AI Guardrails Are Getting in the Way of Legitimate Security Research, Researchers Say — techcrunch.com

Mobileye founder Amnon Shashua to step down as CEO amid robotaxi and robotics push — techcrunch.com

Alexa Plus update expands smart home control and third-party reach — theverge.com

FCC Chairman Brendan Carr faces mounting criticism over free speech record — theverge.com

LG pulls McAfee pop-up ads from monitor app after Microsoft steps in — arstechnica.com

AI Guardrails Are Getting in the Way of Legitimate Security Research, Researchers Say — techcrunch.com

Mobileye founder Amnon Shashua to step down as CEO amid robotaxi and robotics push — techcrunch.com

Alexa Plus update expands smart home control and third-party reach — theverge.com

FCC Chairman Brendan Carr faces mounting criticism over free speech record — theverge.com

LG pulls McAfee pop-up ads from monitor app after Microsoft steps in — arstechnica.com

Editor's note — A thinner-than-usual day for research and open source, but a massive one for silicon. AMD went full-stack at Advancing AI 2026, and the inference chip money keeps flowing.

AMD goes full-stack

Helios is AMD's biggest bet yet — and OpenAI is already on stage for it

AMD didn't just launch a GPU at Advancing AI 2026. It launched an entire rack: 72 MI455X chips with 432GB of HBM4 each, 2nm compute chiplets, and a claimed 2.9 exaflops of FP4 inference per rack. The customer list reads like the frontier lab roster — OpenAI, Anthropic, Meta, Microsoft, Oracle — with Sam Altman personally confirming Helios goes online in Q4 2026. The claim that matters: 30% more inference tokens per dollar than the leading competitor. If that holds in production, the Nvidia tax just got negotiable.

AMD · 2026-07-23 · Read the dispatch →

Etched defies the skeptics: $300M Series C at a $10.3B valuation

The transformer-ASIC bet is paying off. Sequoia led the round with a16z, Jane Street and SK hynix, and Etched claims $1B in booked orders for its inference chips.

TechCrunch · 2026-07-23

xAI's Grok Build can now orchestrate 1,024 parallel agents

Workflows turns plain-language tasks into fan-out agent swarms that verify results and save themselves as reusable slash commands. The agent-orchestration race is moving from frameworks to products.

x.ai · 2026-07-23

The Edge

Vera Rubin NVL72 delivers 5x more tokens per dollar than GB200, says SemiAnalysis

The deep dive numbers are in: a new 3-bit programmable LUT tensor core, $3.57 per GPU-hour TCO, and — the kicker — at 350 tokens/second/user, neither GB200 nor GB300 can serve the workload. Only Rubin can. CoreWeave's own benchmarks go further, claiming 10x throughput per megawatt. The power-constrained datacenter era has its new unit of account.

SemiAnalysis · 2026-07-22

CoreWeave benchmarks Vera Rubin: 10x tokens per megawatt over Blackwell

First independent benchmark results confirm Nvidia's efficiency claims, measured on DeepSeek R1 at matched interactivity. For datacenter operators planning next-gen builds, tokens-per-megawatt is now the metric that decides procurement.

CoreWeave · 2026-07-22

AMD's Kria platform is an open bet on physical AI

CPU, GPU and NPU on a single COM-HPC module with unified memory, targeted at autonomous robotics. While Nvidia sells ecosystems, AMD is selling openness — the industry's first fully integrated open platform for robotics developers.

UC Advanced · 2026-07-23

Corgi's third round in eight weeks: AI insurance at a $4B valuation

Three rounds in eight weeks is either a bubble signature or a signal that AI-native insurance is compounding faster than anyone modeled. $450M projected ARR suggests the latter — for now.

TechCrunch · 2026-07-23

AegisAI raises $36M to fight AI-driven spear phishing with AI

Founded by former Google security execs and backed by Battery, Accel and Foundation Capital. The attack surface is now LLM-generated; the defense has to be too.

TechCrunch · 2026-07-23

Open Pulse

Echo routes across open-weight models to hit Fable-class quality at a third of the cost

An adaptive model built on a pool of open weights — GLM-5.2, Kimi K2.7 and others — that dynamically picks the cheapest adequate model per task. The frontier labs sell intelligence; Echo sells arbitrage on it.

Hacker News · 2026-07-23

Runway's Media Router picks your generative model by cost, speed or quality

As generative media gets crowded, the value moves up the stack to routing. Runway is betting developers would rather delegate model selection than benchmark every new release themselves.

TechCrunch · 2026-07-23

PromptQL wants shared AI threads to replace your Slack channels

From the Hasura team: an AI-native workspace that captures tribal knowledge across databases and SaaS apps, builds task-specific agents on the fly, and compiles a self-writing wiki of how the business actually operates. Ambitious — but the knowledge-compounding angle is the right one.

Product Hunt · 2026-07-24

OneCLI keeps your API keys out of the agent's context window

An open-source credential gateway that intercepts secrets before they reach AI agents. Every team running coding agents against production services needs this layer; almost none have it.

Hacker News · 2026-07-23

OpenCode Superapp: Codex-class coding with local models and voice

A unified coding assistant that runs online or offline, pairing agentic coding with self-hosted models. The local-first coding stack keeps maturing.

Product Hunt · 2026-07-24

In Brief

Source:

Helios is AMD's biggest bet yet — and OpenAI is already on stage for it

Etched defies the skeptics: $300M Series C at a $10.3B valuation

xAI's Grok Build can now orchestrate 1,024 parallel agents

The Edge

Vera Rubin NVL72 delivers 5x more tokens per dollar than GB200, says SemiAnalysis

CoreWeave benchmarks Vera Rubin: 10x tokens per megawatt over Blackwell

AMD's Kria platform is an open bet on physical AI

Corgi's third round in eight weeks: AI insurance at a $4B valuation

AegisAI raises $36M to fight AI-driven spear phishing with AI

Open Pulse

Echo routes across open-weight models to hit Fable-class quality at a third of the cost

Runway's Media Router picks your generative model by cost, speed or quality

PromptQL wants shared AI threads to replace your Slack channels

OneCLI keeps your API keys out of the agent's context window

OpenCode Superapp: Codex-class coding with local models and voice

YouTube Signals

Google's quantum computer now corrects its own errors mid-computation

The Kimi K-3 distillation allegations, and why investors are spooked

Apple is reportedly refreshing every Mac for the AI demand wave

In Brief