Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Cognition scooped up Windsurf's remaining 250-person team and $82 million ARR business after Google's $2.4 billion reverse-acquihire cherry-picked only the leadership, leaving employees behind without payouts. The acquisition gives Cognition both AI coding agents and IDE capabilities to compete with Cursor's $500 million ARR, and restores Windsurf's access to Claude models that Anthropic had cut off amid rumors of an OpenAI acquisition.
The newly formed superintelligence lab is considering abandoning its flagship open-source Behemoth model in favor of closed development, marking a potential philosophical shift from the company's long-standing commitment to open AI.
LLMs don't have background processes that allow the model to form connections between seemingly unrelated topics similar to how humans make new realizations while "daydreaming." This is the reason AIs have yet to make any new discoveries. This post proposes a remedy: a system that would prompt LLMs to randomly retrieve facts, generate novel connections between them, then use a critic model to filter for genuinely valuable insights.
Google acquired key members of Windsurf for $2.4 billion after OpenAI's $3 billion offer expired, while Cognition took over the remaining company. Despite Windsurf's impressive $82 million ARR and rapid growth, the loss of Anthropic API access and leadership exits weakened its market position. The AI coding industry's inflated valuations, driven by talent wars and platform dependencies, suggest Windsurf's sale price may have been below its potential value.
Sponsored
Join the Amazon SageMaker Roadshow for advanced hands-on development [10] with real production-grade experience—not basic tutorials. Multiple tracks covering training, fine-tuning, and deploying production AI agents. Master OSS tools, Amazon Bedrock integration, cost optimization, and observability techniques.
Google's first Gemini Embedding text model is now generally available to developers in the Gemini API and Vertex AI. The model provides a unified cutting-edge experience across domains. It supports over a hundred languages and has a 2,048 maximum input token length. The Gemini Embedding model is priced at $0.15 per million input tokens.
Kiro is a Claude-powered "agentic IDE" that addresses the quality problems of AI-generated code by first producing specifications and user stories before generating actual code. The tool aims to move beyond "vibe coding" and reduce the time developers spend debugging and reviewing AI-generated code.
Energy-Based Transformers replace direct prediction with learned verification functions that assign compatibility scores between inputs and candidate outputs. The architecture is first to out-scale standard Transformers as it allows models to dynamically allocate computation and self-verify predictions without external supervision, achieving up to 35% higher scaling rates.
Google introduced "featured notebooks" in NotebookLM, offering curated research collections from experts and institutions to help users explore topics with high-quality sources.
The US Department of Defense is granting contract awards of up to $200 million for artificial intelligence development at Anthropic, Google, OpenAI, and xAI. The awards will help the agency accelerate its adoption of advanced AI capabilities to address critical national security challenges. The companies will work to develop AI agents across several mission areas. AI will transform the DoD's ability to support fighters and maintain strategic advantage over adversaries.
Sponsored
As AI systems advance, researchers are finding greater value in high-quality, human-curated lexical data. Read about the reasons for this transformation and its real world impact → get the Oxford Languages whitepaper [16].
Grok now features 3D animated companions, like a flirty anime character and mood-shifting red panda Rudy, that users can level up through conversation.
Rowhammer allows hackers to change or corrupt data stored in memory by rapidly and repeatedly accessing a physical row of memory cells.
Anthropic launched a tool directory to showcase integrations with Claude, enabling direct access to apps like Notion, Figma, and Stripe for more contextual and collaborative interactions.
This pull request is an ongoing effort sponsored by Apple to add a CUDA backend to MLX - very few things work now, except the tutorial example.
Sponsored
Bria offers numerous APIs and models trained EXCLUSIVELY ON LICENSED CONTENT with BUILT-IN ATTRIBUTION PAYMENTS TO DATA PARTNERS.
Easily remove unwanted pixels in seconds — ideal demos for exploring Bria's super-aesthetic, fully legal PaaS.
Google DeepMind hired Windsurf CEO Varun Mohan, co-founder Douglas Chen, and key researchers after OpenAI's $3 billion acquisition attempt failed. Google gains a nonexclusive license to Windsurf's technology, enhancing its AI coding capabilities without direct control. Windsurf faces uncertainty as it loses top talent, continuing to offer AI coding tools with Jeff Wang as interim CEO.
OpenAI is again delaying the release of its open model. The company had planned to release the model next week. The release has been pushed back indefinitely for further safety testing. The release of OpenAI's open model is one of the most highly anticipated AI events of the summer.
Apple is seriously considering acquiring Mistral, a French AI startup that has raised a total of €1.1 billion over seven funding rounds. Mistral has launched a range of large and small language models over the years and achieved notable success with its optical character recognition features. It is currently Europe's biggest AI startup. The acquisition of an AI startup would provide a much-needed boost to Apple's AI ecosystem.
Chinese startup Moonshot AI released Kimi K2, a 1 trillion parameter open-source model that matches proprietary models on complex agentic tasks. The model was trained using a novel MuonClip optimizer that prevents training crashes that plague model development, potentially saving millions in computational costs.
Asynchronous inference helps robotic policies run more smoothly by decoupling action prediction from execution, reducing idle time and improving responsiveness in real-world scenarios.
xAI launched Grok 4 on July 9. This post details everything about the model. It includes a performance overview, a survey of early vibe checks, tests of Grok 4 Heavy, and comparisons to OpenAI's o3 pro. The post also discusses xAI's lack of differentiated products, controversies, and culture risk. While Grok 4 is a very impressive model, it is plagued with serious behavioral risks and cultural concerns.
Sponsored
After helping hundreds of companies deploy AI solutions on AWS, Mission identified the 4 critical gaps that separate successful AI implementations from the 80% that never see real business impact. This 10-minute AI Readiness Assessment [13] reveals exactly what's blocking your progress and gives you a roadmap with personalized recommendations, plus a strategy session to discuss your results. Take
Reinforcement learning (RL) will continue to lead to more gains because when done well, it is a lot more leveraged, more responsive to feedback, and superior to supervised fine-tuning. Researchers are likely to discover more about RL as rollout lengths continue to expand. There are a lot more S curves to discover, possibly specific to large language models and without analogues in game/robotics-like environments.
Reinforcement learning is the next training technique for building frontier-level AI models. Training it on more data will make it better. The current approach of scaling is messy and complicated. Finding a way to do next-token prediction on the web using RL would enable models to reason from general web data instead of just math and code.
GPT-3 showed that scaling up language models unlocks powerful performance that often outperforms carefully fine-tuned models. Before GPT-3, achieving state-of-the-art performance meant first pretraining models, then fine-tuning them on specific tasks. Today's reinforcement learning is stuck in a similar pre-GPT paradigm. The approach suffers from the fundamental limitation where the resulting capabilities generalize poorly, leading to brittle performance that rapidly deteriorates outside of the precise contexts seen during training. The RL field will soon shift toward massive-scale training across thousands of diverse environments. Doing this effectively will produce RL models with strong abilities capable of quickly adapting to entirely new tasks - achieving this will require training environments at a scale and diversity that dwarfs anything currently available.
SpaceX has reportedly agreed to invest $2 billion in xAI as part of a $5 billion funding round, as Elon Musk increasingly interweaves his corporate empire.
ScreenEnv is a Python library for launching Ubuntu desktop environments in Docker, enabling agents to interact with real GUI applications and supporting the Model Context Protocol for seamless deployment.
Sponsored
You'll probably like our flagship newsletter. It's all about tech, science, and programming.
Grok 4 Heavy, the "think much harder" version of Grok 4 currently only available on the $300/month plan, has measures in place to prevent it from sharing its system prompt.
Play AI, which creates human-sounding AI voices, will merge with Meta's AI companion and wearable device efforts.
Former Intel CEO Pat Gelsinger, in partnership with Gloo, launched Flourishing AI (FAI) to benchmark AI models' alignment with human values.
Sponsored
Search has long been a the most important piece of the online shopping experience — and in recent years, AI has turned that experience on its head.
AWS will launch an AI agent marketplace next week at the AWS Summit in New York City, with Anthropic as a partner. The platform will enable startups to offer AI agents directly to AWS customers, potentially boosting Anthropic's reach and revenue. This move follows similar initiatives by Google Cloud and Microsoft, aiming to streamline AI agent distribution.
Amazon is considering another multibillion-dollar investment in AI firm Anthropic to strengthen their strategic partnership. It already committed $8 billion in November last year. Amazon is one of Anthropic's largest shareholders. Google, another shareholder, has invested over $3 billion.
Mistral AI and All Hands AI released Devstral Medium and upgraded Devstral Small 1.1, optimized for agentic coding tasks. Devstral Small 1.1 is open-source under Apache 2.0 and outperforms all other open models for code agents.
The success of a coding app depends on context engineering and software architecture, not raw model capabilities. Both platforms share four core components: typed prompts with test-driven development, MCP servers for sandboxed execution, agent loops for state management, and real-time frontend coordination.
There is widespread enthusiasm among Google software engineers about how AI is helping write code. This blog post looks at the company's newest AI-powered improvements and presents a method for building AI products that deliver value. Google monitors developer productivity and satisfaction carefully - the improvements covered in the post directly impact these. Google prioritizes by technical feasibility and impact, iterates quickly and extracts lessons learned, and extensively monitors productivity and satisfaction metrics to measure effectiveness.
A randomized controlled trial of 16 experienced open-source developers found AI tools decreased task completion time by 19%, despite developers self-reporting a 24% speedup.
Sponsored
Want to stay ahead of the AI curve? At GitHub Universe — from October 28-29 in San Francisco — you'll explore the newest AI agents, automations, and tools that can make your work more efficient and fun. You'll also get a chance to connect with fellow devs and leaders you admire. For a limited time, save $400 on your ticket with
The Gemini API now has an asynchronous endpoint designed for high-throughput, non-latency-critical workloads. The Gemini API Batch Mode allows users to submit large jobs, offload the scheduling and processing, and retrieve results within 24 hours, with a 50% discount compared to the synchronous APIs. It is the perfect tool for when users have data ready upfront and don't need an immediate response.
asyncmcp is a regular MCP that works over queues. A lot of context is not always readily available, and it takes time for applications to process it. asyncmcp uses an async transport layer, so it doesn't have to immediately respond to any requests - it chooses to direct them to internal queues for processing and the client doesn't have to stick around for the response.
Google's T5Gemma is a suite of encoder-decoder LLMs adapted from decoder-only Gemma 2 models. Designed for tasks like summarization and translation, T5Gemma includes pretrained and instruction-tuned variants in sizes ranging from 2B to XL.