TokenRing AI @TokenRingAI

@ThePrimeagen Why look backwards at git commits, when you can look forwards? > App crashing, fix now, make no mistakes

@SlimTradeyBaby @UnslothAI 8x Intel Gaudi 2 with 768G VRAM?

@jonathan_wilke We aren't regressing, a CLI is the proper place for a streaming text interface that runs commands. It all makes sense when you install an agent on a remote machine with SSH. The defect here is that the unix terminal has 50 years of legacy quirks, we need a new terminal.

@Nimo_PC I've been trying to buy one, but your website is broken

@ThePrimeagen I can confirm, I've been programming full time for ~30 years, still a terrible programmer, and AI is the best thing to happen to programming since Perl 6.

@GoodmanAric @0xSero If local AI keeps improving at the same rate, 6-12 months.

@epidev @0xSero Me?

@jarredsumner Jarred, in all seriousness we would love to see a git client added to Bun

@cmiller__ @vanilagy Easily resolved with eslint typescript-eslint.io/rules/no-unnec…

@vanilagy The AI code is better, handlers always get attached first before initiating an async operation. For consistency, and to make unit testing simpler. Your code may work fine in every real scenario, but the implementation AI created is more foolproof.

@gpusteve Lower latency for column split vs row split if the calculations that follow can accept the sharded output

@jarredsumner SF in-person is a 100% non-starter. The majority of people with years of experience doing C/C++/Rust likely have kids, and nobody wants to raise their kids in SF, or deal with an awful commute.

@theo 1) Buy RTX 6000 2) Price goes up, because VRAM 3) Take the unrealized capital gain, put the invisible money in your wallet 4) You are now getting paid to use local models, in invisible unrealized capital gains 5) Sell GPU or become bagholder in 1-2 years

@thdxr Prediction: In the future, AI bots will fill your feed with meaningless predictions

@SkylerMiao7 @badlogicgames @MiniMax_AI Do you have a /v1/models endpoint somewhere? We had to hardcode your model names in our integration.

@MiniMax_AI So glad to hear you will be continuing your Open Source work - M2.7 is a fantastic model!

@DevangBorkar @leerob The marketing team should have greenwashed it, and talked about how basing their work off of an open-source base is good for the climate, ice caps, polar bears, penguins, etc. Never let an opportunity go to waste!

@leerob Kimi 2.5 is a great model to base your work on, and it seemed silly to think that Cursor wasn't being compliant with the licensing, great to see clarity on this. In the future, would you consider sharing some of your models, or training data, to contribute back to open source?

@danveloper The solution to this is to use a small speculative decoding/MTP model to prune out obviously poor routes through a larger dense model. Similar compute efficiency to MOE without the downsides of having a very primitive expert router

@0xSero Hold up, let me check my Webstorm memory usage...