Yuxuan Cao

AI Founder and Physicist
PhD candidate at UT Austin

[Elon Musk @elonmusk]

@yunta_tsai A petition is underway to reinstate Jensen and I’m signing it too

[Yuxuan Cao @ycaophysics]

Simple install with claude code:
"""
Install the build-with-agents skill from https://t.co/qE4lv97rzb.

1. If ~/.claude/skills/build-with-agents does not exist, clone the repo there. If it exists and is
a clone of that repo, run `git pull --ff-only` in it instead. If it exists and is anything else,
stop and tell me what is there.
2. Run `bash ~/.claude/skills/build-with-agents/install.sh` (from Git Bash on Windows). It links the
skill into Codex, Cursor and agy where they are installed, and installs quota-cli.
3. Show me what https://t.co/HvXXBeI6D6 printed, and tell me which tools got the skill and which were skipped.

Do not change my Claude Code settings or permissions, and do not log in to any tool for me.
"""

[Yuxuan Cao @ycaophysics]

Stop burning a new login just to reread the repo.

Even with a Claude Pro/Max, a Cursor Pro, and a ChatGPT Plus subscription, my big plan made with gstack still dies the moment a 5h window or context runs out.

So I built "build-with-agents", a Claude Code skill that keeps builds moving across your quotas:

• Parallel worktrees: Paste a plan. Your main session leads while up to 3 workers code at once in separate worktrees on non-overlapping files. Transcripts stay in logs; the lead only reads short reports.
• Model routing: Hard and medium work stays on Opus. Renames, copy edits, and tests from a given table go to Sonnet or other models you have.
• Quota fallback: Checks 5h and weekly windows before each task. At 10% or under, or on a mid-task rate limit, that task moves to Codex, Cursor, or Antigravity (agy).
• Disk handoff: When Claude itself runs out, resume on reset or hand the lead to Codex, Cursor, or agy via the status file on disk.

*Repo + quickstart in the first reply.

How are you carrying big plans across separate logins right now?

[Yuxuan Cao @ycaophysics]

https://t.co/hvpX5RoYBy

[Yuxuan Cao @ycaophysics]

黄仁勋两炮亏掉了一个amd+一个intel

[Yuxuan Cao @ycaophysics]

I’ve ended up paying for (or getting promos for) almost every major AI model: Claude Max+Pro, ChatGPT Pro, Gemini Ultra, Cursor Pro. I also run a couple of big models locally. I don’t come close to using all the allowances.

I use them for coding, scientific research, and some financial workflows with agents that run around the clock. Some subjective notes:

GPT-6 Astra and Claude Fable 5.2 are the two I trust for hard science problems. Fable is also my go-to for financial research and longer workflows. My main complaint with Astra is the usage limit. Even on Pro I hit the five-hour cap fast, which is annoying when I’m in the middle of something.

Claude Opus 5.5 is still what I use most for general stuff: coding, reasoning, random everyday questions. If I had to keep one model for general use, it’d probably be Opus.

Gemini 3.8 Flash vs. Qwen3.8-Flash-Next (Q4, running locally on a Ryzen AI Max+ 395 + Radeon AI PRO R9700, ~42 tok/s gen, ~700 tok/s prompt processing): for the simple stuff my agents do, Gemini is only a bit smarter and the speed is about the same. Kind of wild that the local one holds up that well.

Grok in Cursor with fast mode is good for easy and medium coding tasks, sort of like a smarter DeepSeek. I use 4.6 when I want something quick and cheap. 4.7 is slower and I barely touch it. I wouldn’t use either for science.

I got DeepSeek-V4.1-Flash-1017 running locally but it’s too slow to be useful day to day.

Gemini 4 Carbon came out today. Only tried it briefly: slow, results were fine. Too early to say more. Don't think I can use up the gifted gemini ultra limit.

Overall, which model I pick depends a lot on the job. For research I’ll wait for the best answer. For coding I care most about whether it uses tools reliably. For the 24/7 agents, cost and speed matter more than raw smarts, and a model that hits its limit after a few hours isn’t much use there.

I usually use the agents cross check each other rather than choosing claude or gpt max 20x. Feel like this is the best way to avoid mistakes.

Anyone else running a mix of subscriptions and local models? Curious what you’ve settled on.

#ChatGPT #Claude #Gemini #Grok

[Yuxuan Cao @ycaophysics]

Finally

[Yuxuan Cao @ycaophysics]

Do you really need to spend 20 years mastering one field to do something groundbreaking?

I’ve been questioning this lately. Most successful people push deep specialization, and for a long time, they were right. Human time and brainpower are limited. As fields grow, just getting to the frontier of one discipline takes years, sometimes decades.

Economist Benjamin Jones calls this the "burden of knowledge." As knowledge piles up, we rely more on narrow specialists and big teams just to make progress.

AI changes that.

Up until recently, jumping into a new field meant years of studying, learning the tools, and figuring out what experts already knew. Now you can use AI to pick up unfamiliar concepts, build prototypes, and test ideas much faster.
It doesn’t make you an expert overnight, and it doesn’t replace deep understanding.
What it does is lower the cost of crossing into other fields. I think that’s a much bigger deal than just making existing specialists faster.

A lot of breakthroughs come from connecting ideas that were stuck in separate disciplines. Doing that used to be expensive. You either spent years learning multiple fields yourself or put together a team of specialists. Now, far more people can actually explore those intersections on their own.

Spotting the right problems and connecting knowledge across fields might soon matter just as much as deep expertise in a single area.

Depth still matters, of course. Someone has to know when an idea is wrong, figure out why an approach fails, and check results against reality.

We might just need to flip the order. Instead of spending decades specializing before looking around, we can explore broadly first, find a problem worth solving, and then go deep where it actually counts.

We’ve always traded breadth for depth because human learning time was the bottleneck. AI loosens that tradeoff.

Maybe the next generation of breakthroughs won’t come from knowing more and more about one thing, but from finding connections nobody else thought to explore.

[Yuxuan Cao @ycaophysics]

We pruned 20% of DeepSeek V4.1 Flash. It now fits entirely in home PC VRAM and reaches 88% of an NVIDIA GH200's single-stream generation speed.

Even heavily quantized, the unpruned model slightly exceeded our dual-GPU VRAM (AI max 395 128GB+R9700 windows) capacity, forcing CPU offload that bottlenecked every response.

Because V4.1 Flash is an MoE model, we profiled expert activation across our target domains (math, physics, finance) and pruned the least-used 20%.

What changed:

Zero CPU offload: Shrinking the file by ~20% fit all remaining experts entirely in VRAM.

Speculative decoding: The freed VRAM fit a small draft model alongside the main weights.

22% speedup: Generation rose from 14.5 to 17.7 tok/s.

We ran public evals twice under identical settings (once on the home PC, once on a TACC GH200) and pooled the results against the published unpruned baseline:

Math: 93.8 vs. 95.0

Economics: 87.0 vs. 87.5 (published baseline covers all subjects)

Coding: 78.0 vs. 84.4 (7 of 9 misses hit the max-token limit rather than logic errors)

Science: 79.5 vs. 88.4 (our least trusted score due to biology run variance: 92 vs. 74)

With 13 to 54 questions per subset and a slightly newer quant used in the published baseline, minor gaps here are within noise.

On single-stream decode with the draft model, our home PC hit 17.7 tok/s against the GH200’s 20.2 tok/s. Where the GH200 pulls away is prefill (prompt reading): 395 tok/s vs. our 165 tok/s—though that barely registers on reasoning benchmarks where prompts are a few hundred tokens and answers run into the thousands. Running unassisted across 8 concurrent requests, the GH200 delivers ~9x the total output, but for a single local user, generation speed is remarkably close.

#LocalAI #DeepSeek #AMD

[Yuxuan Cao @ycaophysics]

Texas turning blue the first time in decades

[Yuxuan Cao @ycaophysics]

11/2
Built a Model Context Protocol (MCP) connector to integrate a client's Microsoft Exchange mailbox into their local workflow.

On the research side, I ran several Wake-T simulations to investigate emittance preservation.

Currently implementing a sparse attention head for token processing in ⁠llama.cpp⁠ over ROCm, targeting the DeepSeek V4.1-1017 build.

Wrapped up the day by submitting our application to Y Combinator. Not expecting much.

[Yuxuan Cao @ycaophysics]

10/31
Adding options to my auto trading strategies

Did a few wake-T simulations on electron evolution over distances

Thoughts and ideas:
Ai has been great on the production side of things. However, it has almost improved nothing on the consumption side of capitalism. It has not yet been able to increase or create new demands for consumers. Rather, what it’s doing now is eliminating jobs, which is reducing consumption power. I believe startups should think about how AI would improve on the consumption side of things.

[Yuxuan Cao @ycaophysics]

10/29
Attended AMD AI dev day Austin.
Earned an rocm badge.

Seriously the rocm driver needs serious user experience upgrade. It’s so frustrating to install. Also the llama.cpp performed worse -4% under the most recent rocm 10.1 compiler. AMD adrenalin is also super buggy on windows.

I really like AMD and i do like what’s going on, but please improve your software experience a bit!
#AMD #rocm

[Yuxuan Cao @ycaophysics]

10/28;
Made a more appealing UI UX for a construction contractor tool.

Did some literature search in controlled nuclear fusion.

Built MACD and in my algorithm trading tool.
Btw: my 24/7 AI stock strategy agent sold GOOG two weeks ago, saving me from would-be a $2K loss today.

[Yuxuan Cao @ycaophysics]

I think the three biggest blockers to AI adoption in experimental sciences are: complexity, money, and resistance to change.

1. Complexity

A large experiment can involve thousands, sometimes hundreds of thousands, of interacting components, each with its own settings and degrees of freedom.

It's already difficult to keep everything working reliably, let alone have an AI understand and optimize the entire system.

2. Money

Experimental physics is ridiculously expensive.

Most research funding goes toward equipment and instrumentation. Software, automation, and data infrastructure often get much less attention.

We spend millions building sophisticated experiments, but relatively little on the software needed to make them intelligent.

3. Resistance to change

And this one is probably the hardest.

Lab technicians and experimentalists spend years, sometimes decades, learning the quirks of their machines. Much of that expertise comes from experience and intuition, not textbooks or documentation.

Now imagine someone showing up and claiming that an AI can do their job better.

Why would they trust it? Some might even see AI as a threat to the expertise they've spent their careers developing.

And frankly, their skepticism isn't entirely unreasonable.

Building AI that works is one thing. Convincing people who have spent decades running experiments to actually use it is another.

[Yuxuan Cao @ycaophysics]

For the next 90 days, I’ll post a short log of what I worked on each day.

It won’t be about one project. I’m splitting my time between AI, physics research, and hardware, often switching between them in the same day.

Some things I can’t share publicly, but I’ll document what I can.

[Yuxuan Cao @ycaophysics]

@ThomasBaertl I personally think a 50% increase is very possible. But that only applies to my machine though, not sure about other machines. The issue for me is windows BAR limitation for data transmission between two gpu.

[Yuxuan Cao @ycaophysics]

We actually made the strongest local model (DeepSeek V4.1-1017) run on AMD AI Max+ R9700 and Windows! 🚀

A 748B-parameter MoE (16B active), running locally on one mini PC:
🖥️ Minisforum MS-S1 MAX: Ryzen AI Max+ 395 (Strix Halo, 128 GB unified memory)
🎮 + Radeon AI PRO R9700 32 GB in a USB4v2 eGPU dock
🪟 Windows 11, ROCm 7.2, llama.cpp

Results (Unsloth's 1-bit UD-IQ1_M GGUF 250.4GB):
• 128K context
• ~14 tok/s generation
• ~110–130 tok/s prompt processing
• Loads in ~3.5 min

How it fits:
• R9700: the first 10 layers + the output layer
• Radeon 8060S iGPU (96 GB): layers 10–39 + the context
• CPU: the experts of 4 layers
• The 110 GB Engram n-gram table stays on the NVMe SSD and is read on demand
• V4.1's KV cache is tiny: ~3.75 KB per token, so 128K context takes only ~0.5 GB

The hard part was Windows: the R9700 behind USB4 has no direct GPU-to-GPU access, so every transfer between the two GPUs goes through a pinned host staging buffer. That fix is what keeps the setup stable. Linux shared memory would be much better though. Windows systems takes too much ram by default.

Built on Unsloth's DeepSeek V4.1 PR for llama.cpp. Next we are gonna port it into our Strix Halo + R9700 hybrid fork and add the DSpark drafter for more speed (could be a bit difficult but will try).

#DeepSeek #llamacpp #AMD #ROCm #LocalLLM

[Yuxuan Cao @ycaophysics]

If you build something with AI for a customer and they reject the design, do not immediately keep building on your existing repo!
When requirements change, most people jump straight back into the IDE and try to prompt their way through a pivot. Don't do that. As long as your old design docs are sitting in the project, the agent is going to keep pulling that dead context into the window. You’ll spend hours fighting your own codebase because your prompts say one thing and your repo files say another.
Before you write another line of code:
1. Archive the old design docs so the agent stops reading them.
2. Rewrite ⁠design.md⁠ with what the customer actually wants.
3. Update ⁠claude.md⁠ and ⁠agent.md⁠ immediately so the system rules match the new direction.
Clean up the markdown before you touch the code.

[Yuxuan Cao @ycaophysics]

It’s running 24/7 auto quant research. No crashes on windows (which is a miracle haha)