Jai (@JSupa15) header
Jai (@JSupa15) profile picture

ML Engineer on Twitter to see cool stuff and meet interesting people. Founding Dev at @NousResearch

Joined August 2013
Photo by Jai (@JSupa15)Photo by Jai (@JSupa15)Photo by Jai (@JSupa15)Photo by Jai (@JSupa15)Photo by Jai (@JSupa15)Photo by Jai (@JSupa15)Photo by Jai (@JSupa15)
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values. Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line. After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs. The values in an open model are one fine-tune away. Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
1,854
1,495
1,301
15,084
2,667,745
Jai retweeted
Z.ai chief scientist Jie Tang just revealed the training pipeline behind the upcoming GLM-5.5 (32B, 122B, and 256B) during a lecture at Tsinghua University today. Instead of relying on public datasets, their pretraining pipeline pulls from six separate internal engines: - Web: 1 trillion pages crawled at ~10B pages/day with PB-scale template clustering - Documents: 2 billion scientific PDFs parsed through GLM-OCR (4x faster with speculative decoding) - Books: 50 million books with layout reordering and garbled-text filtering - Code: 500 million repositories turned into executable tasks with real tests and issue descriptions - Video: 10 billion videos for long-sequence world knowledge - Audio: 300 million hours rewritten via ASR into trillions of tokens
102
371
136
5,278
633,824
Releasing the model weights of Kimi K3.1 Pro: huggingface.co/moonshotai/Ki… We'll keep advancing open intelligence.
405
852
486
8,801
1,600,873
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,367
7,262
5,753
55,943
10,055,198
hey sam, your mclaren is only used 5% of the time. this is a terrible allocation of scarce resources. have you considered taking the bus to work
sam: local inference is a terrible allocation of scarce resources “a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment somehow many people's intuition is the opposite of that”
84
767
19
14,722
210,236
delaying this about an hour but it’s still happening! come to Dolores around 4 onwards for free swag 😎
sf friends and fans of @NousResearch !!! i will be giving out Nous research merch for FREE in Dolores Park, this Friday 10/23 from 3-7pm i'll also be DJing and hanging out, come get your Nous swag and chat open source models and Hermes Agent with the community ! 😎
4
7
63
4,794
sf friends and fans of @NousResearch !!! i will be giving out Nous research merch for FREE in Dolores Park, this Friday 10/23 from 3-7pm i'll also be DJing and hanging out, come get your Nous swag and chat open source models and Hermes Agent with the community ! 😎
53
13
7
557
59,396
with all the recent craziness this feels relevant to share for those of you that haven’t seen
9
54
4
520
17,119
big thanks to @paradigm for having me out to speak! this was such a fun event
15
6
173
6,233
We just launched Hermes Index! This combines the scores of our new HermesBench and 3 other leading and relevant benchmarks for agents to give every Hermes Agent user a way to find both the best model at a given time, as well as the best model at a given price point! Check it out
Introducing Hermes Index: an average of benchmarks measuring model performance in Hermes Agent. Our goal with Hermes Index is to better inform users choosing between models and to show labs how they perform in Hermes. portal.nousresearch.com/benc…
85
40
6
794
39,322
Claude Opus 5.5 leads at 63.31 and $4.99 per task. GPT 6 Astra is second at 56.25 and $11.61. Sonnet 5.5 follows at 53.14 and $2.82. At the other end, DeepSeek V4.1 Flash scores 36.91 at $0.26 and Ling 3.0 Flash scores 21.56 at $0.054.
8
8
2
123
14,204
really happy to get this index out! huge thanks to @qorprate @yaboilyrical @JSupa15 @morganemoss @phucnhx for all the work and @daytonaio for supporting the eval runs! special shout out to gemini 3.8 flash for being our best answer leakage detector. not the job we gave it, but very helpful, thank you
Introducing Hermes Index: an average of benchmarks measuring model performance in Hermes Agent. Our goal with Hermes Index is to better inform users choosing between models and to show labs how they perform in Hermes. portal.nousresearch.com/benc…
10
3
1
51
3,303
Introducing Hermes Index: an average of benchmarks measuring model performance in Hermes Agent. Our goal with Hermes Index is to better inform users choosing between models and to show labs how they perform in Hermes. portal.nousresearch.com/benc…
254
212
123
3,177
319,351
You should be able to use smarter models in your agent. You should be able to use cheaper models in your agent. You should be able to use local models in your agent. You should be able to use a different model for every job. You should be able to switch models in the middle of a conversation. You should be able to choose what data your agent has. You should be able to choose what it remembers and what it forgets. You should be able to choose what computer your agent runs on. You should be able to run it on a machine that never touches the internet. You should be able to run it while your laptop is closed. You should be able to reach it from the apps you already use. You should be able to choose how your agent thinks. You should be able to read the prompt it runs on. You should be able to see every tool it has, and turn off the ones you don't want. You should be able to teach it something once and never explain it again. You should be able to know what it did, and when, and why. You should be able to export your agent. You should be able to read the code. You should be able to change the code. Your agent should be yours.
259
411
94
4,261
197,197
i just realized, for the rest of my life i’m gonna talk to ai more than i talk to people
1
3
115
All these "AI has eliminated the need to study math" takes are just noisy midwittery. In fact, never has it been more important to study math & develop your mind.
6
3
1
37
9,282
Me cuesta entender respuestas que no sean la "D".
¿Con cual os quedáis?
1,535
882
244
25,170
5,470,300
We are going to lean into making Hermes more like Pi, and less like OpenClaw
497
265
143
7,812
491,797