Reading

Most Americans want AI development to "slow down."

According to a Quinnipiac University poll released on September 30, 77% of US respondents want to reduce the pace or fully stop AI development until its safety has been verified. Only 5% favor faster progress.

Public trust in AI companies remains low. The survey found that 86% support independent safety standards, even if they hinder innovation; meanwhile, 74% express little or no trust in the leaders of AI companies. Researcher Chetan Jaiswal noted that these results suggest the public wants safeguards that do not depend simply on taking company leaders at their word.

Microsoft and Meta are significantly reducing their internal use of Anthropic’s Claude models, according to a report by The Decoder citing The Information.

At Microsoft, annual Claude spending had been projected at over $1 billion. Executives Scott Guthrie and Jay Parikh have reportedly instructed employees to switch to in-house tools like GitHub Copilot and OpenAI models. The monthly per-employee budget in the cloud division dropped from $100,000 to approximately $10,000, representing a spending cut of more than one-third.

Meta has seen its Claude Code user base fall from roughly 60,000 to 30,000. This decline is partly due to layoffs but also reflects Meta’s push for its own tools, such as Muse Code and MetaCode. In a recent 28-day period, Meta spent over $105 million on Claude Code alone.

This move highlights a broader trend among large tech companies to reduce dependency on external model providers and strengthen proprietary AI infrastructure. For Anthropic, the contraction of its core customer base poses potential challenges to its enterprise revenue growth.

Reflection AI released its first open-weight model, Beam, on October 5, positioning it as a "workhorse" for Western enterprises.

The model uses a sparse Mixture-of-Experts architecture with 501 billion total parameters and 23 billion active. CEO Misha Laskin stated that Beam excels in coding and agentic tasks, requiring three to four times less computing power for reasoning than comparable open models.

Official benchmarks indicate Beam's performance is comparable to China's Z.ai GLM-5.2 and approaching Alibaba's Qwen 3.8-Max. The launch targets businesses and governments seeking sovereign AI options without relying on Chinese infrastructure.

The model is currently undergoing final red-teaming and evaluations, with weights and technical reports scheduled for release later this month.

Amazon AWS hired Pedram Rezaei as a Distinguished Engineer on October 5, joining its Agentic AI and Emerging Technologies group.

Rezaei spent over two decades at Microsoft, most recently serving as Copilot's CTO until June and working on foundation model training within the Superintelligence team. He will initially focus on bolstering Quick, AWS's AI assistant, and Kiro, its AI-powered coding tool. Internal emails note rapid growth for these products: Kiro's user base doubled in Q2, while Quick's monthly AI users grew fourfold.

This hire signals Amazon's strategy to acquire top-tier technical talent from competitors to accelerate the deployment of widely used AI agent products.

CoreWeave announced Model Distillation, RL Rollouts, and a programmable training API on October 5 to close the loop between inference and post-training.

Traditionally, model improvement halts at deployment. These tools allow teams to capture production traffic failures as training data automatically, bypassing manual infrastructure setup.

In a cited example, customer Method used GPT-4o outputs to train a Llama 3.1 8B model for bank IVR systems. After iterating with fresh production data via the new workflow, the updated model won 57.5% of head-to-head evaluations against the previous production version. Another customer, Willow, combined supervised learning and reinforcement learning to improve text styling.

Note that these capabilities are largely in preview or newly launched states. Performance depends on user-specific data quality, and the cited metrics are vendor-reported without independent verification.

CoreWeave Senior Director Harsh Banwait reported that Nvidia's Vera CPU improved Sandboxes startup times by approximately 3x.

Sandboxes provides isolated environments for AI agents to execute tasks during reinforcement learning and inference. Banwait noted this addresses the bottleneck of waiting for environment initialization before agent task execution.

The data comes from CoreWeave's internal testing and has not been published with detailed benchmarks or independently verified.

Ghost announced its emergence from stealth today, completing an $11 million seed round led by Andreessen Horowitz.

Its first product, Core, is a personal computer designed specifically for running AI agents, priced at $3,499. The device includes an Nvidia RTX Pro 4000 SFF Blackwell GPU and comes pre-installed with models like Qwen and Gemma, aiming to let users process data and execute tasks locally without relying on cloud services.

The company states that all source code, logic, and model weights reside on the device, giving users full control. Pre-orders opened Monday, with shipping scheduled for the last week of October.

Instinct announced on October 5 that it is bringing its AI agent into group chats, enabling participation even from friends who do not have an Instinct account.

The feature allows users to add the agent to conversations to help with tasks like travel planning, carpooling, or event coordination. Founder Noah Shinn stated on X that personal agents are siloed from group agents, which cannot access personal data. The system requires explicit user permission before sharing information or taking action.

This move targets the consumer AI market, competing directly with rivals like Meta’s Muse. The feature is currently rolling out to early access users, with a full rollout timeline unspecified.

Having large numbers of phones jointly train one model, with raw data staying on each device and only the learned updates uploaded — this kind of approach is called federated learning. A long 2019 survey listed the problems this approach has yet to solve: unreliable devices, differing data across participants, constrained bandwidth, hard-to-prove privacy — and these are all entangled with one another.

Today, AI offerings in input methods, healthcare, and finance love to pitch with one line: data never leaves the device, so privacy is solved. Later research is still working through the items this survey listed one by one, and no single method solves them all. When you hear that pitch, press with the checklist: what about differing data across participants? How do you prove nothing was actually peeked at?

If you want to compare the merits or performance numbers of specific methods, don't use it to judge — it ran no experiments, only an inventory of problems. And if your scenario is learning without a central server, where devices connect directly to each other, the problems it lists don't match either.

Advances and Open Problems in Federated Learning (2019) | Next review 2027-09-20