TLDR Index
TLDR AI

TLDR AI 2024-08-28

6 sections21 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

01GENERAL

Sponsored

SOME COMPANIES 馃く

馃憢 Say hi to Bland, the world's most realistic AI phone-calling

bland.aiSponsor

agent.

Chances are, you've already spoken to it without even noticing... It can:

bland.ai

02HEADLINES & LAUNCHES

CEREBRAS INFERENCE

Cerebras' chipset has massive unified memory. As a result, it can sidestep bandwidth issues and serve models at thousands of tokens per second.

cerebras.ai7 min read

OPENAI AIMS TO RELEASE NEW AI MODEL, 鈥楽TRAWBERRY', IN FALL

OpenAI is planning to release a new AI product called "Strawberry" in the fall. It will feature advanced reasoning capabilities, such as the ability to solve previously unseen math problems, and can perform high-level tasks like developing market strategies.

pymnts.com2 min read

03RESEARCH & INNOVATION

PHYSICS OF LANGUAGE MODELS: PART 3.3

The 3rd part of a series, this paper explores scaling laws and how many bits are required to store knowledge in a model. The answer seems to be around 2 bits of knowledge per parameter.

arxiv.org27 min read

FIRE-FLYER AI HPC

DeepSeek has released a paper outlining its hardware-software co-design strategy for DL training.

arxiv.org34 min read

ANNOUNCING HIGGS LLAMA V2

Boson AI has introduced Higgs-Llama-3-70B-v2, a new model that excels in dialog and comprehension benchmarks like Arena-Hard and AlpacaEval 2.0. The model reduces response regeneration rates by 21.6% and boosts day 1 retention by 5.3% compared to Claude 3.5 Sonnet. Enhanced by an in-house reward model, Higgs Judger, it ties with Google's Gemini 1.5 Pro on performance.

boson.ai2 min read

04ENGINEERING & RESOURCES

Sponsored

KEYWORD SEARCH IS A DUD

Customers don't always use the words you expect them to, which tends to break keyword-based solutions for knowledge base and in-product search. Luckily, genAI does much better. Make your help center actually helpful with Ask AI's Generative Search [13]. Minutes to

ask-ai.comSponsor

HYBRID MODEL POST TRAINING COOKBOOK

Pre-training hybrid (Mamba style) models is different from pre-training normal Transformers. This post explores how to scale different hyperparameters, data acquisition, and others to get the performance you want.

zyphra.com25 min read

LLAMA DUO: SYSTEM FOR MODEL FALLBACK

This is a framework that fine-tunes small models to serve as a fallback when closed API models go down. It shows how to migrate from a large to a small model smoothly.

github.comGitHub repository

LITSERVE

LitServe is an easy-to-use, flexible serving engine for AI models built on FastAPI. Features like batching, streaming, and GPU autoscaling eliminate the need to rebuild a FastAPI server per model.

github.comGitHub repository

05MISCELLANEOUS

LLAVA BITNET (HUGGING FACE HUB)

Llava BitNet is the first ternary (-1, 0, 1) weight model trained on VLM tasks. The model, weights, and scripts are in the process of being fully open sourced. The technical report will be released soon and suggests the model has promising performance.

huggingface.co

INFERENCE IS FREE AND INSTANT

While large language models (LLMs) may not significantly improve in reasoning capabilities, their decreasing costs and increasing speeds will make them more useful for repetitive tasks. Although these models may lack true understanding, they can still efficiently handle straightforward tasks.

fume.substack.com4 min read

MEMORY-EFFICIENT LLM TRAINING

Online Subspace Descent is a new optimizer that improves the training of LLMs by enhancing memory efficiency.

arxiv.org17 min read