hershy

just trying out random stuff

[hershy @hershyb_]

Favorite SORA video results

Produced by other models 🧵

[hershy @hershyb_]

Mona Lisa

[hershy @hershyb_]

Starry Night by GPT 5.6 Sol

[hershy @hershyb_]

full article: https://t.co/ag7UY5mrTC

[hershy @hershyb_]

$100 AI Music Video: Claude Fable 5 vs GPT 5.6 Sol

[hershy @hershyb_]

Cost vs. speed/reply metrics for grok 4.5 vs other frontier models

[hershy @hershyb_]

opus 4.8

[hershy @hershyb_]

grok 4.5 vs fable 5 creating a gravity sandbox

[hershy @hershyb_]

opus 4.8 👀

[hershy @hershyb_]

grok 4.5 vs fable 5 creating an SVG of a horse riding an astronaut

[hershy @hershyb_]

@KhareAbhishyant holy 🔥🔥

[hershy @hershyb_]

Try TinyAI instead lol

[@levelsio @levelsio]

Tried Gemma 4 ran locally on my iPhone today

I thought it'd be useful in case the apocalypse happens and I need to ask it for survival tips

Like how to make a fire 🔥

I guess I'll freeze to death instead 🫠

[hershy @hershyb_]

https://t.co/YQcLyoVMF3

describe your UI in realtime, share your favorites

[hershy @hershyb_]

@perplexity_ai why,,,

[hershy @hershyb_]

…what

[Garry Tan @garrytan]

Is it possible I used Claude Code so much somehow my USB-C connectors burned out on my MacBook Pro? Two of mine are dead and won’t charge and now the third maxes out at 15 watts

Now I am afraid my code tamagotchi is about to die

[hershy @hershyb_]

Fully agent-based solution, no hard-coded rules 👀

3x speed, took 223.42s, skip to 50s

[Brett Adcock @adcock_brett]

Solve this in under 5 minutes and I’ll offer you $500k/year in cash plus several million in equity

I'm building a Computer-Use team, goal is to use computers better than humans

No experience or PhD needed

Instructions:
1. Solve all 30 challenges on this website in under 5 minutes: https://t.co/7DmmS5OQNl

2. Feel free to use any tools or vibe code it. Provide us a zip folder with instructions on how to run the agent and reproduce your results, as well your run statistics

3. The agent should be able to solve all the challenges, use browser, and provide overall metrics around time taken, token usage and token cost. Your agent must solve this challenge in under 5 minutes

Email your response: [email protected]

If you have any questions about this challenge, feel free to email us

[Lin Qiao @lqiao]

@deepseek_ai just found a surprisingly clean way to make “multi-lane” residuals actually train efficiently, and the benchmark jumps are hard to ignore. Here is the breakdown:

🏁 Primer concept: the residual “highway” is the main path that carries what the model already knows forward through every layer while each layer adds a small tweak on top.

🛣️ Primer idea: split the residual “highway” into multiple parallel lanes so different lanes carry different info, then mix them across layers.

⚠️ Problem: if each layer can freely “amplify” lanes, gains compound → some lanes explode, others go dead → unstable training.

✅ Fix: a simple conservation rule — layers can redistribute/mix across lanes but can’t increase total magnitude (like moving sliders on a mixer board without raising the master volume). This is the gist for mHC.

🎯 Why it matters: activation stays stable even with many stacked layers → predictable training + better quality

This is impressive research. We’re excited to see the next generation of models built on top of this architecture.

[hershy @hershyb_]

Tfw a print statement fixes your code 🥰

[hershy @hershyb_]

If Google goes Gemini 3 mode on Pixel it’s over for Apple

[Lin Qiao @lqiao]

🚀 Eval Protocol is Open Sourced!

Reinforcement fine-tuning is complicated, because there are hundreds of environments and tens of trainers you can pick and choose and integrate with. Even worse, in production, agents don’t live in clean “gyms.” They operate in messy, async environments - flaky APIs, partial observability, conflicting objectives, long feedback loops.

We solve that problem by open sourcing Eval Protocol. The goal is for you to build your production RFT flow without reinventing the wheel of managing such complex integration.
👉 Day 0 support for trainers and environments like TRL (@huggingface), @rllm_project , OpenEnv (@PyTorch),  as well as support for proprietary trainers like @OpenAI RFT and @thinkymachines Tinker. More to come.
👉 Instrument agents in production instead of toy or simulated environments
👉 Move from offline benchmarks to live, continuous improvement

[hershy @hershyb_]

@the_bunny_chen 🚀🚀🚀