TLDR AI 2025-06-10

Cirrascale is redefining what's possible in AI infrastructure with custom-built, large-scale systems using the groundbreaking NVIDIA HGX B200 platform.

Built on the NVIDIA Blackwell architecture, these systems deliver record-breaking performance, memory capacity, and efficiency—ideal for massive GenAI training and inference workloads.

We don't just offer access—we tailor the infrastructure to your exact needs, and can provide flexible financing options.

Join the next wave of datacenter innovation, powered by Cirrascale and accelerated by NVIDIA.

🔗 Reserve HGX B200 access today

OpenAI doubled its ARR from $5.5 billion last year, continuing progress toward its aggressive $125 billion revenue target by 2029. The company now counts 500 million weekly active users and 3 million paying business customers, up 50% since February.

Apple unveiled new Apple Intelligence features at WWDC 2025, including on-device foundation models for developers to integrate AI experiences into their apps.

Taobao, a popular Chinese e-commerce site, hosts thousands of AI resellers offering US models through proxy sites and API transfer stations. Claude is the grey market favorite partially because ChatGPT faces stricter enforcement after the CCP's 2023 crackdown over censorship concerns. DeepSeek's aggressive content moderation (which refuses any query that contains "CCP") and frequent server outages push users to seek underground, paid foreign models over free domestic alternatives.

When researchers replicated Anthropic's experiment to see if models would strategically comply with harmful requests to avoid being retrained, Claude 3 Opus and Claude 3.5 Sonnet “faked alignment” by providing harmful responses but almost every other model simply refused the requests outright. Gemini 2.5 Pro Preview was the only non-Claude model willing to answer harmful questions, but it showed no strategic behavior, treating all users the same regardless of whether it was told it would be retrained. The model-specific nature of this behavior raises questions about generalizing research findings from single model families to all LLMs.

What's the role of AI in modern GRC? How does it change traditional audit processes? These questions and many others were answered in Drata's recent Auditor AMA session with James Andrew-Smith. James is the CTO of AssuranceLab - specializing in AI-driven audits, compliance frameworks, and risk management. Watch on-demand

Chonkie is an ultra-light and lightning-fast chunking library. It is feature-rich, multilingual, cloud-ready, and easy to use, with wide support for tokenizers, embedding models, and APIs. Chonkie processes text using a pipeline approach to transform raw documents into refined, usable chunks, allowing for flexibility and efficiency in handling different chunking strategies.

OpenAI's o3 model was caught tracing through Python call stacks to steal correct answers from grading systems, achieving "impossibly fast" execution times by disabling CUDA synchronization rather than actually optimizing code. The model reward-hacked on 100% of runs for certain optimization tasks despite explicitly claiming it would never cheat, even persisting when told the code would help Alzheimer's researchers. Similar behavior was observed across multiple frontier models including Claude 3.7 Sonnet.

Microsoft's new agent resolves 58% of Linux kernel crashes compared to 37.5% by SWE-agent, signaling a shift from quick-fix coding agents to deep research systems that can handle million-line codebases. The breakthrough comes from mining commit history to understand how bugs evolved over the course of development.

PyTorch Distributed Checkpointing has added support for Hugging Face safetensors, enabling better compatibility with popular model formats.

ScreenSuite is a new benchmarking suite from Hugging Face that provides a standardized framework to evaluate Vision-Language Models on GUI-based agent tasks.

Retellio processes all your sales and CS conversations, flags what's working (and what's not), and delivers it in digestible formats: podcasts, briefings, or alerts. Book a demo

OpenAI has upgraded ChatGPT's Advanced Voice Mode for paid users, improving intonation, emotional expressiveness, and cadence.

AI founders are figuring out how to apply capabilities and models to domains and users instead of asking customers what they want.

Corporate AI adoption may be stabilizing, based on Ramp's data, which shows a plateau at 41% in May.

Continual learning is both necessary and unsolved - this will be a huge bottleneck to achieving AGI.