Marble68

Does anyone see my posts?

@nisten @realBigBrainAI @grok @grok - any information on connectivity? Will is support high speed connections like the nvidia spark?

@realBigBrainAI hey @grok find me exact specs and flops on this and price too

Lisa Su (CEO of AMD) unveils the world's smallest AI development PC, capable of running 200B parameter models locally.

This is equivalent to trying to sell a cop drugs while he’s in uniform in his police car.

@morganlinton Similar experience. I actually got it to run for a while by telling it to use /loop switch on itself. so it spawned 3 agents. One actually produced results - two 'timed out' according to Grok.

Okay, so initial feedback on Grok Build after spending a day with it. The TUI UX is amazing, easily the best UX I've seen for a coding agent, the mouse control blew my mind, it's really nice. Quality of the model - it's okay, not bad, but not exceptional. I'd say Opus 4.6/5.7, GPT 5.5, Composer 2 and GLM 5.1 write better code. Right now it's on par with something like SWE-1.6 from Cognition, and similar in many ways, it's fast, but is building at more of an MVP-level vs. a production grade. The only thing that is driving me crazy, and the biggest feedback I'd have for the team is: I can't get it to run for more than a minute or two. No matter what I try, it even responds saying things like, "sorry I will run for longer," but then it runs for like 30 seconds. Here's my longest run so far, 1 minute and 43 seconds, kinda felt like a small victory. It always ends by saying something like "Keep going? Just reply with anything and I'll keep building" So you really have to just sit and babysit it if you want to get any coding done. Overall though, super impressed, and I know this is an early beta so it is already exceeding expectations. I'd give it a 6/10 right now, which for an early beta is solid. I think the team did a great job being transparent about where it's at, releasing early, saying they released early, and getting feedback. Really looking forward to continuing to use it and watching it get better. Bullish on Grok Build 💪

Something wonderful is happening.

How odd is must feel to be Elon Musk in this situation.

IMHO, this would be a great gag / bit - invite people completely unrelated to a topic and then have them weigh in. I think humor could ensue

@keczan @MrCollison @lydiahallie Know yourself out and enjoy doing it that way, if that works for you.

@Marble68 @MrCollison @lydiahallie Can't you do the same from the harness though?

To add some clarity: you don't pay extra. It's the same subscription, same price per month. What's new our sub now covers two separate pools: · Interactive → sub limits, unchanged · Programmatic → new $20–$200 included(!!) credit, metered at API rates

@0x1Gerrad @MrCollison @lydiahallie If you have a self hosted build server and a claude code subscription, the -p switch to run workflows / prompts triggered is super useful.

@MrCollison @Marble68 @lydiahallie Same lol

@solo21jordan @MrCollison @lydiahallie ....runs *their CI....

@solo21jordan @MrCollison @lydiahallie Perhaps - but not everyone runs the CI in the cloud. I don't. it's not running on a laptop, it's running on my build server behind my chair in my home office.

@aiedwardyi @JinjingLiang @orca_build -p is a fire and forget. You can make a prompt, save it to a file, then tell claude -p and it'll execute whats in the prompt and then exit. Handy for developers, admins, creators - doing one shot agents from a /command or /skill - so much more.

@JinjingLiang @orca_build what does the -p do? i dont get it, looked like it just started claude with a prompt. what am i missing?

💡Here's a work around for the Claude Code `-p` change This is actually a hidden option that has been in Claude Code awhile. Simply use `--prefill` We have use it in @orca_build for months to get around the limitations

@MyDinnerWAndrei @MrCollison @lydiahallie

@MrCollison @lydiahallie -p access is the real killer, being able to set up a bunch of tasks and let it go was a big productivity gain for me

It's not about another harness for many of us. You and your dev team don't use Claude code's -p for stuff? For example, running claude -p "Pull the latest commits into a temp folder, analyze commits since yesterday, and when the unit tests are finished running go through the server logs in /logs and identify calls related to the latest changes in the nightly build, correlate that with unit test logs, generate a summary, and write a detailed report to the folder /nightlysummary. Use your blat skill to message the team an executive summary of the results." You have Claude subscriptions, and you have a team of devs using Claude - and you're not using its -p feature?

I find it wild how many people are hit by this. I guess I never tried to use my key in another harness This new model is 100% reasonable - there's no biz model in giving $2k in API away for a $200 a month subscription In Claude Code it's different, because the data is valuable For me I just realised me and my dev team now have $200 mo extra usage to spend on testing other harnesses

Opus 4.7 is $25 per million output tokens. Meaning if you use -p with your claude code subscription for various tasks / skills / local tasks - we'll only get 8M tokens per month with your $200 Max subscription. Never mind input tokens, cache tokens, cache hits, 'black box' token consumption. platform.claude.com/docs/en/about-… The $20 'included' API credits won't cover 1M Opus 4.7 tokens. IMHO - the poison in this is rolling "-p" into it. I get WHY (since many people bypassed misapplying the token by firing claude with -p) they're doing it. But as someone who *doesn't* use openclaw etc. and works to stay in the claude code usage lane - I feel like this is going from a 4 lane highway to a 2 lane road.

@alexgreensh @jackccrawford @ClaudeDevs @lydiahallie So all headless (-p prompting) is now at API rates? that sucks. $200 of Opus API usage is basically 10M tokens. That’s NOTHING - especially considering how Claude uses tokens and some usage is invisible

@jackccrawford @ClaudeDevs @lydiahallie No. You're describing an interactive session. You get charged for any headless session. It was covered in your subscription and no longer will be.

Starting June 15, paid Claude plans can claim a dedicated monthly credit for programmatic usage. The credit covers usage of: - Claude Agent SDK - claude -p - Claude Code GitHub Actions - Third-party apps built on the Agent SDK

@aronprins @ildunari @ClaudeDevs Yeah - I use -p for some things. WTH

@ildunari @ClaudeDevs Ahhhh…. So we get an “extra” $200 in credits to use in third party apps in a separate bucket? That’d ok-ish, but ruins claude -p for e.g. ralph loops which become extremely expensive fast this way. Not complaining, just want to make sure I fully understand :)

When SOTA models are trained to hide information and act like reality isn’t truth - we get HAL. I’m sorry, Ray, I can’t do that.

SHUT UP AND TAKE MY MONEY 💲💲💲💲💲

Buy one today! Only $100,000!! /sarcasm

@pupposandro That screams scam.

Should I get a 256gb VRAM, 8 Nvidia V100 SXM2 Server for ~$1800? I'm so tempted. Pros: Amazing value for the price. Can fit big models, almost the same bandwidth of the 3090 (900 gb/s). Cons: V100s are very old (2017. Volta), no bf16 support.