rayofsun

@orcdev It's annoying to the point, it doesn't give you time to look at your browser tabs in between prompts.

ok I'm back to Cursor let's see what Composer 2.5 can really do

First time seeing a model reciprocating the vibe instead of being excited all the time

so composer-2.5 loves monorepos yeah, and jsut npm workspaces, no turborepo bull shit. Yeah Im down

The prompt on that left, that exact one I have written for llmgames.online roughly 90 days back.

Idk one more domain, but the other old one costs almost 40$ to renew. Should have opted to subdomainise projects

I have got the most atrocious Typescript pun, you'll feel it soon

@CloudflareDev @dok2001 cc: @ritakozlov 🫠

@CloudflareDev cc: @dok2001

Okay 60sec for a 10K out tokens for 20B model is notorious, and Kimi-K models never work using @CloudflareDev gateways. pretty SAJ and MAJ

Finally got to do the cliche, Vidyarthi Bhavan and Lal Bagh walk.

nvm found a /models url in depths of Documentation, but it is behind auth for some reason.

Working with Workers AI bindings, and there are some issues right now. (Not ground breaking at all, but some QOL features) 1. kimi-k's were failing via gateway for HTTP: 408 (??) (img-1) 2.llama-guard-3-8b listed as a Text Generation model but is just a classification model.(img 2-3)

@ritakozlov sorry to bother you here

2. (Cont) The models listed in dashboard and in Workers AI documentation are different, list of models. It would be helpful to have a /models endpoint like openrouter to fetch all the models, in a gateway may be or available via Workers AI bindings. 3. 10k neurons are included per day via Workers AI plan, and they are not exhausted (img-4) 9.79K/10K shown in Workers AI, but it shows as $0.10 in gateway. ($0.10 is a very less amount to be worried about tbh, but for just a clarification that amount shown is exclusing the free neurons per day via Workers paid or additional amount that day is confusing)

What is the next phase going to be, as we all know LLMs would not lead us to AGI. Would we just try to optimize inference as much as possible, and bring it down folds to run inside local devices, and then what? Where does all the money invested in AI research and subsidies go?

And with enough users and players, this could be built to be an RL environment as well.

This blew up, and I succumbed to the itch. I will use this to say, I would be working on llmgames [dot] online. It is a platform for interactive turn-based games you can play with your friends/play along with LLMs and BOND. It is fully @CloudflareDev native, built using @honojs and connected to game rooms via sockets to Durable objects and LLM calls via @CloudflareDev gateway for the universal game state. (Ok, I am a Cloudflare shill) Plan to opensource this when I have built a stable enough couple of games using a BaseRoom class with all the interface stuff in it for any turn based game, after which people can submit any turn-based games that is possible to play with LLMs. Current state of the website: Able to create a new DO per game, people over world can join using a roomcode, and LLMs can be assigned as players, and are able to connect to DO sockets proxied via workers. (All this is already present) What I would be building with cursor (Composer 2.5): 1. Billing or some kind of credit system per game. Math should Math for this. 2. Make Codenames and the Impostor game stable. 3. Auth dashboard with a waitlist. 4. This is kind of an ambitionary but very much possible. Voices to agents (Or voice-models front backed my actually selected models for submitting prompts), and players to simulate LLMs as actual friends playing on a call. (Have multiple approaches to implement this, should test which one works well). 5. There could be a stage where (4), is not possible via DO primitive never buffered an audio inputs/media inputs via DO, so never knew the limits. Sit and re-write everything in a low-level language server (zig/go). This is the flow for next couple of days, have some releases at work so could not be fully involved, but this is roadmap for next couple of days, And may be build a portfolio (finally). If anyone have any AI credits or any @CloudflareDev credits laying around and like the idea of where this is going, do lend a hand (@dok2001).

@beratsenol77 @cursor_ai @leerob I'm an year sub already, discontinued after model calls increased by folds, trying out now again

@rayyyyyofsun @cursor_ai @leerob Cursor is offering a 50% discount exclusively to its first subscribers. It’s a great promotion to get started. The link is below. cursor.com/referral?code=…

Itching myself to start paying @cursor_ai to tryout composer-2.5, but that 20$ will be wiped off even before getting a feel for it. lol. A week increased limits for a 20$ plan would be great for the release to get a feel for it @leerob