Yash Patil

Building in AI and open source. I track the fastest-rising repos daily at 📡 Sharing what's new in AI, what's worth trying, and what I'm building.

[Yash Patil @Yashpatil]

@RoundtableSpace The hard part for Apple isn't the model, it's trust. An assistant that acts on its own has to touch your messages, payments and apps, and one wrong action on your phone feels a lot worse than a wrong answer in a chat.

[Yash Patil @Yashpatil]

@ai_for_success Parallel agents are great until two of them edit the same file. Curious how Swarm handles that. Does it split the work by files up front, or merge everything at the end?

[Yash Patil @Yashpatil]

@GoogleDeepMind @FryRsquared @Avsecz @adamgayoso The part that clicked for me: the AI doesn't find the cause on its own, it narrows millions of possible DNA changes down to a short list scientists can actually test in the lab. That's where most of the time used to go.

[Yash Patil @Yashpatil]

@kimmonismus Bold, but launching a new model a day before listing is a big risk too. If anything breaks or the evals get picked apart, that becomes the IPO day headline. My guess is they ship right after, not right before.

[Yash Patil @Yashpatil]

@MiniMax_AI It already shows 306K+ downloads on Hugging Face.

One thing to check before you put it in a product: it uses MiniMax's own community license, not Apache or MIT, so read the terms first.

https://t.co/WPuy2dzBPU

[Yash Patil @Yashpatil]

4K images in about 10 seconds, and you can download the model.

MiniMax put out H3-Image as open weights this week. It makes and edits images from text and can follow up to 16 reference pics, so one product or character stays the same across a whole campaign.

Video: @MiniMax_AI

[Yash Patil @Yashpatil]

@kimmonismus Whatever you think of how it ended, he took Nvidia from gaming cards to the chips almost every AI model is trained on. For builders the real question is the stuff we rely on: CUDA, the open Nemotron models and the Hugging Face deal.

[Yash Patil @Yashpatil]

@ClementDelangue Needed to hear this today. Most of what I build sits on open models from the Hub, and that keeps working no matter what happens in any one boardroom. Keep going.

[Yash Patil @Yashpatil]

@benjaminwalker The numbers: 51.5 skill score on fev-bench vs 50.8 for H2O's TimeH2O and 49.0 for Google's TimesFM-3. That's their own chart from public results, so test it on your data.

Open weights, commercial use allowed, runs on CPU or GPU.

https://t.co/MCotDjKqAK

[Yash Patil @Yashpatil]

A lab founded in July just shipped the #1 open model for forecasting.

ci-proto predicts things like power use, sales or sensor data. It can look back over 2M+ past data points, 64x more than other time-series models, so it remembers old patterns.

Video: @benjaminwalker

[Yash Patil @Yashpatil]

@ns123abc The scary part is that no AI hacking was needed. A convincing impersonation got past people who said they'd met him before. Anyone giving agents access to their email should think about this too, because agents are even easier to fool.

[Yash Patil @Yashpatil]

@jerryjliu0 @llama_index Makes sense. Every agent I build ends up stuck on the same step: getting clean data out of PDFs, tables and scans. Most of the time the model isn't the bottleneck, the documents are.

[Yash Patil @Yashpatil]

@chetaslua Banked resets are a smart move. What heavy users hate most is hitting the limit in the middle of a task, and saving a reset for a big day fixes exactly that. Curious if OpenAI answers with something similar.

[Yash Patil @Yashpatil]

@davidvgilmore The catch: those are Rayline's own numbers, and self-hosting isn't smooth yet. Stock vLLM can't load it and their public plugin is unfinished, so most people will use the hosted API for now. The weights are Apache 2.0 on Hugging Face.

https://t.co/YnUCrH4NW3

[Yash Patil @Yashpatil]

Most AI agents use their priciest model for every step. Even the easy ones.

Rayline open-sourced ARC, a tiny 0.8B model that picks the model for each step. Their test: 68.7% on DeepSWE vs 69.9% for Opus 5.5, at about 7% of the cost.

Video: @davidvgilmore

[Yash Patil @Yashpatil]

@ns123abc Closing the rental route matters more than it sounds. For a lot of labs, renting GPUs abroad was the plan B after the chip bans. I'd expect even more push on efficient training and home-grown chips next.

[Yash Patil @Yashpatil]

@RoundtableSpace 22 hours is the interesting number. That's not a one prompt demo, it's a full workday of the agent testing and fixing its own work. The real skill now is giving it a goal clear enough to stay on track that long.

[Yash Patil @Yashpatil]

@ai_for_success Building a tool you use every day is the real test. Demos can break, daily tools have to keep working. Curious how it handled the boring parts, like fixing the pipeline when one of your platforms changes its API.

[Yash Patil @Yashpatil]

@n8n_io What I like: the safety rules live in the server, not the prompt, so an email or web page can't give it orders. Passwords go in through a vault the model never sees. Apache 2.0.

It jumped from 436 to 536 stars in a day on https://t.co/C3iBwBhrL1

https://t.co/PTPRJVwPso

[Yash Patil @Yashpatil]

An AI agent that cancels your free trials for you. And it's open source.

Nod from n8n Labs gets its own computer and browser, so it can chase invoices, prep your calls and cancel subscriptions on real websites. It asks before it sends, pays or deletes anything.

Video: @n8n_io