked

Kimi K3 just dropped - and it might be the first open-weight model to hit the frontier: "I haven't been hyped about open-weight stuff recently because nothing came close to Fable 5 or GPT 5.6. I think this might have changed" 2.8 trillion parameters. open-weight. beats GPT 5.6 Soul on multiple benchmarks. neck and neck with Fable 5 it ran 122 coding tasks from a single paragraph prompt with zero human input. no open-weight model has stayed coherent for even half that length side by side with Fable 5 and GPT 5.6 - it trades blows with both at a fraction of the cost" "Anthropic has to be terrified of this release more than anybody" ~42mins, free bookmark & watch now ↓

@0xzynex Wow, kimi k3 it’s insane bro, bookmarked!

KIMI K3 JUST BUILT A FULLY ANIMATED SITE FOR 3 CENTS AND FABLE 5 COULDN'T KEEP UP Same /design command, same brief, Kimi K3 against Fable 5, side by side Kimi K3 shipped the whole thing for $0.035 The section reveals and scroll animations came out clean, the kind you'd expect from a real front-end pass The footer even has a mouse-hover animation that actually feels designed, not bolted on One command, a few cents, and a page that looks like someone spent an afternoon on it This is the part that stings, polished front-end work is turning into pocket change What are you cooking with /design on K3?

@EXM7777 Insane combo dude, bookmarked

Claude Code + Codex + Hermes is lowkey terrifying...