☕️ Google DeepMind's AlphaEvolve is an AI Agent That Can Invent Brand New Algorithms

Google DeepMind's latest AI agent writes its own algorithms that it then uses to improve itself

Earlier this week, Google's AI research lab DeepMind unveiled what could be one of its most impactful AI-centric releases to date. Dubbed AlphaEvolve, according to Google DeepMind, this AI agent uses an evolutionary approach that allows it to test, refine, and create wholly new computer algorithms. With Google having already deployed this system across its data centers and chip designs, AlphaEvolve is dramatically cutting down on the tech giant's computing costs, all the while improving its own training time.

​

What are the figures, and how exactly does AlphaEvolve work?

AlphaEvolve pairs Google's Gemini models with an evolutionary algorithm engine that iteratively tests and refines complex code. Using Gemini Flash for speed and Gemini Pro for depth, AlphaEvolve generates, evaluates, and evolves solutions using automated feedback loops (called "evaluators") and quality metrics.

This setup lets AlphaEvolve autonomously discover complex yet high-efficiency algorithms, and it has already been seen recovering 0.7% of computing resources in Google's data centers - on Google's scale, that's a staggering efficiency figure. The post additionally claims that AlphaEvolve has cut down on Gemini's overall training time by 1%, made advancements in the frontier of mathematics and science (such as the kissing number problem​), and more.

US tech giants ink multibillion-dollar deals as part of President Trump's $600 billion Saudi deal and trip to the Gulf

​As US President Donald Trump tours the Gulf, tech giants secured multibillion‑dollar AI-related deals in Saudi Arabia, including names like Nvidia and AMD. This will see Nvidia ship an initial 18,000 Blackwell GPUs to Humain, the sovereign fund's new AI startup, as part of plans to deploy "several hundred thousand" of Nvidia tech over five years, while AMD agreed to a $10 billion, 500-megawatt AI infrastructure collaboration. Meanwhile, Saudi firm DataVolt has agreed to invest $20 billion in US AI data centers, while Google, Oracle, Salesforce, AMD, and Uber will commit another $80 billion to joint tech ventures.

​

OpenAI brings its flagship GPT-4.1 model to ChatGPT and makes its mini variant the default for all

​OpenAI recently announced that it will integrate its newest non-reasoning multimodal model, GPT‑4.1, into ChatGPT for all paid account tiers, starting with Plus, Pro, and Team, with Enterprise and Edu users following soon. Simultaneously, GPT‑4o mini has been fully replaced by the more efficient GPT‑4.1 mini across both paid and free accounts as the default option. According to OpenAI, both models excel at coding and instruction-following for everyday use, outperforming GPT‑4o "across the board" and boasting a much larger 1‑million‑token context window, dwarfing their predecessors' 128,000.

​

Meta delays its flagship Llama 4 model release yet again as internal frustration mounts

At a time when Meta is already struggling to keep up with the top, the company delaying its flagship Llama 4 Behemoth multimodel from June to at least fall is another heavy blow, citing that Behemoth isn't making significant enough gains. Although Meta released Llama 4's sister variants, Maverick and Scout, during their debut at LlamaCon, Behemoth has been delayed for a second time. Despite this, Meta claims that the impact on business is limited, as many rely on proprietary cloud models, even as frustration in the company continues to mount over a lack of ROI on the fortune it has poured into its own AI.

​

Google rolls out a slew of new AI and accessibility features to Android and Chrome

​In a recent blog post, Google has rolled out several new AI-powered accessibility upgrades for Android and Chrome. Android's screen reader, TalkBack, now adds more Gemini power to allow users to ask about the images on their screens. Expressive Captions also got an update that enables it to reflect vocal inflections in real time, such as when someone stretches out "no" like "noooo" when expressing sarcasm. On top of that, Google is making it easier for screen readers to read PDFs on desktop Chrome, and Android's Chrome Page Zoom now lets users enlarge text without sending page layouts into chaos.

​

Databricks announces a $1 billion deal to acquire database startup Neon to deploy its AI agents more efficiently

Data and AI analytics giant Databricks made a blog post on Wednesday to announce that it is acquiring open-source database startup Neon for roughly $1 billion. The company stated that it will combine Neon's tech with its own database management system for even more efficient deployment of AI agents for its customers. Neon's database capabilities, from smart automatic scaling to database branching to point-in-recovery and more, make its tech ideal for workloads run by AI agents, says Databricks.

​

Legal AI tech startup Harvey is reportedly looking to raise $250 million at a $5 billion valuation

Harvey is reportedly in discussions to hold a $250 million funding round, led by Kleiner Perkins and Coatue, that would up its valuation to a staggering $5 billion, as reported by Reuters. VC firm Sequoia Capital, which closed Harvey's $300 million Series D just three months prior, is also expected to participate. Investors have been attracted by Harvey's swift market traction and its reported $75 million in annualized run‑rate revenue as of April. This news follows Harvey's recent move to broaden its AI model offerings to now include Anthropic's and Google's offerings alongside OpenAI's.