Aliaksei Zelianouski

@GosuCoder Ralph loop did it first

I feel like collectively we as engineers sound dumber and dumber everyday. This nonsense about writing loops is just nonsense. Why are we acting like it’s the greatest thing ever? Learn to code folks, learn to use AI efficiently, don’t burn tokens for no reason, look at your code, and please always look at the incentives for anyone giving advice. This is coming from someone that is very Pro AI.

@rohit4verse I remember the time when devs were joking about lines of code as a productivity metric. Well... now it's PRs. Why even bother with review? Just implement a Claude Code hook to push straight to main if it feels confident.

boris cherny shipped 259 PRs in 30 days and typed none of them. claude code wrote every line. peter steinberger said the same idea this week and it hit 2.2M views. one built claude code at anthropic. one built openclaw and joined openai. opposite labs, one conclusion: stop prompting agents, write the loops that prompt them. the replies are still arguing about what a loop even looks like in practice. i went looking for the built version. it's a multi-agent workspace manager called gas town: a coordinator agent runs 20-30 workers and a merge queue gates every change before it hits main. the philosophy fit in two tweets. the implementation has 15k stars.

@NathanWilbanks_ Yeah, I'm joking. But the industry needs to come up with a stronger word than 'Overengineering'

@hiper2d actually it ran 6 hours on claude code non stop. the caching plays a big part also used gpt5.5 for the original GB ports x.com/NathanWilbanks…

i have created a master loop harness that allows you to run perpetual mission loops that create goal loops that then spawn agent loops which then run their own workflow loops which finally run the tool loops. i really wish i was joking 🙃

@0xMorlex The 'same model for $20' part is off. The $20 tier taps out mid-session on anything serious - comfortable agentic coding runs ~$100/month now on Claude Code or Codex. Directing it is the real skill, but the top 1% also aren't running it on the cheap plan.

Andrej Karpathy: $20/month buys you the exact same model as the top 1%. almost nobody runs it like they do the gap was never access / one model / one tab / same price for everyone the difference is whether you can actually direct it in 30 min he maps the jump from vibe coding to agentic engineering, the skill that splits that 1% you're not behind on access. you're behind on how you run it

@theallinpod @nikesharora Mythos is great, but the capability isn't unique to it - any frontier model does this. I used Opus to place near the top of a CTF at a security conference. That big companies are only talking about AI and hacking now tells you something.

LIQUIDITY: Palo Alto Networks CEO Nikesh Arora @nikesharora joins The Besties: -- The Claude Mythos hype is real (found 5 years worth of bugs in 6 weeks) -- Analytical SaaS is DEAD, infra software is undervalued -- M&A playbook -- Armchair CEO: Google first to $10T? (0:00) Palo Alto Networks CEO Nikesh Arora joins the Besties! (0:47) Claude Mythos found 5-6 years worth of vulnerabilities in 6 weeks (5:15) Are cyber defenders losing the race against AI attackers? (6:50) Analytical SaaS is dead, so what survives the AI wave? (14:06) If models become a utility, where will the money be made? (20:35) Armchair CEO: Nikesh rates Waymo, Google, and OpenAI (28:22) Palo Alto's M&A playbook and the path to $1 trillion Thanks to our partners for making this possible! EY (@EYnews) - AI ambition isn’t enough. EY.ai Value Blueprints move organizations beyond pilots embedding measurable business value by design. ey.com/en_us/services… NYSE (@NYSE) - Thank you to our partner, the New York Stock Exchange - a modern marketplace and exchange for building the future. It all happens at the NYSE. nyse.com Plaud (@PLAUDAI) - Never miss a moment. Plaud, our official wearable AI note-taking partner at All-In Liquidity Summit, captured every insight. plaud.ai

@pbakaus What happens when two review bots ask for contradicting changes, or one misses the context the others had? AI reviews aren't that deep, and having a dialog via PRs is not the best. Fun way to burn a lot of budget - big tech and their leaderboards would love it.

for a really conceptually simple example of a loop, check out agent-reviews.com a skill + cli combo that continuously polls GitHub PR reviews, fixes bugs, which then triggers re-reviews, and doesn’t stop until every AI code review bot is satisfied.

@Prathkum "Solve the cost problem" assumes the labs want to. They're losing money and the whole dream is pricey enterprise contracts. Every new flagship is more expensive: GPT-5.5 is 2x 5.4, Gemini 3.5 is 3x 3.1, Mythos teased above Opus. Nobody up there is racing the price down.

Hot take: We don't need a more powerful model like Mythos right now. GPT-5.5 (5.6 coming soon), Opus 4.8, and similar models are already more than capable for most use cases. What we need to solve is the cost problem. If AI keeps getting significantly more expensive, 99% of developers won't be able to afford these models at scale and will end up going back to manual coding.

@HumanLoopHQ Love the 'prereqs vs craft keeps shrinking' framing. Had the same ffmpeg moment on a 2-min video - the part I assumed was craft (assembly, captions, PiP) turned out to be prereq Claude just handles. What stays human is which shots earn real motion. Mine: x.com/hiper2d/status…

Set out today to render the Maya talking-head clips and hand them to my editor for the actual Reel assembly. 30 min in, asked Claude 'can you do the whole Reel in ffmpeg?' 4 hours later: finished 31-sec IG Reel. Captions, B-roll PiP, kinetic phrase text. One Python script. "AI does prereqs, humans do craft" keeps shrinking.

@nrqa__ The brain got hands, the catch is the bill. Seedance motion runs 30c-$3 per 5 sec, so 'paste a URL, get a launch video' adds up fast. I run the same loop cheap - ffmpeg over stills, browser demos, motion only on the shots that earn it. Mine: x.com/hiper2d/status…

oh shtt.. SOMEONE JUST GAVE CLAUDE HANDS Higgsfield MCP plugs Claude straight into GPT Image 2 and Seedance 2.0. videos, images, ads, and landing pages get generated in the conversation and land in your working directory. a few things you can ask once it's connected: paste a product URL and get a finished launch video in your folder. drop in a youtube interview and get vertical clips with subtitles ready for TikTok. ask for credible data with a compelling graph, researched and dropped in your project. animate a product photo into a cinematic shot with motion presets, shipped. works inside Claude Cowork, Claude Code, OpenClaw, Hermes, Cursor, and Perplexity workflows. the brain finally got hands. higgsfield.ai/mcp

@jackfriks AI is weirdly best at the ffmpeg end - the exact, cryptic stuff that has one right answer. It's the subjective trivia like logo colours where it hands you 30 options and no decision.

sometimes i forget how ffmpeg is just efficient code and was built without the help of AI and then i remember they also built 500 ft tall pyramids 4500 years ago and yet i spent 30 minutes styling my logo colours with the help of claude but it still looks like s**t

@deedydas Great work, the flashback especially - that takes real budget. I'm on the cheap end: a few gen-AI clips stitched with ffmpeg tricks over static images, browser demos, screen recordings. The generated motion is the only pricey part, so I ration it to the shots that need it.

This is the best scene in Hell Grind, an entirely AI-made movie, the flashback. Watch it and read this analysis on where we are with AI movies today: time, cost, quality. Overall: Phenomenal technical demo by Higgsfield. Mediocre movie. Good graphics, hints of emotion, but superhero movie level quality in certain scenes at best. Too many cuts. That said, 660x fewer man hours, 50x faster and 36x cheaper than the median US film. Time: The 95 min film took 15 people 14 days. The median US theatrical production takes ~200 people ~2yrs. That’s a 660x improvement in man-hours and 50x in calendar time. Economics: It took $500k, 80% of which was compute. The final footage was cut from ~100hrs of footage generated from text to video / image to video models like Bytedance’s Seedance: a 64:1 “curation” ratio. The median US movie takes ~$18M, with even indie films costing $1-5M. Thats 36x cheaper than median. Quality: Average watch *at best*. Way too many cuts between shots, several characters change accents and have “AI” synthetic voices and characters feel like it’s AI too. Movement, editing and blocking feel artificial too. On the plus side, we’ve more or less solved character consistency, camera angles and realism. The reason the movie wasn’t amazing was more about poor directorial choices than innately unusable video models. Hard to put a number on it but maybe we’re at ~90% on quality that is technically achievable. If Scorsese made an AI movie, I reckon it would be quite good. I know the visceral reaction to anything AI is real and well-studied. But I think it’s folly to fight the inevitability of AI film. It’s too cheap and quick to ignore and almost there on quality. Creators with distribution *will* make AI films and shows and just put them on YouTube. This is the worst quality, slowest and most expensive it will ever be. In the end, good content beats “real” content.

@ridd_design Frames are where it gets fun. I had it break my generated clips into frames and diff them with SSIM to check a slow zoom actually landed - caught drift I'd never have seen by eye. Same trick you're describing, just pointed at QC instead of reverse-engineering.

a part of building with AI that still blows my mind is that you can: 1) feed Claude a video 2) tell it to break it into frames with ffmpeg 3) reverse engineer an animation/interaction

@thankyouecom Wow, that's cheap. Last time I tried Veo it was 3 bucks for 5 seconds. I'd landed on Seedance 2.0 as the best on cost and quality, but I'll give Veo Fast a shot. Is it on an API? Would wire it into my pipeline. Meanwhile, if you're curious: x.com/hiper2d/status…

We use Google Flow VEO 3 Fast. $0.10 per 8-second clip on the Ultra plan. Best model available right now and the cheapeast at scale... Ultra plan is just $249.99 a month and we topup once or twice a month extra credits. Much better than costs of paying a creator

@buildwithhassan @thdxr Just made a full 2-minute video this way. I describe the edit - hold on this frame, slow zoom, crossfade into the next - and it writes the filter_complex chain, the zoompan and xfade stuff I would never get right by hand. x.com/hiper2d/status…

@thdxr ffmpeg has always been a strong tool also, nobody wanted to learn the flags. now you just describe what you want and the agent figures it out

a whole bunch of companies that had good primitives but never figured out DX just got saved by AI i'm using all these things that were too rough to use before

Total cost: 45 bucks, no editing background. The best part is the kit of skills it left behind, so the next video starts from there. azelianouski.dev/post/building-…

Gen-AI video is expensive: 30 cents to 3 dollars per 5-second clip. So the whole thing is an economics call. Cheap slideshows carry most of the runtime, generated motion only for the shots that earn it.

I made this 2-minute cinematic AI video by talking to Claude Code. Every image, the narrator, the whole edit, generated. The strange part: the tools to make it didn't exist when I started. It wrote them as it went.