Large Language Models

Today's Stories

1. 1

   Cisco to Give 90,000 Employees AI Agents Starting This July

   Cisco will provide each of its 90,000 employees with a personalized AI assistant beginning in its new fiscal year at the end of July. The system automatically selects the right AI model for each task and is built largely on-site infrastructure. CFO Mark Patterson describes AI as the most significant technology transition in his lifetime. Within Cisco's finance function alone, AI now produces 80% to 90% of the first draft of mandatory financial disclosures, and the company has built AI tools to speed investor relations and forecast business performance—suggesting the rollout could meaningfully reduce routine work across the organization.

   Cisco did not reveal the cost of implementing the AI agents. The company's stock is up about 52% so far this year.

2. 2

   Online Travel Agencies Face New AI Agent Competition

   Online travel agencies (OTAs) are recognizing that artificial intelligence agents—autonomous software that books travel on behalf of users—pose a competitive threat to their traditional business model. The industry is now scrambling to establish trusted relationships with these AI systems to remain relevant in travel bookings. If AI agents become the primary way travelers search for and book trips, OTAs that fail to win the trust of these systems risk being bypassed entirely. This represents a shift from competing for human traveler loyalty to competing for integration into the AI tools travelers actually use.

   The outcome will depend on whether OTAs can position themselves as preferred booking partners within AI agent workflows, or whether they will be replaced as intermediaries in the travel purchase journey.

3. 3

   DeepSeek R1 becomes #2 open-weights reasoning model

   DeepSeek R1 has reached the #2 position among open-weights reasoning models, according to recent rankings. The model is now serving as a strong alternative in the open-source AI landscape. Open-weights reasoning models give developers and organizations access to powerful AI without vendor lock-in or licensing costs. DeepSeek R1's rise signals that capable reasoning capabilities are becoming available beyond closed proprietary systems, which may lower barriers for businesses and researchers building AI applications.

   The competitive positioning of reasoning models continues to shift as open-source alternatives mature. Developers evaluating reasoning AI should monitor how R1 performs on real-world tasks compared to other leading models in this category.

4. 4

   [May 2026 Malware Report] The Mechanism and Countermeasures of LLMShare That Exploits Legitimate Generative AI Sharing Links | Cybersecurity Information Bureau

   [May 2026 Malware Report] The Mechanism and Countermeasures of LLMShare That Exploits Legitimate Generative AI Sharing Links | Cybersecurity Information Bureau

5. 5

   Apple researchers propose efficiency boost for diffusion language models

   Apple researchers published a technical paper describing Residual Context Diffusion (RCD), a method that recycles information from tokens discarded during the decoding process of diffusion language models (a type of AI that generates text in parallel rather than one token at a time). The technique improved accuracy by 5–10 points across benchmarks and reduced computational steps by up to 4–5x on challenging math problems, requiring only ∼1 billion tokens to convert existing models. Diffusion language models promise faster inference than traditional autoregressive models, but current designs waste computation by discarding tokens that still contain useful context. RCD recovers that wasted computation efficiently, which could help make these alternative language models more practical for real-world deployment without adding significant overhead.

   On the most difficult AIME math tasks, RCD nearly doubled baseline accuracy. The method uses a two-stage training pipeline designed to avoid memory bottlenecks, suggesting it may be applicable across a wide range of existing diffusion models.

6. 6

   Apple researchers propose learned policies for diffusion language model sampling

   Apple researchers published work on training sampling procedures for diffusion language models (dLLMs) using reinforcement learning. Instead of relying on manual heuristics like confidence thresholding, they developed a lightweight policy based on a single-layer transformer that decides which tokens to unmask at each step. dLLMs promise efficiency gains during inference by decoding multiple tokens in parallel, but their sampling strategy—which tokens to reveal—has relied on hand-tuned heuristics that require manual adjustment and degrade with larger block sizes. The trained policies match state-of-the-art heuristics in block-wise generation and outperform them in full-diffusion settings, offering a more automated and potentially more scalable approach.

   The work demonstrates that recycling computation from discarded tokens is beneficial, and the researchers note that dLLMs' global planning and iterative refinement features are particularly useful for code generation—a domain where decoding behavior remains under-explored.