TLDR AI 2024-08-26

Selling AI to enterprise means getting serious about security, user identity, and access management — involving features like SSO, SCIM, and RBAC/FGA.

With WorkOS, you can proactively drive enterprise growth with a few lines of code.

→ Flexible and easy-to-use APIs help you quickly ship enterprise features.

→ User management is free up to 1 million MAUs and comes with bot protection, MFA, and roles and permissions.

→ Great user experience all around — from high-quality documentation for your engineers, to self-serve onboarding for your customers.

WorkOS is trusted by the hottest AI startups: Perplexity, Copy.ai, Cursor, Jasper, and many others. Start integrating today.

Microsoft launched three new models in its Phi series: Phi-3.5-mini-instruct, Phi-3.5-MoE-instruct, and Phi-3.5-vision-instruct, each addressing different AI tasks with impressive benchmark performances. These models are open source under the MIT License and available for developers on Hugging Face. Despite their smaller size compared to some counterparts, the Phi models have achieved near-state-of-the-art results, surpassing competitors like GPT-4o and Llama in certain benchmarks.

Google DeepMind's AGI Safety & Alignment team shared a detailed update on their work focused on existential risk from AI. Key areas include amplified oversight, frontier safety, and mechanistic interpretability, with ongoing efforts to refine their approach to technical AGI safety. They highlighted recent achievements, collaborations, and plans to address emerging challenges.

Stephen Wolfram, the creator of Wolfram|Alpha, wrote a great exploration about different phenomena going on in machine learning models. He discovered a set of dynamics about the non-linearities in simple neural networks.

Time Series Diffusion Embedding (TSDE) is the first diffusion-based approach for time series representation learning. TSDE segments time series data and uses dual-orthogonal Transformer encoders with a crossover mechanism to generate informative embeddings.

Spiking Autonomous Driving (SAD) is the first unified Spiking Neural Network (SNN) aimed at addressing the energy challenges of autonomous driving.

Inheritune is a streamlined method for developing smaller base language models from larger ones by inheriting a few transformer blocks and training on a tiny fraction (0.1%) of the original data. This approach enabled the creation of a 1.5B parameter model using only an A6000 GPU in less than half a day, achieving results that compete well against larger models trained on significantly more data.

GenderCARE is a comprehensive framework aimed at assessing and mitigating gender biases. It introduces innovative criteria for evaluating gender bias, emphasizing inclusivity, diversity, and objectivity.

pgvectorscale builds on pgvector with higher performance embedding search and cost-efficient storage for AI applications. It is around 28x faster than competitive and commonly used vector stores.

The Liger Kernel, released surprisingly by LinkedIn, is an efficient set of kernels for language model training. It increases throughput by 20% and decreases memory usage by almost 60% for the popular Llama models. Importantly for practitioners, it only requires 3 lines of code change and integrates with many popular modeling frameworks.

Reach software developers, AI enthusiasts, executives, and other tech professionals reading TLDR every day. TLDR offers 10 interest-based newsletters to help you get in front of your target audience. Learn more about running your first campaign with us.

Generalized SAM (GSAM) is a new method for fine-tuning the Segment Anything Model (SAM) with variable-size images more efficiently.

Traditional base language models are quite good at playing chess, averaging around 1800 elo. However, chat models tend to struggle with a precipitous drop in performance. This post explores how to tune chat models like GPT-4o to play chess through a combination of prompting and fine-tuning.

A vulnerability in Slack AI was discovered where attackers can exfiltrate data from private channels they aren't members of via indirect prompt injection. It allows attackers to use public channel messages to manipulate the LLM into divulging private information, such as API keys, in query responses. This issue, alongside a phishing attack vector, persists even after Slack AI's August 14th update incorporating channel and DM files, significantly expanding the risk surface area for such exploits.

GaussianOcc is a method for fully self-supervised and efficient 3D occupancy estimation using surround views.

Meta has introduced Llama 3.1, an advanced open-source LLM that brings new capabilities like synthetic data generation and model distillation.

Google has released a new SigLIP model that uses a shape optimized vision transformer model architecture.