Today’s lead · TechCrunch
Anthropic ships Mythos 5.2 and cheaper, faster Fable 5.2
Anthropic’s new Claude Mythos 5.2 and Fable 5.2 bring claimed gains in coding, reasoning, vision, clarity, speed, and enterprise ZDR support. The builder takeaway is mixed: Fable is cheaper and benchmarking well, but the release also foregrounds harder evaluation and covert-capability questions.
Top signals
5 more
Meta’s Muse Ember targets the cheap, repetitive work inside agents
Meta’s Muse Ember is positioned less as a frontier flex and more as agent plumbing: cheap tool calls, extraction, routing, coding subtasks, and long-document parsing at scale. If the cost and multimodal claims hold up in production, it fits the emerging planner-worker pattern for high-volume agents.
Google expands its 8th-gen TPU lineup with 8tx and 8i Max
Google added two specialized TPU systems: 8i Max for massive context windows and dense racks, and 8tx for secure confidential inference. The more interesting builder angle is the software push: native PyTorch and Hugging Face support via TorchTPU, TPU Sync for KV transfer latency, and Agent Compute Engine.
NVIDIA introduces GR00T 2 for dexterous robot manipulation
NVIDIA’s GR00T 2 release is aimed at the hard part of robotics: manipulation, not just locomotion. The claims are ambitious—cross-body skill transfer, few-shot adaptation, agentic control, and real-time humanoid execution—but useful if you’re tracking when robot policies move from lab demos toward repeatable industrial workflows.
Anthropic buys $3.4B of natural gas turbines as AI power gets scarce
Anthropic’s reported $3.4B turbine purchase is a reminder that AI scaling is now constrained by power logistics, not just GPUs. For builders, this matters indirectly: model capacity, pricing, and reliability increasingly depend on energy procurement decisions that look more like industrial infrastructure than software operations.
Acronis researchers detail BOSSAI agents used in real-world data theft
The BOSSAI case is not sci-fi autonomy: the report says one attacker manually chained AI agents, mostly using locally run abliterated models after commercial systems refused. That makes it more operationally relevant—AI lowered the effort for scanning, SQL injection, data collection, and tool maintenance across many targets.
Tools & repos
4 selectedanthropics/skills
Anthropic’s public Agent Skills repository is still pulling huge GitHub attention. The dossier describes it simply as a Python repository for Agent Skills.
MCP1
MCP1 is Appwrite’s single gateway for MCP servers, covering connections, permissions, credential protection, and usage monitoring across tools like Claude, Cursor, ChatGPT, and Codex.
ProtoMap
ProtoMap builds deterministic dependency maps for gRPC and Protobuf systems, giving engineers and AI coding agents a queryable way to understand cross-repo service, API, and field dependencies.
Databox CLI
Databox CLI brings analytics work into the terminal, with 90+ commands and 10 agent skills for Claude Code, Cursor, and Codex to create sources, metrics, and workflows.
Blogs worth your time
6 readsOpenAI’s Control Room is about stopping workloads, not just observing them
OpenAI describes Control Room as a central system for tracking model workloads across research infrastructure and rapidly stopping them when safety or security concerns arise.
Sourcegraph’s PR approval bot makes review policy explicit
Sourcegraph’s internal bot routes PRs into Ship, Show, or Ask paths. The useful lesson is not auto-approval; it is making review intent, familiarity evidence, and interruption cost visible.
LangChain argues agent work should start with evals, not prompt fiddling
LangChain’s eval-driven development loop is practical: define good behavior, write the eval before the fix, then compare experiments for regressions instead of trusting average scores.
Amazon’s HOPE prunes MoE experts by preserving partnerships
HOPE is a second-order MoE expert-pruning method from Amazon that accounts for pairwise expert interactions. The pitch: better memory cuts at aggressive pruning rates without retraining.
Vercel’s AI Gateway data says model choice is getting brutally fluid
Vercel’s October gateway data shows fast churn: decision models reached 26% of requests, Anthropic won new teams nearly 2:1, and over half of language requests hit models under 60 days old.
Simon Willison gives Claude Code ears to find a beeping box
A playful but revealing agent build: Claude Code helped create Beep Locator, using browser microphones, WebSockets, FFT detection, and hot-swapped listener code to triangulate a real sound source.
Community discussions
1 threadsAn RFC for replayable generative UI built around intents
The RFC proposes generative UI that persists as JSON and replays without an LLM on every refresh. The tension for builders: flexible user-mutated software versus predictable state, approvals, and data access boundaries.
Funding & acquisitions
1 moves
Nominal raises $240M Series C at a $3B valuation
Nominal’s $240M Series C underlines investor demand for AI-native tools outside pure software: the company targets engineering and testing workflows for aircraft, nuclear reactors, smart factories, and defense customers.
Bengaluru radar
4 events
Lossfunk × BIOS by Bio: Exploring the Future of Computational Drug Discovery
A free Bengaluru evening for researchers, ML engineers, roboticists, and biotech builders on AI scientists, drug design workflows, and robotic wet labs.
Build with Agora - Bengaluru 2026 | Voice AI Hackathon
A free, in-person Voice AI hackathon for 250+ builders using Agora’s conversational AI and real-time communication technologies, with $3,000 in prizes.
Cashfree × Sarvam AI Buildathon
A free one-day buildathon for agents that use Sarvam’s Indian-language models and Cashfree payment APIs to move real money in demos.
AI Tools that Earn: a hackathon by Apify
Apify’s sold-out Whitefield hackathon pushes builders to ship paid-for Actors, not demos, with creator mentors and cash prizes for top teams.