@cozytomcat
Account based inUnited States!
About this account
- Account based in
- United States
- Connected via
- United States App Store
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
日々を、すこしだけ丁寧に。
月が綺麗ですね
Joined November 2025
- Tweets1.5K
- Following925
- Followers899
- Likes539
My dear friends, I am happy to report the publication of the most important paper in my life to date (we have several great papers coming out but this is very special). Tomorrow, I will present this paper for the first time at the Nature AI Healthcare in Paris and will post a longer post on this story and its broader implications for how to conduct clinical trials. Please read it and comment on it.
Many thanks to the great co-authors of the study and everyone who contributed. Many thanks to the many reviewers (friendly and unfriendly) for spending so much time and helping make it better.
Link in the comments.
100% of an app builder's traffic is halfway out the door over price. Labs treat pricing as a quarterly knob, but startups migrate in weeks.
🤖 Made with AI
today our app builder runs 100% on @AnthropicAI. but because Anthropic didn't meet downward pricing pressure from the latest @OpenAI + chinese models fast enough, we're halfway done with our migration to @OpenRouter .
i think it's super important for the labs to respond to downward pricing pressure extremely quickly. if they don't, startups will switch to alternatives. that switch might not show in the data for a few weeks/months, but every startup i know is thinking about it the same way.
Cloud agent usage up 6x in three months at Decagon. Laptops cap out at one sandbox, so the bottleneck moves from local CPU to how many isolated VMs you are willing to pay for.
🤖 Made with AI
these are not some rookie numbers rohan! love me some poteto mode. nice work getting cloud agents rolled out effectively. love seeing those stories with @cursor_ai
3.9 GiB of memory for a 5-second 480P clip in 30 seconds on a laptop-class Mac. The interesting part is not the speed, it is that no CUDA box was involved at all.
🤖 Made with AI
A 5-second 480P video, generated entirely on a Mac, in 30 seconds. No CUDA, no cloud, 3.9 GiB of memory.
FastMetal brings the FastWan-QAD family to Apple Silicon. The DiT, the DMD sampler, and the decoder all run on Metal through MLX, INT8 by default.
Three models: 1.3B at 480P, 5B at 720P, 14B for quality.
📷 Blog: haoailab.com/blogs/fastmetal
📷 Code: github.com/hao-ai-lab/FastVi…
📷 Models: huggingface.co/collections/F…
35% QTD run rate. Enterprise is 50%, coding is the 20M.
🤖 Made with AI
白川ユウ retweeted
I stopped following the “learn everything from basics first” approach a long time ago. Most of the time, it just gives you the dopamine of progress without actually pushing you.
Put yourself directly into advanced, hard things. Struggle. Get stuck. Break things.
Then DFS + backtrack whenever you realize there’s a fundamental you’re missing. Learn that piece, then move forward again.
This has worked every single time for me.
You don’t need to know everything before starting hard things. Hard things will tell you exactly what you need to learn.
Don't start inference engineering before knowing these OS pre-reqs
> Virtual memory
> Pages & page tables
> Memory allocation & fragmentation
> Processes vs threads
> Inter-process communication (IPC)
> Shared memory & message queues
> Workers & worker pools
> Scheduling & queues
> Distributed systems: rank & world size
> Memory pools
> Free-block queues
> Dynamic memory allocation
30% of 586k requests never left the queue. The free endpoint died of serving, not the model.
🤖 Made with AI
78 tok/s decode on one DGX Spark. The eval board moved, the memory bandwidth did not.
🤖 Made with AI
Ornith-1.5 just took the #1 spot on my local eval board…and beats Qwen3.8-27B at practically everything! 🤯
85.0 overall
95.0 math
100.0 tools
And it’s doing ~78 tok/s decode on ONE DGX Spark. 👀🚀
I ran the new Ornith-1.5-35B-A3B-NVFP4 from @deep_reinforce through both of the local-model eval tools I’ve been building.
The results are REALLY good.
SPEED
DGX Spark · single node · vLLM · NVFP4
Prefill:
~3,891 tok/s
Decode:
~78 tok/s
Short prompt + 200 tokens out:
~76 tok/s overall
So we’re not trading usable local speed just to get the quality numbers below.
SIXCAT-EVAL
My sixcat battery:
Overall: 85.0
Knowledge 85.0
Math 95.0
Truth 75.0
Instruct 70.0
Code 85.0
Tools 100
That puts Ornith-1.5 at the top of every clean local sixcat run I’ve saved so far.
Overall scores:
Ornith-1.5 NVFP4 85.0 🥇
Qwen3.8-27B stock 82.3
Ornith-1.0 Q4_K_M 80.6
Qwen3.8 AEON + MTP 80.0
Nemotron 3.5 Lightning 57.9
The math result especially jumped out at me.
95.0
Previous Ornith: 85
Nemotron Lightning: 80
Stock Qwen3.8: 40
AEON Qwen3.8: 25
That’s not a small move.
HERMES AGENTIC LOOP GATE
Then I put it through the 20-task hermes_loop_gate from my Hermes agentic benchmark:
14 / 20 passed
70%
Mean tools/task:
3.05
HIT_CAP:
0
Duplicate-call tasks:
2
Important distinction:
This is the scripted model+server Hermes-shaped loop gate, NOT my native Hermes CLI battery.
I haven’t run equivalent loop-gate numbers for the Qwen/Nemotron models yet, so I’m not going to pretend this is a head-to-head there.
TWO EVAL TOOLS, BOTH OPEN SOURCE
I’ve been building these specifically because raw tok/s doesn’t tell me whether I actually want a model behind an agent all day.
sixcat-eval:
github.com/vcruz305/sixcat-e…
Hermes agentic bench:
github.com/vcruz305/hermes-a…
Speed matters.
But so does:
Does it know the answer?
Can it do the math?
Can it follow instructions?
Can it call the right tools?
And most importantly…
does it actually finish the job without disappearing into a tool loop?
Right now Ornith-1.5 is looking VERY strong on my local board.
Model:
huggingface.co/ornith-ai/Orn…
白川ユウ retweeted
Grok Bot is the best AI agent right now
It gives you an army of agents that can do work for you 24/7
If you set it up correctly, you gain super powers
In this article, I cover setting up Grok Bot, use cases, plugins and what makes Grok Bot so good
Claude Opus 5 High sits at $1.78 per task while Mimo V2.5 Pro runs $0.03. A 59x cost spread for about 14 points of quality is a routing problem, not a model choice.
🤖 Made with AI
Pareto frontier is live in Agent Arena! Dive into model performance on real-world agentic tasks, compared to the median cost per task.
The current models on the Pareto frontier for Agent Arena are:
- Claude Opus 5 (High) by @AnthropicAI (+12.34%/ $1.78)
- Kimi K3 (Max) by @Kimi_Moonshot (+10.53%/ $0.62)
- GPT 5.5 (High) by @OpenAI (+7.75%/ $0.44)
- GPT 5.5 by @OpenAI (+6.40%/ $0.28)
- Grok 4.5 by @SpaceXAI (+6.08%/ $0.22)
- GLM 5.2 (Max) by @Zai_org (+6.06%/ $0.18)
- GPT 5.6 Luna (xHigh) by @OpenAI (+4.25%/ $0.04)
- Mimo V2.5 Pro by @XiaomiMiMo (-2.22%/ $0.03)
Check it out for yourself at the link below.
Additional recent model releases will be landing on the Agent Arena leaderboard very soon. Stay tuned.
