TLDR AI 2025-05-19

Google Cloud's survey of 500+ global technology leaders validates what everyone is feeling: generative AI is now mainstream (sorry AI hipsters), and it's present at the vast majority of larger organizations.

However, not all is smooth sailing, and many are still figuring out the best way to build an AI-ready cloud that's robust, secure, and cost-effective.

The 2025 State of AI Infrastructure report offers a glimpse into how enterprises are preparing for the next tidal wave of AI adoption.

Get a free copy right here 👉

OpenAI just launched Codex, an autonomous coding agent that writes features, fixes bugs, and submits pull requests in isolated cloud environments. Companies like Cisco and Temporal are already using it to handle entire codebases while their engineers focus on higher-level work. The agent can tackle multiple tasks simultaneously, run tests, and provide detailed citations of every change it makes.

Anthropic has secured a new line of credit to further accelerate its growth, adding to the $18.2 billion already raised from investors. The company reported that its annualized revenue doubled to $2 billion in the last 6 months. The move mirrors OpenAI's strategy, which established a $4 billion revolving credit line last October.

President Trump's Middle East trip has yielded massive AI deals, including Nvidia shipping 18,000 Blackwell chips to Saudi Arabia and plans for the largest non-US AI data center in Abu Dhabi. Critics warn these partnerships could enable China to access restricted US AI technology through its extensive economic ties with Gulf states prompting new bipartisan legislation to prevent such transfers.

Claude 3.5 Sonnet significantly outperformed human persuaders in a controlled experiment where both attempted to influence participants' answers on quiz questions. Claude achieved 7.6% higher success rates when trying to convince participants to select specific answers and was more effective both at promoting correct answers (+12.2%) and incorrect answers (-15.1%).

Superhuman coding by AI will likely occur by 2033, later than AI Futures' estimates of 2028-2030. This delay is due to factors like handling engineering complexity, working without feedback loops, and achieving cost and speed. Unforeseen challenges, such as geopolitical issues or changes in lab priorities, could further push out timelines.

Major algorithmic breakthroughs for LLMs, like the transformer architecture, multi-query attention, and mixture-of-experts systems, only demonstrate substantial benefits (10-50x performance gains) when implemented with massive amounts of compute. This challenges predictions of rapid AI self-improvement since hardware constraints like chip export controls, power consumption limits, and cooling infrastructure bottlenecks would significantly limit any "intelligence explosion" scenario.

The world's biggest event for AI builders is in SF from June 3-5. Speakers include leaders from OpenAI, DeepMind, Anthropic, and more, click here to get 30% off your ticket.

WavReward is an audio-language model-based evaluator for assessing spoken dialogue systems on both cognitive and emotional metrics. It is trained on ChatReward-30K, a dataset containing diverse preference-labeled audio interactions.

ParScale has introduced a third LLM scaling paradigm by leveraging parallel computation at both training and inference time.

OpenAlpha_Evolve is an open-source Python framework inspired by the recently released technical paper for DeepMind's AlphaEvolve.

Organizations are prioritizing generative AI over security spending for 2025. Companies are creating leadership roles like Chief AI Officer and adopting aggressive hiring and internal development strategies for AI talent. Many are using a hybrid model, combining off-the-shelf AI models with custom applications using proprietary data.

The challenge invites competitors to use o3, o4-mini, and GPT-4.1 models with open-source data to pinpoint undiscovered archaeological sites in the Amazon by June 29. Winners will receive funding to verify their findings in the field alongside archaeologists.

The "AI button" UX pattern is limiting and creates unnecessary distinctions between AI-assisted and manual workflows. Better alternatives, like integrating AI as a "shadow teammate" within workflows, enhance collaboration without fragmenting user experience.

Cohere has acquired Ottogrid, a Canadian startup focused on automating enterprise market research workflows.

Stability AI is open-sourcing Stable Audio Open Small, a 341 million parameter text-to-audio model optimized to run on Arm CPUs.

U.S. lawmakers raised concerns over Apple's reported partnership with Alibaba, questioning the handling of user data and transparency amid rising U.S.-China tech tensions.