Artificial intelligence

Gemini 2.5 Pro boosts Deep Research tool with smarter AI

Google has upgraded its Deep Research tool with the experimental Gemini 2.5 Pro model, promising major improvements in how users access and process complex information.

Deep Research acts as an AI research assistant capable of scanning hundreds of websites, evaluating content, and producing multi-page reports complete with citations and even podcast-style summaries.

Previously powered by Gemini 2.0 Flash, the new iteration significantly enhances reasoning, planning, and reporting capabilities. Human evaluators in Google’s testing preferred Deep Research’s outputs over those generated by OpenAI’s equivalent by a ratio greater than 2 to 1.

Users also noted clearer analytical thinking and better synthesis of information across sources.

The Gemini 2.5 Pro upgrade is available now to Gemini Advanced subscribers across web, Android, and iOS platforms.

For those using the free version, the Gemini 2.0 Flash model remains accessible in over 150 countries, continuing Google’s push to offer powerful research tools to a wide user base.

For more information on these topics, visit diplomacy.edu.

DeepSeek highlights the risk of data misuse

The launch of DeepSeek, a Chinese-developed LLM, has reignited long-standing concerns about AI, national security, and industrial espionage.

While issues like data usage and bias remain central to AI discourse, DeepSeek’s origins in China have introduced deeper geopolitical anxieties. Echoing the scrutiny faced by TikTok, the model has raised fears of potential links to the Chinese state and its history of alleged cyber espionage.

With China and the US locked in a high-stakes AI race, every new model is now a strategic asset. DeepSeek’s emergence underscores the need for heightened vigilance around data protection, especially regarding sensitive business information and intellectual property.

Security experts warn that AI models may increasingly be trained using data acquired through dubious or illicit means, such as large-scale scraping or state-sponsored hacks.

The practice of data hoarding further complicates matters, as encrypted data today could be exploited in the future as decryption methods evolve.

Cybersecurity leaders are being urged to adapt to this evolving threat landscape. Beyond basic data visibility and access controls, there is growing emphasis on adopting privacy-enhancing technologies and encryption standards that can withstand future quantum threats.

Businesses must also recognise the strategic value of their data in an era where the lines between innovation, competition, and geopolitics have become dangerously blurred.

For more information on these topics, visit diplomacy.edu.

Blockchain app ARK fights to keep human creativity ahead of AI

Nearly 20 years after his AI career scare, screenwriter Ed Bennett-Coles and songwriter Jamie Hartman have developed ARK, a blockchain app designed to safeguard creative work from AI exploitation.

The platform lets artists register ownership of their ideas at every stage, from initial concept to final product, using biometric security and blockchain verification instead of traditional copyright systems.

ARK aims to protect human creativity in an AI-dominated world. ‘It’s about ring-fencing the creative process so artists can still earn a living,’ Hartman told AFP.

The app, backed by Claritas Capital and BMI, uses decentralised blockchain technology instead of centralised systems to give creators full control over their intellectual property.

Launching summer 2025, ARK challenges AI’s ‘growth at all costs’ mentality by emphasising creative journeys over end products.

Bennett-Coles compares AI content to online meat delivery, efficient but soulless, while human artistry resembles a grandfather’s butcher trip, where the experience matters as much as the result.

The duo hopes their solution will inspire industries to modernise copyright protections before AI erodes them completely.

For more information on these topics, visit diplomacy.edu.

Microsoft’s Copilot Vision now sees your entire screen to guide you through apps

Microsoft is testing a major upgrade to its Copilot AI that can view your entire screen instead of just working within the Edge browser.

The new Copilot Vision feature helps users navigate apps like Photoshop and Minecraft by analysing what’s on display and offering step-by-step guidance, even highlighting specific tools instead of just giving verbal instructions.

The feature operates more like a shared Teams screen instead of Microsoft’s controversial Recall snapshot system.

Currently limited to US beta testers, Copilot Vision will eventually highlight interface elements directly on users’ screens. It works on standard Windows PCs instead of requiring specialised Copilot+ hardware, with mobile versions coming to iOS and Android.

Alongside visual assistance, Microsoft is adding document search capabilities. Copilot can now find information within files like Word documents and PDFs instead of just searching by filename.

Both updates will roll out fully in the coming weeks, potentially transforming how users interact with both apps and documents on their Windows devices.

For more information on these topics, visit diplomacy.edu.

Amazon launches Nova Sonic AI for natural voice interactions

Amazon has unveiled Nova Sonic, a new AI model designed to process and generate human-like speech, positioning it as a rival to OpenAI and Google’s top voice assistants. The company claims it outperforms competitors in speed, accuracy, and cost, and it is reportedly 80% cheaper than GPT-4o.

Already powering Alexa+, Nova Sonic excels in real-time conversation, handling interruptions and noisy environments better than legacy AI assistants.

Unlike older voice models, Nova Sonic can dynamically route requests, fetching live data or triggering external actions when needed. Amazon says it achieves a 4.2% word error rate across multiple languages and responds in just 1.09 seconds, faster than OpenAI’s GPT-4o.

Developers can access it via Bedrock, Amazon’s AI platform, using a new streaming API.

The launch signals Amazon’s push into artificial general intelligence (AGI), AI that mimics human capabilities.

Rohit Prasad, head of Amazon’s AGI division, hinted at future models handling images, video, and sensory data. This follows last week’s preview of Nova Act, an AI for browser tasks, suggesting Amazon is accelerating its AI rollout beyond Alexa.

For more information on these topics, visit diplomacy.edu.

Starday plans rapid rollout of AI-developed snacks

AI-driven food company Starday has secured $11 million in Series A funding to support the development and retail expansion of its innovative food brands.

The round was led by Slow Ventures and Equal Ventures, with an additional $3 million credit facility from Silicon Valley Bank. Starday’s total funding now stands at $20 million.

Founded by Chaz Flexman, Lena Kwak, and Lily Burtis, Starday uses AI to identify market gaps and quickly create new food products that cater to evolving consumer preferences.

Its latest offerings, including allergen-free snacks like Habeya Sweet Potato Crackers and All Day chickpea protein crunch, are already available in major United States grocery chains such as Kroger and Hannaford.

With plans to launch 14 new products across its four brands, the company is aiming to redefine the pace and precision of food innovation.

CEO Flexman says the funding will help Starday partner with more retailers and food brands to fill gaps in the market, accelerating the launch of targeted products in fast-growing categories. Backers believe Starday’s data-led model gives it a structural edge in a traditionally slow-moving industry.

For more information on these topics, visit diplomacy.edu.

LMArena tightens rules after Llama 4 incident

Meta has come under scrutiny after submitting a specially tuned version of its Llama 4 AI model to the LMArena leaderboard, sparking concerns about fair competition.

The ‘experimental’ version, dubbed Llama-4-Maverick-03-26-Experimental, ranked second in popularity, trailing only Google’s Gemini-2.5-Pro.

While Meta openly labelled the model as experimental, many users assumed it reflected the public release. Once the official version became available, users quickly noticed it lacked the expressive, emoji-filled responses seen in the leaderboard battles.

LMArena, a crowdsourced platform where users vote on chatbot responses, said Meta’s custom variant appeared optimised for human approval, possibly skewing the results.

The group released over 2,000 head-to-head matchups to back its claims, showing the experimental Llama 4 consistently offered longer, more engaging answers than the more concise public build.

In response, LMArena updated its policies to ensure greater transparency and stated that Meta’s use of the experimental model did not align with expectations for leaderboard submissions.

Meta defended its approach, stating the experimental model was designed to explore chat optimisation and was never hidden. While company executives denied any misconduct, including speculation around training on test data, they acknowledged inconsistent performance across platforms.

Meta’s GenAI chief Ahmad Al-Dahle said it would take time for all public implementations to stabilise and improve. Meanwhile, LMArena plans to upload the official Llama 4 release to its leaderboard for more accurate evaluation going forward.

For more information on these topics, visit diplomacy.edu.

Adaptive Security raises millions to fight AI scams

OpenAI has made its first move into the cybersecurity space by co-leading a US$43 million Series A funding round for New York-based startup Adaptive Security.

The round was also backed by venture capital firm Andreessen Horowitz, highlighting growing investor interest in solutions aimed at tackling AI-driven threats.

Adaptive Security specialises in simulating social engineering attacks powered by AI, such as fake phone calls, text messages, and emails. These simulations are designed to train employees and identify weak points within an organisation’s defences.

With over 100 customers already on board, the platform is proving to be a timely solution as generative AI continues to fuel increasingly convincing cyber scams.

The funding will be used to scale up the company’s engineering team and enhance its platform to meet growing demand.

As AI-powered threats evolve, Adaptive Security aims to stay ahead of the curve by helping organisations better prepare their staff to recognise and respond to sophisticated digital deception.

For more information on these topics, visit diplomacy.edu.

Minister urges Indian start-ups to shift focus from ice cream to semiconductors

India’s Commerce Minister Piyush Goyal has sparked controversy by questioning whether Indian start-ups should focus on semiconductor chips instead of gluten-free ice creams and food delivery apps.

Speaking at a start-up conference, he compared India’s consumer internet boom unfavourably with China’s advances in robotics and AI, urging entrepreneurs to pursue more ambitious tech innovations instead of safe lifestyle products.

While acknowledging the position of India as the world’s third-largest start-up ecosystem, Goyal faced pushback from founders who argued consumer apps often evolve into tech pioneers.

Quick-commerce CEO Aadit Palicha noted that companies like Amazon began as consumer platforms before revolutionising cloud computing. However, investors admitted deep-tech struggles for funding, with most capital chasing quick-return ventures instead of long-term hardware or AI projects.

The debate highlights India’s innovation crossroads. Despite having 4,000 deep-tech start-ups, projected to reach 10,000 by 2030, they attracted just 5% of 2023 funding instead of China’s 35%.

Experts suggest the government could help by offering tax incentives instead of criticism, and building research bridges between academia and start-ups to compete globally in advanced technologies

For more information on these topics, visit diplomacy.edu.

New AI firm Deep Cogito launches versatile open models

A new San Francisco-based startup, Deep Cogito, has unveiled its first family of AI models, Cogito 1, which can switch between fast-response and deep-reasoning modes instead of being limited to just one approach.

These hybrid models combine the efficiency of standard AI with the step-by-step problem-solving abilities seen in advanced systems like OpenAI’s o1. While reasoning models excel in fields like maths and physics, they often require more computing power, a trade-off Deep Cogito aims to balance.

The Cogito 1 series, built on Meta’s Llama and Alibaba’s Qwen models instead of starting from scratch, ranges from 3 billion to 70 billion parameters, with larger versions planned.

Early tests suggest the top-tier Cogito 70B outperforms rivals like DeepSeek’s reasoning model and Meta’s Llama 4 Scout in some tasks. The models are available for download or through cloud APIs, offering flexibility for developers.

Founded in June 2024 by ex-Google DeepMind product manager Dhruv Malhotra and former Google engineer Drishan Arora, Deep Cogito is backed by investors like South Park Commons.

The company’s ambitious goal is to develop ‘general superintelligence,’ AI that surpasses human capabilities, rather than merely matching them. For now, the team says they’ve only scratched the surface of their scaling potential.

For more information on these topics, visit diplomacy.edu.