This year's hottest gift for businesses is BILL

Beat the Black Friday rush and get your business what it really wants for the holidays—streamlined financial operations with BILL Spend & Expense.

Choose BILL to help your business:

• Control spending with customizable budget controls.

• Get real-time visibility into your finances.

• Automate expense reports to save time and focus on strategy.

Take a demo and get what YOU really want for the holidays—choose from a Nintendo Switch, Apple AirPods Pro, Samsung 50" TV, or Xbox Series S¹.

Request a demo

¹Terms and Conditions apply. See offer page for more details.
BILL Divvy Card is issued by Cross River Bank, Member FDIC, and is not a deposit product.

Today:

• OpenAI Swarm Framework Sparks Debate

• AI Struggles with Math Reasoning

• Intel Partners for AI Hardware

• Khosla Raises $405M for OpenAI

• New AI Scam Targets Gmail

OpenAI Swarm Framework Sparks Debate

OpenAI’s new experimental framework, Swarm, allows developers to create networks of AI agents that collaborate to tackle complex tasks. While the technology is not yet an official product, it provides a blueprint for developing multi-agent systems in enterprise settings. Swarm could lead to highly automated business operations, raising concerns about job displacement, security risks, and bias. 

The framework has sparked debates on ethical and societal impacts as AI becomes more integrated into workplaces. Despite these challenges, Swarm gives businesses a glimpse into the future of AI-driven automation and decision-making.

AI Struggles with Math Reasoning

A recent study by Apple researchers reveals that large language models (LLMs) struggle with math problems when trivial changes are introduced. While LLMs can solve basic arithmetic, they falter when faced with irrelevant details, like small kiwis in a math problem. This suggests that LLMs don’t genuinely understand problems but instead replicate patterns seen in their training data. 

The paper raises concerns about AI’s reasoning abilities, as models fail when complex reasoning or distractions are involved. The research underscores the limitations of LLMs and questions their true capabilities in tasks that require logical thinking.

Intel Partners for AI Hardware

Intel Corporation (INTC) is partnering with Inflection AI to develop new hardware aimed at supporting large language models (LLMs). This collaboration comes amid a surge in artificial intelligence (AI) innovations, with companies racing to monetize their AI technologies. 

Investors are closely watching how major firms are integrating AI to drive profits, especially in the tech and communication sectors, which are expected to show the strongest growth. 

Khosla Raises $405M for OpenAI

Khosla Ventures has raised $405 million for OpenAI through a special purpose vehicle (SPV), which pools capital from other investors. Although the filing suggests Khosla’s stake in OpenAI is around 6% of the recent $6.6 billion funding round, it’s unclear how much, if any, of Khosla’s own capital was used. 

The firm has a history with OpenAI, having invested $50 million in 2019. OpenAI’s valuation has surged to $157 billion, up from around $29 billion in 2023 when Microsoft made a significant investment. Khosla Ventures declined to comment on the specifics of the recent funding.

New AI Scam Targets Gmail

A new AI-driven scam is targeting Gmail users, leveraging sophisticated techniques that can deceive even experienced professionals. Hackers use fake account recovery notifications and follow-up calls, posing as Google support to steal credentials. These scams often employ AI-generated voices and legitimate-looking documents, such as Google Forms, to create trust and trick victims. 

Google has responded by launching the Global Signal Exchange, a new platform aimed at fighting scams in real-time by sharing cybercrime data. Additionally, Gmail users are encouraged to enroll in Google’s Advanced Protection Program to safeguard against phishing and account takeovers.

🧠RESEARCH

Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis

Meissonic is a text-to-image model that improves on previous approaches by using non-autoregressive masked image modeling. It achieves high-quality, high-resolution image generation comparable to state-of-the-art models like SDXL. Innovations in architecture, positional encoding, and training data improve both performance and efficiency.

Pixtral 12B

Pixtral-12B is a 12-billion-parameter multimodal language model designed to excel in both text and image understanding. It outperforms larger models in multimodal tasks while maintaining strong natural language capabilities. Pixtral's unique vision encoder processes images at natural resolutions, and it handles long contexts of up to 128K tokens.

MathCoder2: Better Math Reasoning from Continued Pretraining on Model-translated Mathematical Code

MathCoder2 improves mathematical reasoning in large language models by generating mathematical code paired with reasoning steps. Using a 19.2 billion-token dataset, MathCode-Pile, which includes math-related web data and textbooks, the model enhances precision and accuracy. This method significantly boosts the mathematical abilities of various models, with all training and data open-sourced.

WALL-E: World Alignment by Rule Learning Improves World Model-based LLM Agents

WALL-E is a language model-based agent that improves world modeling through rule learning. By aligning an LLM with its environment using a few rules, WALL-E enhances prediction accuracy and efficiency. Tested on Minecraft and ALFWorld, WALL-E achieves higher success rates and lower costs compared to existing methods, improving performance by 15-30%.

Agent S: An Open Agentic Framework that Uses Computers Like a Human

Agent S is an open framework designed to automate complex, multi-step computer tasks through graphical interfaces, mimicking human interaction. It tackles challenges like domain-specific knowledge and dynamic interfaces using experience-augmented hierarchical planning. Agent S outperforms previous models by 9.37% in success rate and shows broad generalizability across operating systems. Code is open-sourced.

🛠️TOP TOOLS

Latitude - Open-source prompt engineering platform to build, evaluate, and refine your prompts with AI

Fyli - Create AI agents to automatically find your ideal customers online, post engaging content, and get more high intent leads who are ready to buy your product

Volamail - Write email templates with the help of AI and send them with an API call

Height.app - Autonomous project management tool

Landing Pro AI - Website copy that sells, built in minutes

📲SOCIAL MEDIA

🗞️MORE NEWS

• AI21 CEO Ari Goshen argues that transformer models aren't ideal for AI agents due to their high error rates and costs. He promotes alternative architectures like Mamba and Jamba, which offer better efficiency and affordability for enterprise AI.

• Apple's iOS 18.1 beta, featuring AI-powered text summaries, unexpectedly summarized a breakup message for user Nick Spreen. The feature, aimed at improving productivity, sparked concern about its emotional sensitivity, prompting hopes for future adjustments.

• Chinese AI firm Moonshot AI updated its chatbot, Kimi Chat Explore, enhancing problem-solving and multi-step task capabilities, similar to OpenAI's o1 model. The chatbot can now analyze 500 web pages per query and handle complex tasks efficiently.

• Researchers used AI to identify 70,500 previously unknown RNA viruses, many found in extreme environments like hydrothermal vents. This breakthrough highlights AI's potential in discovering viruses that may explain mystery illnesses.