Large Language Models

Today's Stories

1. 1

   Priceline integrates Anthropic's Claude AI into its travel assistant Penny, enabling users to book trips in a single conversation and cutting planning time by an average of ten minutes.

   On June 3, Priceline—a subsidiary of Booking Holdings—upgraded its AI travel assistant Penny by integrating Anthropic's Claude. The new system works as a coordinated set of specialized agents that guide users from initial travel ideas through to final booking in one continuous conversation, replacing traditional search filters with an interactive map-based experience. Travel planning typically involves juggling multiple trade-offs and searches. By combining Claude's reasoning capability with Priceline's travel data and booking inventory, the tool aims to simplify that complexity and increase the likelihood that browsers actually complete a purchase. Early testing shows users save an average of ten minutes per trip, suggesting the integration meaningfully streamlines the decision-making process.

   The system uses a preference layer that learns from traveler history and trip goals (budget, location, purpose) to deliver personalized recommendations, including features called "Penny's Pick" and "Penny's Take" (in beta), which help users evaluate trade-offs using real-time pricing and inventory data.

2. 2

   Interactive Brokers has integrated ChatGPT and Grok into its trading platform alongside existing Claude support, and expanded the types of assets clients can trade.

   Interactive Brokers added ChatGPT and Grok to its AI integrations, which already included Claude. The broker has also expanded the range of supported asset classes to include options and futures trading. The integrations let clients connect their Interactive Brokers accounts to these AI platforms to research investments, analyse portfolios, and generate trading instructions. This broadens the tools available to traders for decision-making.

   The expansion now covers options and futures trading alongside the existing capabilities, giving clients access to multiple AI assistants for research and portfolio analysis.

3. 3

   Anthropic launches Claude Tag, a Slack-based AI assistant designed to work like a shared team employee, as the company courts enterprise customers ahead of a likely IPO.

   Anthropic released Claude Tag, a new version of its Claude chatbot that operates within Slack and breaks down tasks into stages to complete them independently. The tool lets all employees in a company collaborate with the same Claude identity and hand off partially finished work to each other. Within Anthropic's own product team, Claude Tag is already approving and incorporating 65% of code changes submitted. Large organizations struggle to embed AI into their day-to-day workflows due to data silos, employee training gaps, and concerns about sensitive data leaks. Claude Tag addresses these friction points by embedding AI where teams already work (Slack) and letting administrators tightly control which tools, information, and channels each Claude instance can access—for example, preventing HR data from reaching engineering teams. According to Ramp's May AI Index, Anthropic has just pulled ahead of OpenAI in business adoption for the first time, with 34.4% of firms paying for its services compared to OpenAI's 32.3%, driven largely by its agentic coding tool Claude Code.

   Claude Tag will initially roll out as a research preview on Slack for Claude Enterprise and Claude Team users, with plans to expand access later. Administrators can also set spending limits on tokens both per channel and organization-wide, and the product includes an ambient behavior feature that allows the bot to proactively update employees and follow up on forgotten tasks.

4. 4

   Nearly half of Harvard seniors admit to cheating, but the problem long predates AI—and colleges are only now starting to address it seriously.

   A 2024 Harvard Crimson study found that 47% of 850 surveyed senior students said they had cheated. The article traces cheating prevalence across U.S. high schools (51–95% depending on the study) and notes that reported academic misconduct cases at Ohio State University increased by 57% between 2014 and 2018. Princeton recently abandoned its 133-year-old prohibition against proctoring exams to address academic integrity violations, and Oberlin changed its honor code to allow professors to proctor tests. Cheating is not new—it predates AI and runs much deeper than current technology. The article shows that many college students arrive already accustomed to academic misconduct from high school, and instructors often look the other way rather than report violations. Few colleges have developed an intellectual integrity curriculum to counter cheating over a student's four years, meaning the problem persists institutionally. The issue signals a gap between colleges' stated policies and their enforcement.

   The article suggests that building commitment to academic integrity requires faculty to weave discussions of intellectual integrity throughout their courses and help students think about who they want to be. Without structural change—including clear, severe consequences when cheating is caught and sustained support programs—colleges are unlikely to curb the pervasiveness of the problem.

5. 5

   DeepSeek releases its largest open-source reasoning model, claiming competitive performance with fewer computational resources than rivals.

   DeepSeek introduced DeepSeek-R1, an open-source reasoning model that the company says achieves performance comparable to OpenAI's o1 while using 27% of FLOPs compared with DeepSeek-V3.2 and 79 GiB of memory compared with DeepSeek-V3.2's 83.9 GiB. The release makes a high-capability reasoning model freely available to researchers and developers, rather than locked behind a commercial API. This may shift how organizations build AI applications, since they can now run comparable reasoning performance on their own infrastructure without paying per-query fees to a vendor.

   DeepSeek-R1 is available in open-source form, meaning anyone can download and run it. The model's efficiency claims (lower memory, fewer computational operations) will determine whether smaller organizations can actually deploy it compared to commercial alternatives.

6. 6

   Hallu, a new web framework, lets developers use an LLM to automatically generate application code instead of writing it manually.

   A web framework called Hallu has been created and shared on GitHub that uses an LLM (AI that understands and generates text) to generate application code. The project is available at https://github.com/alehlopeh/hallu and has been posted to Hacker News. This approach flips the traditional development model — instead of developers writing code line by line, they can have an AI generate it, potentially speeding up the creation of web applications. Whether this model is practical at scale remains to be tested by the developer community.

   The project has generated minimal initial engagement (1 point, 0 comments on Hacker News at the time of posting), so adoption and real-world feedback from developers will be the key indicator of whether this method of building apps gains traction.