AskNews

Who's in Zürich for GenAI Zürich? Stop by the Innovation Stage at 10:45 am on April 1st for our talk, "Building a News Sleuth for the Deep Research Paradigm" Or catch me for a coffee outside :)

I had the pleasure of speaking with the indomitable Robert Caulk about all things AI. We talked about deep research with Qdrant, Agents, grounding, data management, and much more. Robert and AskNews are launching DeepNews. DeepNews is the result of years of open-source work and applied research—spanning entity extraction, graph construction, geo-coding, multilingual narrative tracking, visualization, bias mitigation, and more. Check it out at Newsplunker.com Full video on Qdrant's youtube now! https://lnkd.in/gXe55u5B

Let's. Go.

Do you want good answers? Or do you want the *best* answers? High risk decision making demands the highest quality data, and that is what we prioritize. The Metaculus forecasting tournament is a real-time benchmark of LLMs, data accessibility, and data quality. It covers political forecasts, financial forecasts, technology forecasts, and much more. The questions are live, the answers are unknown until resolved by real-time news. In other words - this benchmark is as real as it gets. This quarter, the Metaculus team created the same bot with OpenAI's o1, except replaced the data source for each bot, using AskNews/Qdrant, Perplexity, and Exa. So, if you are building a serious decision making assistant, would you rather have the 36th place decision maker, 9th place decision maker, or the 1st place decision maker at your finger tips? AskNews sits on an API, ready to infuse any application with high-quality, high-accuracy, news. How does AskNews hold 1st place like this every quarter? It all boils down to our dedication to quality, accuracy, latency, and scalability. We have state of the art news processing research, entity extraction, and knowledge graph construction, all built on top of the most performant and scalable vector database on the planet - Qdrant. Wanna find out more about how we distribute this data source to LLMs? Join us at the GenAI Zürich conference on April 1st, for our talk titled "Building a News Sleuth for the Deep Research Paradigm".

OpenAI wants to ban DeepSeek AI, but are they protecting national security—or just their monopoly? And what's the potential impact on newsrooms? ?? That's the topic of today's featured story on TVNewsCheck. ?? ARTICLE LINK: https://lnkd.in/eSTpUSsX OpenAI is urging 18 allied nations to ban Chinese ???? AI startup DeepSeek, citing security risks ??. Critics, however, argue this could be less about safety and more about stifling a rising competitor that is offering cheaper, open-source models. ?? The OpenAI vs DeepSeek Conflict in 3 Quick Bullet Points: ? ? OpenAI’s claim: DeepSeek is “state-controlled,” making its AI models (like the viral R1 reasoning engine) a security threat. ? DeepSeek’s defense: Their open-source models are transparent, auditable, and often hosted on Western infrastructure (AWS, HuggingFace), reducing data risks. ? Media industry impact: Broadcasters and newsrooms increasingly rely on affordable AI for content creation, analytics, and audience engagement. A ban could limit innovative cost-effective tools. Where do you stand? Is OpenAI’s proposal a legitimate security safeguard—or a bid to crush competition? How should media orgs balance AI innovation with risk? ?? Shoutouts to Joshua Brandau (Nota), Robert Caulk (AskNews) and Michael Newman (Graham Media Group) for their insights. Also thanks to "KO" for his "deep background" for this story. #aiinnovation #airisks #AI #MediaTech #NationalSecurity #OpenSource #Innovation #ArtificialIntelligence #FutureOfMedia #TVnewscheck #ainews #deepseek #openAI #openaivsdeepseek

If you haven't seen Columbia Journalism Review's analysis of AI search and how bad it is at finding specific articles, it's worth a look, but it has a major flaw: The queries it used aren't representative of how people actually use search engines. Basically, they fed excerpts of articles into AI services like Perplexity and Microsoft Copilot to find those articles and summarize them. Still, it was suprising just how terrible a job most of the AI search engines did. Despite the nature of the queries, there are some good takeaways from this exercise: 1?? AI search needs to be a system:?The study demonstrates hallucinations seriously degrade AI search outputs at a scale that's significant. There's a meta level to the task that asks, "Did I get this right?" that's obviously not happening, so it's put on the human. That's not a good experience. AI search would benefit from the approach OpenAI is taking with GPT-5, which is going to simply pick the right model for you. In the same way, an AI search engine should be able to kick a task (or parts of a task) to a regular search engine when it makes sense. It also needs a thinking model's reality check. We're gonna need more compute. 2?? There is plenty of room to innovate in AI search, both with the tech and overall approach: Microsoft Copilot, for example, clearly has decided that, in many cases, no answer is better than a wrong answer. AskNews, for example?surfaces contradictions among coverage rather than trying to synthesize them incorrectly. 3?? We need a better understanding of AI blocking:?The study showed that articles from sites that were explicitly blocking specific AI search engines still appeared on those search engines. There are myriad reasons for this: syndication, reblogs, 3rd-party crawlers, and more. But the fact that it's happening indicates that AI companies and media companies are playing with two different sets of rules Finally, disclaimers like ""ChatGPT can make mistakes," make the whole enterprise of generative AI feel like a long beta, with everyone waiting for the day they can be taken off and ChatGPT, generally,?*won't*?make mistakes. I don't know if that aligns with the current definition of AGI, but the CJR study is the latest indicator that hallucinations will continue to be a serious issue in AI for a long time.

?? Optimizing RAG with Qdrant, Ollama, and Neo4j Retrieval works best when it balances speed, relevance, and structure. This project integrates Qdrant for high-performance vector search, Neo4j for entity-aware reasoning, and Ollama for LLM-driven insights, creating a retrieval system that efficiently surfaces information while maintaining context. Vector search finds relevant data, graph traversal refines relationships, and LLMs bridge structured and unstructured knowledge. Check out this GraphRAG by project Athos G: https://lnkd.in/ezSNtDfj

Come to GenAI Zürich to watch our C*O, Robert Caulk, pull the curtain back on how Qdrant is powering the new paradigm of investigative news reasoning ??.

“Our customers lost money because of this issue” People often ask why we focus the majority of our data processing pipeline on open source models. This is why. OpenAI screwed up big time yesterday - gpt4o stopped following instructions, even on a version pinned model! Yikes! OpenAI has yet to admit to anything or release a statement. Stable open weights never looked so appealing. I hope the people in this thread (link in comments ????) learned their lesson - OpenAI is not reliable, they care more about milking cash out of their infra than maintaining a stable product for your business!

Do you think you will need a license to deploy AI in the future? Full video with Robert Caulk on the way... Qdrant AskNews