Introducing ambientGPT: an open-source and multimodal MacOS foundation model GUI
Run GPT-4o and open-source models with full ambient knowledge of your screen.
Foundation models have long been confined to the browser. With ambientGPT, your screen context is directly inferred as
Introducing Gemma with a 10M context window
We feature:
• 1250x context length of base Gemma
• Requires less than 32GB of memory
• Infini-attention + activation compression
Check us out on:
• 🤗:
• GitHub:
• Technical
Re Llama3V: Firstly, we want to apologize to the original authors of MiniCPM.
@AkshGarg03
and I posted Llama3V with
@mustafaaljadery
. Mustafa wrote the code for the project. Aksh and I were both excited about multimodal models and liked the architectural extensions on top of
Shocked! Llama3-V project from a Stanford team plagiarized a lot from MiniCPM-Llama3-V 2.5!
its code is a reformatting of MiniCPM-Llama3-V 2.5, and the model's behavior is highly similar to a noised version of MiniCPM-Llama3-V 2.5 checkpoint.
Evidence:
Mustafa (
@maxaljadery
) and I are excited to announce MLXserver: a Python endpoint for downloading and performing inference with open-source models optimized for Apple metal ⚙️
Docs:
Building on top of last week’s release, we introduce mlxcli. Build on top of Apple MLX (
@awnihannun
) and 🤗 (
@reach_vb
@julien_c
) with mlxcli.
Usage: pip install mlxcli
Docs:
MLXcli achieves over 20+ tok/sec on M2 Mac’s
Need a way to keep up with the Cambrian explosion in AI research? Built a better way to find and bookmark the latest papers with
@apatwa7
- . We enable rapid search and QA for over 240,000 ML papers.
Built with
@supabase
,
@LangChainAI
,
@pinecone
,
@vercel
Don't think in USD, think in dollars per hour for 8x NVIDIA H100 SXM with 80 GB VRAM/GPU, 208 vCPUs, 1800 GiB RAM, and 26 TiB SSD storage.
For the price of a vanilla latte, you could probably get some gradient updates.
We're excited to announce $80M in seed and Series A funding co-led by
@sequoia
and
@lightspeedvp
to further our mission of orchestrating the world’s compute capacity, making it universally accessible and useful.
How we can help 👇
Unlike OpenAI’s desktop app where you must provide a screenshot or upload a file, the context from your screen is automatically parsed. We also provide the ability to run secure local models like Gemma and Phi-3 multimodal from our interface. Due to the local model sizes, at
We're changing how people discover, understand, and share AI research. 🪄
Now with folders, bookmarks, chat-w-paper, highlight and explain, friends, inbox, and more, check out Cambrian 2.0: 🦕
We’d like to thank the folks at Meta for their work in ensuring open-source is here to stay. We also wanted to shout out the authors of LLaVA-UHD as our methods are directly inspired by their intuition when it comes to image splitting and prepending the latents to the text.
1/ Some NEWS–– unveiling Lux 8 → ♾️…
• Early-stage (from inception to expansion)
• From founders first $100k to their last $100m
• Lux’s largest ever fund ($1B+)—now with $5B+ AUM
.
@maxaljadery
and I authored a short textbook on RL to explain the concepts with our own style of pedagogy. We understand best by teaching and there's no better way to learn than trying to convey nuanced topics with as much simplicity as possible.
We’re excited about the future of open-source models and would love to hear any thoughts and/or suggestions on how we can take long-context further.
Shoutout to the Gemma team for their awesome work in building these open-source models!
@OriolVinyalsML
@clmt
@JeffDean
@koraykv
ambientGPT is open-source and we plan to integrate vllm and ollama to provide more extensive inference hosting abilities with our multimodal GUI. We also aim to release ambientGPT on the apple app store soon.
Hey all, last weekend
@maxaljadery
and I built a rapid indexing and 80x+ faster visualization layer to understand how groups of neurons (features) activate and cluster based on the latest data from
@AnthropicAI
's mechanistic interpretability research. 🧵
ML research is the most "empirical" science to ever exist. we test things and when they work or look promising, we double click ... so a lot of major results are child nodes of a tree of trial and error
i love coding. it makes me think about the highways and mountain roads - someone constructed them with intention. in software, the same effect is just as wild - someone wrote the compiler, built the text editor, and designed the GUI.
people building off one another is how the
Amazing news!
@graceisford
is the best mentor I could ever ask for. She's made my time at
@Lux_Capital
super memorable and full of amazing learnings. There's no one who deserves it more.
Having worked at
@AWS
, and with
@AnthropicAI
recently adding Claude to Bedrock,
@maxaljadery
and I built Python and typescript SDKs to interact with Anthropic’s models on AWS Bedrock. It makes it really easy to do all of the AWS auth and use Anthropic models in production. The
Hey folks -
@maxaljadery
and I are excited to launch : an open-source infrastructure for labeling multimodal data while enabling RLHF tagging and augmenting your existing training data at no cost.
Today I’m thrilled to announce
@Lux_Capital
's NYC AI Directory & NYC AI Map - 2 resources for the burgeoning AI talent ecosystem
READ MORE👇
NYC AI Directory:
NYC AI Map:
we're collecting project ideas for D3N to try out!! have interesting devin projects you want to try out? reply below and we'll select the most upvoted/interesting ideas to send to devin
Bonus points if the ideas are naturally distributed or parallelizable
Great work is when you get in a flow state and genuinely enjoy the process of what you're creating, not when you fully pander to an institutional ranking or external arbiter.
It feels surreal that we're talking about academic conferences and submitting papers and bad reviewers in 2023.
We're literally living in a sci-fi future. You could choose to do anything. But you're choosing to play a snobby cargo-cult game for fake internet points.
It's an
Run Apple MLX from your menu bar.
Introducing Pico MLX Server, a graphical frontend to download and start multiple(!) AI models locally on your Mac.
You can use it with any chat client you like (e.g.
@PicoGPT
) that uses the OpenAI API standard.
I'm very excited to share that I've started as an Assistant Professor of Computer Science at Stanford University! My lab will focus on self-improving AI methodologies and systems. Looking forward to working with the next generation of researchers! ()
Good lesson in software development today. Never have your product totally rely on third-party APIs/services.
For Cardinal Compass, we happened to suffer from 3-4 hours of lost traffic on the day of course enrollment due to pinecone being down.
Better be safe than sorry.
@sama
I’m super excited to have you join as CEO of this new group, Sam, setting a new pace for innovation. We’ve learned a lot over the years about how to give founders and innovators space to build independent identities and cultures within Microsoft, including GitHub, Mojang Studios,
When I was in middle school I qualified for Nationals at MathCounts
and I remember distinctly watching
@0xShitTrader
(CEO of Ellipsis), absolutely destroy in the Countdown round
That was when I realized I was very very good at math, but I was not Eugene
If you like solving problems, the good news is that you'll never run out. Either you fail to solve a problem, and need to keep trying, or you succeed, in which case you've created new problems that need solving.
# explaining llm.c in layman terms
Training Large Language Models (LLMs), like ChatGPT, involves a large amount of code and complexity.
For example, a typical LLM training project might use the PyTorch deep learning library. PyTorch is quite complex because it implements a very
Novak is the greatest all-around player and winningest of all time. Rafa is the greatest clay court player and fighter of all time. Roger played the most beautiful tennis of all time with the most grass-court dominance.
Novak is on top but all three are GOATs in their own right.
A mentor shared this article with me several months back ... the lessons and beauty of the article are unmatched for anyone who enjoys writing software:
1/
@SohamGovande
,
@jameszhou02
,
@jzhou891
and I spent the weekend building PodPlex: A platform for distributed training & serverless inference at scale
I'm very glad to say that we left $10,000 GPU credits richer and 36 hours of sleep poorer
more details in 🧵
If someone could encode the entirety of YouTube, that would be the world's ultimate reflector of humanity: GPT-6 if you will.
TLDR: Video data extends the richness of images by another dimension and YouTube's 1B+ videos enable this at an unprecedented scale.
something like an IAEA for advanced AI is worth considering, and the shape of the tech may make it feasible:
(and to make this harder to willfully misinterpret: it's important that any such regulation not constrain AI below a high capability threshold)
We have a new version of Speed Insights coming.
The key evolution: it's an intelligent "Kanban Board" of what pages you should be optimizing.
… instead of maintaining TODO lists and issue trackers, let
@vercel
do it. Oh, and the issues "close" as data comes in in realtime 😁
Unless you’re an incumbent provider (ex. AWS, Nvidia, Azure, etc.) or model builder/source who can pretrain at a scale that less than 8-10 companies can operate at (inflection, OAI, anthropic, etc.) - you are most likely ngmi
The most underrated startup paradigm in my eyes:
Take an AWS service --> understand its customers and potential shortcomings --> make it as intuitive and straightforward as possible --> distribute/market and capture similar horizontal opportunities (consolidate)
Favorite Driver of all time: Ayrton Senna
Driver I dislike: Max Verstappen
Driver that grew on me: Charles Leclerc
Most overrated Driver: Fernando Alonso
Most underrated Driver: Nico Rosberg
The GOAT of F1: Michael Schumacher
Favorite Driver of all time:
Driver I dislike:
Driver that grew on me:
Most overrated Driver:
Most underrated Driver:
The GOAT of F1:
Comment/quote with your answers
Deep learning is unique in that the field has such high promise when customer-facing/in-production but we know so little about why certain optimizations create certain effects.
.
@_sholtodouglas
poses a challenge.
In the spirit of
@natfriedman
(whose Vesuvius Challenge was solved by a listener of my podcast -
@LukeFarritor
).
Can you figure out what the experts in a Mixture of Experts model are each specialized in?
"A wonderful research project to do:
Love letter to
@obsdmd
to which I very happily switched to for my personal notes. My primary interest in Obsidian is not even for note taking specifically, it is that Obsidian is around the state of the art of a philosophy of software and what it could be.
- Your notes are
Google announces Stealing Part of a Production Language Model
We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like OpenAI's ChatGPT or Google's PaLM-2. Specifically, our attack recovers the
Introducing ambientGPT: an open-source and multimodal MacOS foundation model GUI
Run GPT-4o and open-source models with full ambient knowledge of your screen.
Foundation models have long been confined to the browser. With ambientGPT, your screen context is directly inferred as
1/ As promised, here's my thesis on the future of decentralized training of foundation models. Covers:
1) why decentralized makes sense from scaling, margins, and marketplace lenses
2) challenges
3) exciting enabling research shifts
In long form at:
A lot of people ask about : why do we exist? Google would just eat you and do the same. How dare you think you have better engineers than Google? Well, we’ve never ever said we have better engineering. I’m a fan of every OG Googler and the founders. The