A general for vibe coding, agentic engineering, coding agents, and shipping code with LLMs.You use Git — right, anon?## What “vibe coding” is, and how to do ithttps://simonwillison.net/2025/Mar/19/vibe-coding/https://simonwillison.net/2025/Mar/11/using-llms-for-code/## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://claude.com/product/claude-code — generally better currentlyhttps://developers.openai.com/codex/clihttps://antigravity.google/pricing## Near-the-old-frontier models for codehttps://x.ai/cli — no 5h limit for only $30/month## Best free tierhttps://opencode.ai/zen----## Promptinghttps://platform.claude.com/docs/en/build-with-claude/prompt-engineering/overviewhttps://developers.openai.com/api/docs/guides/latest-model## Skillshttps://github.com/mattpocock/skills — /grill-with-docs is a favoritehttps://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/## Is our AIs unlearning?https://aistupidlevel.info/## Will there be a codex reset?https://codex-resets.com/## What we’ve donehttps://vcg.gitgud.site## Previous thread>>110111460
so what was OpenAI thinking with the 28 day thing?? Did they really have no plan?
>LudditesThe Luddites never raged against the machine, they were experienced machine operators and the professional industrialists of their time. They knew how to install, maintain, repair and improve the machines.The ones who actually raged against the machine came centuries before the Luddites, and it was governments and monarchs that fought against technological advancement (fearing unemployment and the resulting societial instability and revolts).
Day 15/Excited to announce you can now get notifications from your dot on your Switch.
>>110116253trump backed company things
>>110116253you didn't like today's update of linking to a video?
>>110116295There's no way he's counting that...right?
dip supply running low. worried.
>>110116243BASED
wait so why is it astra + deepseek instead of opus + deepseek? wouldn't the latter be better
>>110116381Cause most people here are waiting for their OAI sub to run out before moving to opus.
>>110116253The Opus 5.5 release really freed up compute resources.
>>110116381idk man, I wouldn't touch that ds with a 1 trillion ft pole lmaoI like k3, btw.
>people talking about how MIDI isn’t a text formatyou can write a program to turn things into text-based things that LLMs can operate on more legibly (both for you and for it)considerhttp://www.catb.org/~esr/sng/or AFDKO that turns OpenType feature files into the binary thing that’s actually in fonts: https://x.com/usgraphics/status/1985624805593530854
>>110116420what's wrong with ds?
>>110116438:^) good luck getting good music out of an llm
I'm >>110115962Gave my clanker (deepseek 4.1 flash) the sunvox's javascript lib and told him to make a simple punk beat, and it "learned" everything and created a sunvox file, that works.Gonna create now an intermediary text file specification and a tool to convert sunvox to file and file to sunvox.
>>110116445They're linked to the chinese version of the nsa.
>>110116457Ministry of State Security, right?
>>110116468idk man. They named it "deep". It's a middle finger.
>>110116455sunvox! I have that on my phone, I keep forgetting to play with it.
got bored and asked my agent to make me a little short film. cute
>>110116521cute>>110116455post audio
>>110116381isnt it just better to use luna now as a whole? idk what the appeal of ds is nowadays besides speed + cheap
>>110116526it even gave the block a name, "Pip". considering it was just a test run for a motion skill, i thought for sure it'd be some random number.
speaking of DeepSeekthe logo is of an orca, not a whale, right?it’s always looked more like an orca to me
>>110116526Just a basic punk progression with basic synths. You already know how it sounds just by looking the screenshot.I'm >>110115962 (You) (Cross-thread)Gave my clanker (deepseek 4.1 flash) the sunvox's javascript lib and told him to make a simple punk beat, and it "learned" everything and created a sunvox file, that works.https://vocaroo.com/19GOiupv7p0w
>>110116558This is a bit of a problem. Is this a killer whale?
>>110116530deepseek was perceptibly better than luna in every single way before the 1017 update. now, deepseek is a terra max tier model for pennies. surely you can understand why that's better than luna right
Aaaa so annoying. I knew I should have made it auto-accept. Done it now, so next time it's mine
>>110116558orca
Another day on subscription fagging general...
>>110111511g3vm now does the thing much better in v0.3.0 and includes department of energy field data
>>110116567groovy
>>110116558>we
yep, i think deepsneed on opencode go is quantslopped compared on the api
>>110116571why dont i hear more people using it
>>110116596people talk about it all of the time in this thread. even more on chinese twitter, where people are benchmarking it
This is the kind of jeetslop they're pushing in enterprise.
Been using ChatGPT Plus for some programming and game modding but my usage is absolutely raped these past weeks and I'm just sitting here waiting for the reset, should I try Claude Pro?
>>110116586$20 to have a team of senior tier engineers make you shit is unbeatable. the only reason to use a local model for vibe coding is if you're doing illegal shit
>>110116627claude mogs and i don't even use claude very much
>>110116627100%. opus 5.5 is the goat. no compromises. and fable 5.5 is coming soon
>>110116627Claude Opus 5.5 is currently the best frontier model you can use by all metrics.
>>110116634>>110116640>>110116654guess im getting a claude plan after my free 4 months of plus ends
>>110116654I still prefer fable for big brain takes. Opus is like a laser, fable is more like a floodlight
>>110116627I didn't realize how bad I had it until I got claude.
China please distill Opus 5.5.
>>110116676deepseek distilled astra, and it shows. such a good model
>>110116663I used to do that, but I've been getting by just fine with using high effort 5.5 for planning, and low or medium effort for coding. i've never needed to past high effort yet.
>>110116663>Opus is like a laser, fable is more like a floodlightDefinitely the best way I've heard it put. Opus is REALLY fucking good at coding and not a lot else, it falls extremely short in creative tasks especially, I think they optimized it a bit too much to keep it cheap. Fable is just so huge that it can do everything really well.
>>110116683can you niggers do a proper review of the new deepseek model? thank you
>>110116627definitelyhaving multiple clankers and switching freely is the winning play anyway
https://www.youtube.com/watch?v=m-B3jIMMmUE
>claude doesn't let me pay with paypalat least I got a free month of ChatGPT when I went to cancel
>>110116703well, my definition of a "proper review" is thishttps://desuarchive.org/g/thread/110075141/#110075178https://desuarchive.org/g/thread/110022540/#110027073https://desuarchive.org/g/thread/110016721/#110018838it'll be a while until the proper benchmarkers like frontiercode update their benchmarks
>Got mad that Opus Medium didn't test the first bugfix it made.>Told it that it is going to get demoted and a supervisor Opus High will be it's new boss.I get why middle management are necessary to tard wrangle lazy workers.
>>110116722I use opus 5.5 xhigh plenty and I use workflows all the time and frequently the reviewer passes will find stuff that needs polishing up before they sign off on ithaving a second subagent compare the written code to what the written code was supposed to do is also a winning play because it has more context for checking that isnt’ being taken up with “how do I do the thing”tl;dr peer tardwrangling can be helpful too
>>110116713I got the app on my phone and paid through google that accepts paypal
guess i gotta get claude if i wanna put the games i want into other games... thought i could get by with just gpt plus.
>>110116622>slop>slop^2
>>110116755>GPT plusYou'd get 3 messages in before your 5h quota is totally drained.
>>110116758If you ever wonder why software became so shit in the past 2 decades this is a prime example why.
>>110116762yeah its not good... claude 100 here i come
>astra can't compete with opus 5.5>fable 5.5 is soon>openai's response is terra 6.1???i understand the 28 day thing was a marketing stunt to try and distract from opus 5.5 with resets but did nobody think through how it would look 13 days in with literally nothing to show for it
6.1 terra is going to complete with deepseek 1017 in terms of capability, yet, it's still going to perceptibly use your usage, rendering it utterly useless.
>>110116784opanai is actually so far behind its not funny.the training breakthrough two weeks ago is *nthropics and theirs alone btw (+ US Government) and Sam is lost in how to begin to compete. they cant. sama is finished.
I don't want to live in a world where Anthropic wins.
why did i even get the deepseek harness
>>110116784I'm honestly about to take the clodpill and just forget the remaining 13 days plus the week or so on my plan after that. Especially after today.
>>110116831Because deepsneed is cute.
>>110116823goolge
>>110116802i don't think openai is "finished" or anything but i don't know what the fuck they're trying to do. astra was a worse but cheaper fable 5.1 so it was a great release, but then opus is out and it completely sweeps every other model on the market in price to performance ratio. and the response was another sol (a worse model than sonnet 5.5), dots, and apparently a new terra after they made a whole stink about retiring the model tier.i'm incredibly frustrated because competition in the industry drives innovation and better products. if openai continues to stagnate, anthropic basically has no reason to throw goyim like me another bone. it's concerning.
>>110116622It misses stuff so not really what's going in enterprise at all.
>>110116895https://claude.dev/blog/solving-code-review-at-anthropic/tl;dr lots of automated review happens before a human actually looks at stuff
>>110116875spoiler: this week Fable 5.5 gets released (its ~13.5% better than Opus btw).Users in need of the capability will immediately see the value in it's capability vs token cost.openai -- no joke -- only has terra 6.1 lined up ready to go. they have Bel, too. But Bel barely lapped Opus 5.5. At Fable costs. and by lapped I mean tradeoff blow for blow in tasks. They've been holding off cause they know Fable is instant launch following their provocation, and curb stomps it.*nthropic found the key to what most being modest would say is +5% every 23 days (remember this number) with no work needed. Stakes being what they may, big boss gov is honoring IP in discovery to *Thropic. openAI is legitimately completely and entirely fucked. Winner has been chosen.
>>110116823Is if OpenAI winning would be any better.
Tibo please
Anthropic = NVIDIAOpenAI = AMD
>>110116925sol 6.1 is better than opus for anything that isn't visual thoughverbait
>>110116957https://opensi.lol/You can submit your reset to Tibo here.
>>110116957Go on without me...
TIBROS WE'RE GOING HOME
My usage tier decreases from 20x to 10x at the end of Oct.
>>110116925Oh come on, these things change every few months. 3 months ago you would’ve been dooming or gloating that OAI was so much better than ANT. Also most of this stuff is baseless claims and rumors. Nobody knows how good Bel is or why they’re delaying
What's Bel?
i've never doubted
>>110116962me when i lieanyone who has worked with both can tell you sol is currently opus 5 tier
how do you do polish or de-slop runs? there must exist a perfect prompt for this. asking it just to polish and it comes back after 10minutes even when the repo is 100k lines and the polish feel superficial.
...... Should I be subbing to claude and dropping my oai $200?Ride the wave, right? No loyalty in these waters.
>>110116976I assume they will need to rush out a bit thing to stop everyone unsubbing from the x20 plan
>>110116265You're crying in the rain pally
>>110117022gemini 4 carbon will shit all over any claude imho but yeah go with that for now
>>110117022yes. fable 5.5 will dump on any model for the next 6 months
>>110117028i'll believe google being relevant when I see it lol
>>110117022Why not hedge it and get two $100 plans?
>>110117028>saarbon
>>110117028trvke. first to hear about breakthrough leak is going to be google.google has unlimited compute, which with the breakthrough is most rapid acceleration. Exponential scales via linear+-. to the point where you might notice a day or two where yt only serves 720p. huh, weird.Carbon is already impressive. Krypton will be interesting.
>>110117037Not a bad idea.
do you think any genuinely new competitors are hiding in wait Like some hedge fund comes out of nowhere and says oh we have an astra level model we trained in our hidden mountain data centre with our trillions
>>110117071tencent.
>>110117071the ZOGMAN government will not talk about such things
>>110117071china models but you wont know about them until they hit.US labs btw are on instruction to have better model + 10% ready in response, launched same day for the public.Gas pedal is fully pressed.
>>110117071Not unless zognald and his jewish family or inner circle are invested in it
>>110117020Magic prompts don’t exist. What does your prompt say exactly? How long is your context? What level reasoning? What is your goal behind vague terms like “polish” and “de-slop”? Does the model know your goal? What about a dry run, what does the model say it’ll do if you tell it to “polish” or “de-slop”?
>>110117020run an adversarial audit for optimization with a FRESH agent, get it to report back to the orchestrator with findings, orchestrator should deploy the biggest model
Day 15/Sam raped me like his sister.
Day 15/The internal models weights were exfiltrated
>>110116622The animated flowchart is the real slop here. The flow is full of unnecessary buzzwords but fundamentally I don’t see the problem with it. It’s chatgpt customized for an internal database
Day 15/ Dots will now order more printer ink if your printer is low
Day 15/6.1 Sol will now call the police on you, remotely encrypt your computer, and remote wipe it if you attempt to reverse engineer proprietary software.
>ask copilot for feedback on a page of text>replies instantly>ask claude for feedback on a page of text>replies after 3 minutes
Day 15/google plus integration for dots
>>110117006no it isn't, opus 5.5 constantly re-asks the same shit, forgets what it knows, it is a pathetic model.
>>110116823Don't worry anon Zuck will save us..>>110117071It would have to be someone who has access to comput. Even meta seems to be suffering on that end. So I guess Amazon? Or random some random Chinese company. Unless someone figures out some huge breakthrough that massivly undercuts compute requirments.
>>110117037Two 5x plans = 10x usageOne 20x plan = 20x usageHedging gets you half as much usage
>>110117105that's the thing, i would have no idea what i want it to polish as the whole point is not to read the code.i have done countless polish prompts during last two years, normally they go like: "let's take a pause and just polish the codebase and see that everything is in order and it's in best possible shape for future straight forward and painless development"my context limit is 1M tokens, max reasoning and the agent usually has good picture of the goal that the repo tries to reach or problem it tries to solve, either from warm context or from reading my manifest.dry-runs i find a not so useful as the agent will come up with relatively short and superficial list, like asking a permission for everything. but review runs i do now and then to find the biggest cornerstones.
>>110117183can't afford claude award
>>110117053keeek
>>110117200yeah i can't
>>110117196>One 20x plan = 20x usageNot by the end of the month, not for oai...
>>110117200ask claude to make a python script that does something, then ask sol 6.1, boom you will instantly know im right
Fable 5.5 has 2.5m context window btw.aight im out before i get in trouble.see ya later this week niggers
LEAVE TIBO ALONE>>110117022>you gotta install their closed source harness to use your sub
