Tommy Zhao

building @

[Tommy Zhao @tommyyzhao]

gave 5.6 Sol and Opus 5 the same engineering prompt - Sol seems to provide more thorough, ideal, "theoretically complete" proposals by default, whereas Opus gives more practical and executable proposals. interesting.

[Tommy Zhao @tommyyzhao]

"Physical prompt engineer" is about to become a thing

Can't wait to see the physical benchmarks

[Joe Harris @_joe_harris_]

i don't think people realize what's happening right now

heres why this is might be the "GPT-3" moment for robotics:

1. "physical prompting" is real
show GEN-1.5 just 3 to 12 seconds of one demonstration. no fine-tuning or gradient updates required. you show it once and it figures out the rest.

2. in-context learning emerges from scale
generalist says this capability was not explicitly trained. but it emerged from pretraining on physical data at scale.

3. it shows “improvisational intelligence”
teach it to sweep with a brush, then give it a banana and it turns the banana into an improvised tool 🍌. give it a dustpan and it invents a new two-handed strategy.

4. skills become composable
show it two tasks and it can chain them together, including intermediate motions it was never shown.

5. prompts cross worlds and bodies 🤯
demonstrate a task in simulation and it can execute it in reality. demonstrate it with human hands and it can reproduce it with robot hands.

the tasks are still simple and success rates are not yet high, but these are very promising signs.

robot programming is starting to look a bit like "watch and learn" kiddo 🤖

[Tommy Zhao @tommyyzhao]

@andrewqu @vercel 🏋️🧑‍💻🏋️

[Tommy Zhao @tommyyzhao]

@MalwareJake How wobbly is the top I wonder

[Tommy Zhao @tommyyzhao]

@PhiladelphiaGov pls do the fireworks already 🙏 people need to sleep

[Tommy Zhao @tommyyzhao]

a aura desse jogo vai entrar pra história

[Dan @odanabreeu]

que domínio de respeito do vozinha #ARGxCBV

[Tommy Zhao @tommyyzhao]

@psychiel This a fascinating and terrifying. The forfeiture of moral relevance would be entirely consensual. People would line up for these Z-pills. In fact some might say they already are. More and more people will sign up to become "meat puppets" or "human api" for their AI agent/master.

[Tommy Zhao @tommyyzhao]

are we looking at web 4.0?

[Zain Shah @zan2434]

Imagine every pixel on your screen, streamed live directly from a model. No HTML, no layout engine, no code. Just exactly what you want to see.

@eddiejiao_obj, @drewocarr and I built a prototype to see how this could actually work, and set out to make it real. We're calling it Flipbook. (1/5)

[Tommy Zhao @tommyyzhao]

his last name is literally Shipper 😂

[Tommy Zhao @tommyyzhao]

@chaodoze Google's WebMCP might be a first step in that direction

[Tommy Zhao @tommyyzhao]

https://t.co/Aod1TSaC1e

[Tommy Zhao @tommyyzhao]

@chaodoze I wonder how useful agent-to-agent would really be, socially. One could imagine a social platform where instead of DMs, everyone has a "public agent" and other agents have to first "convince" the public agent to get a message/request to the human, before the human sees anything

[Tommy Zhao @tommyyzhao]

need a gom jabbar level interview question

[Tommy Zhao @tommyyzhao]

@jianxliao @tinyagicompany might go well with https://t.co/D8qgiKOpin and https://t.co/w9VUti7vz9

[Tommy Zhao @tommyyzhao]

👀

[Patrick Hsu @pdhsu]

Evo 2, our fully open-source biological foundation model trained on trillions of DNA tokens spanning the entire tree of life, is out in @Nature today

We & the scientific community have done a lot with this @arcinstitute @nvidia model in the last year! 🧵👇

[Tommy Zhao @tommyyzhao]

TLDR:
• Supervised Fine-Tuning (SFT) is broken for agents. SFT only teaches the happy path. When APIs timeout or searches fail, SFT models hallucinate because they lack error-recovery data. You cannot imitate your way through a dynamic state space; Reinforcement Fine-Tuning (trial and error) is a necessity.

• Absolute scoring is a stupid requirement. Legacy RL required training reward models to score outputs absolutely (e.g., 0.8/1.0). This demands massive human labeling and fails because LLMs are fundamentally bad at isolated grading.

• Relative ranking is the unlock. LLMs suck at absolute grading but are excellent at comparing options. GRPO + RULER works by generating multiple attempts, having a frontier model rank them, and updating weights based purely on relative ordering. Zero hand-written reward functions needed.

• RL architecture had to be unbundled. Legacy frameworks assumed simple, closed-loop chatbots. Real agents are messy and take time to execute. Modern setups completely decouple the agent's environment execution from the backend weight updates, allowing the system to hot-swap new LoRAs continuously.

• API mastery via self-play is the new baseline. Point a cheap 3B model at an MCP server, auto-generate synthetic tasks from the tool definitions, and let it run continuous trial-and-error loops until it masters the API without any human oversight. Stop hand-writing test suites for edge cases.

[Akshay 🚀 @akshay_pachaar]

https://t.co/VhUJPU8alY

[Tommy Zhao @tommyyzhao]

stablecoin issuers and DeFi liquidity pools present alternatives to the state-bank monopoly over money creation—though still a small, early-access club

[Tommy Zhao @tommyyzhao]

human attention is zero-sum yet infinite money can be loaned into existence

[Tommy Zhao @tommyyzhao]

the future contains buildings filled entirely with mice, machines, petri dishes, drugs, and the occasional human

[Tommy Zhao @tommyyzhao]

@penguin2716 oh-my-opencode (or oh-my-claudecode for Claude Coders) has basically already done this