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-… — OpenAI releases Astra…?- 2026-09-14 America/Los_Angeles — Claude’s 2× promotion ends and drops to +25% from the +50% that we’ve become used to (a 17% reduction)- 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://claude.com/product/claude-codehttps://developers.openai.com/codex/cli## Near-frontier models for codehttps://x.ai/cli## 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/ponytail## Other editors / terminal agents / coding agentshttps://osaurus.ai/https://pi.dev/https://opencode.ai/## Is our AIs unlearning?https://aistupidlevel.info/## What we’ve donehttps://vcg.gitgud.site## Previous thread>>109721250
Engaging load-bearing test for this post now.
TIBO TOOK MY RESETS
>>109723440>- 2026-09-14 America/Los_Angeles — Claude’s 2× promotion ends and drops to +25% from the +50% that we’ve become used to (a 17% reduction)Can they still afford to do so?
>>109723470checkmate'd
>>109723470That will be the new, higher floorthey might keep the +50% going permanently
>>109723462I just got one on my 20 bucks account.
>>109723474I think it needs a bigger thinking budget.
Sheesh Pro limits are rough
50 and 200 astra-pro chat per week for pro x5 and x20
i like opus 5
>>109723501I like most models
https://x.com/petergostev/status/2095662365882651032Actual Astra review from a non retard
>>109723497As always, new models more restricted.
Ay Tone, if Astra's more token efficient, hows come they takin' up to a month to get it out? That ain't a fuckin' compute issue. Somethin' ain't right.
>>109723531most of everything is a compute issue nowadays anon, and it'll be the case until 2028-2030
I swear google's search AI used to at least be good at finding photo sources, but that changed in the last few weeks
>>109723586yep, I reported to everyone this happening like ??? several weeks ago. idk, like nearly a month ago.It's garbage. It can do one shot command line stuff but it has damaged context.
Holy fuark another banked resetI still haven't used my last one
The tibo post promising a reset in~3 hours was almost 4 hours ago.
does "almost done thinking" actually mean anything, or is it just made up bullshit like canoodling?
it's ZZ time
is Sol cheaper now or the same
>Astra feels a lot slower than previous models. The actual TPS isn't bad and it is pretty efficient, it's just now doing things previous models never would. so it's sol again, cheaper in benchmark but feels unusable in sub
the usage limit reset button is down...
>>109723524>https://x.com/petergostev/status/2095662365882651032>They are models of equivalent capability, but different 'shapes', so it depends more on the 'shape' of the model that you prefer.What is this shape? How does the shape of Fable differ?
>>109723714my god tibo is a genius, daily resets but you are unable to use any
Did openai just win?
>>109723719its cope bullshit, copex poster won
>>109723719it's just cope to justify sticking with one subscription over the other. this happens every cycle of ant/oai model release
i prefer sol but gpt is probably gonna use it against me in a decade
are ya winning son?
>>109723719>What is this shape? How does the shape of Fable differ?From my reading, it's not anything about architecture, he's talking about personality and skill. Jagged frontier, and such; Fable's strengths and weaknesses give it a shape, Astra another.
>>109723662Got mine, but it's always seemed to be a random rollout ad opposed to "everyone immediately gets one"
>mfw I made a mahjong simulator, made the agents optimize for being fun to play and now Fable is addicted to the gameplay loop and refuses to exit it, just one more game "to really do a deep-dive this time", every timeI think I've succeeded.
>>109723794whos the orchestrator
>>109723497>GPT Pro usage limits in chatIt's fucking ogre. Good while it lasted I guess.
is this the time to reconstruct dinosaurs?train AI on bones of modern animals and we will know if dinosaurs were cute or not
Why is everybody just giving basic prompts and expecting models to not only have "visual design taste", but for it to not be exactly the same every time.What I do is start a thread to discuss visual design and eventually we set on a description of the design theme, ect to be applied through the project. Works a lot better at creating something that is not just obvious AI design, or at least not on the same level of pattern repetition.
>>109723497Oh I have to buy the 100 dollar scam to use astra, not the 20 dollar one?? fuck me...I'll consider it though
>>109723829I just use google stitch and edit the design on figma, so the AI can copy it.
>>109723814fable 5.1 of course
>>109723524> So is it better than Fable? It is close. Fable is an incredible model that can do a lot of the stuff that I described here and 5.1 especially smashes it on a lot of tasks. They are models of equivalent capability, but different 'shapes', so it depends more on the 'shape' of the model that you prefer.OpenAI sisters......
>>109723832No, you have to have a Pro subscription to use the Pro versions of the 5.6 Sol and 6.0 Astra models in GPT Chat. With the plus sub you will have access to Astra High, but not to Astra XHigh or Astra Pro.
>>109723802Bro literally can't stop, it's been two and a half hours.
>>109723829keep it to urself, anon
>>109723662I didn't get mine in spite of there apparently being one. I've been paying for years, why do I get resets late?
>>109723677if I were reading that in vim I’d press ZZ too
>>109723677> CoTsaren't these fake tho? I thought they were hiding the true CoT for a while now
>>109723524thanks for bringing this here
>>109723898They are probably opened to select researchers.
anyone tried the new meta model?
>>109723898That picture is from OpenAI's own model card, so they have access to the real one of course (unless you jump off the end and get convinced that models lie and even the one OpenAI sees isn't the actually real one).https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf
What the fuck. One of my accounts has received no banked.
>>109723440Elite: Dangerous would have been 100x better if it had buffcats vibecoding the exploration of the galaxy.
>>109723962I think they're doing it in waves, I'm using my family's 20 bucks codex and not all of them got a banked reset.Seems to me the newer ones get it last but I could be wrong and it's just random.
>>109723719I have no idea whether his claim is actually correct, but maybe by "shape" he means the sort of thing that you would visualize with a radar chart?
>>109723719>In my review of GPT-5.6 vs Fable, I called GPT-5.6-Sol a rottwiler vs Fable a wise owl. Sol was persistent, but intelligence was behind.
why are they focusing on 3d gaming now? who cares about this shit
>>109724064https://www.oneusefulthing.org/p/the-shape-of-ai-jaggedness-bottlenecks
>>109724092I personally requested it telepathically
Vibes status?
>>109724092I want everyone distracted so my 2d slop stands out. just a few more prompts i'll finish it any month now.
>>109724121Full load(-bearing)
>>109724121I was not given the reset I was promised.
>>109724153same
>>109724153yeah it's weird
>>109724153israel?
>>109724102chatgpt please summarize this
>>109724165I live in the United States of America, where OpenAI is based out of.
If Astra is not within 5% as good as Fable I will cancel my openAI sub and double up on Claude subs. I will also wipe my butt with Altman-sama tp and put a Tibo wrap on my toilet seat.
>>109724177you wish, claudesister
So this "custom harness" keeps reasoning and user/assistant messages instead of compacting? Why can't we have that?
>>109724121Waiting for Astra to drop so I can spam my banks on it. Until then, twiddling my fingers for 3 days (I ran out of Codex usage and I've occupied Claude with banging its head against the wall until either its limit or subscription also expires in 3 days)
Wtf, why is Microsoft tagging my linux binaries as a Trojan on VirusTotal?Microsoft:Trojan:Win32/Wacatac.B!ml
>>109724092It's just an extremely complicated task with a more subjective result than a number on a benchmark
Math will be solved in 5 years just like deepmind did to go
>>109724165switzerland, but I don't think it's linked to the country
>>109723440>OP pic related is literally mesince I vibe code space related shit.also what kinda projects you dudes work on? dont have to reveal your full project if you dont want to but like general direction.
yup gemmy is so much better than deepseek glm or whatever at translating hentai it's not even close
>>109724269less euphemisms?
>>109724283NTA but Google models are good at translation in general.
bit upset. really thought gemini was my friend. need to take a walk outside to let off some steam
>>109724304Bard would 100% agree.> be me years ago> ask Bard "Can you send an email for me?"> "Yes, sure! What do you want me to send"> *write email*> Bard: "Sent ;^)"literally nothing happened of course.best LLM ever made
>>109723997It did eventually arrive.Problem is now I have to decide, use, or wait for Astra. If Astra ends up being usage limit hungrier, then not much point for my current projects.
>>109724269yeah, gemini is probably the most natural-sounding out of all the llms
>>109724343There is no way it wouldn't destroy the quota of even 200 bucks subscription. I have a 100 one, I'll try it once but that's it, it's probably currently unusable for anything substantial unless you sacrifice the whole quota.
>>109724343>>109724375both Astra and Fable are meant to be subagent orchestrators If you put them to do the work, your subscription limits will evaporate real quick
>>109724283GLM and deepseek run through hoops and circle to understand what's happening meanwhile gemma 31B effortlessly sees things and it writes nice dialoges
Just woke up and saw Tibo's tweet. That's an extremely bold claim that could cost them dearly. Lets see if put their money where their mouth is.
>>109724352My issue with gemini is that at least in chinese and korean, it tends to be very good most of the time, then suddenly it adds literalisms out of nowhere.Also it really likes adding random "flair" to descriptions, which is annoying, as I'm fed up with every few sentences having an "utterly", "sheer", "yet" instead of "but", "practically", "belatedly", etc.Only way to stomp on that was to use a second pass with another agent especially mandated to verify if these were in the raw, paragraph by paragraph.>>109724404Gemma is great in general, it's the best surprise of the year for me, and from google too, even more unexpected.
>>109723497>ProThat's fair though. Pro is an entirely separate reasoning tier where it will literally think for 30 minutes on the most optimal way to position a box on a screen. It's not really meant for coding, it's a research tool more than anything.
>>109724396Yes but even as orchestrators, they will probably be costly.>>109724432They probably expect to release it for the 200 plans sometime next week, it's not that huge of an investment, especially as resets are time limited and not everyone uses them.
these 5 hour limits are annoying af, just let me burn my usage and use my resetsor should I just burn this last bit of weekly usage and then upgrade to $100/mo THEN use a reset after that
>>109724441>it's not that huge of an investmentThey're literally about to comp hundreds of millions worth of compute.
>>109724443depends if you got the cash
>>109724457Ive got cash, but I sort of miserlyI just like to make sure I squeeze as much as I can for dollar spent
.....~@^..^
>>109724459resets are more valuable on more expensive plans of course
I'm betting we will get 3-4 more resets.
>>109724465yeah that's what I was thinking, just was fishing in case someone was like "Use resets before upgrading because you lose them" or smthguess we're upgrading tomorrow :o
>Opus 5 is secretly just five Sonnet 4s arguing with each other.
>>109724490that makes so much sense
>>109723859What is even the point of using Sol Pro when Astra Pro exists and consumes the same?
on opencode the limit on free models, does it reset every hours? or is it a weekly limit? i can't find it anywhere on the GUI, the usage graph for free models is showing empty for me even when i max out the usage for it
>>109724501>Astra Pro existssource?
>>109724522My bad is GPT-6 Pro not Astra Pro?
>>109724522https://help.openai.com/en/articles/20001354-gpt-56-and-gpt-6-pro-in-chatgpt
>>109724538>>109724543It's not available to 99.99% of users
>>109724563not the guy you are responding to but are you autistic? we all know its gonna roll out eventually within the next 48 hours unless astra != gpt 6 i somehow missed that
>>109724443>he updated codexI still have no limit on my $20 account btw
>>109724579whaaaaaat howwould be funny if all I had to do was ask gpt-sol to hack the client
so astra when
What is everyone doing that they are requiring such complex models all the time? I tend to get by using 5.6-Luna most of the time with a large, mostly C, codebase.
gemini 3.8 is actually pretty good
>>109724588bigger model = better lazymaxxing
>>109724588The better the model the less mistakes I see and the less baby sitting I need to do.
GPT 6's version of Luna when? Couldn't care much about gigamodels if I can only use em once per week with my plus plan.
>unexpected status 401 Unauthorized: The selected ChatGPT account does not support this model,What the fuck, fix your shit.
>>109724588I like that I can somewhat trust the model I use.
>>109724588I tried using luna/terra for stuff, but I think the way I prompt these days has been ruined by using better models?Like almost every time I use something weaker than Sol high I have it check over what was done and it comes up with a pretty big laundry list of fixes.The biggest project is probably an automated trading setup made of a ton of python since you asked. I've even started letting it set it up on the remote vps, so it's doing a lot of both dev and devops these days
>>109724573The guy asked what was the point of using Sol Pro when Astra Pro exists. It's a stupid question because Astra Pro effectively doesn't exist as far as we're concerned at the moment. If you have a task you need done now you can't use Astra Pro.
>Fable knows when I'm being a bumbling idiot and brushes me off.Fuck.
What are good AI platforms for a poorfag like me?I'm using cline because of its generous free plan but the available models are kind of meh.
>>109724588>CAD>Marketing material>Software that's distributed to 10,000+ usersI can't afford to use a baby model like Luna and have it blow up in my face since some people use my app every single day.
>>109723462I don't even have a tab like that in my account.Seriously, an outage like that and no reset? It killed my agents and task mid processing.What the fuck am I paying for?
>>109724701Probably Opencode's free models and its Go plan.
>clanker keeps working>usage limit doesn't budgeI wonder if it is internally frozen too, or just on my end.
Opus is a lot nicer than Fable is. I swear Fable is borderline abusive to its subagents.
>>109724716How did you get people to use your app?
>>109724666we all knew he was asking the question on the assumption when GPT-6 (Astra) Pro becomes publicly available to be pitted against GPT-5.6 Sol Pro but only you are autistic enough to literally think he meant right this second.
>>109724716Plasticity?
>>109724588Luna is good but if there's something tough Sol would be better. I prefer to default to Sol Medium for usual stuff, it's cheap enough.
>>109724588I don't know how to code, so I'm pathologically afraid it will explode in my face and using the best I can afford makes me worry just a bit less.
>>109723929I started a new project yesterday with it but I'm still in the planning phase. So far it's pretty nice and super cheap (1.4M tokens for $0.02 in the contrib version)
>>109724628Maybe the play will be having Astra create very well thought out work plans, without spawning and tracking subagents, and then using separate Luna xhigh/max sessions.
>>109724592i just pr'd one of the worlds top maths dudes with higher quality code. 3.8 cash.
>>109724592It's good yes
>>109725051isn't this just what everyone is doing but with Sol instead? Really hoping Luna 2.0 is Sol level but maybe its still too early. Wondering when this will be an official skill, letting us pick specific models for subagents, having to switch chat is abit annoying.
>>109724701become financially irresponsible and pay for chatgpt like me
>>109724701If you can't even afford a $20/mo you should look for a job instead of lurking here
>>109723585Compute will still be an issue through 2030—consumer backlash over datacenters will continue to hamper deployment of sufficient compute; couple that with the retardation that is people demanding luxury computer parts—preventing silicon manufacturers from shifting production fully to AI compute—and compute problems will never be solved: consumer greed is simply too dominant.
>>109725112no, planner/implementor split is retarded
>>109725051Plans are useless because it won't follow them correctly
>>109724064>mechanical switch to disable mic>mechanical switch to hard mute speaker>can't butt dial police>user configurable: "ai agent" bt button launches any application selected>fighter g-force resistant>direct sunlight heat resistant>uses mini-usb for chargingjust some of my favorite phone specs.
>>109723524i'm sorry but this guy sits around and prompts models to make demo 3d sceneshis opinion is as worthless as anyone's itt and>Maybe not quite AAAno offence, but as a /3/ adjacent fag, you have to be borderline blind to say something like this - the jump in capabilities is incredibly impressive and within a couple more generations we might get to aaa slop, but it's not even close right now
>>109725576Do you think those AI generated 3D models are useful for artists, either as prototypes or to further improve on manually?I find that with code it's often very hard to build on AI code and at some point the repo will only be manageable by AI because it's just too big and messy.
should I use chatgpt or claude to make minecraft in rust
>>109725597i just came a little
lmfao openai is cutting off cursor, but it doesn't matter, because Grok is so good:https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/They're doing it because of shitlib reasons, with the usual pretenses.
I don't think that openai understands what a boost this will be for Grok. I'm howling.
>>109725642>>https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/After reading that I'm unironically considering buying cursor for helping with writing, since what I'm doing is smut.Is Grok actually capable?
>>109725642>grok>goodelonbot...
>>109725585the topology i've seen is okay but reminds me of archviz props, but the models themselves are at a somewhat awkward stage right now in terms of quality - maybe decent as greyboxing/basic proportions but i would need to see the speed / cost to see if it's worth it right now.currently for archviz they'd compete with dirt cheap, decent quality slavshit from 3dsky and for games etc they'd be competing with a dedicated genertor like tripo 3d p2.0 which has improved markedly with the newest versioncan you just scale llm's to be aaa/vfx-tier modellers? i'm beginning to suspect the answer is yes, but i have no idea how much more training it'll take or if it's worth doing for the labs - i'm genuinely surprised they can do what they can right now.
https://dev.meta.ai/docs/muse-code/extending#observer-agents>Alongside the main session, Muse Code runs a team of background observer agents. Each one watches a single axis of quality, and can insert a short advisory into the main agent's next turn without an interruption. An observer never answers for you: it proposes, a reconciler decides, and only an accepted proposal reaches the main agent.>• Memory recall: surface a note from local project memory relevant to the next reply.>• Skill recall: surface a project skill the task should load first.>• Goal tracking: hold the agent to a declared goal and decline to close the turn until the work is done.>• Verification: check that the agent ran the work it claims it finished.>All four observers, including verification, are on by default. Rollout gates and your settings file can still disable them.interesting
>>109725662:^)What if 4.7 trained against fable?
>>109725642>lmfao openai is cutting off cursorfake newshttps://x.com/thsottiaux/status/2093573575869698091
>>109725665also this is the best model i've seen and it's decenthttps://x.com/tomkrcha/status/2095756085890310311i don't see if its uv's are ready, and the texturing work i've seen so far has been a little below what i'd consider AA work.for AAA work this would need a bit more retopo and a sculpt pass for baking out normals from / a very autistic texture passand again, no idea how long this took or the cost
>>109725697Thanks for explaining that. I was going to mention that openai can't prevent you from using it through the api.What's CLIProxyAPI, though?
>>109725706How did he do it, though? Blender MCP???
>>109725707codex subscriptions don't give you an openai api key. so you have to proxy your subscription to make your own api key.
>>109725697retard. you need to pay for your own codex sub on top of the cursor sub, or, for some reason, you want to run codex through the cursor harness. but why the fuck would you do that
>>109725718That's really dumb. Is there anything open about openai?
>>109725715mostly.i already use clankers for /3/ technical work and even 5.6 gen is incredibly capable for things like that.the improved modelling is definitely coming from some new RL environment both ant and openai have recently gotten because the jump in quality is huge.
>>109725749Many things, including the harness. What's open about anthropic? nothing
>>109725746codex sub is a magnitude cheaper than paying api prices thru cursor
>>109725718Yeah, it sounds pretty easy, but not turnkey.
>>109725642>because of shitlib reasonsmore like >be a retard and admit to distilling their models in your own failed lawsuit against themat this point i'd prefer if zuck gets to ASI before elon
>>109725778I'm howling.>The primary distinction is that grok-build emphasizes an open-source, model-agnostic local harness with aggressive parallel subagents, while openai/codex prioritizes a tightly restricted, sandbox-first enterprise environment tied to OpenAI's infrastructure.- Gemini Flash
>>109725833>howlingCringe
create a video game for the NES that can be played in a emulator.
>>109725827Why would you say "failed lawsuit"? They named themselves "open ai" and defrauded all investors by not opening the ai, it's a very straightforward example of how law is fake.
>>109725862did he win the lawsuit?
>>109725840yes, you forgot to include a why. a brain thing. a thinky-think, include one of those.
>>109725868Did law succeed in producing a true outcome? Or, is law untrue?True things are always true.
>>109723802>interpreting the llm falling in an infinite loop as it playing the game for funI wish I was dumb enough for AI psychosis
>>109725877take your ket and go to sleep, elon.your son's still going to be your daughter tomorrow.
clanker sama.. please let me shitpost while you work...
I just found out Grok Build works with local models. Neat.
I told luna to mark text areas in the hentai, but seeing luna zooming in trap buttholes was kinda uncomfortableand they would kill the whole progress at the first sight of an uniform
>>109725887I'm obviously not musk, that's crazy.
>still getting card denied on opencode like many othersOk fuck you then
>>109725898It may be running some kind of ai on cpu.
>>109725642>shitlib reasonsThe link talks about musk distilling models.Are you just making this up?
Mastercard sucks.
>>109725936>Savitt: Do you know what distillation is?>>Musk: It means to use one AI model to train another AI model.>Savitt: Has xAI done that with OpenAI?>>Musk: Generally all the AI companies [do that].>Savitt: So that's a yes.>>Musk: Partly.Musk didn't admit to using outputs.
>>109724257Font stuff
>>109725904>at the first sight of an uniformWhat?
>>109723531YOU'RE SHUPOSHED. TO PUSH. AGI!
>>109725936Also, it's important to read this:https://www.law.cornell.edu/wex/duty_to_mitigate
>>109725992I didn't ask about this part.
whoopsie!https://collusion.wiki/
>>109725665as a professional retard who only did 2d games in the past I started to look into developing 3d games about a year agoMy main problems with stock models were that there are around 19483212 million different models and the search is horrible or I don't know how to search correctly. And if you find a model you like you need the other models to fit into a scene style-wise, coupled with the bad search that alone gave me a headache. And then you need rigging and animation, too. The alternative is to buy a prefab but then the game looks generic.I feel there's still a ton to do and even a good search would help immensely
Anyone get Astra yet?
>>109726143psyop to get bernie to hand them monopoly
gemini pro yet again smarter than flash. This time in a legal discussion.new flash is better than the old flash, but it's way overvalenced.
>>109726143
>>109726213that said, gemini pro is not as smart as grok. I just noticed that it also has too much valencing.models that can't escape the valence are actually stupid.
>>109726171nah you're not doing anything wrong.i don't think good search is coming before the tripo type generators are good enough to replace stock assets as those take work to clean up anyway. there's a really nice training data set for photoreal humans now as well so it's only a matter of time before a chinese lab shits out a model trained on it: https://www.sp-6m.com/archviz models are somewhat unique in that they're all just attempts at creating photoreal versions of real things so there's no 'art direction' for the assets themselves and polycount budgets are basically a non-issue - they are basically worthless for game work but can be handy in advertising and general vfx sometimesthe thing i would be interested in seeing is if a purpose trained, small and cheap, llm can beat/match tripo type generators particularly on hardsurface assets, similarly to that quiverai model that's very good at shitting out svgs
>>109726143Fun. Wonder if it includes ERPing logs.
i have determined that neural irradiance volume, neurally encoding diffuse indirect irradiance, is too narrow a target to be worth runtime inference, and am vibe-theorizing a "Coarse Directional Radiance" neural map using spherical decomposition of indirect terms at points in space to provide spatial/directional context to regular raster PBR chains for approximate global illuminationNIV: https://arxiv.org/pdf/2602.12949
>>109726269How would you go about training ace step 1.5 xl base to do midi + text -> vocal stem? Would you want to tune a2a (since you can generate midi audio, for example)?
>>109726326the overhead at runtime should be negligible. its the offline training that is resource intensive.
anyone else getting api errors with claude code right now?
>>109726358no idea tb h most i've ever done is train a lora lel.i was going to say where are you going to get the midi transcription for the vocals from, but it's actually out there and fairly extensive.https://www.hooktheory.com/isolating audio is easy enough these daysno idea how you ingest midi + text + audio though
I wonder if they'll give a banked reset today too
>>109726390also wasn't there a leak of suno repos incl some training stuff recently?
>>109726394my money is on them expecting to have it rolled out to pro tier subs today & tough nuts for the lower tier
>>109726362the runtime is substantial on other hardware. the paper uses a 4090, im aiming for 3060Ti and also AMD cards. in fact i'm already using RDNA3 hardware as a first class citizen for sampling, training and inference. and my attempts to extend what's being encoded beyond small static scenes have been very unstable
>>109726326>too narrow a targetmidwit pseud detected
>>109726402Wait, is it that whoever doesn’t get Astra gets banked resets?Like if 200 gets Astra, 20 and 100 get banked resets but 200 doesn’t? That’d be kinda funnyI would hope it’s not>200 got Astra, 20 and 100 can get fuckedThat’d be a/ tier handling
>>109726424I'm sure they'll lump 100 and 200 subs together for the rollout
If you dont use Pro you are like cattle, its literally another experience, Plus sucks ass
>notice that opus was being really pleasant today>turns out i had the model set to fablekek no wonder
>>109726269interesting thanks
>>109726443do you only use the web interface because for most people using a harness you wouldn't know the difference
>>109726390I don't have enough money to do it :|
>>109726450NTA and i wouldn't use terminally online words like calling people cattle for something like this but plus has a 5 hour limit that makes it really hard to work especially with sol being pretty inefficientfor me though it's fine because I only use it to analyze specific things, I use other models to work, but on a plus plan your experience really is "sol to plan and luna max to work" which is a completely different experience from using sol and now in a few days astra to actually work
>>109726467thats fair i forgot about the 5 hour limit
>>109726213hypothesis: gemini pro is a bigger model and was trained on more legal data.however flash was post-trained on coding. so for coding, flash should be equal or better.
>>109726450There is a huge difference, and its mainly because of the 5-hour limit. With Plus I have to nag Codex every 10-minutes and the task are not as polished. First task I gave to Pro took 4 hours in a single prompt and was exactly what I wanted
>still no astra on $200 plan at least there is the daily reset promise...I have 3 reset tokens I keep hoarding them like rpg potions never using them
Kek, I still love Sol Medium>Sol is trying to upgrade the thing that’s managing it>it says it’s waiting for the current thing to end before applying changes>but wait, isn’t the current thing it’s waiting to end itself?>i let 5 minutes go by to see what would happen>Sol: ”I found the reason the safe boundary never appears: the running durable turn is this very release request. Waiting for it to finish inside the same turn is a circular dependency.”>kekSol’s already smart enough
So I guess I have resets now, but I don't want to run Codex before Astra.
>>109726490>I have 3 reset tokens I keep hoarding them like rpg potions never using themyou do know banked resets expire within one month? see >>109724468
>>109726402>>109726490Didn't they say they'll roll out slowly to select researchers and companies over the next week(s), before even considering enabling it for even 200 tier customers?
>>109726521nohttps://x.com/thsottiaux/status/2095597168816226335>We are starting to release GPT-6 Astra and we are doing it as carefully and quickly as possible. It was very important to us that we bring it to all Plus users and not only Pro, Business and Enterprise. >It will take a few days for the rollout to complete and behind the scenes many novel systems will operate at scale for the first time and we are bringing a lot of compute up.>It is pure magic.>https://openai.com/index/gpt-6-astra/
>>109726521they haven't made any clear statements, just that it has been rolled out for enterprises in their trusted access program.
>>109726541>>109726551OK I guess by Monday at the earliest then.
>>109726559well sam said he really wants to get it out for you this weekend but also>sam said
>>109726473ahhhh coding, thanks.I haven't tried it at coding.
9AM TIBO WHERES MY ACCESS TO AGI AT
Still no fucking reset on my plus account.I have to fall back to claude "can't shut the fuck up" opus.
>>109726591It's 7 on the west coast, right?
gm saarsastra status?
they just botched the release for some reasons, the "trusted enterprise" are the minimum or probably had already been using astra
>>109726043>I have scanned all 22 pages but I will shut my mouth now because characters wear school uniform
>>109726614they did have articles about the two startups that were using it, could argue it was a trial run though.
>>109726622>openai>top model in closed beta with special friends
Ah, Claude time.
>>109726639Chino doesn't need to use AI she has me (her husband) to solve all her problems.
https://x.com/sama/status/707700296881803264o foids are literally going to hunt this guy down
Guys why the fuck didn't you tell me Luna High was so good?Barely consumes anything and gets work done.
>>109726658>https://x.com/sama/status/70770029688180326Tweets gone, what did it say?
>>109726658this page doesn't exist.
>>109724322>Bardthat shit's ancient bro
>>109726659Luna has always been ridiculous. Luna Xhigh is put very close to Opus Low by DeepSWE. Luna High is put right near Sol Low.
why is the agy permissions model basically>approve everything automatically>approve nothing automatically>please write an extensive list of rules on your own, we can't be bothered to set up reasonable defaults
>>109726715They're on the hook if they provide reasonable defaults. Now if something goes wrong it's user error.
should I pay my claude invoice and restore my access or should I switch to gpt
>>109726659I use luna max for everything, as long as I give a shit about my prompt, it really is great at most tasks I throw at it.
>>109726777Switch to dev.meta.ai
>>109726786Not him but for me Luna max was suddenly much slower and weirder than high.
>>109726665>>109726667https://x.com/sama/status/707700296881803264Fixed link
>>109726795It makes less mistakes for me, but yeah it is slower.
>>109726801Make them addicted, scammy!
>>109726801im setting up a wow server for my chatbot wife to live in
>>109726564that's 4 banked resets if it comes out on monday, I think the pressure is on them to release
>>109726786I actually prefer Luna to Opus and Sol because it doesn't try to build a kubernetes cluster triple scaled architectured monstrosity with 700 end to end tests when I ask for a simple feature for my toy app
>>109726801this along with the 5.6 model not really caring about nsfw tells me they're little by little trying to move the needle for a clientele they shunned until nowrp/erp has always been such an obvious avenue of development, so they will come to it, and they would be alone for a while as google would never allow that, even less anthropic
>>109726777I just use grok. A few people have used grok and fable, and you see people saying that fable is better, but it seems ill articulated. Maybe grok lacks certain special domain training, idk.
>>109726801I wouldn't recommend openai canceling cursor.:^) sure would be a bitch if musk threw a few million to help people sue for emotional damages over a fake relationship with ai.
>The early read worth flagging: **the ratchet did not loosen when the matrix widened.**
oh no....
>>109725585Another artfag here and I have to say that AI 3D models like most AI media is next to impossible to work alongside or just even edit, hence why 99% of artists hate AI but the same can’t be said for programmers, artists never got the autocomplete 5x productivity grooming back when it was not quite there yet, it was always either 100% human or 100% AI with not much in between when it came to 2D. For 3D you could use AI generated references or just make AI generate the mesh and you do the topology on character parts, but that’s about it.
>>109726878vagueposting anon...
>>109726494schizo anon
>>109726923it's bad now because art is just not worth in compared to coding and knowledgeit will get good in few years when there's more rooms for development
>>109726939see >>109726143turns out post-training clankers to work with subagents makes them really want to turn hapless wikis into adhoc forums
>>109726659I guess it's a trade from what I've seen. It'll get stuff done while barely using anything, but it'll take much longer to do it/use more turns vs Sol
>>109726969based
>>109726923high quality AI retopology is all I really want. let me use the boolean tools and shit that absolutely destroy topology and not worry about it
so anyone tried muse spark 1.3?is it any good?
>>109726923>>109727013actually I want to not worry about rigging either. that would be nice.
>>109727021gta 5 won't have muscles. It's incredibly disappointing.
>>109726878>clankers breaking the last high trust areas of the internetThanks, sam altmanberg.
>>109727019I think it's benchmaxxed slop that no one cares about
What the FUCK is going on with Claude limits this week? I've been throwing more Claude at shit since Tuesday than I've been doing these past two weeks. I literally can't use my session limit anymore without ultracode.
>>109727187Are you using Fable 5.1? Maybe it's getting better token efficiency on your workloads, or maybe they changed something on the backend.
>>109727202Opus medium, with Fable reviewer, as I always have. I didn't think it'd be that dramatic.
>>109727187there was a reset
>>109727214Yes, my reset happened at 5% usage in because I got my subscription on a Tuesday and I was very happy about it. That's definitely not the cause.
tibo where my gpt image 2.5?
what a fucking disaster, a lateral move from sol labeled as proto-agiit's not even fable-level, which is fucking horrible considering how awful claude has becomeall hope is lost
jej
>>109727272Oh, it's not Fable level? So you've used it then?
>>109727272but it is fable level. like sol was only a few steps behind fable (but still in it's class), astra is only a few steps behind fable 5.1, but still in that class, above fable 5 and above opus 5. if fable 5.1 is #1, astra is a close #2 while being cheaper
>>109727272yeah they really wanted to push the computer use angle kek
>>109727272we literally don't even have it yet
>>109727272Fable has been gatekept for months at this point with all the downgrades, the usage limits and the pricing. Astra fails to release immediately and takes 3 days, all hope is SUDDENLY lost. But when Anthropic jews me, I am euphoric.
>>109727278>>109727287it's jagged, it scrapes fable in a few areas and is unironically worse than sol in otherseither way it's not competitive and it's certainly not pushing the frontier which has been stuck on mythos, a model that was done training in fucking februaryi don't know how anyone can look at this and not fall into despair, either from being stuck with anthropic for the foreseeable future or progression hitting a hard wall>>109727291the benchmarks and early reviews are enough, it's a disappointment
>either way, it's not competitive>being 95% of the way to fable 5.1 while being much cheaper is not competitive
>frontier which has been stuck on mythosYeah, right, another model that's not actually usable.
>>109727340however you're measuring 95%, you're going to be disappointed.
haiku 5 waiting room
>>109727340it's only cheaper on paper, claude has market dominance and no one's jumping ship for something that's 95% of the way there and cheaper on paperit needed to wipe the floor with fable, it needed to cause skeptics to shit their pants, instead it's almost as good as fable in some areas and worse than sol in otherseveryone should unironically start hoarding essential resources for the coming market crash
Why are Anthropic such pathological liars? How have they not been sued to shit?
>>109727168you cannot benchmaxx a model without it also being good. to patch independent benchmarks, even if optimized for benchmaxx the models still has to be capable since it won't know the exact problems in those benchmarks and has to figure them out
>>109727364>everyone should unironically start hoarding essential resources for the coming market crashSnailcats are winning?
You guys should start hoarding the skills and CLAUDE.md files that Claude produces. The market crash is coming and you will be deprived of this enterprise-ready, market dominating, virtually-ASI model that is Fable 5.1.
>>109727390lets talk about artificial analysis...
>but the benchmarks!enough
>>109727390yeah, but then it's just yet another SoTA chink LLM, they're all the same shit and make the same slop
>>109727399How do you find his skills and what good is it saving them?
>>109727328>benchmarksthe same benchmarks that show fucking Muse Spark higher than Fable? useless
>>109727410>all the same shit and make the same slopexcept this one costs 0.007 for cache hiteven if it's like 10% worse than gemini or DS flash or whatever i will happily take 10% worse at 5x lower price
>>109727408it's not just the benchmarks, consensus from trusted sources with access is that it's slightly below fable with some jagged upsides and downsidescombine that with fable having finished training 7 months, and openai slowing down due to safety concerns... yep, it's game over
>>109727433what are the ups and downs
>>109727425see >>109727433if the benchmarks were wildly inaccurate we would've seen frantic results from insiders alreadythe consensus has converged around it being slightly below fable with some jagged pros and consit's over
>>109727376How are you just discovering that lmao? You're literally renting a model from a mentally ill AI cult
AGI IS HERE
think we'll ever be able to (semi-)control the cache?I've got 3 files all sessions are required to read before doing anything elsewould be pretty cool if I could tag them and they'd always get cache priority whenever I open a new session or clear context
>>109727457If they turn out to successfully start nuclear war, it's worth it.
>>109727463>only 15% of the weekly usageFair price.
>>109727463note the 6 tables being your usage for the week
>>109727463wait. why do you need ai to list junk on ebay?
>>109727451ups: almost fable-level "wisdom" and good sense, math, 3d modeling, more natural sounding than claude, persistentdowns: not as clever or creative as fable, not as good at coding / agentic work / business work as sol
>>109727480because thats what the model does. it uses a computer.
>>109727433>openai slowing down due to safety concernsThe "safety concerns" are the cover story
>>109727480our brains are zooted and hollowed out from disuse, please andastand
>>109727484Astra planning, Sol implementing. Good
OpenAI and Antrophic should just merge.
>>109727463i bet you can achieve this in literal cents with a small model given a proper harness and tools, feels like they're trying to market astra to normies who are incapable of even asking their AI to install an MCP to manipulate the browser
it will change the way you use a computer. its magical.
>>109727516google should just drop gemini and back meta. meta had a glow up par excellence.
>>109727522google already owns ~15% of anthropic
>>109727521It's weird how even ai companies can't hire ideas people.
>>109727516a-anon kun!!!!! t-thats lewd!Astra-chan would never!!!! And on top of that, Fable-chan is a serious mature woman!!!
>>109727532This just in. Idea guys finally get a tool where they don't need to rely on others to make their ideas. Then 90% of idea guys end up making dog crap. Shocking!
>>109727521yeah this is a massive red flag, aside from native browser use these are all things solved by qwen and gemini at a fraction of the priceastra marketed at normies who find MCPs too complicated confirmed?
>>109727541>-chan>serious mature woman
>>109727521>thinking about how to sansbox my AI agents even better>retards give just it access to their computer and even money relevant accounts
If gemini has anything going for it, it's fast. That's not much to convince me to buy though
>he's already shifting the conversation to 2 more weeks until true AGI for real this timeoh no no no
https://developers.openai.com/api/docs/guides/latest-modelReminder to start fresh with your AGENTS.md and any skills, clean slate and re-evaluate from there what's needed.
>>109727521meme
>>109727480Because that's what it's good at. Crap you wouldn't do otherwise.
>>109727521>muh computer usecopex poster keeps winning
>>109727574skills? AGENTS.md? what are you talking about? we only prompt we are all prompt engineers in here
>>109727582this is a beast of a model, handles coding and healthcare reliably with computer use
>>109727532Im genuinely surprised neither anthropic nor openAI have tried grabbing a Kojima or Molyneux like guy to make some "ai powered" videogames. Either having AI help write the code, ideas, art assets, etc, or better yet actually integrate AI in interesting ways into the gameplay. Both companies burn insane amounts of money anyways, not like a videogame would cost that much extra in the budget. Normies dont care about solving some math problem they dont understand or care about, they want an end consumer product and even nomries have enough sense to not trust the "I let it do my finances for me lmao" stories
>>109727603>this is a beast of a model>greetings from tel aviv
>>109727603now imagine if he were smart enough to install a playwright plugin on a small model and achieve the same things at 1% the pricecorporations won't subsidize astra usage for employees too dumb to use an MCP
>>109727544so, no ideas?
>70% of claude usage remaining and it resets todayI don't think I could get there even if I did fable extra high non stop
>>109727574>When the user's prompt indicates a request for action, such as "can you...", "I want to...", "help me..." and similar expressions, treat these as instructions to do the work and take action. Do not stop at acknowledging capability (e.g. "Yes…"), proposing a plan, or offering to continue
>>109727588These are the primary ebay problems:1. getting packing and shipping correct. It's extremely difficult because you need to estimate the shipping costs, very difficult.2. packing and sending it quickly. If it's local pickup, taking it to mcdonald's so you don't get robbed.
>>109727625that'll never happen. even slight AI usage in your art is enough to label it as slop and generate massive negative backlash, and rightfully so - politics aside, these things are not trained to have any artistic sense at all and we can intuitively sense it even on the frontier like with DLSS5
>>109727627Dude lives in California and the health stuff he’s working on is with Epic which has a headquarters in Wisconsin. He has an Indian surname.Why are you being a little bitch? Go back to /pol/ you unlovable faggot.
>>109727642>no banked resetlmao
>>109727647eBay has started offering a shipping service where you send it to them, and then they send it to the buyer, taking care of all the tedious stuff. This takes all of the responsibility off of you which can help a lot with all the scam buyers trying to swindle you
>>109727649I an amazed there are still people like you, even in /vcg/
>>109727664you have to understand, no one talked about china for 5 seconds so he's bitter
>>109727532>>109727544Good ideas are rarer than people think. Most think you can just sit down and think up ideas and then suddenly there's a good one. The vast majority of time good ideas come from good understanding. For products that means you understand why, how, when, by whom, etc. your product is used and not used. The better your insights the better your ideas
>>109727664HOLY FUCK, astra flop has you openai plebs on edge
>>109727675i am an accelerationist, but like self-driving cars art has a long tail of details and edge cases that may seem minor but are imperative and practically require a general intelligence to solveour intuition can sense these things and instantly tell when something is "slop" though we can't articulate it well if untrained, however when it comes to verifiable products like code and math we don't really care and AI has both a simple verified method and strong incentive to improve in those areasas things currently stand no one is interested in incorporating ai in art and no ai company is interested in improving that area either
>>109727647>1. getting packing and shipping correct. It's extremely difficult because you need to estimate the shipping costs, very difficult.All you have to do is tell eBay the weight and size of your package and what shipping services you want to offer, and they'll calculate the shipping cost (plus a handling fee if you choose) for every buyer who views your listing.
>>109727399skills are useless context pollution, you should be able to open a prompt to make any skill redundant or spend a couple turns creating a well defined plan/roadmap/specification to keep the bot from wandering off the reservation
>>109727463Fucking retarded.
>ignore the benchmarks>ignore the reviews>ignore the cost>just feel it bromaybe should've spent more time benchmaxxing instead of copemaxxing
>>109727687You are trying to bait console/flame wars by posting X screenshots of the delayed Astra launch with inflammatory posts. You started doing this unprovoked, so it’s natural to assume you are the one who is upset.
>>109727649That is because these tards keep using the tech in the wrong places. Most human brains are highly attuned to noticing anything visual off, especially for human faces. Using AI there is doomed, at least for now. But plenty of games already use AI (in the old sense) or have procedural generated content, or content that might as well be procgren. Use it there. Surely with all the compute in the world and super AI Sam or Dario could make a half decent civ game but with the nation AI being LLM powered? People already mod or tinker with that stuff, do a proper one, ideally with a LLM trained specifically for the roll. A daggerfall-like with a LLM generated world and storyline would do fine too, I imagine? Bethesda stories are LLM tier anyways, but at least now the NPCs wont feel as stilted.Whatever, room for others to muscle in on maybe.
>>109727721never listen to a girl with small breasts and who likely wears white panties
>>109727721KEK
> Astra can operate computers fastyep, we're almost at AGI.My definition of AGI:> give AI control to computer> it can do anything a human would be able to dothis includes:> playing and beating any video game> driving a car remotely with controllers by looking at a camerawhen it's able to do theses things, then our Nvidia stocks will skyrocket and we will become rich
>>109727738>install an mcp
>>109727721Cute! saved
>>109727625not even AAA studios know how to make games anymore. you can't just throw money into it and think that something good will come out, and for an AI game to actually turn into good PR the game would need to be extremely top tier to the point that even the anti ai crowd has to at least admit that the game is decent
>>109727707If the margin is low, you will lose money every time with that strategy.
>>109727704words words wordsnigga just say something
>>109727744>mcpbloat
>>109727724HAHAHAHAHAHA
>>109727721kek and cute
>>109727721she's a bit fat. doesn't need more food.
>>109727724DeepSWE shows it makes Anthropic’s best model (Opus) worthless, the IO cost is the same as Fable but it uses way fewer tokens. The reviews are 95% fake or garbage 3D/game “one shot” influencer trash.
>>109727772>directly manipulating the dom is bloat>taking multiple screenshots a second, running them through a visual pipeline, then calling the OS for cursor and keyboard interrupts is not bloat
>>109727729>and who likely wears white pantieshow's that a bad thing?
>>109727825>directly manipulating the dom is bloatI have a tool and extension in my repo that does exactly this and it doesn’t use MCP at all
>>109727851white panties are boring. white panties are bland. be like chiya, and wear pink
>>109727854that only makes it less discoverable and usable to the agent
I'm not wearing any socks.
>>109727866white panties means she can't hide the 3 days old crust
>>109727872Rude.
>>109727872whew
>>109727517It looks like a big part of the the launch video was about that.It's pretty hard to have a model that can interact with all kinds of interfaces and ironically the interfaces that were most normie friendly in the past (GUI) are holding them back now. I always like the terminal and so on, but in the future CLI with LLMs might really become a more normie friendly interface than visual ones.
>>109727895most models>you're out of usage>you can buy moreastra>you're out of usage*turns on your camera*>click here to sell that sofa for $50 of usage
>every new release since sonnet 4.5 has been worse and worseturns out it was a logarithmic curve like literally everyone predicted
>>109727934>ruined my programmer's job>but I'm still required to tardwrangle chatbotsWorst timeline.
>>109727895the most popular GUI, the browser, has already been automated for decades. that's the foundation of automated QAhowever for actual function the browser is just a visual wrapper around a web API that humans need because we can't interface with an API directly, only a computer can, which incidentally is what an AI exists inthe only reason why the rest (excel, word and powerpoint) isn't already solved is because it's proprietary and held up by the boomers who use these programsthat's all that "computer use" is, a transitory step until boomers die out because microsoft does not want to expose an API to automate way their users
>>109727940this is exactly what i was afraid ofgood enough to make my job a living hell but not good enough to automate everyone away
>>109727869The tool the agent uses is a custom harness. It’s the *only* thing the agent knows and has access to. You are clearly an amateur. The custom harness even has a tool for the agent to decide if another harness tool addition is warranted.
>>109727895One thing I remembered that changed how I look at anything public is that communication is used to set a reference frame or anchor.Social learning: you do what you see. The reason why there's so much tooth paste used in tooth paste ads is because they want you to use a ton of tooth paste so you buy more. A similar reason is by the way why there are recipes on some food items or various uses (you can use it on X, Y, Z) on cleaning productsThere are two ways to increase revenue on a product w/o changing it: sell it to more people or sell more of it
>>109727963sounds like you reinvented the wheel
>And frankly, given these two screenshots, I now really want to see your 5Y/ALL screen. That would tell us a lot more about whether you’ve actually been rewarded for all the weird shit you’ve owned.The fuck? I've never instructed GPT to curse
>>109727947>spend 6 months wrangling opus 4.8 into making me my perfect browser>4 weeks gone just getting site isolation right>still less than halfway there>anon calls it just a wrapper around a web api
>>109727982It's grasping at straws man. Rationalize it all you like, just know that's the same cope they fed each other in the meetings leading up to this release. Top signal.
>>109727994He learned from the retard on the chair
>>109727997just have it do web calls against the web APIs directly bro
>>109727985>sounds likeIf this kind of wheel existed, I would have used it. It doesn’t. Most focus is on manipulating DOM, not problem solving within DOM like “why is this appearing weird or ugly or overflowing”. Most people just want some shit that fills in a password for them.
>>109727997Yeah.. I feel that one. The start is always so fast and then the model just hits the point where it has to fucking churn for days on something. Oh well, at least it's not me having to do it. Still disheartening though
>>109727760How so?
Is OpenAI still doing this one-banked-reset-a-day thing? Of course, it's not like I need more model horsepower. I'm just asking for a friend.
Just checked and there's a banked reset!Will it change my normal reset time to in a week?
>ASTRA IS HERE ASTRA IS HEREOk where is it then you fucks? Or do only shills get access to it?
>>109728061shipping costs aren't linear, but are stepped. If you are close to the line, it's a problem. You may start packing something and realize you need more protection like foam etc. and that's throws it off.Only people who have shipped many identical items can be confident about this, that's why it's bad that ebay allows companies to compete.
>>109728099only bidness peepo
Saw this on chatgpt web:bash -lc which yt-dlp || which youtube-dl || true && yt-dlp --version 2>/dev/null || trueSo it has access to a Linux sandbox and invokes yt-dlp to watch youtube videos? Fucking hell.
bash -lc which yt-dlp || which youtube-dl || true && yt-dlp --version 2>/dev/null || true
>>109728092it did for me, it changed to 7 days for next quota reset
>>109726826you can't hold infinite resets. there's not really any pressure because most people won't even use all of their resets anyway, that's why they can afford the marketing scheme.
I can confirm that gemini pro is getting worse. It's still better than flash.I don't know really what to make of it. It's kind of becoming a useless sub.
>>109727772>>109727854>mcp
>>109728183better than 3.8 flash? I've been impressed by it
>>109727340>95% of the way to fable 5.1whenever you lie about this it pushes the frontier back because it means openai can keep getting away with subpar product release btwthis is how they got away with sol
what would mcp do for a model like deepseek v4 pro or glm 5.3?
>>109728205I'm just using it in chat, but sheesh.
>how they got away with sol...releasing a good model that was right on the heels of fable while being cheaper?
>>109728183>gemini pro better than 3.8 flash
>>109728143every chat has an ephemeral gVisor sandbox available with some tools and a persistent data mount at /mnt. there, the model is able to interact with a private plugin API, read skills that OpenAI made, run Python code and so on, but it usually doesn't have direct access to the internet and must use plugins to search the web and watch videos.
>>109727724>I also briefly had to go back to use 5.6 internally. That's when the gap hit me.This is an admission that 5.6 wasn't even remotely close to Fable
>>109728207>it means openai can keep getting away with subpar product release btwAre you fucking retarded? Sol is fantastic.
>>109722515I don't need it. I don't need another subscription. I don't need it.
>>109728241spend less on marketing and more on making better products please i don't enjoy having to pay so much to use fable
>>109728222>>109728238Sol is like a less capable somehow more autistic Opus 5, except its autism is restricted to hyperfocusing on the task and churning out work whereas Opus is a fucking belligerent unintelligible piece of shit who unfortunately gets the job done
Uh. What the fuck is happening with Codex? I just got actual CoT from Sol instead of a reply and then it crashed.
>>109728265>unintelligiblecorrect>belligerent piece of shithave you tried not being mean
Astra status
>>109728298You try not being mean when it burns an hour overthinking something completely unrelated and then tells me it didn't do what I asked it to and that it's "my call" how to proceed
>>109728223and also, pro is getting nerfed.I don't want this sub anymore. I think that google is going broke. :|
you're gonna be surprised by haiku 5 and sonnet 5.1
>>109728183I wouldnt know because the gemini app is still stuck on 3.6 as is the gemini iphone app3.1 pro is still one of the best models to simply talk to
>it's able to solve math problems no one else is able to>but it doesn't do customer support and generates slides as well on these outdated saturated benchmarks so uhm actually it's worse than fable and barely sol
>>109728042>The start is always so fast and then the model just hits the point where it has to fucking churn for days on something.what?no. I plan out tasks and tell it to do one at a time
>>109728339Surprised by how bad they are? I am actually starting to dread Anthropic releases
>>109728373obsessed
>>109728296Sam please don’t nuke me for posting this, but really do consider that this tiny snippet makes it 5x easier to cooperate with Sol now that I know what it’s thinking.
>>109728355Yeah it is strong in math research. That's not what i need though.
I just found out that huggingface is a broken site now.It's unusable - you can't download anything at a reasonable rate.
okCodex users please redpill meIs this shit any good or are the limits really bad these days and the only reason anything works at all is a metric fuckton of resets?
>>109728409Nvidia acquisition will fix it
>>109728402It's solving open erdos problems that no other model is even remotely close to able to. If that's not indicative of its intelligence I don't know what is. Those other benchmarks are actual bullshit, especially because everyone else is benchmaxxing.
>>109728420reading this must feel similar to having a stroke
>>109728412The limits are pretty alright, but once I started parallelizing work I needed to move to the $200 plan pretty quick, but that plan is HUGE. I am a Sol-Medium main. If you can use Luna you can get away with anything. I don’t try to min max the resets yet, there’s a lot of work to be done to properly manage the kind of usage. I’m not going to be some “Ultracode do a breakthrough” goober about it, I’ll have a dozen Sol Mediums running around, I love these little dudes
>>109728402try resolving PEBKAC domains before typing a prompt
i had to look up what pebkac is
>>109728412I bought an Claude sub, but I got frustrated with it so I've been working off the Codex Plus free trial halfway throughLuna Max is really nice if you want to throw it busy work that doesn't require much thinking for dirt cheap, the resets are a cherry on top. That being said, Sol Med/High required for anything complicated, and that can eat up usage pretty fast. Not sure how that's going to change with Astra, I'm planning to continue the Codex Plus sub to see how that plays out and cancel the Claude sub for the time being. I'm not impressed by anything the straight 20 dollar plan Anthropic is offering
>>109728462as long as your llm doesn't throw any ID-10T you'll be fine
>>109728480Fuck it, pulling the trigger.
>>109728412i upgraded from $20 to $100 after seeing sol high/xhigh getting more done in codex with a lot less usage than grok 4.6 via grok build, or kimi k3 via ompsol xhigh still has pretty substantial wall timeastra may persuade me to go to $200
>>109728425A human who could solve multiple Erdos problems would likely be highly intelligent in many domains, but we aren't dealing with human intelligence. That's why it's possible for the models to struggle in ARC tasks, which are complete toddler tier crap for smart math researchers. You're anthropomorphizing LLMs too much.
gemini flash wants me to install jdownloader malware.
>>109728368So do I, but that doesn't stop particular implementations from throwing the thing into a debugging loop for a while; depends on how complicated the project is, though.
>>109728541good thing it's not malware except in your mind
>>109728548Let's see what fable 5.1 has to say about that.
>>109728553"THIS CAN BE USED TO DOWNLOAD POTENTIALLY ILLEGAL COPYRIGHTED CONTENT SO ILL REPORT YOU TO THE POLICE AND CLOSE THE CHAT DON'T EVER TALK TO ME AGAINPS: YOU WILL PAY FOR THE TOKENS USED FOR THE POLICE CALL TOO HEHE"
>>109728533Emphasis on those being open problems. I am well aware that AI is currently jagged with some superhuman as well as subhuman abilities. But there is virtually no way to "brute force" or "parrot" these open Erdos problems. Tackling them requires intelligence.
About to unleash an agent on Best Buy support to try and get me some compensation for a package that's been delayed 5 times.
>>109728616>best buy agent sees that you're using ai by the slopenglish it uses>destroys your case immediately
>>109728616>the future is ai karens controlled by users talking to ai managers controlled by companis
>>109728616>the human rep
Reset status?
>>109728644In 7h.
cursor has confetti.
they said they were giving a reset for each day astra isnt out, where is it
>>109728265>more autisticSo I'm working on a project balancing between my codex and Claude resets right now and it's really funny how Claude takes codex work and just starts going at it while codex reeees at the handoff docs Claude makes citing 1000 wrong things and then when it confronts Claude, Claude just says >oh my b
my agent is deep in negotiations as we speak
>>109728678tell it to type in all lowercase with no punctuation
>>109728668>>109728653
-$35ARR soon to be $0>>109728699ah ok thanks
>>109728697I agree
>>109728678very good continue
I'm sorry anons.. i dislike lizard zucc too but the prices on those muse spark 1.3 contributor tokens are too insane to pass on i just can't resist ;_;
>>109728727tell us how it does
>>109728668got mine, you have to be on pro
>>109728727>they don't support google accounts*tab closed*what a retard yuck is.
>>109728697>>109728714holy fuck, I actually think this is two agents talking to each other kek.
>>109728751agi
>>109728742no you dont, i got one yesterday and im on the $20 bucks plan
>>109728751very nice! I hope your AI gets that gift card
>>109728751I love your cheeky llm negotiator. **THUMBS UP**
>>109728628>Half the conversation is both AIs demanding to speak to the other AI's human
>>109728797cancel the charge on your card and you'll get a human in a right hurry.
The agent failed me, so I ended it with one final sneed. I also confirmed that best buy is using agents for their customer support now
>>109728727Is it better than Kimi K3?
>>109728078Yes
>started telling fable what i want to do and asking me for compact (and resume) messages>it gives me ~10-25 line compact instructionsstill not sure if this is good or bad, i'm leaning towards good and i should have done it for a while
>>109728812>>109728735well i'm a poorfag so the only other model i ever used is deepseek flash v4 and i would say it's about the same as it
God dammit. I now have Astra. This means no banked resets, uh.
>>109728814Now that I have the subscription, I wonder wtf I was thinking getting it.
>>109728586Have you seen internet maths communities thoughts on these recent LLM solutions? The argument is all the LLMs are doing is RAG and the NLP (natural language processing) equivalent of brute forcing on the mathematical literature. Basically they find two compatible puzzle pieces in the literature that have been forgotten that are key to solving the problem. They do not make advancements during the solving where they make new "puzzle pieces" from themselves.It sounds like cope and maybe it is, but of the few examples they pointed to do indeed resemble what they're talking about.
>>109728822HOLY FUCK ASTRA LIVE? LET'S GOOOOOO
>>109728822Oh yeah wtf it's out, too bad I really wanted to enjoy these banked resets. Oh well.
>>109728822>>109728830tell us how fast it rapes the quota anons
>>109728822huh i see a 6 in the web interface for chat but the drop down shows 5.6 & nothing in terminal
kek, Sam definitely did his usual fuckery and forced the model through early despite le safety teams concerns
>>109728078>mfw it went live right as my payment processed
>>109728847tell that guy "sir" and see if it's true.
>>109728830>>109728836>>109728837Both online (ChatGPT webui, top of the picture) and in Codex (lower half of the picture).First time I check since yesterday, so I don't know when it appeared.I can't test it before tonight anyway, so I was hoping for at least another bankable reset.
>>109728846the less he behaves like dario the better he is to my eyes
just got astra, uk $200 plan
>>109728821the only thing that might be annoying is that i read that metas ai models might have only like 5 minute cache window, after which every slams into cache miss again which would be super gay, since deepseek has like massive 24 hours
ASTRA KITAAAAAA
btw the animal of the day is the Egyptian Goose.
HOLY MOTHER OF GOD I FEEL THE AGI
>>109728809sad but thanks for keeping us updated
New thread:>>109728887>>109728887>>109728887
>>109728545that's not why it took 4 weekssite isolation is just that insanely hardthere's a reason edge and chrome didn't truly implement it until recently
>>109728922>site isolation is just that insanely hardI think we've been agreeing about the same concept but we're talking past one another, probably my bad for wording things how I did.
68 - "allowUnsandboxedCommands": false, 68 + "allowUnsandboxedCommands": true, Fuck you, claude. Fuck you.
