Jesse White

Taking things apart to figure out how everything works

[Jesse White @byjlw_]

@karpathy Definitely lines up with my experience.
Claude sonnet 3.5 is still my go to model. Doesn't matter what the benchmarks or the arena says. That model outperforms all others consistently

[Python Trending 🇺🇦 @pythontrending]

executorch - On-device AI across mobile, embedded and edge for PyTorch https://t.co/7lzoykDmyZ

[Jesse White @byjlw_]

ExecuTorch is way easier to use now. Fastest framework for LLMs on Mobile and embedded devices

[Executorch: Open-source ondevice AI from @pytorch @Executorch]

ExecuTorch v0.6.0 is now live
This release includes:

* Significant usability and stability fixes
* Windows support
* Ready Made Packages for iOS and Android
* Native Object-C and Swift APIs
* New OpenVino backend

The full details here
https://t.co/iLlll1uV9w

[Jesse White @byjlw_]

folks hanging out around a local max almost always resist when someone pushes the group towards the global max

[Jesse White @byjlw_]

@zerohedge But why?

[Jesse White @byjlw_]

Before people get carried away.....
it's just a handful of genetic edits. Cool, but as per usual, the headlines are a bit misleading

[Alina Chan @Ayjchan]

Colossal founder told @techreview they’re concerned about Game of Thrones people coming to check out the so-called dire wolves.

Maybe state the facts that these are not real dire wolves but whitened grey wolves with 15 genetic edits with yet unclear changes.

[Jesse White @byjlw_]

@karpathy AI agents with eth are the way to go. If you know someone who'd fund it, I'll quit what I'm doing and go build it.

[Jesse White @byjlw_]

@karpathy Yep, my repositories all have markdown files just for LLMs now.

[Jesse White @byjlw_]

It's a protocol.
The internet was cool and exciting in the early 1990s but not particularly useful because there wasn't a standard way for how data should look like between servers and browsers. Then came HTTP, a protocol with a list of rules and guidelines.
MCP is the same thing between LLMs and services that want to interact with it. Now things can really take off since everyone has the same set of expectations and rules.

MCP servers are APIs that use the same rules as everyone else.

[Jesse White @byjlw_]

People who use reasoning models to solve R1 find it solves a lot of their problems pretty well compared to 01 mini and 01. It can't solve the hardest problems 01 solves but for a fraction of the price and open weights, it's the leader for most applications from a total token standpoint

[Jesse White @byjlw_]

@ID_AA_Carmack Same, and by how many programmers who don't use them.

[Jesse White @byjlw_]

@ID_AA_Carmack Also looking

[Jesse White @byjlw_]

Take my money
https://t.co/DNkeycSqKn
#ai #tinysupercomputer #nvidia

[Jesse White @byjlw_]

@realGeorgeHotz @TimSweeneyEpic @AMD @LisaSu My cap ex is higher with H100s but I have a power problem with a stack of 4090s.

None the less. I agree with you, all AMD needs to do to start closing the gap with NVIDIA is to give us a good end to end software stack for their cards, mirroring CUDA.

[Jesse White @byjlw_]

@noahgwok @aigclink Search: creating a text index of the videos for easy searching
Assisted: Helping folks vision issues, such as blindness, figure out what the video is about
Data Automation: Processing large amounts of videos

[Jesse White @byjlw_]

@daniel_mac8 @karpathy If the cognitive load is high, LLMs will struggle, but developers can use LLMs as they build systems and write code to reduce the cognitive load of future developers. A bonus is that it make it easier to use LLMs to add improvements and fix bugs in the future. A win win.

[Jesse White @byjlw_]

A couple weeks ago I released a GitHub project for describing videos using whisper, clip and Llama 3.2 Vision and it's starting to get a bit of traction. The prompts still need some tuning to improve the output but it's staring to work pretty well.
https://t.co/1MqAaczWLw

[AIGCLINK @aigclink]

一款AI视频分析工具：video-analyzer，基于Llama3.2 Vision和OpenAI Whisper构建，可本地运行

自动提取关键帧，智能识别画面内容，可处理长视频
生成自然语言描述 ，可输出详细的JSON格式分析结果，包括元数据、音频转录、逐帧分析和视频描述等

#AI视频分析工具 #视频分析助手

[AIGCLINK @aigclink]

一款AI视频分析工具：video-analyzer，基于Llama3.2 Vision和OpenAI Whisper构建，可本地运行

自动提取关键帧，智能识别画面内容，可处理长视频
生成自然语言描述 ，可输出详细的JSON格式分析结果，包括元数据、音频转录、逐帧分析和视频描述等

#AI视频分析工具 #视频分析助手

[Software Mansion @swmansion]

We've been actively exploring AI on mobile devices & we are excited to introduce React Native ExecuTorch 🚀!

A new library that enables the execution of AI models with just a few lines of JavaScript, without the need for API calls or external infrastructure. 🤩

More in 🧵👇

[Jesse White @byjlw_]

It's late and i spent most my day chasing a bug. It was a dumb bug but it felt good in some odd. I spend so much time managing things now that I rarely get to be technical. Miss the days when i often got to put my head down and just build. #torchchat #llama