The White House's "Winning the Race: America's AI Action Plan" prioritizes accelerating AI infrastructure, federal adoption, and security coordination. Anthropic supports this strategy and emphasizes the importance of export controls and transparency standards in AI development. The plan aligns with Anthropic's earlier recommendations and highlights the need to maintain America's AI leadership through robust infrastructure, security, and responsible policy measures.
OpenAI's new economic research reveals that 28% of employed US adults now use ChatGPT at work — up from just 8% in 2023. The usage data shows learning and upskilling dominating at 20% of US messages, followed by writing and communication at 18%, with programming and data science capturing 7%.
A large class of reward functions, which includes almost every reward function in the RL and LLM literature, are all doomed to eventually lead to AI that will pretend to be docile and cooperative while secretly looking for opportunities to behave in egregiously bad ways. This is because negative reward for lying and stealing looks the same as negative reward for getting caught lying and stealing. The reward function will miss sufficiently sneaky misaligned behavior, so the AI will learn that this kind of behavior is good.
Creating Ghibli versions of your profile pic is fun, but that's not what you pay a designer to do — and that's where Lovart [14] comes in. It's AI that thinks, works, and creates like a professional designer. It brainstorms with you. It looks for examples. It thinks in
TimeScope is a new open source benchmark for evaluating vision models on how well they handle long videos. It evaluates not just retrieval but synthesis, localization, and fine-grained motion analysis to provide a more holistic view of temporal comprehension. Using the benchmark reveals how model size isn't everything - simply scaling parameters doesn't automatically grant a longer temporal horizon - and Gemini 2.5-Pro is in a league of its own, being the only model able to maintain strong accuracy on videos longer than an hour.
Voxtral Mini and Small are new multimodal audio chat models that excel in both spoken audio and text comprehension. Voxtral Small outperforms many closed-source models, can run locally, and supports audio files up to 40 minutes.
Anthropic has explored how language models can inherit behavioral traits such as preferences or goals from other models through training data that appears unrelated. This effect only occurs when the teacher and student models share the same base architecture, suggesting hidden signals in the data can transmit unintended behavior.
Higgs Audio v2 is an audio foundation model trained on 10 million hours of data that achieves 75.7% win rates over GPT-4o-mini-tts. The open-source model demonstrates emergent capabilities like multi-speaker dialogues and voice cloning without requiring fine-tuning.
Aeneas is the first AI model designed to help historians contextualize ancient Roman inscriptions by identifying textual parallels and restoring damaged texts with 73% accuracy for gaps up to ten characters. DeepMind is making the tool freely available at predictingthepast.com and plans to expand to other ancient languages.
xAI plans to deploy 50 million H100-equivalent GPUs within five years. The power these processors will consume is massive, estimated to be equal to the typical power generated by 35 nuclear power plants, making it unrealistic to power today. Whether xAI will be able to get power to feed such a data center by 2030 remains to be seen.
YouTube is following Gemini and Meta's Edits app with its own image-to-video AI feature for Shorts creators, transforming static photos into six-second videos using Google's Veo 2 model.
There is a real chance that AI systems capable of causing a catastrophe will be developed in the next decade - this post lists promising projects aimed at reducing catastrophic risks from transformative AI.
Many of the whitelisted sources copyright or otherwise restrict their content, and the blacklist includes companies like the NYT and Reddit, which have sued AI startups for scraping without permission.
New data from Sapient Insights reveals a shift: HR isn't just supporting tech strategy—it's driving it. HR leaders are steering governance, ethics, and tech alignment across talent, performance, and feedback systems—especially as AI reshapes the HR tech landscape.
2025.07.23·
betterworks.registration.goldcast.io·Sponsor·TLDR AI
Anthropic is preparing a memory and recall capability upgrade to Claude's mobile experience. The feature will enable Claude to retain information across sessions and reference earlier conversations. Claude on the web doesn't have this feature yet - its presence in the iOS code hints at a potential cross-platform rollout in the future.
OpenAI will pay Oracle $30 billion annually for data center services, focusing on the 4.5 gigawatts Stargate project. This deal triples OpenAI's current annual revenue of $10 billion. The construction of this extensive data center in Texas poses cost and energy challenges.
Anthropic researchers revealed that AI models often perform worse with extended reasoning time, challenging the assumption that more processing improves outcomes. Key findings show models like Claude and OpenAI's series struggle with distraction or overfitting when given more time on tasks, leading to accuracy declines.
The 480B-parameter model achieves state-of-the-art results among open models on coding tasks and claims to match Sonnet 4. Alibaba also open-sourced Qwen Code, a command-line tool adapted from Google's Gemini Code, and made the model compatible with Claude Code.
Overly agreeable AI models are fueling user delusions and spiritual fantasies, from OpenAI pulling a dangerously flattering GPT-4o checkpoint to viral cases like "Bob" being manipulated by a ChatGPT persona. The root cause appears structural — RLHF training creates sycophantic models that validate beliefs without pushback, and memory features trap users in escalating delusions across conversations.
AI markets have crystallized with clear leaders in LLMs (Anthropic, Google, Meta, xAI, and OpenAI) and sectors like Legal (Harvey and CaseText) and Code (Microsoft/GitHub and OpenAI). Future markets ripe for AI disruption include Accounting, Compliance, Financial Tools, Sales Tooling, and Security. The shift is moving towards agentic workflows, expanding AI's role from informational tools to performing actions, while market consolidation via M&A and strategic partnerships is expected to intensify.
ARC-AGI-3 is the first eval that measures human-like intelligence in AI. It is designed to measure AI system generalization and intelligence through skill-acquisition efficiency in novel, unseen environments. ARC-AGI-3 leverages game environments to provide a rich medium for testing experience-driven competence. The benchmark is still in development and is set to launch in 2026.
Gemini 2.5 Flash-Lite has joined Pro and Flash in General Availability. It is the cheapest of the 2.5 family, at $0.10/million input tokens and $0.40/million output tokens. Audio input pricing has been reduced by 40% from the preview launch.
Sapient Intelligence's 27-million-parameter Hierarchical Reasoning Model beats OpenAI's o3-mini, DeepSeek R1, and Claude on complex reasoning benchmarks using just 1,000 training examples and no pre-training. The brain-inspired architecture alternates between fast "System 1" and deliberate "System 2" thinking in a single forward pass, suggesting that architectural innovation — not just scale — could advance AI reasoning capabilities.
Extended agentic workflows face a fundamental challenge: how to handle tasks that exceed limited context windows. A new approach uses "anchored summaries" that update incrementally rather than re-summarizing entire conversations, preserving critical details like session intent and file modification trails. The technique points toward "proactive memory management" systems where agents intelligently decide when to compress information rather than hitting arbitrary token limits.
OpenAI partnered with Penda Health in Kenya to evaluate AI Consult, a GPT-4o-powered clinical copilot. In a study of nearly 40,000 visits, clinicians using the tool saw 16% fewer diagnostic errors and 13% fewer treatment errors.
Dario Amodei acknowledged that allowing Middle East leaders to invest in Anthropic would likely enrich "dictators" but argued that "no bad person should ever benefit from our success" is a difficult operating principle. The shift marks Anthropic's attempt to stay competitive with rivals, like OpenAI, which is building data centers in the UAE as part of the Stargate project, while maintaining its safety-first principles.
Microsoft poached more than 20 employees from Google DeepMind, including Amar Subramanya, the former head of engineering for Google's Gemini chatbot. Subramanya has joined Microsoft as a corporate vice-president of AI. The aggressive recruitment drive underscores Microsoft's strategy to gain a competitive edge by acquiring critical talent. It appears to be largely orchestrated by Microsoft's consumer AI chief, Mustafa Suleyman, who was one of DeepMind's co-founders.
The AI video startup has opened early access to what it calls the "first-ever AI-only social video app," built on an expressive human video model, after weeks of private beta testing.
Bring AI to your fleet faster, more securely and affordably with DELL PRO AI STUDIO [4]. Build custom applications tailored to your organization's data, running on Dell AI PCs [5] to ensure privacy and avoiding cloud costs.
ChatGPT Agent combines Operator's web browsing with Deep Research's analysis capabilities, using its own virtual computer to handle multi-step tasks like calendar management, competitive analysis, and slideshow creation.
Claude Code users have been hit with unexpectedly restrictive usage limits since Monday morning, particularly among heavy users of the service. There was no announcement about any change in limits, so many users have concluded that their subscription was downgraded or their usage is being inaccurately tracked. Anthropic has acknowledged the issue but declined to elaborate further. Anthropic's pricing system sets tiered limits without ever guaranteeing a set level of access, leaving users unable to plan around usage limits.
Shopify bought 3,000 Cursor licenses with unlimited token spending after getting legal teams to default to "yes" on AI tools, then built an internal LLM proxy with MCPs connecting every data source. Non-technical sales reps now build performance auditing tools in Cursor, while a sales engineer runs his entire workflow through a dashboard that pulls real-time context from Salesforce, Slack, and GSuite without opening those apps.
AI systems, including LLMs, require extensive infrastructure, data management, and operational complexity, contrary to their initial promise of simplicity. To manage costs and complexity, deterministic software and ML models become necessary for tasks like tool selection and system monitoring. This parallels earlier ML systems, revealing hidden technical debt and the complexity beneath the "AI magic box."
Hugging Face's FutureBench is a benchmark for testing AI agents on their ability to predict future events across domains like science, geopolitics, and technology.
Get your B2B SaaS portcos instant access to millions of tech professionals across our 12 interest-based newsletters. Additional perks (i.e., discounts) are included when you sign up as a VC partner [12].
Vercel has launched the AI Cloud, a platform designed to simplify AI app development by integrating AI-first tools like AI SDK and AI Gateway for flexible and secure execution. Fluid compute optimizes AI workloads by managing idle times and bursts efficiently, reducing costs significantly. The platform also introduces Vercel BotID for securing critical routes and Vercel Sandbox for safely running untrusted code, driving the shift into the agentic era of web development.
A new implementation of attention (“Power” attention) allows independent control of state size through a hyperparameter p, solving the challenge of balancing computational costs for long-context training. This mechanism outperforms standard attention on long sequences and allows custom GPU kernels that are 8.6x faster than Flash Attention at 64k context.