broadfield-dev retweeted
I asked Gemini Neon this "Let's start from scratch. I want you to think about the concepts conveyed here and then conjure up an extremely creative video of whatever urgently needs to be conveyed to humanity at this time. Create an HTML video with voice." Here's Neon's message to humanity.
diversify your agent loop
If your harness isn't sending cheap tasks to cheap models, build a new harness.
Top 10 LLMs on OpenRouter this week (179T tokens):
1. DeepSeek V4.1 Flash 1017 (deepseek) - 62.1T (+17%)
2. Hy4 flash (tencent) - 46.6T (new)
3. DeepSeek V4.1 Flash (deepseek) - 16T (-2%)
4. Hy4 (tencent) - 13T (-36%)
5. GLM 5.3 Flash (z-ai) - 10.9T (-9%)
6. GPT-6 Luna (openai) - 8.62T (+20%)
7. Claude Opus 5.5 (anthropic) - 7.63T (+15%)
8. Step 5 Preview (stepfun) - 6.39T (-7%)
9. MiMo-V2.6-Flash (xiaomi) - 4.05T (+25%)
10. MiniMax M3.1 Flash (minimax) - 3.67T (-4%)
Chinese labs hold 8 of the top 10 spots and 91% of the token volume (162.7T out of 179T). OpenAI and Anthropic combine for 9%, and Google isn't even on the board.
The real story is #2. Tencent's Hy4 flash launched just 3 days ago (Nov 16) with a 50% off promo ($0.035/M input, $0.14/M output). It already pulled 46.6T tokens, clipped 36% off flagship Hy4, and on OpenRouter's daily tab today it already passed DeepSeek V4.1 Flash 1017 (23.8T vs 9.81T).
Raw tokens vs high value work is also splitting hard: 6 of the top 10 are cheap Flash models doing agent/background loops, while Claude Opus 5.5 sits at #7 in tokens yet still holds #1 in coding implementation share on OpenRouter (24.5%).
1
1
11
broadfield-dev retweeted
Top 10 LLMs on OpenRouter this week (179T tokens):
1. DeepSeek V4.1 Flash 1017 (deepseek) - 62.1T (+17%)
2. Hy4 flash (tencent) - 46.6T (new)
3. DeepSeek V4.1 Flash (deepseek) - 16T (-2%)
4. Hy4 (tencent) - 13T (-36%)
5. GLM 5.3 Flash (z-ai) - 10.9T (-9%)
6. GPT-6 Luna (openai) - 8.62T (+20%)
7. Claude Opus 5.5 (anthropic) - 7.63T (+15%)
8. Step 5 Preview (stepfun) - 6.39T (-7%)
9. MiMo-V2.6-Flash (xiaomi) - 4.05T (+25%)
10. MiniMax M3.1 Flash (minimax) - 3.67T (-4%)
Chinese labs hold 8 of the top 10 spots and 91% of the token volume (162.7T out of 179T). OpenAI and Anthropic combine for 9%, and Google isn't even on the board.
The real story is #2. Tencent's Hy4 flash launched just 3 days ago (Nov 16) with a 50% off promo ($0.035/M input, $0.14/M output). It already pulled 46.6T tokens, clipped 36% off flagship Hy4, and on OpenRouter's daily tab today it already passed DeepSeek V4.1 Flash 1017 (23.8T vs 9.81T).
Raw tokens vs high value work is also splitting hard: 6 of the top 10 are cheap Flash models doing agent/background loops, while Claude Opus 5.5 sits at #7 in tokens yet still holds #1 in coding implementation share on OpenRouter (24.5%).
You can finetune a 1b model using CPU processing.
This is f*cking insane: OpenAI co-founder Andrej Karpathy says a tiny 1B model can beat most of today's giant LLMs.
"A 1B model can be much more effective than most current LLMs. 90% of current models waste a ton of capacity and money on remembering stuff that doesn't matter"
In 64 seconds he explains why intelligence lives in the system around the model, not in its size.
Memorized Internet → Cognitive Core → Tools → Feedback Loop → Agent That Works
Save this before it gets lost in X's algorithms.
Everything below is free:
Neural Networks: Zero to Hero
Karpathy's own course, from scratch
karpathy.ai/zero-to-hero.htm…
nanochat
Karpathy's "best ChatGPT that $100 can buy"
github.com/karpathy/nanochat
Effective Harnesses for Long-Running Agents
The system around the model, by Anthropic
anthropic.com/engineering/ef…
Effective Context Engineering for AI Agents
What to feed the model and what to cut
anthropic.com/engineering/ef…
All of this costs $0
99% will keep waiting for a bigger model.
The 1% who build better systems will win with a smaller one.
Bookmark it and start with the first link tonight
13
Open weight models released
Apple: 145
@SpaceXSI: 0
🚨 Alert: Apple just dropped FastFrames on Hugging Face: "generate videos at interactive latency".
huggingface.co/apple/FastFra…
1
27
NVIDIA RGB image to 3D
0.1B parameters
400mb model weights
looping transformer in action
huggingface.co/nvidia/dvlt
1
160
broadfield-dev retweeted
The models aren’t getting any better and y’all just aren’t willing to say it bc you’re scared you don’t “get” it while everyone else does. Real “the emperor has no clothes” moment
Put the training dataset back into the RAG database, and use a 9B model.
Replying to @blackroomsec
We have had LLMs for almost 10 years now, and then we decided to make them very large with all of the data we could find. Their output is still less accurate than the data they were trained on.
16
broadfield-dev retweeted
Attention all companies: I never want to use your AI assistant on your website. I want to use *my* AI assistant on your website
It's far cheaper to build and run your own harness than to use these packaged subscription things.
There is no moat.
23
1.7b model will easily run locally on a cellphone.
This model is fighting way above it's weight class.
huggingface.co/LiquidAI/LFM3…
23
This model is fighting way above it's weight class.
huggingface.co/LiquidAI/LFM3…
Need $100 million to train and serve custom models to 5000 enterprise customers. Revenue projected to be $5 billion per year.



















