OpenAI Is Growing Fast and Burning Through Piles of Money (5 minute read)
OpenAI's recent financial documents reveal $300 million in monthly revenue, a 1,700% increase since early 2023, with annual sales projected to reach $3.7 billion. Despite this growth, the company expects to lose approximately $5 billion this year due to high operational costs. OpenAI is seeking $7 billion in a new funding round, which will value the company at $150 billion.
|
Altman reportedly trying to sell Biden on a slew of AI DCs (4 minute read)
OpenAI CEO Sam Altman is urging the Biden administration to build AI data centers consuming up to five gigawatts of power to maintain US technological leadership over China. The plan details constructing multiple data centers in the US. Other tech giants, like Microsoft and Amazon, are also securing nuclear power agreements to support their AI operations.
|
|
Emu 3 open early fusion multimodal model (6 minute read)
Emu 3 is a next token prediction model that outperforms SDXL on image synthesis, LlaVa-1.6 on image understanding, and OpenSora 2 on Video generation. It is a 9B parameter model trained on all these tasks in an interleaved manner, similar to Gemini.
|
Improved adaption of pretrained priors (12 minute read)
Using a pretrained diffusion model for tasks like depth estimation is extremely popular and powerful. This work shows how some of the previous methods were slightly wrong and improves performance while dramatically simplifying the modeling process.
|
|
Lean RL from PyTorch (GitHub Repo)
A fork of CleanRL that has been optimized to use PyTorch's newest performance and stability features. It is dramatically faster while also being simpler to understand and extend.
|
New Benchmark for Video Language Models (2 minute read)
E.T. Bench is a new benchmark designed to evaluate video language models on fine-grained, event-level tasks. Unlike previous benchmarks that focus on video-level questions, E.T. Bench covers a range of time-sensitive tasks across multiple domains.
|
|
Table Extraction using LLMs: Unlocking Structured Data from Documents (30 minute read)
This article highlights how large language models (LLMs) are revolutionizing table extraction from complex documents, overcoming the limitations of traditional methods like OCR, rule-based systems, and machine learning. LLMs demonstrate flexibility and contextual understanding, notably enhancing accuracy in diverse and intricate table structures. Despite challenges like hallucination and high resource demands, combining traditional techniques with LLMs is currently the most effective strategy for automated table extraction.
|
The Other Bubble (40 minute read)
Microsoft considered reallocating its US-based server power to GPUs for AI but ultimately scrapped the plan. Big Tech companies, including Microsoft, Google, and Amazon, heavily invest in AI but primarily see underwhelming returns in generative AI applications. The industry's reliance on SaaS and the integration of AI tools, which often add little genuine utility while incurring high costs, highlights a growing desperation to maintain growth amidst a slowing market.
|
|
Wispr Flow (Product Launch)
Wispr Flow is an AI dictation app that lets you speak naturally and write in your style across every application.
|
|
|
Love TLDR? Tell your friends and get rewards!
|
|
Share your referral link below with friends to get free TLDR swag!
|
|
|
|
|
Track your referrals here.
|
|
Want to advertise in TLDR? 📰
|
If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us.
If you have any comments or feedback, just respond to this email!
Thanks for reading,
Andrew Tan & Andrew Carr
|
|
|
|