Search the archive

NOUSE HERMES 3

Nous Research has released its flagship model. Trained on top of Llama 3, the model has strong performance and a great personality like many of the company's original models.

CALIFORNIA AI BILL SB 1047 AIMS TO PREVENT AI DISASTERS, BUT SILICON VALLEY WARNS IT WILL CAUSE ONE

California's SB 1047, designed to prevent "critical harms" from large AI models, faces pushback from Silicon Valley, with stakeholders divided over its potential impact on innovation. Major industry figures and companies debate the bill's merits and consequences on AI development and safety. The bill, which requires AI model safety protocols and third-party audits, with enforcement and hefty fines outlined for non-compliance, is heading to a final Senate vote.

SOFTBANK'S INTEL AI PROCESSOR PLANS IN DOUBT AS INSIDERS SAY IT IS NOW CONSIDERING A TSMC PARTNERSHIP

Intel failed to produce AI processors for SoftBank's Project Izanagi, leading SoftBank to explore a partnership with TSMC. Despite setbacks, SoftBank remains committed to challenging major AI players with its own hardware and data center ecosystem, potentially backed by significant investment from global partners. The move could strain SoftBank's relationship with Arm clients as it risks direct competition.

ANCHORED PREFERENCE OPTIMIZATION

Tuning models to align with human preferences usually requires some post training. However, when training these models it is not clear why one example is better than another. APO is a way for models to degrade existing examples and then use that degraded example to anchor the difference in preference.

MUTUAL REASONING MAKES SMALL MODELS SMARTER

Tree search is an extremely active area of research for inference time computation for language models. This paper from Microsoft shows a particularly compelling case that enables small models to improve dramatically on mathematics problems.

SEMANTIC SEGMENTATION WITH METASEG

MetaSeg is a powerful semantic segmentation network built on the MetaFormer architecture that enhances both the backbone and decoder of the network.

⭐ Star the repo

LONG CONTEXT RAG

This post explores how long context models perform on various RAG tasks. Improving the number of examples can help. These models often fail in strange but predictable ways.

UNIVERSAL WAVEFORM GENERATION

PeriodWave is a new model for universal waveform generation that excels in capturing the natural periodic features of high-resolution waveform signals.

SECURITY RISKS IN MODEL MERGING

As Model Merging (MM) becomes a popular method for combining fine-tuned models without additional training, new security risks emerge. This paper introduces BadMerging, the first backdoor attack specifically targeting MM.

A COMPREHENSIVE SURVEY ON MODEL MERGING TECHNIQUES

This survey provides an in-depth review of model merging techniques, an increasingly popular method in machine learning that doesn't require raw training data or expensive computation.

BRANDS SHOULD AVOID “AI”

A recent study found that labeling products as "AI-powered" can decrease consumer intent to purchase due to trust issues and fears of the unknown. The research suggests that both cognitive and emotional trust are factors, with people wary of AI's inner workings and risks, especially regarding personal data privacy. Companies are advised to focus on clear messaging about AI benefits rather than using "AI" as a buzzword.

TRANSFORMER EXPLAINER

Beautiful visualization of the inner workings of a Transformer.

ANOTHER APPLE SMART RING PATENT, INCLUDES CONTROL OF SMART GLASSES

Apple's newly granted patent details a smart ring capable of health monitoring and controlling other Apple devices, hinting at possible integration with AR/VR headsets and smart glasses.

14% OF PCS SHIPPED GLOBALLY IN Q2 2024 WERE AI-CAPABLE

There was a significant increase in AI-capable PC shipments in Q2 2024 with 8.8 million units, representing 14% of total PCs shipped.

CONVERSATIONAL REPLICAS BY TAVUS (PRODUCT)

With CVI, developers can build live video conversations with digital twins that feel real.

Image

* Use lightweight container images to minimize the AI attack surface * Deploy AI frameworks like PyTorch with 0 CVEs from day one

Join the Learning Labs live (08/27)

GROK 2

xAI has released its newest model, Grok 2, a frontier class model capable of reasoning, code, and mathematics. It is collaborating with Black Forest Labs to bring FLUX to X users.

PROMPT CACHING WITH CLAUDE

Anthropic has introduced prompt caching for its Claude models, allowing developers to cache frequently used context, significantly reducing costs and latency, with early users like Notion already benefiting from faster and more efficient AI-powered features.

PRUNE AND DISTILLING LLAMA 3.1 8B INTO A 4B MODEL

Nvidia has released its Llama 3.1 minitron 4B model. The model scored 16% better on MMLU compared with training from scratch by using knowledge distillation and pruning and required 40x fewer tokens.

CAUSAL REASONING IN LLMS

Causal Agent is an agent framework equipped with tools, memory, and reasoning modules to handle causal problems.

RECTIFYING BIAS IN CLIP

Biases in CLIP can hinder its performance in unsupervised semantic segmentation, a task where images lack annotations. This paper proposes a method that explicitly models and corrects these biases.

→

PRACTITIONER'S GUIDE TO RUNNING LARGE GPU CLUSTERS

Together AI has put together a great guide for operating large clusters of computers for generative AI model training.

ULTRA-LONG TEXT GENERATION IN LLMS

AgentWrite is a method that breaks down lengthy tasks into smaller parts, allowing models to produce coherent outputs exceeding 20,000 words.

AI-POWERED SCIENTIFIC DISCOVERY

OpenResearcher is a new AI-powered platform that aims to answer a wide range of research-related questions.

AMERICA & CHINA'S CHIP RACE

The U.S. is pursuing aggressive measures to strengthen domestic semiconductor production through the CHIPS Act and sanctions aimed at crippling China's tech advancement. Despite these efforts, China's semiconductor buildout is surging with near-record imports of manufacturing equipment and growing domestic chip production. This intensifying race suggests a continuous geopolitical tug-of-war over semiconductor supply chain dominance.

OPENAI UPDATES CHATGPT TO NEW GPT-4O MODEL BASED ON USER FEEDBACK

OpenAI quietly released an updated GPT-4o model for ChatGPT, with improvements based on user feedback, but not introducing a new reasoning style. Users speculated enhanced multi-step reasoning and image generation capabilities, although OpenAI clarified there's no fundamental change in how the model reasons. The company also noted that the latest ChatGPT variant may differ from the API version to optimize developer experiences.

14 NEW THINGS YOU CAN DO WITH PIXEL THANKS TO AI

MIT RELEASES COMPREHENSIVE DATABASE OF AI RISKS

MIT and other institutions have launched the AI Risk Repository, a comprehensive database of over 700 documented AI risks, to help organizations and researchers assess and mitigate evolving AI risks using a two-dimensional classification system and regularly updated information.

UNIVERSAL MUSIC AND META ANNOUNCE ‘EXPANDED GLOBAL AGREEMENT' FOR AI, MONETIZATION, AND MORE

Meta and Universal Music Group have renewed their multi-year licensing agreement to expand monetization and evolve creative possibilities for UMG's artists on platforms including Facebook, Instagram, and now WhatsApp, focusing on fair compensation and addressing unauthorized AI content issues.

GAS PIPELINE OPERATORS IN TALKS TO FUEL AI DATA CENTER DEMAND

Pipeline companies like Energy Transfer LP and Williams Companies are in discussions to supply natural gas directly to data centers, as the AI industry's power demands surge.

agent [4].

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

Pretty wild, huh? Here, give it a call yourself [4].

Or TLDR readers can get exclusive access to an even more advanced

version here [5]...

SAKANA LAUNCHES AI SCIENTIST

Sakana, a Japanese AI firm, has released a system that can autonomously perform research by suggesting hypotheses, performing experiments, writing code, and summarizing the results in well-reasoned papers. The company has provided examples of the papers the system wrote along with an open sourced version of the system.

AS ALEXA TURNS 10, AMAZON LOOKS TO GENERATIVE AI

Amazon's Alexa division incurred a $10 billion loss in 2022 and laid off staff, highlighting the unsustainability of its loss leader strategy despite the high household penetration. As enthusiasm for smart assistants like Siri and Google Assistant also wanes, Amazon is banking on generative AI to reinvigorate Alexa's capabilities and user engagement. The company's focus is on enhancing conversational interactions and overcoming the "smart timer" limitation.

REPLIKA CEO EUGENIA KUYDA SAYS THE FUTURE OF AI MIGHT MEAN FRIENDSHIP AND MARRIAGE WITH CHATBOTS

Replika CEO Eugenia Kuyda recently discussed her vision for AI companions in human relationships, highlighting the app's role in offering friendship, therapy, or romance through avatars. With the evolution of LLMs, Replika aims to complement, not replace, human interaction, creating a new category of relationships. Despite controversies, such as temporary restrictions on adult content, the app's intent remains to improve users' emotional well-being. With a user base in the millions and a team of 50-60, Replika is planning a significant relaunch to enhance interactivity and realism in conversations.

SWE-BENCH VERIFIED

OpenAI has introduced a subset of SWE-bench that is easier and more in line with what humans and AI can solve today. It is a good benchmark for validating and working towards before running the entire original benchmark.

SMALL BUT MIGHTY: INTRODUCING ANSWERAI-COLBERT-SMALL

ColBERT is an extremely powerful model for retrieval. This new model has only 33m parameters but achieves amazing performance on a number of benchmarks. This post explores how to train a similar model and what tricks led to strong performance.

OBJECT MASK DISCOVERY FOR SEMANTIC SEGMENTATION

Join August 27th and learn the cutting-edge approaches for:

AI TOOLKIT

A nice collection of scripts and notebooks for AI things. It has a strong focus on image synthesis and tuning.