Hello Tuners,

This week, we examine some significant developments in AI and government contracts. The Department of Defense has awarded its first generative AI defense contract to Jericho Security, marking a significant step forward for military cybersecurity. Meanwhile, Meta and Anthropic are forging partnerships with U.S. government agencies, securing their place in the expanding defense sector.

On the tech front, we also explore the controversy surrounding Perplexity’s CEO Aravind Srinivas’ support for the New York Times during its strike. We also discuss the mixed reception of OpenAI’s SearchGPT, which has struggled to meet the lofty expectations set by its fanfare. The growing influence of AI in government and industry continues to raise important questions about responsibility, innovation, and the future of work.

• Department of Defense Awards First AI Defence Contract

• Content Creation After Content Kleptocracy

• Hey Meta and Anthropic, Will You Invite GPT to the Party?

• SearchGPT, Still Searching for Purpose…

Department of Defense Awards First AI Defence Contract

The Department of Defense has awarded its first generative AI defense contract to Jericho Security, a New York-based cybersecurity startup. Through a $1.8 million STTR Phase II contract with AFWERX, Jericho is set to develop advanced cybersecurity tools for the Department of the Air Force. This partnership marks a significant step for military cybersecurity, focusing on using AI to simulate real-world, multi-channel phishing attacks, from text and phone scams to video deepfakes, that increasingly target military personnel. CEO Sage Wohns explains that their generative AI platform capitalizes on human vulnerability, offering personalized training programs based on individual risk profiles to prepare personnel for complex, evolving threats.

Jericho’s unique “predator and prey” model is designed to adapt to new attack patterns rather than merely reacting to existing ones. The platform gathers real-time data to refine offensive and defensive AI capabilities by continuously simulating and analyzing attacks. This contract is a giant leap for Jericho, allowing the startup to move into the government sector, where cybersecurity threats are rising alongside budgets. Through AFWERX, the Air Force’s innovation arm, the Defense Department aims to accelerate private-sector tech adoption to keep U.S. defenses agile against emerging digital threats, positioning Jericho as a key player in next-generation military cybersecurity.

Content Creation After Content Kleptocracy

Aravind Srinivas, CEO of AI search company Perplexity, stirred controversy by offering support to the New York Times amid a strike by its tech workers. The NYT Tech Guild, representing software support and data analysis employees, called the strike after months of stalled negotiations for a 2.5% wage increase and a clear two-day in-office policy. Srinivas responded on X (formerly Twitter), suggesting Perplexity could help NYT manage high traffic during critical election coverage. This prompted NYT publisher AG Sulzberger to criticize the strike's timing, highlighting the public’s dependence on the Times’s election coverage.

Srinivas’s offer met swift backlash, with critics accusing him of “scabbing” replacing striking workers in a way that could weaken their bargaining power. Responding to inquiries, Srinivas clarified that Perplexity only offered tech infrastructure support, not a replacement for journalism or engineering roles. However, this isn’t the first tension between NYT and Perplexity; last month, NYT issued a cease-and-desist to Perplexity over article scraping for AI models. Though intentions may vary, Srinivas’s offer fueled a delicate labour and tech debate amid high-profile election coverage.

Hey Meta and Anthropic, Will You Invite GPT to the Party?

Meta and Anthropic are expanding their AI offerings to U.S. government agencies, particularly in defense and national security. Meta is rolling out its open-source Llama models through partnerships with companies like AWS, Lockheed Martin, and Palantir, addressing critical needs in aircraft maintenance, national security missions, and business process optimization. Hosted on secure platforms like AWS and Microsoft Azure, Llama aims to solidify the U.S.'s technological leadership while fostering public-private collaborations to drive innovation in education, energy, and small business development. By emphasizing responsible AI deployment, Meta aims to ensure advancements align with international law and ethical standards while countering competition from nations like China in the global AI race.

Meanwhile, Anthropic has partnered with Palantir and AWS to deploy its Claude models in highly secure U.S. defense and intelligence environments. Integrated into Palantir’s IL6 environment, Claude will support sensitive data analysis and critical decision-making. Anthropic’s cautious approach to AI use aligns with growing trends of AI companies targeting government contracts, positioning them as major players in the defense sector. Backed by significant investors like Amazon, Anthropic’s expansion into public-sector services highlights its commitment to responsible AI while navigating the challenges of the government and defense markets, such as concerns over pricing for its latest model, Claude Haiku 3.5.

SearchGPT, Still Searching for Purpose…

After all the hype, OpenAI’s SearchGPT falls short of expectations, especially regarding short, navigational queries like "Nuggets score" or "San Francisco weather." Despite OpenAI positioning the tool as a potential Google killer, the results were often inaccurate or hallucinated, with vague responses and broken links. While it works better for long-form questions, it simply doesn’t provide the quick, reliable answers that users need for everyday searches.

What’s most frustrating is the gap between OpenAI’s grand launch and the reality of the product. The fanfare around ChatGPT Search promised innovation, but the tool still lacks the efficiency and depth required to challenge Google’s dominance. Until OpenAI addresses these issues, ChatGPT Search remains more of a curiosity than a genuine replacement for traditional search engines.

Weekly Research Spotlight 🔍

AFlow: Automating Agentic Workflow Generation

AFlow is making waves by automating the process of workflow optimization. Traditionally, LLMs have relied on manually crafted workflows to tackle complex tasks, which limits scalability and generalizability. AFlow, however, takes a significant step forward by reframing workflow optimization as a search problem over code-represented workflows. Using Monte Carlo Tree Search (MCTS), the framework efficiently explores the space of possible workflows, refining them through iterative code modifications and feedback loops. This approach allows for optimizing workflows with minimal human input, opening the door for more scalable LLM applications.

Empirical tests across six benchmark datasets highlight AFlow's impressive performance, demonstrating a 5.7% improvement over current state-of-the-art methods. In addition to its performance gains, AFlow provides a unique cost advantage, enabling smaller models to outperform GPT-4 on specific tasks at just 4.55% of its inference cost. This combination of enhanced performance and reduced cost makes AFlow a compelling tool for developers leveraging LLMs more efficiently. The release of AFlow’s code promises to be a game-changer for optimizing LLM workflows and reducing the resources needed for complex tasks.

LLM Of The Week

Oasis

Oasis is the first real-time, open-world AI model, enabling players to interact with a dynamically generated game world through keyboard inputs. Unlike traditional games, Oasis doesn't rely on a game engine but instead uses a foundational AI model to generate gameplay, physics, and graphics in real-time. The team has released Oasis's code, including the weights of a 500M parameter model for local use, along with a demo showcasing a larger checkpoint. This breakthrough, powered by Decart's inference engine and Etched's transformer ASIC, demonstrates the potential of generative video and fast transformer inference.

The model understands complex game mechanics such as building, lighting physics, and inventory management. Oasis can generate diverse environments, from dark space-like settings to lively worlds with animals, proving its versatility. By combining a spatial autoencoder and latent diffusion backbone, Oasis ensures stable scaling and fast inference. It also solves temporal stability issues with innovations like dynamic noising, ensuring consistent, high-quality outputs over long time horizons. This research paves the way for new interactive worlds controlled by text, audio, or other modalities.

Best Prompt of the Week 🎨

Today's Goal: Try new things 🧪

Acting as a Content Strategy Planner

Prompt: I want you to act as a content strategy planner. You will create a structured daily plan specifically designed to help an aspiring lifestyle content creator establish a strong presence on social media. You will identify key steps for defining personal brand style, develop strategies for content creation and engagement, select tools for effective social media management, and outline additional activities needed to grow and connect with their audience. My first suggestion request is: "I need help creating a daily activity plan for someone who is planning to become a social media lifestyle content creator."

This Week’s Must-Watch Gem 💎

This Week's Must Read Gem 💎