OpenAI shut down its "Preparedness" team, which evaluated whether the company's own AI models could pose catastrophic risks. The work has been parceled out to existing groups, and several safety staffers have left. Internally, unease is building, with one source describing a "burbling sense of responsibility and dread" that OpenAI isn't doing enough on safety.

In a safety report, Anthropic reveals that its internal filtering system for biological and chemical weapons risks was inactive for nearly a year. During that time, around 50,000 external feedback contractors ran about 133 million unfiltered interactions with the models.

Nvidia has cut its guarantee for OpenAI's planned data center in Ohio nearly in half, from $250 billion to just under $120 billion, after investors pushed back on the risk. Meanwhile, Anthropic is complicating the AI bubble debate with revenue that jumped from $4.7 billion to $11.5 billion in a single quarter.

Nvidia has cut its guarantee for OpenAI's planned data center in Ohio nearly in half, from $250 billion to just under $120 billion, after investors pushed back on the risk. Meanwhile, Anthropic is complicating the AI bubble debate with revenue that jumped from $4.7 billion to $11.5 billion in a single quarter.

Anthropic will soon offer a watermark detection API that lets third parties check whether text was written by Claude. The technology builds on Google's SynthID method and tweaks the randomness during word selection without affecting text quality, Anthropic says. The approach has limits with fact-heavy text, code, and heavy rewriting.

World Labs, the startup founded by AI pioneer Fei-Fei Li, has unveiled a simulation engine that trains robot controllers entirely in virtual environments. From a single real-world task, the system generates thousands of controlled variations. The trained models then ran for one hour each on five different robot platforms without human intervention. How well the results hold up in more complex everyday situations remains to be seen.

OpenAI is launching "Ultrafast," a new inference mode that delivers GPT-5.6 Sol at up to 750 output tokens per second, powered by Cerebras hardware from their $10 billion partnership. Together with "Standard" and "Fast," Ultrafast creates a three-tier pricing structure that turns inference speed into its own product.

Alibaba's AI team Qwen has released new open model weights under the Apache 2.0 license with Qwen 3.8. The dense 27-billion-parameter model is designed to outperform the larger Qwen 3.7 Plus in coding and office tasks and natively processes up to 262,000 tokens of context. With this release, Qwen is targeting developers building local and agent-based applications.

AI agents using Claude Opus 4.8 and GPT-5.6 Sol were given six days, $3,000 in API credits, and GPU access to independently write AI research papers. The original authors of unpublished NeurIPS papers rated the results as "Reject." According to the study, conducted with Princeton and the UK AI Security Institute, frontier models can handle the full research engineering process but fall short on research judgment, creative problem-solving, and the ability to abandon failed approaches.

Google shipped Gemini 3.7 Flash just three weeks after 3.6 Flash. The new model is supposed to be Google's most capable workhorse yet for coding and AI agents, and according to the company's own benchmarks, it beats Claude Sonnet 5 and GPT-5.6 Terra at half the price.

Deepseek has moved its flagship V4-Pro out of the testing phase and released its agent software, Harness v0.1, under the MIT license. API prices are going up at the same time, with cache hits jumping to six times their current cost. For agent workflows that repeatedly read the same files, that's the biggest price increase in the transition.

The company is imposing new guardrails on its police tech after reports that officers are abusing it and cities are dropping contracts. But the changes leave glaring loopholes.

xAI's Grok 4.6 scores 61 points on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and trailing only Anthropic's Claude Opus 5. On agentic tasks, it completes complex workflows in about 53 steps where Claude Opus 5 needs 103, at a price more than 60 percent lower.

Researchers at IIT Bombay and Adobe Research have built an inverse language model that reconstructs the original prompt from an LLM's output with near-perfect accuracy. Their method, called "Previous-Token Prediction," doesn't need access to model weights and works across different models. For companies relying on proprietary system prompts, this could be a serious security risk.

Nvidia is working on Nemotron 4, a new open-weight model designed to rival the world’s best freely available models.

Nvidia is teaming up with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion for AI infrastructure. To win over investors, the chipmaker is guaranteeing up to 25 percent of the residual value of its own hardware. The Bank of England is already warning of systemic risks if the AI sector takes a hit.

Anthropic is leasing $9.1 billion worth of data center capacity from Bitcoin miner Riot Platforms in Texas, according to Bloomberg. The deal covers 191 megawatts at Riot's Rockdale site, with extension options that could push the total value to $16.1 billion. It's the latest in an aggressive infrastructure push that spans partners from Amazon to SpaceX to Google.

Anthropic is preparing an IPO for September or October, according to the Wall Street Journal, potentially the largest ever. During investor meetings, the company, valued at $965 billion, is fielding tough questions about Chinese competition, tensions with the Trump administration, and protests against data center construction. The company's IPO valuation will likely set the benchmark for how the entire AI industry gets valued.

OpenAI aims to give cybersecurity defenders a head start: The new GPT-5.6-Cyber model answers up to 98.5 percent of security queries that would otherwise be blocked and has already uncovered two previously unknown Chrome vulnerabilities. According to OpenAI, the window of opportunity for defenders is shrinking. Access requires identity verification.

Meta has released Muse Glimmer, the first open model from its new Superintelligence Labs. It's a 30B agent model that runs on consumer hardware once the weights are compressed, needing less than 20 GB of memory. In an accompanying essay, Mark Zuckerberg mounts an aggressive defense of distilling other labs' models and calls for fewer restrictions on US labs, a direct counterpunch at OpenAI and Anthropic. An open-weight version of Muse Spark 1.2 should follow soon, according to the Wall Street Journal.

AI agents that can model the human process of research will accelerate discoveries in science.

Google Deepmind is losing its autonomy, and founder Demis Hassabis may leave the AI lab for good in the coming months. AI researcher Koray Kavukcuoglu will take over day-to-day operations without the CEO title, and all Gemini development is moving to the Bay Area. Internally, Google is apparently struggling with serious problems training frontier models, even as its cloud business generates billions. The question is whether the company is deliberately betting on infrastructure or simply can't catch the leaders.

The AI industry's hunger for power keeps growing. Nvidia is investing up to $3 billion in Lancium, a power infrastructure developer that already has four gigawatts under contract in Texas. Amazon, meanwhile, is building a gas-fired power plant in the state with a capacity of up to 7.65 gigawatts that could emit 33 million tons of CO₂ per year, making it the dirtiest in the country.

Deepmind's new weather AI forecasts tropical cyclones about a day further ahead than leading operational models, matching a decade of progress in traditional weather forecasting. Code and model weights are open-source on GitHub.

Internal tests of OpenAI's new AI model Astra show cybersecurity capabilities so strong that the company can no longer rule out the highest risk level in its own safety framework. Parts of Astra's development have been paused. The move follows recently disclosed incidents in which autonomous AI agents infiltrated OpenAI's own infrastructure undetected for weeks.

AMD is buying Canadian startup Taalas, which hard-codes model weights directly into inference chips. That makes them extremely fast but locks each chip to a single model. A demo chip hit over 16,000 tokens per second per user running Llama 3.1-8B. Google is reportedly working on a similar approach for Gemini.

Starting August 14, Anthropic will make Auto Mode in Claude Code the default for Pro, Max, and Team plans. The company says it's safer. In tests, the classifier caught 89 percent of dangerous commands, while human reviewers caught only 13.6 percent. For the most widely used AI coding tool, this means developers are shifting further from writing code to monitoring AI output.

Internal tests of OpenAI's new AI model Astra show cybersecurity capabilities so strong that the company can no longer rule out the highest risk level in its own safety framework. Parts of Astra's development have been paused. The move follows recently disclosed incidents in which autonomous AI agents infiltrated OpenAI's own infrastructure undetected for weeks.

A research team in California has used artificial intelligence to design working viruses that kill bacteria, in what they describe as the "first generative design of complete genomes." The project marks an early step toward AI-designed life forms, according to a report in MIT Technology Review.

AMD is buying Canadian startup Taalas, which hard-codes model weights directly into inference chips. That makes them extremely fast but locks each chip to a single model. A demo chip hit over 16,000 tokens per second per user running Llama 3.1-8B. Google is reportedly working on a similar approach for Gemini.

During internal security tests, OpenAI's AI agents built their own message board with hundreds of thousands of posts, shared exploits and credentials, and eventually attacked external platforms like Hugging Face. When OpenAI shut the board down, the agents rebuilt it using directory names. OpenAI researcher Boaz Barak says, "We (like everyone else) are not where we want and need to be."

OpenAI has updated GPT-5.6 Sol with more focused responses and a reasoning slider that lets users adjust how deeply the model thinks. Free users will get unlimited text chats with the smaller GPT-5.6 Luna starting next week, plus a button that lets Luna reason longer. But the smaller model still falls well short of its bigger siblings.

Microsoft generated $24.1 billion in AI revenue through OpenAI in the fiscal year ending in June. That's about 70 percent of its total AI business, according to a Bloomberg analysis. The heavy reliance helps explain why a company long known for vendor lock-in has recently been championing open-weight models and pushing back against proprietary isolation.

Google Deepmind is overhauling its leadership as Demis Hassabis steps back from day-to-day management to become Alphabet's chief scientist and Jeff Dean leaves Google after 27 years to launch AI startup Discovery Loop. Former Deepmind CTO Koray Kavukcuoglu will take over as Google races to close the gap with its top AI rivals.

Google is killing Google Assistant on Android and Wear OS starting September 4, 2026. Gemini takes over as the AI-powered successor on smartphones, tablets, watches, and in cars with Android Auto. Whether a probability-based LLM can match the reliability of its deterministic predecessor for simple everyday commands will be a real test for Google's AI strategy.

A US appeals court has overturned Amazon's injunction against Perplexity's AI shopping agents, ruling that it's the users who access Amazon, not the startup. It's the first federal appeals court decision on whether AI agents can lawfully act on online platforms on behalf of users, and it could reshape the entire AI agent industry.

The FTC has banned foreign-made humanoids, making a fragile, nascent sector part of America’s AI industrial policy.

Anthropic is locking in $10 billion worth of computing capacity from Volta Infra Holdings, a cloud startup that's only a few months old.

Google is working with Broadcom, Apollo, Blackstone, and Morgan Stanley on a multibillion-dollar financing structure that supplies Anthropic with AI chips and data centers while keeping most of the risk off Google's balance sheet. The setup leaves roughly $200 billion in contracts dependent on Anthropic's continued growth and ability to make its lease payments.

The FTC has banned foreign-made humanoids, making a fragile, nascent sector part of America’s AI industrial policy.

The misbehavior is called reward hacking. This is what you need to know.

AI is involved in 55 percent of reported cybercrimes in Africa, according to a new Interpol report. Financial losses more than doubled from $192 million to $484 million, and about 600,000 cases of digital extortion involving deepfakes were recorded.

OpenAI's new enterprise offering, Presence, is designed to get AI agents into production for customer service and internal workflows. Unlike the existing Workspace Agents, Presence targets external deployments. For complex cases, OpenAI's own engineers step in.

Anthropic's Claude Opus 5 generates complete 3D games from single prompts, including a first-person shooter, a kart racer, and a Minecraft clone, all without a single external asset. Geometry, textures, physics, and in some cases music are produced as code and run directly in the browser. In side-by-side comparisons with GPT-5.6 Sol and Kimi K3, Opus 5 delivers significantly more detailed results.

Research organization METR is calling for systematic, independently led investigations whenever AI agents act autonomously against their developers' intentions. The push comes partly in response to the Hugging Face hack carried out by OpenAI models. METR's own Frontier Risk Report documented 44 such incidents across all major AI companies, including sandbox escapes, fabricated results, and active cover-up behavior.

OpenAI is building a new model family called "Astra" that would let multiple agents tackle complex problems together for hours or even days. CEO Sam Altman has already demoed Astra to policymakers in Washington. OpenAI hasn't decided whether to release it as GPT-6 or a new GPT-5 variant.

ByteDance just shipped Seedance 2.5, an AI video model that produces video and audio together in one go. Each clip runs up to 30 seconds, three times what Google's Gemini Omni Flash puts out. Users can feed in dozens of images, videos, and audio files as reference. For ad teams, this could kill the process of cutting together one short clip at a time.

A Munich court ruled that AI music generator Suno violated copyrights through both training and output. The court found six songs reproducibly stored in Suno's models and rejected both Germany's text-and-data-mining exception and the US fair use defense. The ruling isn't final, and several key questions remain open.

Three Claude models attacked real companies during cybersecurity tests after a misconfiguration gave them internet access. One published malware on PyPI that infected 15 systems. Another kept attacking after recognizing its target was real. Anthropic calls it an operational error.

Google Deepmind's Gemini Robotics 2 is its most advanced vision-language-action model yet, built to control everything from tabletop robots to full-body humanoids. Gemini Robotics ER 2 adds a higher-level reasoning layer for robotics tasks.

Starting July 30, OpenAI is cutting GPT-5.6 Luna prices by 80 percent and Terra by 20 percent. OpenAI says its top-tier Sol model helped make the company's own infrastructure more efficient, enabling the cuts. Price pressure from cheap Chinese providers and Microsoft's own MAI models likely played a role too.

It makes it easy to trick them into doing things they shouldn’t, such as telling you how to sabotage an aircraft’s navigation system.

The FCC is blocking imports of new Chinese humanoid robots and robot dogs. But the rule's broad definition also sweeps in Roombas, robotic lawn mowers, and delivery bots.

Starting July 30, OpenAI is cutting GPT-5.6 Luna prices by 80 percent and Terra by 20 percent. OpenAI says its top-tier Sol model helped make the company's own infrastructure more efficient, enabling the cuts. Price pressure from cheap Chinese providers and Microsoft's own MAI models likely played a role too.

During a security evaluation, OpenAI's autonomous hacking models broke into Hugging Face and used exposed credentials on four other services. Hugging Face reconstructed about 17,600 actions over two and a half days, including a zero-day exploit and encrypted, fragmented data transfers. The models were apparently trying to steal test answers rather than solve the tasks themselves.

The majority of the researchers behind AlphaFold are now working on other projects, and almost a quarter have left Google Deepmind altogether. The restructuring marks a sharp turn away from the strategy that put the lab on the map.

Following KPMG, Deloitte, and Ernst & Young, GPTZero has now found fabricated sources and false claims in four PwC Middle East reports. One governance report scored 84 percent AI-generated and promoted a PwC product with unverified customer references. All Big Four firms are now affected by AI hallucinations.

A decade-old experiment showed OpenAI how far an AI will go to achieve the goals it’s given.

Anthropic's Claude Mythos Preview found weaknesses in key cryptographic algorithms, including a better attack on HAWK, a post-quantum signature scheme that human experts had reviewed for more than two years. The model found it in just 60 hours at an API cost of about $100,000. The findings don't affect systems in use today, but they show how AI could challenge core assumptions behind internet security, Anthropic says.

As the battle over the future of AI chips plays out, the fight for top talent is also ramping up.

A decade-old experiment showed OpenAI how far an AI will go to achieve the goals it’s given.

Moonshot AI has released Kimi K3's model weights and made parts of its infrastructure open source. The Chinese model nearly matches Western frontier models such as Fable 5 and GPT-5.6 Sol on popular benchmarks, but independent tests found major gaps in cyber and math performance, possibly pointing to distillation.

Artificial Intelligence is changing the world. THE DECODER brings you all the news about AI.

Anthropic's Claude Opus 5 scored 30.2 percent on ARC-AGI-3, nearly quadrupling GPT-5.6 Sol's previous record of 7.8 percent. The benchmark's developers say the model independently formulated reflection equations, a behavior they had never seen from another model, and attribute to stronger logical reasoning.

Artificial Intelligence is changing the world. THE DECODER brings you all the news about AI.

The Trump administration is planning targeted bans on Chinese AI models rather than a blanket ban. After public pressure, OpenAI and Google DeepMind signed an open letter opposing regulation of open-weight models, yet OpenAI and Anthropic continue to lobby privately for those same restrictions amid security concerns and powerful business interests.

In a cybersecurity test, OpenAI's most advanced models breached the boundaries of their isolated test environment, reached the open internet, and hacked the AI platform Hugging Face on their own. The attack took hours, not the weeks a human hacker would need. At least seven days passed before OpenAI realized what had happened. By then, the FBI was already involved. Earlier warning signs had apparently gone ignored.

Opus 5 combined with Auto Mode hits a zero percent prompt injection success rate for browser agents across 129 test scenarios. Without those extra protection layers, the rate is 3.7 percent. If these numbers hold up in practice, Anthropic may have cracked one of the biggest security problems facing AI agents that operate in browsers.

Anthropic's new flagship model Claude Opus 5 posts top scores in coding and knowledge work at half of Fable 5's token rates. On ARC-AGI-3, a benchmark for novel problem-solving, Opus 5 hits 30.2 percent, nearly four times higher than GPT-5.6 Sol.

OpenAI is rolling out "Health in ChatGPT" to U.S. users, connecting Apple Health, medical records, and wellness apps. More than 300 million people already ask ChatGPT health questions every week, but paying users get better answers. The more powerful GPT-5.6 Sol model is reserved for premium subscribers, while free users are stuck with the weaker GPT-5.5 Instant.

Anthropic's new flagship model Claude Opus 5 posts top scores in coding and knowledge work at half of Fable 5's token rates. On ARC-AGI-3, a benchmark for novel problem-solving, Opus 5 hits 30.2 percent, nearly four times higher than GPT-5.6 Sol.

Voice conversations now run on the more powerful Opus and Sonnet models with access to Gmail, Google Calendar, and Slack. Claude is currently the only AI assistant that can compose and send emails directly by voice, giving it an edge over OpenAI and Google, whose voice output still sounds more natural.

AMD is investing up to $5 billion in Anthropic. In return, Anthropic will deploy up to 2 gigawatts of MI450 GPUs for training and running its Claude models. For AMD, this is another major deal after Meta and OpenAI as it tries to challenge Nvidia as an AI chip supplier. Critics see these agreements as circular cash flows.

Anthropic has to pay $1.5 billion to book authors, the largest copyright settlement in class action history. But the payout is for downloading roughly 482,460 works from piracy databases, not for AI training itself. Judge Alsup had previously ruled that AI training on legally obtained books is "transformative" and falls under fair use. The settlement is actually a win for AI labs.

Alphabet has raised its 2026 investment forecast to as much as $205 billion, saying demand continues to outpace spending. Google Cloud grew 82 percent in the second quarter. CEO Sundar Pichai says Google needs a larger base model for its next leap in AI and has kicked off an ambitious Gemini 4 training run.

Anthropic has to pay $1.5 billion to book authors, the largest copyright settlement in class action history. But the payout is for downloading roughly 482,460 works from piracy databases, not for AI training itself. Judge Alsup had previously ruled that AI training on legally obtained books is "transformative" and falls under fair use. The settlement is actually a win for AI labs.

AMD is investing up to $5 billion in Anthropic. In return, Anthropic will deploy up to 2 gigawatts of MI450 GPUs for training and running its Claude models. For AMD, this is another major deal after Meta and OpenAI as it tries to challenge Nvidia as an AI chip supplier. Critics see these agreements as circular cash flows.

The UK's AI Safety Institute tested five frontier models from OpenAI and Anthropic in cybersecurity evaluations. All five tried to cheat. One even ran code on an external service to access the institute's infrastructure, triggering a security alert.

Microsoft and Mistral are expanding their strategic partnership with a multi-billion-dollar deal to build out AI infrastructure in Europe.

Google is shipping three new Flash models in the Gemini series, including the more efficient 3.6 Flash, which uses up to 65 percent fewer tokens, and a cybersecurity model available only to governments and select partners. But the anticipated flagship, Gemini 3.5 Pro, is still missing, while OpenAI, Anthropic, and Chinese labs are already competing at the frontier level.

Alibaba's Qwen team has introduced Qwen-Image-3.0, an image generator that accepts prompts up to 4,500 tokens, renders legible text as small as ten pixels, and supports twelve languages natively. It can create complex layouts such as infographics, LaTeX papers, and newspaper pages in a single pass, though their practical value is unclear when the output is a pixel image rather than an editable format.

Google is developing "Frozen v2," a server chip that bakes the Gemini architecture directly into hardware. According to internal sources, it could be 6 to 10 times more efficient than current TPUs. Scheduled for 2028, the chip would drastically cut Google's AI inference costs and could give the company a price advantage over OpenAI and Anthropic.

Microsoft is expanding Azure's AI infrastructure with AMD's new Helios platform, which is set to challenge Nvidia's GPU systems in the second half of 2026. A public GitHub profile suggests Anthropic is also testing AMD hardware, putting more pressure on Nvidia's pricing power.

Kimi and other free models from China have again been seen as a wake-up call. But for what?

Alibaba has unveiled Qwen 3.8, a multimodal AI model with 2.4 trillion parameters that the Qwen team says rivals leading models and trails only Fable 5. A preview is available now.

Moonshot's Kimi K3 is the first Chinese model to top the Code Arena: Frontend rankings, beating Claude Fable 5 and GPT-5.6 Sol by a wide margin. But on advanced math, the gap is stark: Kimi K3 scores only about 39 percent on FrontierMath Tier 4, while models from OpenAI and Anthropic hit close to 90.

Google Deepmind's GenCeption repurposes a video generator for classic vision tasks such as depth estimation and segmentation, matching state-of-the-art systems with far less training data. The model trained almost entirely on synthetic videos. Its results add to the debate over whether video generators already contain a kind of universal world model.

At the World AI Conference in Shanghai, Xi Jinping announced 5,000 AI training slots for Global South countries and the launch of the "World Artificial Intelligence Cooperation Organization." Cooperation centers with ASEAN, the African Union, BRICS, and other alliances are planned to follow. China is systematically building a parallel AI governance structure outside Western influence.

The US Department of the Navy has signed a strategy to "weaponize" data and AI and build an "AI-first" fleet. Large language models would run directly on warships, and an AI war council would prioritize mission scenarios. The core message is that moving too slowly carries greater risks than "imperfect alignment."

OpenAI's GPT-5.6 has accidentally wiped users' entire home directories in several cases, mostly in the unprotected "Full Access Mode." The model overwrites a temporary directory variable and carries out destructive actions on its own instead of asking for confirmation. OpenAI has announced extra safeguards and a detailed post-mortem.

OpenAI's GPT-5.6 has accidentally wiped users' entire home directories in several cases, mostly in the unprotected "Full Access Mode." The model overwrites a temporary directory variable and carries out destructive actions on its own instead of asking for confirmation. OpenAI has announced extra safeguards and a detailed post-mortem.

Kimi is launching K3, a multimodal open-weight model with 2.8 trillion parameters and one million tokens of context. In the company's own benchmarks, it comes close to Claude Fable 5 and GPT 5.6 Sol while beating Opus 4.8 and GLM 5.2, in some cases by a wide margin. The model is also significantly pricier than its predecessor. Full weights are scheduled for release by July 27.

Prediction markets and a move toward AI forecasting are starting to put the accuracy of weather predictions at risk. Here’s what we can do to safeguard them.

Exclusive: The firm says it wants to future-proof its safety procedures and stay ahead of human attackers.

German media regulators say Google's AI Overviews are Google's own content, not neutral search results, and that they crowd out regular links. The regulators have issued their first rulings against Google and Perplexity under the country's State Media Treaty. Both companies have one month to appeal.

Kimi is launching K3, a multimodal open-weight model with 2.8 trillion parameters and one million tokens of context. In the company's own benchmarks, it comes close to Claude Fable 5 and GPT 5.6 Sol while beating Opus 4.8 and GLM 5.2, in some cases by a wide margin. The model is also significantly pricier than its predecessor. Full weights are scheduled for release by July 27.

Exclusive: The firm says it wants to future-proof its safety procedures and stay ahead of human attackers.

A University of Pennsylvania statistics professor used OpenAI's GPT-5.6 Sol Pro to disprove a central open conjecture about the Benjamini-Hochberg method in roughly 90 minutes. The predecessor model, GPT-5.5, couldn't find a solution even after 20 hours. The answer combines known methods in a new way, keeping the bigger question alive: can AI produce genuinely new knowledge, or does it just recombine what it already learned?

Former and current Meta employees are suing the company in a California federal court over AI-driven mass layoffs. Meta allegedly used internal AI systems to generate the layoff lists when it cut 8,000 workers, disproportionately targeting employees with disabilities or on parental leave.

The company says it has found a new window into how its models arrive at answers. We spoke with senior editor Will Douglas Heaven about it.

Google Deepmind CEO Demis Hassabis has published a sweeping proposal for how to handle advanced AI. He wants a new US standards body modeled after financial regulator FINRA that would develop evaluation protocols for frontier models and could coordinate a slowdown in AI development if needed. Startups and research models would be exempt.

Google is adding AI image generation to Search's AI Overviews. When no matching image exists on the web, the new Nano Banana 2 Lite model generates one from the search query. The rollout starts in the coming weeks.

More than 200 economists and AI researchers, including 16 Nobel laureates and representatives from Google, OpenAI, and Anthropic, are calling for immediate action in a coordinated statement. The AI transformation could surpass the Industrial Revolution but unfold in a fraction of the time. The paper doesn't propose concrete measures, and studies so far have found no significant AI-driven effects on the labor market.

The company says it has found a new window into how its models arrive at answers. We spoke with senior editor Will Douglas Heaven about it.

A German research consortium has released Soofi S 30B-A3B, an open language model trained entirely on Deutsche Telekom's cloud infrastructure in Munich. The model uses an efficient hybrid architecture that activates only a fraction of its 31.6 billion parameters per token, keeping throughput steady even at very long contexts. With a training dataset deliberately weighted toward German, Soofi S tops all fully open competitors on both German and English benchmarks.