This week, the U.S. FTC has opened an inquiry into major chatbot makers like Alphabet, Meta, and OpenAI. Regulators want to know how these companies test, evaluate, and mitigate the risks of their AI systems. It’s not regulation yet, but the move signals a tightening focus on safety, transparency, and accountability in large language models. For developers and product teams, it’s a reminder that robust evaluation and monitoring practices aren’t optional — they’re fast becoming table stakes.

GitHub’s Spec-Kit brings spec-driven development into mainstream AI coding workflows, giving developers a structured way to generate more reliable and maintainable code with AI tools, while Google launches VaultGemma, a privacy-preserving model that matches benchmark performance of larger peers but with strong differential privacy guarantees and open-source weights.

And… Chinese scientists claim their SpikingBrain1.0 model achieves 100× efficiency gains over ChatGPT using brain-inspired processing.

U.S. regulators probe major chatbot makers over safety, testing, and harm-mitigation practices

The U.S. Federal Trade Commission (FTC) has initiated an inquiry (begun September 11, 2025) into how companies that build consumer-facing chatbots (including big names like Alphabet, Meta, and OpenAI) evaluate, test, and monitor negative impacts of their AI tools.

AI regulation is tightening, especially around safety, transparency, and harms in large language model tools. This inquiry signals that regulators want more accountability and oversight in how these models are deployed and used.

Beyond innovation, companies will increasingly need strong internal processes for harm assessment, monitoring, and safety. Expect pressure on transparency of evaluations, model performance, possibly even audit trails.

“The FTC…seeking information from seven companies including Alphabet, Meta and OpenAI that provide consumer-facing AI-powered chatbots, on how these firms measure, test, and monitor potentially negative impacts.”

This is an inquiry, not yet regulation. The process may take time. The outcome might result in guidelines, enforcement, or new standards — depending on what the FTC finds and how companies respond.

The new research initiative “Defeating Nondeterminism in LLM Inference” aims to make LLM outputs deterministic: given the same input, the model gives the same output every time. This would improve reliability and trust in AI systems.

A model called SpikingBrain1.0 mimics more “local” neuron-style processing (only nearby words in context, etc.), which could be far more efficient. It reportedly uses Chinese hardware (MetaX chips), trained with much less data, for big speed/efficiency gains.

Meta has set up a research unit (TBD Lab) of “a few dozen” researchers focused on next-gen foundation models, pushing further into large-scale capabilities.

Analysis covering submissions to AACR journals shows rising amounts of text likely generated by LLMs in abstracts, methods, and peer-review submissions. Also notes that most authors did not disclose AI use even when required.

A hands-on look at GitHub’s new Spec-Kit, an open-source toolkit for spec-driven development. The review covers how it helps developers move away from vague prompts toward structured specifications, ensuring AI-generated code is more reliable, maintainable, and aligned with project goals. It also highlights Spec-Kit’s CLI, integration with popular AI coding assistants, and some early limitations.

Google introduces VaultGemma, a differentially private version of its Gemma architecture. Comparable performance to some non-private models on benchmarks, while offering stronger privacy protections. Open-sourced weights and code.

Another week of separating AI signal from noise. If we saved you from a demo that would've crashed prod, we've done our job.

📧 Got a story? Reply with your AI tool wins, fails, or war crimes. Best stories get featured (with credit).

📤 Share the skepticism: Forward to an engineer who needs saved from the hype. They'll thank you.

✍️ Who's behind this? The Augment Code team—we build AI agents that ship real code. Started this newsletter because we're tired of the BS too.

🚀 Try Augment: Ready for AI that gets your whole codebase?