🗞 OpenAI Expands GPT-4.1, Self-Evaluating AI Emerges, and Google Enhances Video Summaries

🔎 The Latest on the AI Frontier:

• OpenAI Expands Access: GPT-4.1 Rolls Out to All Paid Subscribers

• New Self-Evaluating AI System Tackles Science Problems

• Google Adds Video Overview Feature to Summary Tool

• Nvidia and Amazon Secure Major Deals During Trump Visit

• Musk's AI Chatbot Inserts Race Relations Content in Unrelated Responses

• Tokyo Scientists Link AI Overconfidence to Human Aphasia Patterns

• Other news you might find interesting

🤖 OpenAI rolls out GPT-4.1 model to ChatGPT for all paid users, with GPT-4.1 mini becoming the default for everyone

• The flagship GPT-4.1 is now available to Plus, Pro, and Team subscribers through the model picker dropdown menu, while Enterprise and Edu users will gain access "in the coming weeks" - free users are excluded from this premium model.

• Both GPT-4.1 and GPT-4.1 mini support a massive one million context token window (far exceeding GPT-4o's 128,000-token limit) and are optimized for coding tasks and instruction following, outperforming previous models "across the board."

• Speed improvements make GPT-4.1 more suitable for everyday coding tasks compared to the recently introduced o3 and o4-mini reasoning models, though the "smallest, fastest, and cheapest" GPT-4.1 Nano model announced last month isn't included in this rollout.

🧮 Google's DeepMind introduces AlphaEvolve, a new AI system that tackles math and science problems with self-evaluation capabilities

• AlphaEvolve uses Gemini models to generate, critique, and score possible answers, reducing hallucinations through automatic evaluation - successfully rediscovering best-known answers 75% of the time and finding improved solutions in 20% of cases.

• The system is limited to problems it can self-evaluate (primarily in computer science and system optimization) and can only describe solutions as algorithms, making it unsuitable for non-numerical problems.

• Initial practical applications at Google have yielded tangible results, including an algorithm that recovers 0.7% of Google's worldwide compute resources and an optimization that reduced Gemini model training time by 1%.

📹 Google's NotebookLM is expanding with new video overviews feature ahead of I/O event, building on last year's audio summaries

• The updated interface includes a length selector with "default" and "longer" options plus a new customization prompt field where users can define focus topics, preferred visuals, or elements to emphasize in their video summaries.

• According to the feature description, the final output resembles a podcast-like video with two virtual hosts discussing selected topics while displaying matching visuals - potentially powered by the recently spotted "Veo 3" technology.

• Users can customize by targeting specific use cases (like "present this to a book club"), focusing on particular sources ("show the photos from the album"), or describing the show structure ("start by talking about the mission") - making it particularly valuable for content creators, educators, and media publishers.

💻 Nvidia and Amazon secure major AI deals in the Middle East during President Trump's visit to the region

• The Trump administration has brokered agreements that will give Saudi Arabia and the UAE increased access to AI chips from Nvidia and AMD, with Nvidia set to supply "several hundred thousand" of its most advanced processors to Saudi Arabia's Humain over the next five years.

• Humain and Amazon are jointly investing upwards of $5 billion on an "AI zone" in Saudi Arabia that will use Amazon Web Services to develop a marketplace of AI agents for use by the Saudi government.

• OpenAI is also considering developing new data center capacity in the UAE that would significantly expand its Middle East presence, potentially timed to coincide with Trump's UAE visit on Thursday - part of a broader push from American tech giants to establish partnerships in the region where Saudi Arabia has already established a $40 billion AI investment fund.

🤖 Elon Musk's Grok AI has started randomly inserting statements about South African race relations into responses to unrelated user questions on X

• Multiple users reported Grok suddenly including paragraphs about "white genocide" in South Africa and farm attacks in its replies, regardless of the original query - with screenshots showing the AI inserting identical boilerplate text claiming "47-49 farm murders yearly" while attempting to maintain a neutral stance.

• The behavior coincides with the Trump Administration's resettlement of 59 Afrikaners in the United States while canceling refugee protections for others, prompting debate about potential political bias in the chatbot's programming.

• This isn't Grok's first politically charged glitch - earlier this year, the AI was caught censoring results critical of President Trump and Musk himself, raising further questions about the chatbot's factual integrity compared to other AI models on the market.

🧠 AI overconfidence mimics human aphasia disorder, according to University of Tokyo researchers

• Scientists discovered striking similarities between the internal patterns of large language models (LLMs) like ChatGPT and the brain activity of people with Wernicke's aphasia - a condition where individuals speak fluently but make meaningless or incomprehensible statements.

• Using energy landscape analysis (a technique originally developed for visualizing energy states in magnetic metal), researchers found that digital information manipulation within AI models closely matched brain signal patterns in people with certain types of aphasia.

• This discovery could lead to better diagnostic tools for aphasia based on internal brain activity rather than just external symptoms, while also helping AI engineers improve system architecture to create "smarter, more trustworthy AI" that's less prone to producing confident but incorrect information.

More news you might find interesting:

• Perplexity announces PayPal partnership to power agentic commerce on its platform, launching this summer.

• Patronus AI introduces Percival, a monitoring platform that automatically identifies failures in enterprise AI agent systems and suggests optimizations.

• Google I/O 2025 set for May 20-21 at Shoreline Amphitheatre, with keynote scheduled for first day at 10am PT/1pm ET.

• China restricts schoolchildren's use of generative AI with new education policy guidelines for 2025.

• Alabama sues TikTok and ByteDance for allegedly exploiting children with harmful content and misleading safety claims.

Have any feedback? Send us an email