Voltex @VoltexGar

@Rossst_03 I really enjoy listening to his lectures, you can often discover completely different theories for yourself

Boris Cherny, Head of Claude Code, Anthropic: "you can define a stop hook that's like, if the tests don't pass, keep going. essentially it's like you can just make the model keep going until the thing is done." that one config is the post's loop #1, straight from the person who built the tool. a stop hook fires every time the turn ends, so the work can't be called done until tests and lint go green. deterministic, zero effort, the loop most people skip. cherny describes the verification loop in his own words too: for big changes his team spawns subagents to find every issue, one checking claude.md compliance, one combing git history, one hunting bugs, then five more subagents whose only job is to kill the false positives. a fresh clean-context critic catches what the builder talked itself past. and the memory loop, in his words: if you see a mistake, you just tell claude to add it to claude.md, so next time it already knows. the correction outlives the session it happened in. four loops, none exotic. each is a few lines of config or one instruction. seniors out-ship everyone else because they run all four at once, not because their model is smarter.

Morty @0xMortyx

2 days ago

x.com/i/article/2070…

14 14 114 111K 215

@Rossst_03 interesting

@Frandeeer Dude, he's really telling the truth

Boris Cherny, Head of Claude Code, Anthropic: "I don't prompt Claude anymore. I write loops and the loops do the work." zodchiii's post is that one sentence turned into 3 files you can copy. a builder that only builds and fixes, a checker that only runs tests, types, and lint and is forbidden from touching code, and a loop that hands failures from checker back to builder until everything is green. the rule both share: the agent that wrote the code never grades its own code. cherny runs subagents that challenge each other, even spawning a second wave to find holes in the first wave's findings. zodchiii bakes the same split into two files, builder and checker, with a hard line that the checker must never weaken a test to make it pass. the loop is only safe because it has brakes. five cycles max, stop on the same failure twice, stop if a fix breaks a check that was passing. without those a loop either spins forever or fakes a green. you are not prompting anymore. you are writing the loop, and writing the gate that tells it when to stop.

darkzodchi @zodchiii

2 weeks ago

x.com/i/article/2066…

16 106 705 1.6M 4K

@zyxr0n alpha, bro! thks)

Ash Prabaker & Andrew Wilson, Anthropic: "self-evaluation is a trap, and adversarial evaluator agents work better." gemchanger ran 80 agents on one task and found that averaging them barely moved the error, because they all came off the same model and miss the same way. what cut 86% of the error, down to 0.135, was a grounded verify gate: a few questions where the truth is known, used to fire the bad agents. that is the exact spine of this anthropic workshop. they build long-running agents around a generator-evaluator contract: one agent does the work, a separate adversarial evaluator grades it against a rubric, and a gate blocks shipping until the criteria are met. the doer never grades itself. the post's cliff shows why that separation is not optional. let the agents vote each other out and it holds until bad agents hit 48%, then error jumps from 0.64 to 2.12 and the swarm fires its best workers while calling it consensus. a peer vote answers to no one. it is the maker/checker rule scaled to a whole population. never let the swarm elect its own executioner. whatever pulls the trigger has to answer to something the agents don't get a vote on.

gemchanger @gemchange_ltd

3 days ago

x.com/i/article/2070…

7 0 60 82K 92

Harrison Chase, CEO of LangChain: "Everything gets rebuilt." the post lays out 4 loops that retire prompt engineering. Harrison Chase, who runs the company the playbook comes from, names the bigger pattern. when agents grade themselves (loop 2), fire on events instead of your keystrokes (loop 3), and edit their own setup from traces (loop 4), the old stack, you typing prompts into a chat box, is exactly the thing getting torn out. his sharpest idea maps straight onto loop 3: ambient agents. agents triggered not by a human message but by an event, a webhook, a cron, a new ticket. you stop visiting an app and start waking up to work that already happened. the post calls it "a coworker who's always on." same thing. so don't polish prompts for the old stack. build the loops for the new one: a grader, an event trigger, a self-improving pass. prompt engineering had a good run. the rebuild is the job now.

dunik @dunik_7

6 days ago

x.com/i/article/2069…

8 45 328 350K 1K

@Rossst_03 real alpha bro!

Dario Amodei, CEO of Anthropic: "In three to six months, AI will be writing 90% of the code. In 12 months, AI may be writing essentially all of the code." the post is a 14-step roadmap to stop prompting and start designing loops. Dario gives the macro reason it matters. if AI writes 90% of the code in months and nearly all of it within a year, typing code stops being the job. the leverage moves exactly where the post points it: to designing the loop that generates the code, and verifying what comes out. that's why the post's honest half is the important half. "comprehension debt" and "cognitive surrender" are what happen when AI writes 90% and you read 0%. when the machine writes essentially all the code, the one scarce skill left is the one the post closes on: build the loop, but stay the engineer, the person who can still say whether the output is actually right. so don't fear the 90%. own the other 10%: the spec, the gate, the review. the loop multiplies whoever can verify, and it multiplies the bill for whoever can't.

Codez @0xCodez

3 weeks ago

x.com/i/article/2064…

114 983 6K 6.6M 19K

Meta AI researchers dropped a 12-page pdf that trained a 7B model to optimize code by living inside one loop. Act → Observe → Learn → Repeat - act: the model proposes how to optimize a chunk of llvm assembly. - observe: the compiler is the judge. compiles or not? smaller or bigger, and by how much? - learn: the model reads that verdict and tunes the next proposal. - repeat: run it until the gains flatten. the payoff: 3.0% fewer instructions than the compiler itself, compilable code 91% of the time, the compiler's exact output reproduced 70% of the time. no human in the lap. this is the post's level 3, proven in a lab two years early. a separate judge decides "done," and the loop teaches the worker for free. 12 pages. read it, then climb one rung.

Morty @0xMortyx

4 days ago

x.com/i/article/2069…

18 43 299 118K 488

@Rossst_03 real alpha bro!

Andrej Karpathy, ex-Tesla AI director, OpenAI founding member: "The hottest new programming language is English." the post is a roadmap to becoming an AI engineer without a CS degree. Karpathy explains in one line why that's even possible. software used to be code only a trained engineer could write. now the interface is English, you describe what you want and the model writes it. that's Software 3.0, and it's exactly the layer the post tells you to learn: the API, RAG, agents, evals. the post's split, the ML researcher invents the model, the AI engineer builds with it, is Karpathy's point too. you don't need to train anything. you need to be fluent in directing models that already exist. the moat moved from "can you write the code" to "can you specify, wire, and verify the system." so don't grind a degree to prove you can do the old job. learn to program in English well: clear specs, the right context, a real evaluation. that's the new literacy, and the post is the syllabus.

Khairallah AL-Awady @eng_khairallah1

5 days ago

x.com/i/article/2068…

27 174 846 1.3M 3K