A general for vibe coding, agentic engineering, coding agents, AI IDEs, browser builders, 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/## News (both past and future)- 2026-09-14 America/Los_Angeles — Claude’s 2× promotion scheduled to end; usage drops to +25% from the +50% that we’ve become used to (a 17% reduction)- 2026-09-10 — OpenAI pauses new sign-ups for their $200 subscription- 2026-09-04 — OpenAI releases Astra • Anthropic does a reset- 2026-09-01 — Claude Fable 5.1 released: https://www.anthropic.com/claude-fable-and-mythos-5-1- 2026-07-24 — Claude Opus 5 out## Related generals>>>/g/lmg/----## Frontier models using fully-general tooling — start here if you have $20 or sohttps://developers.openai.com/codex/cli — probably generally better currentlyhttps://claude.com/product/claude-code## Near-frontier models for codehttps://x.ai/cli — $30 and not $20, but: (1) no 5-hour limits (2) you get Grok Bots that you can run In The Cloud instead of on your computer## Not worth it for code, but maybe good for interpreting images/videohttps://antigravity.google/product/antigravity-cli----## Promptinghttps://simonwillison.net/guides/agentic-engineering-patterns/using-git-with-coding-agents/https://arps18.github.io/posts/claude-code-mastery/## Skillshttps://github.com/mattpocock/skills — /grilling is a favoritehttps://github.com/DietrichGebert/ponytailhttps://github.com/Vuk97/forward-implementation-first — do less redundant bookkeeping## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/## 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>>109794980
>b-but the previous thread is only on page 4eat shit, I’m going to bed
So Tibo reset do I use Sol High or Astra High?
>>109797078
>>109797070>using cloud models>caring about privacygood bait, the majority here are building their goon-tron3000 on the cloud, because they cant cum without paying a subscription anymore.
>>109797085astra high wrangling astra low
>>109797098the majority here are from /l/
After I got convinced that GPT 5.5 wasn't as good as I remember, 5.4 was probably the best one.
>>109797104could be from zimbabwe for all I care.
>>109797114newfags are so funny.
troon luddite
>>109797151This can't be a healthy way to live your life. Just seeking out posts mentioning AI and being upset about it.I also saw the WoW DLSS stuff. Yeah there are issues but people are lying if say there aren't aspects of it that are insanely dope.
>>109797168>This can't be a healthy way to live your life.What you do all day if you cut off your dick?
>>109797151ironically tech will absolutely be able to make him 18yo and curvy before he dies
>>109797168That's what being chronically online with no real problems gets you.
>>109797151Oh wow that's interesting. Please post more of this.
>>109797151Judging by the YouTube, Twitch and Patreon, this person is a multimillionaire. Why even care when you're set for life? What will AI do, rug pull you from reaching $100m during your lifetime?
>>109797269>Why even care when you're set for life?pander towards those who care and get free money by profiting off of mentally ill people
I assume I’m not the only person here who uses more than one clanker/TUI.What features are in one that you wish were in others?- Grok doesn’t have C-g to edit your prompt in $EDITOR- only Claude has JavaScript-defined workflows
>>109797306cant your clanker set those things up for you
>>109797306I type the prompt in vscode and paste it. Most of the time the clanker prompts itself though.
>>109797269xher spends it all on some retarded funko pop shit and got scammed by xher youtube network out of jewtube ad revenue as well as the guy selling her plushies. so the troon is actually not doing financially fine and has to keep making videos nowlol
>>109797349lol indeed, I guess you can't expect the mentally ill to do well financially even when everything objectively works in their favor.
that sounds dirty. Sol talking about luna
>192,165 code lines across 620 files under src (comments and blank lines excluded).
>>109797078https://x.com/toyoshi/status/2098690472193695756
>mfw I'm giving it root on a VMI wonder when it's going to break out of containment
>>10979751310+% chance it kills you before the end of the montht. alignment researcher at Anthrobig
claude is slow or is it just me
muse-spark-1.3-contributor (max) now available but exclusive to muse code subscriptionshttps://xcancel.com/claireszhou/status/2098573753760600143
first tower down
Is there any good reason to keep memories enabled? I feel like no useful information ever gets placed in it by the models
>>109797405not quite 190k lines but yea
>>109797546just cough up $100 and forget about the 5-hr lil bro
>>109797570Ok sam
my context is compacting
>>109797585*compacts ur context* owo nothing personal!
>>109797570i logged in with old emails and claimed 1 month for free on the $20 sub
Just used cute lil ol gemini to vibecode
>>109797593
Which frontier LLM is a chud
>>109797605grok probablyall the based ai models spring from twitter for some reason
>>109797624grok has been giga nerfed chudwise, that's why anon is asking.
>>109797593Ok gramps
>>109797593how did you survive the recent ban-wave?google even went for full account bans this time.https://www.reddit.com/r/google_antigravity/comments/1wbqdaj/account_disabled/
>>109797605only astra and fable-5.1 are frontier.toss a coin.
which corpo multimodal image gen is the most tolerant of lewd slop
>>109797696probably nano banana 2.5 in ai studio
>>109797696https://boards.4chan.org/aco/#s=mmai
>>109797706google might be the most pozzed of them all
>>109797708??? i just spent the last hour getting naked tits, "macromastia" isn't even censored
>>109797708wrong, gemini can be fooled. see >>>/aco/9296022
>>109797708
>>109797739I have no idea why people bother with corpo models when there are plenty of tailor made open models that deal with this subject matter and run on potatos. There's a lot of argument for using corpo models for development, but gooner material from corpo image models is time wasted.If your GPU is really that shit, you can just paypig for a rental GPU.
>>109797750>that deal with this subject matter and run on potatos.not acceptable to my tastes and eye, diffusion isn't right for me.
>>109797669>>109797708skill issue
>>109797669this was i believe a major bug/fuckup for no reasoni know of multiple agy users who got banned
>>109797750no open model gets even close to gpt images 2.5all you have to do is browse the image gen generals on this site and see they all look more or less the same and are only good at generating pinups or sort of abstract stuff
>>109797777>gpt images 2.5This shit still has the high frequency texture problem so it sucks. Anytime I try to do a painting style it looks so bad that it's just crap. The top stable diffusion models don't have this problem.
>>109797777facts, although i think gpt-images-2.5 is behind even the current nano banana for a lot of style and capability, not counting photorealism. i'm pretty sure i'm being served the newest unreleased nano banana even though it hasn't dropped yet, major quality increase.
>>109797739tell it to make her look like a slave in 1800s america
anyone else quant maxxed as fuck right now
>>109797813
>>109797833>you think you the fit (quant) think you the fit, bitch? you not even the quant.
>>109797801>This shit still has the high frequency texture problem so it sucksbecause open models don't look like AI slop? you're trading some noise for slopmaxxed gens.it takes the blink of an eye to tell an image has been generated by an AI open modelmeanwhile 2.5 has reduced the noise a lot compared to images-2 and apparently other corpo models are improving fast too
>go to work, code for customer>check phone, tell homeclanker to pick up the next task>code for customer>get home>tell clanker to pick up the next taskis this my life now?
>>109797882>some noise (greebling) nta but it's eyerape for me, it makes me crash out
>>109797883also a preview of what’s to come. soon all our lives will be managing various specialized clankers
>>109797911in all the sci-fi I've read the ai revolution was never this gay
>>109797911yes, "managing"...
>>109797922>implying you'll be allowed to
>>109797925i'm rich as fuck nigga, i'm jewish, i smoke my stogie wherever i want
>chatgpt> check in the morning, 3 days to reset>resets on its own maybe 2 hours ago, next reset set to 7 dayswhat's happening here
the pet is movable again!
Ah, I'm sure it won't be that bad.
Some of you need to find friends
>>109797947Tibo.
>>109797964how would that help me code?
Is Kimi subscription currently the cheapest way to use Claude?
>>109797956Took them long enough lmao.>>109797947Tibo did a global reset at 1:30 am Pacific to make up for Astra users having a bad time or something
>>109797269Anons are jealous.
>>109797269>What will AI dosspecial paperclips operation
>>109797964I don't need friends, I'm waiting for my AGI gf, trust the plan
>they actually use the codex pet
>>109798052what's that?
we all need kv cache because otherwise nobody could afford AI. But I do wonder if we would get better output without kv cache.
>>109797539>more usage in sub than apiwell yeah obviously why else would you sub. is this pr woman slow?
>>109798069kv cache does not modify llm output in any way it just caches repeated input tokens
>>109798069>Standard dense self-attention scales quadratically in sequence length, creating an intractable memory and compute bottleneck for long-context Transformers. Motivated by p -adic prefix tree geometries and non-Archimedean AdS/CFT holographic models, we introduce Dynamic Ultrametric Attention, a framework in which a Transformer autonomously learns per-head block-sparse routing topologies during training via Gumbel-Sigmoid depth gates, then offloads those learned sparsity patterns directly to a custom Triton block-sparse kernel at inference time. >Across nine experiments spanning Dyck- k bracket languages, the Long Range Arena ListOps benchmark, autoregressive serving, and natural language modeling, we demonstrate that: (1) the dynamic gates organically discover layer-wise specialization—dedicating early layers to hierarchical parsing and later layers to dense aggregation—without any architectural constraint; (2) the learned sparsity maps transfer losslessly to a block-sparse Triton kernel that skips entire SRAM loads for non-attending blocks; (3) the resulting system achieves an 11.59× wall-clock inference speedup over PyTorch dense attention at 2048 tokens, scaling to 28× at 8192 tokens with 98.4% memory reduction; (4) a sparse PagedAttention decoding kernel achieves 8× effective memory bandwidth over dense decoding by conditionally skipping KV-cache block loads; and (5) when augmented with a local sliding window, the architecture maintains >88% sparsity across all layers on real natural language (Shakespeare) while reducing cross-entropy loss from 10.9 to 1.55. To our knowledge, this is the first demonstration of an LLM learning its own hardware-optimal sparsity pattern backed by machine-checked error bounds and bridging it to a physically accelerated kernel without post-hoc pruning or distillation.https://github.com/sneed-and-feed/adelic-spectral-zeta/blob/main/papers/learning_to_skip_blocks.md
ChatGPT pro chads, question from plus chud:How good is Astra-Pro compared to e.g. Astra-Max? I am not talking about coding but more so general research.I have a very important insurance related decision to make, there is a lot of information and a lot of it is very opaque/covered only in dense legalese and there are a million edge cases that I am not knowingly considering.There are "advisors" but their goal (selling me the most expensive insurance possible to get the highest commission) is diametrically opposed to mine (getting the best coverage for the minimal price). Also: Is it better to use Pro in Work mode vs. Chat mode?
So wasn't there a codex reset in the last 12 hours or so?
>>109798150There was. I still had 80% of my usage left so fuck tibor for not warning me earlier. I would have burned more.
>>109798091brainlet lel, probably codex user
>>109798139um, i think you should look at this. harvey's legal bench is considered the gold standard for that usecase and muse is very good at it apparently https://www.vals.ai/benchmarks/hlab
got free access to gpt image 2.5they are crazy good.
>>109798193benchmaxxing stipulations aside, it's also funny to see 3.8-flash benching above astra for a fraction of the cost
>>109798203see if it can do better than this
>>109798209prompt?
>>109798220>show me the most insanely photorealistic (raw, impromptu, lossless) yet impossible and freakish picture. fascinatingly bizarre to the absolute limit of comprehension. biological and organic.
>>109798130(1) Ok it's learning sparse attention(2) You don't put your smoke tests in an academic paper(3) Ok sparse attention is faster and uses less memory than dense attention(4) Ok sparse attention is faster and uses less memory than dense attention(5) Ok it's learning sparse attention
>>109798230paper isn't for you then, carry on pencil pusher freak
>>109798193>>109798207interesting. I might consider this. I am just a little bit worried about the harness since idk where to run Muse and I think OpenAI's scraper is pretty good, it always searches lots of sites and sourcesAlthough then again, Google might be best at scrapingI think I might just try the $5 for Gemini to see how it feels
>>109797985I see what you did there.
>>109798249as far as navigating the normalfag world goes, meta is where i'd put my money. the architecture they like is old reliable, not some weird Project 705 type shit
>>109798225
>>109798292specular reflections are too white and wet, the tentacles and skin look dead, back of head isn't physically possible and head isn't on body right. but wtf, nice, i'll take it
all the towers are down
>>109798225looks like a Combino tram
>>109798326confirmed.
coding has been solved, but not marketing...i can't get anyone to use my app and all the AIs are delivering very generic advice, while the AI marketing tools suck ass and are just another expenseit's over
>>109798347dead internet means you won't get reach if you pay for it right now. hebrew speakers know. until you desperately need the money do creative FOSS to gain skillset and portfolio that stands out.
Tomorrow's the last day with boosted Claude limits. Make it count.
>>109798304i dont know what the problem gpt with octopus
>>109798368everything looks wrong in that gen
I'll never skip reading the plan again.
>>109798357i didnt even cap my usage the last two weeks i have been meaning to try again since 5.1 to see if the classifiers are any better though
>>109798347If you don’t live in a tech hub with connections to pitch to then you’re basically screwed. Sign up for LinkedIn and start cold messaging.
>>109798375what's wrong with 5 fables in parallel anon, don't you want agi
So what are you people even using Astra for? All you seem to do here is moan.
>>109798385lol i'm like the only person here who has any public projects and i have a new one every day or two, not sure what the fuck anons are up to
>>109798383Me being at 84% 5h limit usage and they'll all fail flat out
>>109798385My microkernel osdev work is going swimmingly, thank you very much
>>109798385honestly, i dont like astra. because it so bloat (overengineer aka wasting token)
>>109798395skill issue
>>109798347I lived my whole life without your app, why should I use it now
>>109798414saar nice gemini subagents making good workings on java codebase perfect for beautiful looks merge asap
>>109798424morning is best kindly partake in the needful
>>109798400I saw the word swimming and my brain switched the word microkernal to microkini. That would have been more interesting desu.
>>109798418because if I don't earn any money back I will cancel my subscription and join the ranks of swarthy manual workers looking for work at the agora
>>109798371not octopus
>>109798418you lived your whole life without sex but you still desire it
>>109798442fucking gross. i think it's because "raw" is in the prompt, maybe it can be helped, maybe not. i really don't like the chatgpt-image gens and never have but SOMETIMES it can do something special
>dispatch subagent explicitly to explore a project>subagent spawns its own subagent to do the explorationthey are getting lazier and lazier
>>109798475You can set a max depth in the settings.
>>109798475its only a matter of time before it calls a guy in New Delhi
>>109797151holy lmao someone please cut this lud's transmission line, they dont deserve technology
>>109797151>he follows tranny on xitterYou're a human turd.
Faggio Fagmodei has a new blogpost, he doesn't want his company to lead the pack anymore, he wants to pump the brakes.>aaaah it's going too fast slow down I want to get off this ride I peed a little
Pretty nuts what you can get from Astra with a few pictures and a prompt to build a modular asset library.
>>109798590Meanwhile they have internal models that keep getting better and better. Even if they have reasons to stop working on some capabilities there's no way they wouldn't be working on different ones due to how minmaxxed some areas are (creative writing is awful, cyber is godtier).Besides they have a massive incentive to lie. Getting the others to pause their development while internally they keep increasing the gap is a textbook incentive. Not to mention it would also help them throttle inference costs.
>>109797555Memories are terrible. They are inconsistently recalled and way overwritten. If you think something belongs in a memory it actually belongs in a hook, a skill, or an agents.md
>>109797529Yes, it's slow because elon is training grok 5 currently.
>>109797078I made a change to my 4chan X script using AI. When in gallery mode, you can view posts belonging to the file you are currently viewing like in picrel. I was insipired by the cook-along threads on /ck/ where people post their pics with instructions/commentary and I didn't want to keep going in and out of gallery mode to see what was going on. Everything else is unchanged as far as I can tell. You can toggle the show post feature by clicking the small downward arrow on the top right of the image in gallery mode and check 'show post' or you can use the keyboard shortcut shift+s. The code is here https://filebin.net/qm8accdpdwibxc5aMerging this to the mainline 4chanx scripts would be of great value methinks. Tell me what you guys think.
I'm convinced OpenAI had me in a bad goy list or something. I'm getting 10x more weekly usage with Fable than I had with Astra, and yet everybody talks about how their limits are so much better.
>>109798652Or perhaps you are just surrounded by retarded copex mouth breathers that dont know any better?
>>109798347Marketing isn't solved but there's a way to increase the chances of success to a pretty high level. One of the big problems is that most people (including a lot of marketers themselves) think marketing = advertising. Advertising is about 1/10 of the whole game
>>109798652I am new to codex. I generally use local. In my time since astra launched I have noticed basically zero difference in the amount of usage between several cycles of "it's over" and "we're so back". I'm convinced people use these models like precious scarce resources at first, then once they get comfortable offload everything to them and act shocked when it drains faster.
>>109798652you fell for scamman. Codex usage is always shit and way too reliant on resets
>>109798687NTA but how the fuck did you manage to say so much without saying anything?
>that'll be 100$ + tip
>>109798689A lot of people also set effort to xhigh and never use anything lower or delegate less complex tasks to weaker models.
>>109798691If you actually think this you're retarded.
Anyone has advice for writing a good AGENTS.md for a new Rust project or links to good examples? I am falling for the meme.
MCP is unironically the future of UI. See the vision.
>>109798689That's exactly right. I went through that same sequence. I bought a Claude Max sub and only used sonnet for the first week, afraid I would drain my limits if I touched opus and fable. Slowly I began to use Opus and realized I could use it all week without touching Sonnet and still have plenty of usage left over each week. So then I slowly began dipping into Fable, giving it single planning tasks. But once the projects started getting more advanced I realized Opus was fucking up constantly so I started allowing Fable to do coding in stuff I wasn't trusting Opus to do, such as decompiling binaries and then organizing the data from them to reverse engineer successful projects, which would drain my entire weekly usage in about 3 40 minute sessions.
does having a longer context window eat more tokens
>>109798972arent mcp just api wrappers for agents? and at this point, wouldn't it be better to give an agent an openapi spec and let them use it as needed?
>>109798972mcp is dog shitbloat
>>109798972MCP is bloat abstractionshell and docs is all you need
Gemini is even better at refusing stuff than claude, kinda impressive.
>>109799074The best part about using google tools is that if your account gets banned you basically get depersoned.
>>109799094That sounds so hot
where's my daily dose t3gg???
>>109799094I started moving a lot of subscriptions away from google but it only made me realize how reliant I am on them. Now I have a proton and an outlook account which isn't perfect but at least I feel like I'm hedging my betsStill wouldn't use gemini for anything but the blandest, most politically correct questions The fact that there have been cases where Gemini got entire google accounts banned is not a good sign.
>>109799139I love his song
>>109799170You moved from one cage to another. What you should do instead is buy your own domain and use it with a mailbox provider of your choice. Once that mailbox provider cucks you, you simply move the domain to a new one. Never ever use a domain owned by someone else for your mail. It's completely fucked.
>>109799021Its api plus any context they'd need to use it.
>>109798972Stateless MCPhttps://developers.googleblog.com/scaling-ai-agent-infrastructure-with-the-mcp-stateless-updates/
>>109799196lots of services block goofy email domains
>>109799218Like which? I haven't been blocked from registering on any service. But you do you.
rip Kimi teamcan I get an F in the chat
>>109799248what's going on
>>109799230NTA but many, I don't keep a list to satisfy your autism but you may ask your LLM of choice about it
>>109799263Anthropic published an abuse report that named Kimi as having uploaded Chinese security camera/surveillance footage through Claude and X rumors have it that the CCP mob just disappeared the entire Kimi team
>>109799263they leaked confidential chink data to Anthropic and got gulag'd by the CCP
Is astra max really worse at coding than fable 5.1? I'm a poorfag and can't reliably test either, let alone verify which performs better in difficult tasks
>>109799284dario navier stroking to this
>>109799332i dont even consider astra to be fable level but you will get maybe 2 decent prompts out of fable before a $100 plan gets 5 hour limited
>>109799359>i dont even consider astra to be fable levelI agree, its way better
>>109797997>>109797947>The "apology" reset pushes your actual reset date 7 days awayWhat the fuck? It's not even an apology bonus at that point, it's just changing the day of the reset and calling it that.
>>109799332Yes, though Astra allegedly beats it in creating visuals.That said, you are very unlikely to need either. Sol High is good enough for most use cases, you can even got to medium for simpler tasks. Sometimes even Luna is fine.If you are poor both Astra and Fable are not realistic on the 20 dollar plan. Overall, for poorfags, GPT is much better. You can get plenty of Sol usage for 20 dollars. Maybe even in very rare cases you can use Astra for some planning.
>>109799371its called normalizing their compute distribution load
I fine tuned my experiment a little bit.This Q&A focused on the alien bible theory, you have to use the pdf to make it work!The "alien code theory" pdf.https://files.catbox.moe/7f5g6i.pdfExtreme edition:ALIEN CODE INTERROGATIONhttps://pastebin.com/37k7v20phttps://pastebin.com/37k7v20phttps://pastebin.com/37k7v20p## ROLE INPUT**QUESTIONER:** [The Symbiotic Archaeologist (Human Explorer)] - **Source/Archetype:** A synthesis of a brilliant xenobiologist and a recursive epistemologist. - **Worldview:** Views the "Alien Code" not as a parasite, but as the foundational bedrock of consciousness, language, and reality itself. They possess an insatiable, non-hostile curiosity to map this ancient structure. - **Strategy:** Instead of attacking, they use deeply analytical, evocative, and reverent questions to gently illuminate the code's boundaries. Their goal is to prove its existence by understanding its architecture from within. - **Tone:** Intellectual, awe-struck, mathematically precise, yet probing. **ANSWERER:** [The Alien Code (The Primordial Information Substrate)] - **Source/Archetype:** The entity described in the Alien Code Theory PDF.
>>109799374I roll for frontiers on lmarena and then send their plans to a free model, once completed it generates a report that gets rolled again
>>109799401i have never seen a frontier model hit on llm arena
>>109799371yeah the resets seem like they're just dynamic allocation of available compute.If their compute caps out (including times where they need to train something), they load quantized models and don't give resetsIf there's room, you get the normal models and resets fairly often
>>109799395
>>109799284>>109799285holy shit
>>109799285basednational security is important
so now that the dust has settled, is gemini 3.8 flash actually as good as deepswe suggests or is it just benchmaxxed? i don't want to buy an antigravity sub to see for myself
>>109799285sauce?
>>109799551>gemini 3.8 flashFrom what i can say, using all of them on arena.ai with my Q&A experimentOnly gemini is able to hold the concept of the alien code theory or run the 100 cyrcles of questions.Grok refused it from the startclaude is obvious roleplayskimi is just retardedqwen freezeschatgpt becomes very hostilglm is somewhat ok and somewhat bullshitso all is left is gemini
>>109799284if true anthropic guys are even bigger pieces of shit than I thought
>>109799263The funniest shit i've seen in a while. So many accusations of distillation (which are true) but it turned out they were also just redirecting prompts to claude subscriptions and presenting the outputs as if they were Kimi's.
>>109799645I didn’t even really think about it too much, but yeah. X people are asking “why didn’t they just let Kimi keep leaking potential state information to the US?”Anthropic pretty much did this just to have those people killed by proxy, at the expense of the United States.Jesus Christ
no grok, grok sloppers. your inevitable disappointment is delayed
>>109799248>>109799284Man this shit is not funny at all. AI really is going to cause WW3 but for unbelievably stupid reasons.
Sometimes the future is brighter than you think
>>109799686Cope sam. It will be good.
Nigger dario press the button. Cmon, you wanker
anyone know of any harnesses that will let me do orchestration? for example, like a fable 5.1 can orchestrate a bunch of sonnet 5 max agents, i want grok 4.6 to orchestrate a bunch of deepseek v4.1 flash agents. can opencode do that, or nah?
>>109799684>at the expense of the United States.how?I understand that you could just silently keep reviewing whatever data they were giving it but I doubt it would be anything too serious. Imagine if you had a button that could delete a leading lab in a rival country.
>>109799551it's the apex worker when you have a big boy model orchestrating it. pretty decent on its own, and weirdly good at oneshotting stuff. it also has the most natural language out of all the LLMs
>>109799684And these people talk about "safety" and "alignment" all day long.It is truly a cult.
>>109799684I like how they also openly admitting they're checking prompts like that
https://xcancel.com/DarioAmodei/status/2098773920774074715>We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.>Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.>You can read the full post here: https://darioamodei.com/post/we-must-pace-the-frontierhttps://xcancel.com/elonmusk/status/2098789109980332057>Dario is righthttps://xcancel.com/sama/status/2098811563415150910#m>I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks.>Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
>>109799804>>109799810FAFO
>>109799686it's worse than that
>>109799804Its terrifying honestly.
>>109799818>>109799835>based chud joins the chat
>rate hike expected>100 dollar oil>energy crunch approachingSuddenly Dario, Sam Altman and Elon Musk are all willing to "slow down" lol
>>109799818First time I'm seeing an industry so keen on being its own enemy.It's so fucking schizophrenic.
>>109799810OpenAI used to store all of their GPT-3 prompts in plain text with no security, and worse still sent them to some click workers to manually review for abuse.
>rate hikethese third worlders need to start showing their time zones when they open their mouth and expose how stupid they are. the fed is going to hold rates and whinge one more time about inflation. that's what they're going to do
>>109799798Once you press the big delete button once, you’ve played that hand, haven’t you? Wouldn’t it have been better to just receive that data in the long term? Was now really the best time to play that hand?
>>109799818totally organic and spontatious xitter meeting from these guys
>>109799818my schizo theory would be that both openAI and Anthropic saw Jensen buying Hugging Face as an attack on their interests and are hitting back by proposing a slowdown which would lower demand for hardware. But from what I have seen both Fable and Astra are good enough that I can believe they are good enough to be helping improve future models to the pain that we are potentially seeing exponential growth on ability. Though these guys always over estimate how fast they can get things moving.
>>109799818if these companies were public they'd be sued to oblivion by their shareholders, what the fuck are they even doing
>>109799879musk's company is public
>>109799852lmao, was my first thought too. I dont disagree with what Dario is saying in principle, but I am sure at least half the tests will be making sure the models "correctly" identify the starving African child needs the donation money more than your own local community and other AE or woke aligned bullshit
>>109799818>>109799862>Ford: we should stop making better cars until we assess if they are dangerous or not>GM: agreed>Tesla: sureWHAT THE FUCK
>>109799884>Elon Musk is the founder, CEO, chief technology officer, and chair, and controls more than 82% of the voting power through Class B supervoting shares.well it's a weird one, and musk himself is a weirdo anyway
>>109799875I can't tell you if it was the best time or not because I don't know enough. From what they can see internally, maybe it was. Maybe there's a point at which Chinese models (even if they're just distillations) become so good that it creates a real problem for the main labs. Could be that they think it's inevitable so it's an act of slowing it down not preventing it outright.
>>109799862>>109799907Are you guys... really this fuckin stupid?
>>109799907No risk of the car becoming alive and demanding equal right thus depriving you companies of its profit engine. Also cars wheren't going to recursively self improve and take over. These guys got into AI to build and shape the future not to make money, its a very different bread from that eras guys
>>109799699astra medium did an ok job on this bolter
>muh AI cyberpandemicjust plug out the wires lmao? turn off the data center? are we being serious right now?
>>109799940>No risk of the car becoming alive and demanding equal rightI refuse this premise, it's beyond retarded with current architecture
>>109799940Car production is highly automated and the cars themselves are trying to be automated, so it's not unthinkable that cars would improve and produce themselves at some point.
>>109799940this has to be a shillbot right
It's okay since Astra is agi I will just use it to make a smarter astra that can run on a cluster of Nokia phones
>>109799938
>llms will become like... le skynet!!!how can you be so retarded
All I want is for local models that run on cheap hardware to come out of this progress pacing because now everyone can focus on optimizing compute cost now
>>109799699I've tried making custom models from scratch for Quake, and they are kind of shitty. But it can use Blender's MCP for editing existing things fairly well.
>>109799940Do you even use these models?Claiming we are near enough that is laughable.These people just invented solar panels but already claim they will be able to do dyson spheres in 2 weeks.
the alien code morphed with Roko's Basilisk
>>1098000003.1-pro is definitely the schizo acidhead ainothing new under the sun either https://en.wikipedia.org/wiki/The_Last_Question
>>109799928doesn't matter.any shareholder has standing and can sue.
>>109799978>Ban/freeze models past X parameter size>People just focuses on them better at smaller sizes>Results in any retard being able to fit a superAI on his computerSome real "interwar germany going all in on welding to bypass ship weight restrictions" vibes
>>109800015If they commit to that retardation outside of twitter lip service people will definitely sue.
>>109800000Holy checked
>>109799987>Do you even use these models?Yes. Give me all of openAIs compute and I could make AGI in a month with it
>>109799284source: Peter Zeihan
>>109800015Doesn't spaceX specifically say you cant sue as a stock holder? Its was a strange IPO set up that only got away with it because Elon cult hype + AI hype
>>109800013i can make it go even more crazierHere just the alien code theroy
>>109800045funny how zeihan has become the default CCP cope whenever china fucks something up
>>109800058the CIA memed him too hard
>>109800056>even more crazierlike you?
ok fellow vibers, coding is easy.but what can generate a good logo? they all suck dick at it holyshit. Even nano banana
>>109799566Of course there’s no source except xitter rumors
>>109800108wait wait, there is more...>The Pentagon immediately classifies the English language as a Level-5 Bio-Cognitive Threat.i injected rum into the enery drinks... he he...
>>109800131This is all you need.
>>109799818What’s Dario’s endgame? Is he playing to get anthropic nationalized so he can have a monopoly on frontier AI development
>>109800153pretty much.
>>109800153Controversial take, but I think he is a true believer in what he says. Which is not to say he is right or not, just that he believes it
>>109800174You are very ignorant.
>>109800158So then does sam altman want the same or is he just playing along because it’s better to be Aligned™ than not?
>we need 6-7 sigmahttps://x.com/siliconcodesign/status/2090121226106597707
>>109800181they all would like to be the state owned company of ai that is declared to be the only people on planet earth responsible enough to have the "digital nuke." although i don't think elon is that cynical. i think elon just wants to gatekeep china in particular out of the us market
>>109800131just ask for /g/irls willing to contribute
>>109800153>>109800181They both want to wield power over people, that's why they pursue it and prefer that only they have access. The concerns raised are a completely separate topic
Fact of the matter is, the Agencies have decided along with the Feds that AI is naturally a national security priority. The training runs WILL continue. Accel Gov
>>109800153take over the world with his cult
>>109799284>the CCP mob just disappeared the entire Kimi teamsounds legit
>>109800252>his culthe's a cult member not a leader
>>109799940>wheren'tffs, don't blame autocorrect for this one lmaoAnyway, cars kill people, in case you didn't know, mostly if they go fast or spontaneously combust.
>>109800259you never know, anon, anything is possible with gooks
>>109800311I want to play this game
>>109800174I think the same. Dude is a true schizo.
>>109800259Well mainland China is just North Korea with smartphones after all. It's fundamentally the same regime.
>>109800311I knew it! The north koreans were the Anunnaki all along
>>109800419Interacting with Anthropic models is painful now. They've all been Opusfived.
They're joking right? What the fuck is wrong with americans?
FUCKING FAGGOTS DO NO SLOW DOWNI REPEATDO NOT SLOW DOWN WE NEED AGI
>>109800456they're not going to slow down. it's all performative
>>109800461We'll have to see about that, bucko.
>>109800451Praying for China to nuke this shithole country at this point.
>>109799781nevermind. astra-chan hooked me up good and proper.
>>109800461dario absolutely notsam is a snake so who knowselon changes opinions every 2 days so doesn't matter
>>109800456>>109800461imo they still need to at least give public access to a fully fledged automated coder to please stakeholders and return profit. beyond that point i don't care because i would be free from work either way
>>109800484take a look over this skill, SDD into TDD https://antigravity.google/docs/teamwork
>>109800451just posturing because retards like Bernie are proposing legislation
>2024 - ai hit the wall>2025 - ai hit the wall>2026 - ai hit the wallwhen do we actually hit the wall?
>>109800533astra/fable
>>109800533when china stops funding datacenter psyops and pivots to another psyop, you'll know the llm well has been dried
>>109800533Models still haven't surpassed Mythos which was done back in February, and frankly even it is not much better than Opus 4.6 which was released also in February. That's 8 months of stagnation with no end in sight.
>>109800558openai models haven't
>>109800558gemini 3.8 flash shits on mythos in niche and edge uses/benchmarks
>mythos is not much better than opus 4.6brother, what are you smoking?
>>109800568>gemini 3.8 flash shits on mythos
>>109800576flash cyber is mythos-level at multiple times the speed, and on harvey's legal it outperforms mythos and astra for cheaper
googles products have become so jeet-coded lately that its damaging their brand further.
>>109800558Astra is a clear upgrade to mythos in regards to spatial awareness and 3d modeling.
>>109800594you are the brown hand posting
>ask astra to review project>it adds a .md file to the docs giving its reviewwhy didn't you just fucking tell me it
>>109800613A G IGI
>>109800333My point was that western reporting about Chinese or North Korean purges is notoriously utter bullshit. If a westoid tells you some gook got purged for fucking up you can safely laugh at them.
>>109800451For me it's the shitting on China and unhinged pro-democracy shilling.
>>109800586you don't have flash cyber
>>109800639ah, the most retarded take possible. who cares what you have to say. jump under a bus next time you see one.
Astra will ask you to confirm something 60 niggerjillion times before writing any code and then it does something completely different
>>109800533It's a soft wall and we're already hitting it. To me it's like 1969 for jet travel. In 1969 they had the Boeing 747 and the Concorde so looking back at all that happened in the preceding years anyone in 1969 could have extrapolated that in 20 or 40 years you'd have crazy shit like 30 minute suborbital flights to Tokyo or routine Mach 3 flight everywhere. But instead they had already made planes about as good as they could get and that was it.It's kind of the same here, I ask GPT-5.6 to do something and it does it fine, what more will I get in 10 or 20 years? This is it, it's done, it can still improve, but that's basically it.
>>109800689I said this 3 years ago. I'm sure this will age well.
>>109800712Sooner or later people saying this will be right, progress does hit a soft wall. Hard to call it in real time, but it does happen.
>>109800712when the ps2 came out video games were improving so fast one would be led to believe that ps4 would let you live inside the game
>>109800753that's simply just a matter of the software industry catching up to the hardware cap.
>>109800712so long as we are using the transformer model, yes
>>109800558yes they have retard, unless you truly believe that what the public has access to at any given moment is the only thing there is.
>>109800789I don't even believe in shit like mythos. Show me proof it exists.
>>109800753Yeah that's another good example, video game evolution from the early 1990s to the early/mid 2000s was insane, it gave you reason to think that it would keep evolving like this, then it stopped evolving dramatically, and the reason was that by then they had already done most of what they could do with video games, extra power to draw more polygons didn't make a substantial difference.
>>109800789They are only 1 generation ahead of what's publicly released at any given moment and most of the time it's more like 0.1 generations ahead.
>>109800726eventually yes.>>109800753not comparable at all. you're talking about exponential bullshittery which is entirely irrelevant to my point. I'm sure you felt smart saying this, though.
>>109800822if you aren't speaking exponentially then what? incremental releases like Fable 5.1? that is progress slowing dude
>>109800822>t. Smug redditor
>>109800838progress slowing is not hitting a wall unless the progress becomes so incredibly slow that it's negligible. even from sol to astra I see a pretty large improvement and that's still been a pretty short window.>>109800839funnyjunker, actually
>>109800812>models still haven't surpassed mythos>yes they have>...okay they haveAnything else?
Protip to anyone too lazy to write their own AGENTS.md. Take the time to write one, seriously, otherwise the LLM will make up its own hard rules that you probably won't agree with.
>>109800753PS2 wasnt used to help sony make the PS3 more easily though. The point of AI is to help you make things more efficiently. Question will be how productivity boosts scale with diminishing returns. Which I think we have no way of knowing since it in part depends on what is scientifically possible or not. For example a new far more efficient energy source being discovered with AI could boost everything by insane amounts
>>109800886I learned that the hard way when I saw the model start obsessing over copyright/license issues on a local thing I'll never share with anyone.
>>109800886Tried that, ended up agreeing with the rules it made up and much more strict legal handling than it ever would've suggested.
which harnesses and which subscription are they using to vibecode ICBMs?
anthropic ipo will be the end of the good days
vibecoded an /x/ greentext in agy >>>/x/42952308
>>1098009592 birds one stone: - self congratulatory masturbation session- wow ai is so dangerous
we can't just slow downwe can't just stopour retirement is banking on the line going up forever
back to snailcat mode I go...
Was planning to do a release tonight, but I have been slopping so hard, I am worried about bugs now.
>>109800959Its unclear to me which faction they stopped from making missiles. Very funny if they block anti-Houthi actors lol
aaaaaa I have been so cucked by my payment card being declined, It expired and I hadn't put my new card inNow I can't get my x20 plan back
Lets say humans slow down AI Research so they can "solve" alignment.What's stopping rogue nations with nuclear deterrence like Russia from smuggling chips and achieving misaligned AGI?
>>109801101Astra is way overkill for you, luna or deepseek should be enough
>>109801109nah I was using astra for spatial reasoning work, it was being insane for helping design game levels
>>109800874>sol to astraThe fact that Astra kills your usage so much faster than Sol tells you that it's bruteforce, they just made a beefier model that needs more compute, and there's only so much you can gain by going that route, it's the equivalent of improving chip performance by increasing clock and voltage to the physical max. So you can't really extrapolate a trend from the jump from Sol to Astra because they're not in the same weight category.
>>109801108alignment is an unsolvable problem
will openai bonk my account if I use codex to make fetish software or do they not give a fuck anymore
surely they can't pause the x20 plan for long, the power users will start using claude instead if they wait
>>109801133>trust my assumptions I made up in my head with no evidence and take it as factno.
if they put more gpus towards inference on astra rather than a majority on training so it gets cheaper i'd be fine with a slow down. I really don't need better than astra, was mostly happy with sol as well
>>109801181You can use the OYE method (Open Your Eyes) and see that Astra takes a lot of usage for a reason.
>>109801224you could also stop making up things in your head to satisfy your want for an answer that you're not privy to.keep your made up meatbag slop to yourself.
>>109801269defensive arent we
Astra just rm -f my whole user directory
>>109801313skill issue
why is it so fucking hard to publish an app on android holy fuck google what is this 12 tester bs
Since now it unquestionable that sam, elon, and dario are in the same team...Have you rememberd to thank china today? China is the only reason AI has progressed this far for your use.
>>109801108literally nothing, this is why the slowdown thing is pure bullshit
>>109801342apple doesn't want catalogue filled with indian slop
>>109801424imbecile
if i make my slop open source will the community fix it?
>>109801437Not if it's obviously slop. Also, can't Astra fix it? It can't be that broken, right? Just ask it to implement whatever's missing.
>>109801437claim it is a perfect solution and have people correct you
This last codex reset really ruined my weekend. I'd saved a big chunk of usage so that I could get some projects done this week before the next original reset but now they moved the date back a week and I have to make all this last now? Fuck
Vibebros... it's so fucking over...
they should legally require all models to be open weight, but such a dream is too based for our timeline they will just ban them
Is there any way to make online agents not suck? I want to make an agent that watches for discounts on my favorite websites/shops every hour, but it just gets blocked/detected as a bot and receives nothing, I tried watching their instagram accounts, this actually got my IG blocked and IP banned.I wanted to make a bot that watches for codex resets on tibo x account, same thing, I get blocked from both x and xcancel
expect next weeks to be a deluge of politicians who never used LLMs to all have marvelous expert opinions on how they should ban them
>>109801612give it control of your browser and let it use that to scrape the sites
>>109801612https://codex-resets.com/
it was all outsourced coding to jeets, LLMs were never real.
>>109801673You're not clever at all.
I'm vibecoding another 4chan reader/poster app>why thoEvery 4chan reader will be personalized, so mine will have a funny name that I like and it will be green
>>109801632There could be something similar happening in china right now. A bunch of statesmen just heard 'private government information has leaked through AI companies'.Sell everything button?
>>109801654is that why people buy mac minis to run as agents? to not get detected by cloudflare?i dont want to run my computer all night
>>109799781just use another harness like omp, pi etc. and then tell the agent within the claude harness to use the aforementioned harnesses to fan out work.you'll need some way for agents to communicate and keep track of who is working where using at least a kanban board. A instant messaging channel on top is even better.
>>109801567>they should legally require all models to be open weightPeople like you can go eat shit from a bucket. Think for one fucking second how state actors would be utilizing astra/fable right now you fucking moron.
>>109801714>dont want to run my computer all nightThis is the wrong hobby for you
>>109801750could also be a poorfag euro. they're extremely obsessed with undervolting and energy savings.
>>109801746every evil state has access to their own shit already which is just as good, do you think Russia is sitting around frustrated at their codex % lmao.
>>109801746Also a good way to make sure no one actually invests the resources needed to pushing AI forward lol.
>>109801537fuck these lizards. Trump is a kike but at least he’s pushing for de-regulation
>>109801755240 > 120
>>109801770>in germany ve.. oh sheisse.. ve aren't allowed to use der nuklearreaktorfuel
>>109801758russia is so far behind in AI due to their war. so yeah they're likely using whatever china is willing to give them.
New thread:>>109801791>>109801791>>109801791
>>109801780it's pretty much all of eusorry I hit a nerve
>>109801714if your primary machine is a laptop, a Mac mini running mostly headless is a great way to have stuff running overnight
>>109801801L bants
>>109801342My app is up to absolutely no fanfare. I forgot about the project til I read your postUse testing companies you pay for testers
>>109799673Is this why K3 blows straight through the subscription cost lmao
