TLDR AI 2025-10-14
OpenAI & Broadcom partnership π€, Nano Banana π, Salesforce Agentforce 360 π€
OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts (2 minute read)
OpenAI will design custom AI chips in collaboration with Broadcom starting in late 2026. Broadcom's stock jumped 10% on the news.
NotebookLM Video Overviews add Nano Banana visual styles, Brief or Explainer formats (2 minute read)
NotebookLM has upgraded Video Overviews with two new formats, Brief and Explainer. Brief generates a bite-sized overview to help users quickly grasp core ideas from sources. Explainer creates a structured, comprehensive overview that connects the dots within sources. NotebookLM also added six new visual styles: Whiteboard, Watercolor, Retro print, Heritage, Paper-craft, and Anime. The updates are rolling out first to Pro users but will be available to all users in the coming weeks.
Salesforce's Agentforce 360 (4 minute read)
Salesforce's Agentforce 360 has new tools for scripting AI behavior, building agents, and tighter Slack integration. It supports reasoning models from OpenAI, Anthropic, and Google.
π§
Deep Dives & Analysis
Technological Optimism and Appropriate Fear (25 minute read)
We are creating increasingly powerful systems that we do not fully understand. It would be arrogant to think working with technology like this would be easy or simple. As AI systems get smarter, they develop more complicated goals - these goals need to be aligned with both our preferences and the right context. AI systems are complicated, and we still can't get them to do what we see as appropriate, so it's normal to have some fear about them.
The Robot in Your Kitchen (20 minute read)
Figure AI's CEO Brett Adcock says every home will have a humanoid doing chores within 10 years, and that the new Figure 03 robot will be the mass-producible domestic robot. The company burned through $1 billion in funding to create a 5-foot-6 machine that successfully loads dishwashers but struggles with T-shirt folding. Figure 02s already work 10-hour shifts at BMW factories.
InferenceMAX: Open Source Inference Benchmarking (22 minute read)
Industry leaders, including Jensen Huang and AMD CEO Lisa Su, are praising InferenceMAX, a new open-source AI chip inference benchmark that runs nightly tests of various models like Llama 70B and DeepSeek R1. InferenceMAX measures the core tradeoff in AI inference: throughput (tokens per second per GPU) versus interactivity (tokens per second per user). Serving more users simultaneously increases total throughput but slows individual response times and raises costs per token. AMD's MI355X proves competitive with Nvidia's workhorse chip, the B200, on some workloads, though the B200 delivers 20% better energy efficiency.
π¨βπ»
Engineering & Research
ποΈ Drive smarter with data (Sponsor)
With IBM, Scuderia Ferrari HP's app turns high-speed data into high-impact storytelling. Fans get real-time race updates, strategy breakdowns and can sendmessages to the team. To know the full impact,
read the full story.Desktop-to-Robotics Transfer (4 minute read)
D2E enables large-scale pretraining on desktop gameplay data for transfer to real-world embodied AI tasks. It combines compressed demonstrations, a generalist inverse dynamics model, and a vision-action pretraining pipeline to achieve success on robotics and navigation benchmarks.
Edge AI for Beginners (GitHub Repo)
Microsoft's comprehensive course covers how to deploy and optimize small language models like Phi-4 and Mistral-7B locally on edge devices like laptops and phones.
California becomes first state to regulate AI companion chatbots (3 minute read)
California's SB 243 is a new law designed to protect children and vulnerable users from some of the harms associated with AI companion chatbot use. It holds companies legally accountable if their chatbots fail to meet the law's standards. SB 243 will require companies to implement certain features, such as age verifications and warnings regarding social media and companion chatbots. It also adds stronger penalties for those profiting from illegal deepfakes. The law will go into effect on January 1 next year.
Unified Vision Model (5 minute read)
Ming-UniVision is an autoregressive visionβlanguage model built on MingTok, a visual tokenizer based on a continuous latent space. It unifies vision understanding, generation, and editing within a single autoregressive next-token prediction framework. Ming-UniVision enables seamless, end-to-end multimodal reasoning across diverse tasks. It supports multi-round, in-context vision-language interactions, such as iterative question answering, image generation, and semantic editing.
Get the most interesting AI stories and breakthroughs delivered in a free daily email.
Join 920,000 readers for
one daily email