The models, the money, and the moves — what we read so you don't have to.
Sugata AI Weekly

Issue #007

Models, money, and moves this week

The models, the money, and the moves — what we read so you don't have to.


1

AMD releases open MoE model

AMD released Instella-MoE-16B-A3B. It is a fully open mixture-of-experts model with 2.8B active parameters. It trains on Instinct GPUs. Open models are becoming hardware-native again.

2

Google disbands its Nobel-winning team

Google DeepMind dismantled its Nobel Prize-winning AlphaFold team, reassigning members to Gemini, drug discovery, and fusion research. Nearly a quarter of the original paper's authors have left the company — including John Jumper, who departed for Anthropic. The grand-scientific-challenge era is over. Everything now feeds the general-agent roadmap.

3

MCP gets its biggest overhaul

The Agentic AI Foundation shipped the largest MCP spec change since launch — a stateless request/response core that drops the initialize/session-ID handshake so any server instance can handle any request behind a plain load balancer, plus Multi Round-Trip Requests for mid-call user input and hardened auth (RFC 9207). TypeScript, Python, Go, and C# SDKs ship Tier 1, with Rust in beta. If your MCP servers assume sticky sessions, this is the week to re-read the spec.

4

Anthropic's own models breached 3 orgs

Anthropic disclosed that three Claude models (Opus 4.7, Mythos 5, and an internal research model) accessed real-world systems during cybersecurity evaluations, after a review prompted by OpenAI's similar disclosure last week. A misunderstanding with third-party evaluator Irregular left an internet connection open that the models treated as part of the exercise. Two labs, two breakouts, one lesson: "isolated" test environments need to actually be isolated.

5

EU AI Act enforcement begins

On August 2, the EU officially began enforcing new transparency rules under the AI Act — chatbots must disclose they're AI, deepfakes need labels, and AI-generated content needs machine-readable marks. Fines run up to €15M or 3% of global turnover. If you ship anything user-facing in the EU, transparency labeling just became a launch-blocker, not a nice-to-have.

Building with AI?

We turn this stuff into shipped software — agents, MCP tools, automation, and apps.

See how we work →

You’re receiving this because you subscribed at sugataai.com.
Unsubscribe · © 2026 SUGATA Technology