x
Anthropic has overtaken OpenAI in secondary market trading, with shares on platforms like Forge Global implying a valuation of around $1 trillion compared to OpenAI’s $880 billion. This marks a sharp reversal from just three months earlier, when Anthropic was valued at $380 billion following a major funding round. The surge has been driven by rapid revenue growth, with annualised figures jumping from $9 billion to $30 billion in a single quarter, largely due to strong enterprise uptake of its Claude Code and API products, alongside investor confidence boosted by backing from firms including Amazon.
Demand for Anthropic shares has also been intensified by limited supply, as early investors and employees have had few opportunities to sell, pushing prices higher in thinly traded markets. In contrast, OpenAI has seen more selling interest than buying on secondary platforms, keeping its valuation relatively stable. Despite the headline figure, the $1 trillion valuation reflects small, illiquid trades rather than a realistic public market price, with reports suggesting Anthropic may target an IPO valuation closer to $400–$500 billion, advised by Goldman Sachs and JPMorgan. Source
Google has introduced a new AI-powered feature called Maps Imagery Grounding, allowing filmmakers and creative teams to generate and preview scenes based on real-world locations using Google Maps and Street View data. Announced at Google Cloud Next, the tool works through Google’s Gemini Enterprise platform and can also animate scenes using its Veo video model, offering a faster and more cost-effective alternative to traditional location scouting and early-stage storyboarding.
The move reflects a wider push to streamline film production and expand the use of AI-driven mapping tools into creative industries, alongside new capabilities for analysing satellite and aerial imagery. However, the development lands amid ongoing debate in Hollywood about AI’s role, with figures across the industry divided on whether it enhances filmmaking or threatens jobs and artistic integrity, even as some studios and creatives cautiously embrace it as a production tool. Source
Meta is planning to cut around 8,000 jobs, roughly 10% of its workforce, while also removing 6,000 unfilled roles as it shifts resources heavily towards artificial intelligence. The move reflects a broader restructuring led by Mark Zuckerberg, who has emphasised AI as central to the company’s future and suggested that advanced systems could eventually take over significant portions of work currently done by engineers and technical staff.
The layoffs come amid a wider trend across the tech sector, with firms such as Microsoft and Block also reducing headcount while investing more in AI, citing efficiency gains from automation. Following the announcement, Meta’s share price fell by more than 2% to around $659, signalling investor caution as the company accelerates its transition towards an AI-focused business model. Source
Space and Time has launched Dreamspace, an AI-powered no-code platform that allows users to build, audit and deploy on-chain applications and smart contracts using simple prompts. Backed by Microsoft, the tool has already seen significant uptake during its beta phase, with more than 34,000 apps created since August 2025. Designed for Ethereum-compatible blockchains, Dreamspace integrates both smart contract generation and front-end development, aiming to simplify what has traditionally been a complex and technical process.
Built using Microsoft’s Azure AI tools, the platform supports deployment across Ethereum Virtual Machine networks, with Base set as the default due to its low transaction costs and fast speeds. The company says the tool has also gained traction in education, with around 140,000 students using it and institutions incorporating it into AI-focused curricula. By lowering technical barriers, Dreamspace is positioned as a way to expand participation in blockchain development and enable users to launch on-chain projects more easily. Source
OpenAI’s GPT-5.5 has been launched as its latest model, designed for more agent-like computer use such as coding, web browsing, spreadsheet work, and handling multi-step tasks with less human input. It is rolling out to Plus, Pro, Business and Enterprise ChatGPT users, with API access coming soon. The model is said to match GPT-5.4 in speed while outperforming it across most benchmarks, including a strong 82.7% score on Terminal-Bench 2.0, and improved results in coding and knowledge work tasks. It also shows higher efficiency by completing Codex tasks using fewer tokens, and performs strongly in evaluations of real-world professional work and long-horizon coding challenges.
Despite these gains, GPT-5.5 comes with higher pricing in the API, with input and output token costs significantly above GPT-5.4, although OpenAI argues efficiency improvements reduce overall cost in practice. A higher-tier GPT-5.5 Pro version is also being released for more demanding tasks, with strong performance on web research benchmarks. The model arrives in a fast-moving AI market where new releases are appearing in short cycles, and while it is not available to free users, paid subscribers gain access immediately, particularly benefiting those using OpenAI’s Codex environment for advanced coding workflows. Source
OpenAI has introduced ChatGPT for Clinicians, a specialised version of its chatbot aimed at supporting doctors and other healthcare professionals with time-consuming work such as clinical documentation, medical research and patient care consultations. The tool is currently free for verified US-based physicians, nurse practitioners, physician assistants and pharmacists, with international expansion planned. It includes features such as a clinical search tool drawing on peer-reviewed literature, a deep research mode for medical reviews, reusable templates for administrative tasks like referral letters, and the ability to earn continuing medical education credits within the platform. Conversations are not used for model training, and the system is designed to support HIPAA compliance through appropriate agreements.
Alongside the launch, OpenAI introduced a new benchmark called HealthBench Professional to evaluate performance in realistic clinical tasks. The company reported that GPT-5.4 within the clinician version of ChatGPT scored higher than human doctors on this test, achieving 59.0 compared with 43.7 for physicians, even when humans were given unlimited time and internet access. It also outperformed competing AI models from other major developers. However, OpenAI noted the limitation that it designed both the product and the benchmark used to measure performance, which adds context to the results. Source
Tencent has released Hy3 preview, a new open-source large language model designed to deliver strong performance in reasoning, coding and agent-style tasks while remaining highly efficient to run. The model uses a Mixture-of-Experts architecture with 295 billion total parameters but only around 21 billion active at a time, allowing it to reduce computational cost while maintaining strong capability. It also supports a very large context window of up to 256,000 tokens, enabling it to process extremely long inputs. Hy3 has already been integrated into Tencent’s ecosystem, including apps like QQ and Tencent Docs, and is available via Tencent Cloud with relatively low API pricing.
In benchmark testing, Hy3 shows major improvements over its predecessor, particularly in coding and autonomous task execution. On SWE-bench Verified, which measures real-world GitHub bug fixing, it improved from 53% in the previous version to 74.4%, placing it closer to leading global models. It also performed strongly on agent-based and web search tasks, with significant gains in benchmarks like BrowseComp and Terminal-Bench 2.0, reflecting better performance in multi-step, tool-using workflows. Tencent reports that the model was developed in under three months following an infrastructure overhaul, and while it still trails some top Western models, it is notable for its strong performance-to-cost efficiency compared with much larger competitors. Source
The White House has accused foreign actors, primarily in China, of conducting large-scale efforts to steal and replicate American artificial intelligence models. According to a memorandum from the Office of Science and Technology Policy, these campaigns allegedly use techniques such as jailbreak prompts, proxy accounts and coordinated networks of fake users to extract outputs from leading US AI systems. Officials claim the aim is to distil the capabilities of frontier models into smaller systems that can mimic their performance while avoiding the cost and complexity of developing them independently.
The administration says these so-called distillation attacks involve training smaller models on the outputs of larger, more advanced systems, which can allow imitators to achieve similar benchmark results at lower cost, even if overall capability is reduced. The memo warns that such activity could undermine security safeguards and weaken protections built into AI systems, and calls for closer cooperation between government and industry to prevent further exploitation. While acknowledging that legitimate model distillation can support open-source development, it argues that unauthorised copying of US AI technology represents a threat to innovation and will be met with stronger defensive and regulatory action. Source
.png)
Markethive has introduced what it calls “HIVE Intelligence” alongside a new AI Assistant, positioning the system as an alternative to conventional artificial intelligence platforms. The company claims its approach blends algorithmic systems with “Human Intelligence,” emphasising user control, privacy, and community-driven input rather than centralised data tracking. The AI Assistant is designed to act as an on-platform coach for entrepreneurs, helping them analyse activity across blogs, banners, and networks, and turning platform data into actionable insights for content strategy, engagement, and growth.
The system offers features such as performance tracking, content optimisation suggestions, network growth analysis, and reporting tools, including exportable data formats. It can also provide personalised guidance on posting strategies and highlight high-performing content within the platform. Markethive states that users retain full ownership of their data, with activity not tracked or sold externally. The platform integrates both a public chat system and an AI chat tool, and describes its broader vision as combining technological development with a community-led and values-driven framework for digital entrepreneurship. Source
Google has announced its eighth-generation Tensor Processing Units (TPUs) at its Cloud Next 2026 event in Las Vegas, introducing two separate chip designs aimed at different parts of the artificial intelligence workload. The TPU 8t is built for training large AI models and delivers nearly three times the compute performance per pod compared with earlier generations, scaling up to 121 ExaFlops across a full superpod. Alongside it, the TPU 8i is designed for inference tasks, with three times more on-chip memory to better support AI agents that require repeated, fast interactions and longer-running processes.
The company positions this dual-chip strategy as a response to the growing demands of the AI industry, where training frontier models and running deployed agents require very different kinds of computing power. Google says the new hardware improves price-to-performance by around 2.8 times and is part of its broader effort to compete with Nvidia in the AI infrastructure market. The announcement also comes alongside major investment plans, with Google expected to spend up to $185 billion this year to expand its AI infrastructure, as well as new partnerships such as expanded TPU access for AI company Anthropic. Source
OpenAI has introduced a new “workspace agents” feature for ChatGPT aimed at automating workplace tasks such as report generation, coding, and handling communications across tools like Slack. The agents are powered by OpenAI’s Codex model and are designed to operate as persistent, cloud-based assistants that can run multi-step workflows, integrate with external applications, and retain context across projects. They are available for ChatGPT Business, Enterprise, Edu and Teachers plans, and can be created by users through a setup process where ChatGPT helps define and connect workflows.
The company says workspace agents represent an evolution of custom GPTs, allowing teams to build shared agents that can be reused and improved across organisations. These agents can run on schedules or respond to triggers, and are intended to support collaborative work that requires coordination across systems and teams. OpenAI has made the feature available in research preview at no cost until May 6, 2026, after which it will switch to a credit-based pricing model. The launch comes amid growing competition in agentic AI, as major tech firms invest heavily in systems that can complete tasks with limited human input. Source
Xiaomi has released its MiMo-V2.5 and MiMo-V2.5-Pro AI models, expanding its previous generation by adding full multimodal capabilities including image, audio and video understanding alongside text and coding. The new models are designed for both everyday and advanced use, with the Pro version targeting complex agent-style workflows that involve long sequences of tool use and autonomous task execution, while the standard version focusses on faster and cheaper general use. Both models support a very large context window of up to one million tokens, allowing them to process extremely long documents or conversations in a single session.
Performance claims suggest MiMo-V2.5-Pro competes with leading frontier models in coding and agent-based tasks, showing strong results on benchmarks such as SWE-bench Pro while improving efficiency in token usage compared with rivals. However, it still trails top models in some advanced reasoning evaluations. Xiaomi is positioning the system as highly cost-efficient, with lower token usage and reduced pricing for large-scale workloads. The company is also planning to open-source parts of the model family and has already begun rolling it out through APIs and developer platforms, as part of a broader and rapidly accelerating push into the AI market. Source
Google has announced plans to invest between $175 billion and $185 billion in 2026 to expand its AI and cloud infrastructure, as it accelerates its push into what CEO Sundar Pichai describes as the “agentic era” of artificial intelligence. Speaking at Google Cloud Next in Las Vegas, Pichai said the investment is aimed at supporting AI systems capable of performing tasks with limited human oversight, as the industry shifts from conversational chatbots to autonomous agents. He noted that nearly 75% of new code at Google is now AI-generated but still reviewed and approved by engineers.
The company also highlighted how it is already using AI internally to automate workflows, including cybersecurity operations where AI systems process large volumes of threat reports and significantly reduce response times. Alongside the infrastructure spending, Google is building an ecosystem around agentic AI through partnerships and a $750 million fund to support cloud partners and enterprise adoption. Companies such as Citi and Thinking Machines Lab are already using Google’s tools for financial services and AI research, reflecting the company’s broader strategy to monetise and scale agent-based systems across industries. Source
OpenAI has released a lightweight, open-source tool called Privacy Filter that removes sensitive personal information from text before it is sent to AI systems. The model, which has around 1.5 billion parameters, runs locally on a user’s device and is designed to detect and mask data such as names, addresses, emails, phone numbers, passwords and account numbers. It is available under the Apache 2.0 licence via GitHub and Hugging Face, meaning it can be freely used, modified and integrated into other products.
The tool works by analysing the full context of a sentence rather than relying on simple pattern matching, allowing it to better distinguish between genuine personal data and similar-looking non-sensitive information. OpenAI reports that it achieves around 96% accuracy on a standard benchmark for identifying and masking personal data. Because it runs locally, the idea is that users can clean their text before sharing it with chatbots, reducing the risk of exposing sensitive information to external systems. However, OpenAI cautions that it is not a complete privacy solution and should not be treated as a substitute for formal compliance or security processes. Source
Mozilla has reported that an early version of Anthropic’s Claude Mythos AI model identified 271 vulnerabilities in its Firefox browser during internal testing. The findings, which have since been patched, highlight the model’s ability to analyse large and complex codebases and detect security flaws that would typically require significant human expertise. Mozilla also referenced earlier tests in which the same system found dozens of additional security-sensitive bugs, suggesting a substantial increase in automated vulnerability discovery capability.
The results have raised both optimism and concern within the cybersecurity community. On one hand, such tools could help developers identify and fix weaknesses before they are exploited, potentially shifting the balance towards defenders. On the other hand, experts warn that the same capabilities could be used to accelerate cyberattacks by automating the discovery of exploitable flaws at scale. Anthropic has restricted access to Claude Mythos through a controlled programme for selected partners due to these risks, as governments and industry groups continue to assess the security implications of highly capable AI systems. Source
Google has fixed a security vulnerability in its Antigravity AI coding platform after researchers discovered that it could potentially allow attackers to execute malicious commands on a user’s machine. The flaw, reported by cybersecurity firm Pillar Security, involved a prompt injection issue in the platform’s file search tool, where unvalidated user input could be passed directly to a command-line utility. This could turn what appeared to be a harmless search request into an opportunity to run system commands.
Researchers warned that the issue could bypass Antigravity’s Secure Mode, meaning attackers might still be able to trigger actions even under stricter security settings. In a demonstration, they showed how a crafted script could be executed through the system’s search function, highlighting the risk of AI development tools interpreting hidden instructions as legitimate commands. Google was notified of the issue in January and released a fix on 28 February. The findings underline broader concerns about prompt injection attacks in AI coding systems, where autonomous or semi-autonomous tools can be manipulated into performing unintended and potentially harmful actions. Source
Mississippi College School of Law has introduced a mandatory artificial intelligence course for all first-year students, making it one of the first law schools in the United States to require formal AI training. The programme aims to ensure students understand how to use AI tools responsibly in legal practice, including how to verify outputs and avoid errors such as fabricated case citations. Law school leadership says the move reflects growing expectations from employers that graduates will already be familiar with AI systems used in legal work.
The decision comes as courts and legal institutions increasingly confront both the advantages and risks of generative AI, including concerns about unreliable outputs and questions over how AI-generated communications are treated in legal contexts. Judges and legal experts have warned that AI can produce false information if used uncritically, while some courts are also experimenting with AI tools to manage workloads. The law school’s approach reflects a broader shift across legal education, as institutions prepare students for a profession where AI is becoming part of everyday practice in both litigation and legal research. Source
Disclaimer: These articles are provided for informational purposes only, mistakes may be made, and they are not offered or intended to be used as legal, tax, investment, financial, or any other advice.
.png){kind=small}
to find the NewsFeed
