anish chhaparwal

@bindureddy can we have gpt-4o back on chat llm please?

Introducing Genie 2: our AI model that can create an endless variety of playable 3D worlds - all from a single image. 🖼️ These types of large-scale foundation world models could enable future agents to be trained and evaluated in an endless number of virtual environments. → goo.gle/4f48y2e

@jeremyphoward This should help

For folks that don't use conda, how do you debug problems that only occur on some particular Python or CUDA version? With conda, I have a few envs set up with various python and CUDA versions, so I just activate one and get to work. With something like uv, what do you do?

And Just Like That - Tencent just dropped an excellent open-source video model! IT LOOKS AWESOME The final gap that we had in open-source AI - namely a video-gen model - is now closed 😀 AS USUAL, CHINA RULES

Most generative models predict pixels. Predicting a 3D scene instead has many benefits: the scene won’t change if you look away and come back, and it obeys the basic physical rules of 3D geometry. The simplest way to visualize the 3D scene is a depth map, where each pixel is colored by its distance to the camera. 4/n

@alexalbert__ @alexalbert__ you guys can start an official api repository mpc that can provide Claude with context via /llms.txt for popular standard libraries.

It's been great to see the community's response to MCP! This launch is just the starting point - the team is already working on fast-followup improvements to the docs and Claude Desktop app experience based on your initial feedback. MCP's true potential will unfold in the coming months, but we wanted to get it into devs' hands early so you could help shape its future. Your critiques, suggestions, and use cases are incredibly valuable in ensuring MCP becomes a protocol that works for everyone. Keep the feedback coming!

OH WOW! The Whale aka @deepseek_ai is BACK!! New model, with complete reasoning outputs and a gracious FREE TIER too! 🔥 Here's a quick snippet of it searching the web for the right documentation, creating the JS files plus the necessary HTML all whilst handling Auth too ⚡ I really hope they Open release the model checkpoints too!

@yacineMTB My bro, parsing them into tree. 🫠

decorators are an antipattern and should be avoided

Love Google’s podcast style paper discussion generator, Illuminate! You can try it at: illuminate.google.com/library?pli=1&…

What is the performance limit when scaling LLM inference? Sky's the limit. We have mathematically proven that transformers can solve any problem, provided they are allowed to generate as many intermediate reasoning tokens as needed. Remarkably, constant depth is sufficient. arxiv.org/abs/2402.12875 (ICLR 2024)

@jeremyphoward Windows : ctrl + h

Anyone know of a MacOS or Windows app that does real time voice recognition (i.e hold down a hotkey, and your speech is converted to text and placed into the active text area), BUT can work with a custom model? I'd like to fine-tune whisper for my voice and use it with my PCs.

@tom_doerr do you use any tool to post x? Love the content!

@tom_doerr Try github.com/jimmc414/1file…! It works on webpages too

As context windows get longer, should we just add the complete repo for coding?

@tom_doerr How would you apply for one then?

I feel bad for people applying to jobs on LinkedIn themselves

@rohanpaul_ai @exolabs_ @__tinygrad__ Is there an example script?

Quite incredible setup for running your own AI cluster at home with everyday devices. @exolabs_ and @__tinygrad__ backend. The key is that exo optimally splits up models based on the current network topology and device resources available. See the code in image shows a partition strategy. 👉 Topology: So, the `Topology` object represents the entire network of available devices, their capabilities, and their connections. It's a flexible representation of how the devices in the network are connected and what their capabilities are. The partitioning strategy uses this to make decisions about how to divide the model across the available devices. To make sense of the "Ring" in the name, you can think of it like devices are conceptually arranged in a ring, where each device is connected to two neighbors. This allows for efficient communication between partitions. 👉 Memory Weighted: The strategy weighs each device's contribution based on its available memory. Devices with more memory get larger portions of the model. 👉 Here's what the code actually does: - It sorts the nodes primarily by memory capacity (in descending order) and secondarily by node ID. - It then allocates partitions to these nodes based on their proportional memory capacity. - The partitions are represented as continuous ranges from 0 to 1

@LXIXthenumber Love the thread. Has me inspired. Thanks for letting me know it’s possible!

@Prathkum 👋

I want to collaborate with someone with whom I can build a SaaS and run it for at least a couple of months to generate some revenue. Drop a 👋 if you are interested.

@alexalbert__ Can you make it most to delete multiple chats at once. Checkbox mechanism.

We've got some exciting things coming up in the pipeline but we want to ship even more features that people want. What do you wish we added/fixed on claude dot ai or the API?

@svpino @riverarodrigoa @hridayakuhara I think people will read your content if you tell them it’s promoted in the first line too. At least for the first few times. The quality of products you promote will determine if they continue to do so. Not a fan of dubiously being subjected to promotion.

Will you read the content if I tell you in the first line is promoted? I’m sure the answer is no. You don’t trust the kind of selection I make, so you won’t have any reason to trust it will be good. Therefore, that’s the trade off. I make the selection, you read and decide for yourself. If you don’t like what I bring, you can ignore it. I think it’s a fair trade.

If something could kill Jupyter notebooks, this would be it. Zerve is a web-based IDE. It's block-based, like a notebook, but it follows a different philosophy, and the experience is very different. And they just released a killer feature! You can now drag and drop any model from HuggingFace, a dataset, and fine-tune it using your infrastructure and GPUs. No privacy concerns, no data leaks, and no need to manage any infrastructure. Fine-tune your model with your data and host it in your environment without leaving the tool. A lot is happening in the attached screenshot, but that's an end-to-end process to fine-tune Gemma 2B. This is huge, but that's not all: 1. Zerve is web-based, so you can access your work anywhere 2. Collaboration is a first-class citizen 3. You can deploy your code directly 4. It promotes code modularity by using blocks 5. You can use multiple languages together 6. You can run code blocks in parallel Zerve is new, but it has the potential to change how data scientists and machine learning practitioners write code. I'm rooting for them. The tool is free. You can start right now by visiting. Thanks to their team for collaborating with me on this post.

What if you could send preliminary AI radiology report to the referring physicians within minutes of CT scan being done? That's exactly what Dr. Arunkumar and his team at @AarthiScans are doing in this partnership with us! Ground breaking stuff!