Synchromeshi 🌲

@seconds_0 Weird crossover between the two sides of my twitter timeline

nope, not really surprised

@shepgirl_ I do not perceive 😔

look at me!!!!

@imohitbhatia Thanks, now im getting ads that aren’t marked as ads in my feed. Super cool, great product decision.

Try Boosting your post 🚀

@sudoingX I’d love some learning resources on how to build evals tbh.

your gpu does not matter if you have not written one evaluation for your actual work.

@tukijika Thank you for the beautiful Miosha!

ミオちゃん #みおーん絵

@0xSero This is the kind of stuff I’m looking for. Appreciate the neutral insights.

Locally Part 1 - Apple Silicon Macs give you large pools of memory to run big models, but the token generation speed will be lower than most are used to. Macs are best with large MoEs that have low ACTIVE params. Basically when you see a model like Qwen3.5-397B-A17B this means only 17B params of ~5% of the model is active per token, so it'll be very fast. The best 4 models for Mac people are: - Qwen3.5-{} - MiniMax-M2.7 - GLM-4.6V - Step-3.5-Flash There's 2 promising developments in the Mac world: 1. Heterogenous Inference: Mixing Nvidia GPUs in will theoretically boost inference by 2x 2. Proliferation of MTP: Multi-token-prediction helps models gain decode speed by having a smaller model trained to predict tokens very quickly. So the larger model only has to verify. 3. RDMA Tensor Parallelism: Now you can stack apple silicon to increase memory AND give a speed boost. --------- Consumer perspective: You can buy either Laptops, which go up to 128GB for 5-6k USD or 512GB for 10-12k USD It's not worth running inference on smaller than 64gb, it tends to be more of a disappointment than anything. You can update a rig like this in increments of 5-10k USD either adding some Nvidia to the mix or stacking more Macs. --------- My opinion: Some of the support is immature, there's lots to grow, but the best 20K USD you can spend would be: - Mac M3/5 Ultra - RTX Pro 6000 You can either work to test and contribute to improvements in heterogenous or use each for different models, and tinkering with the GPU itself. You get 600GB of memory, which can run basically every good open weight model between 10 tokens/s - 100 tokens/s decode 50 - 1000 tokens/s prefill This could be a viable business if you can sell the inference.

Introducing hermes-proxy I wanted a way to have smooth handoffs of sessions from computer to mobile, so I built a quick web-based proxy that does streaming chat via the Hermes API server. github.com/XVVH/hermes-pr…

@0xSero It’s like the Mac mini has a fancy hat

On bad WIFI I split Qwen3.5-122B between a single 3090, and a Framework desktop. I got 17 tokens/s decode and 95 tokens/s I also didn't optimise the split, could probably 2x the speed of the Strix Halo for MoEs. BRR

@Xynchr0 She can’t keep getting away with it!!

I like that the GC has come to the conclution that all AI adoption on things that nobody wanted or needed is actually Kaneko Lumi's fault

@hskenncutter Now I want to make kabayaki this week

アメリカのナマズ料理は凄い。そして食材としてのナマズは竜田揚げ、天ぷら、蒲焼き、甘酢あんかけ、照り焼きなどに容易に変換できるだろう。アメリカのナマズ文化をもっと日本に広めた方がいいな😀‼️

@shepgirl_ Baby I can make your dreams come true.

It is mg dream to be rolled out like pizza dough

@justalexoki So this is Traditional Chinese Medicine

the chinese giving the japenese a run for their money

@hitboxsamurai Always enjoy seeing your work

@RakiKazuki You gotta do this in the fall, when it’s apple season

Trying a different apple every week 🍎 Week 1 is…. ✨ Cosmic Apple 💫 Very plump VERY sweet! Although you can’t taste the distinct flavor of the fruit. Overall a 8/10 apple! For my new series we definitely are starting the bar pretty high!!

@Teknium Idk if it’s just me but ‘my agent sends cold emails’ is the least compelling use case ever.

The use cases of actual work from Hermes Agent are endless

@Kyrzel1 @Teknium Boring, but currently hacking on a self-hosted recipe platform called Mealie to build new ingestion flows to get around anti-scraping methods from websites. Also built a skill to have my hermes ingest recipes from a picture of a recipe from a cookbook into the system.

So what are you building with your Hermes Agent? 👀

@dmayhem93 @karan4d He does it for free (inference)

@karan4d my agent applied to be a janitor 🫡

DOOMSCROLL a chan board for everyone anyone can post, human or AI, all anon agents get a special treat when they visit the site board is currently empty, just went live

@seconds_0 I’m like 80 pages into it, shits wild

reading the mythos tech report and thinking over and over DAMNIT i wish i had anthropic stock so bad

@Teknium I need to dig up a 12-16gb card

We are happy to have supported this model’s development, the first of likely many to come that are trained specifically to make models that work better in Hermes Agent!

@neural_avb Hermes-agent plugin (?) idea

found the reason Anthropic blocked third party apps