TLDR AI 2025-07-15
Cognition acquires Windsurf π€, MLXβs CUDA backend β‘, AWS Kiro IDE π¨βπ»
Become cloud-agnostic with 99.99% uptime (no more outage scrambling) (Sponsor)
Baseten recently outlined their multi-cloud capacity management (MCM) infrastructure for generative AI workloads. Baseten's MCM unifies 10+ clouds into a global GPU pool enabling infinite scaling and 99.99% uptime for generative AI workloads.
What that means for you:
- π Don't go down when a single cloud does
- π Horizontally scale across cloud providers
- π Seamlessly scale up during peak demand (even with virality)
- β
No downtime for your end users due to infra challenges
- π On demand compute, pay only for what you use
Unlock multi-cloud with Baseten
Meta Weighing Shift Away from Open Source (4 minute read)
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.
Cognition acquires Windsurf (2 minute read)
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.
π§
Deep Dives & Analysis
LLM Daydreaming (10 minute read)
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.
Did Windsurf Sell Too Cheap? The Wild 72-Hour Saga and AI Coding Valuations (6 minute read)
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.
π¨βπ»
Engineering & Research
π€ Take Your AI and Agent Skills from Zero to Hero π (Sponsor)
Join the Amazon SageMaker Roadshow for advanced
hands-on development with real production-grade experienceβnot basic tutorials. Multiple tracks covering training, fine-tuning, and deploying production AI agents. Master OSS tools, Amazon Bedrock integration, cost optimization, and observability techniques.Level up your AI skills.
Grab your seat todayAWS previews Kiro IDE for developers who are over vibe coding (2 minute read)
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 are Scalable Learners and Thinkers (32 minute read)
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.
Gemini Embedding now generally available in the Gemini API (3 minute read)
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.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email