SSI gets Nvidia compute at $32B | Claude beats benchmark

Stop typing what you could say in 10 seconds.

Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.

Try Wispr Flow free

In today's issue

01 · Business

Safe Superintelligence lands Nvidia deal for order-of-magnitude compute boost

Safe Superintelligence, the two-year-old research lab founded by Ilya Sutskever, has secured a multi-billion-dollar investment from Nvidia and access to its next-generation Vera Rubin GPU platform. The deal will increase SSI's computing power by roughly 10 times, enabling the lab to scale its research into AI safety and reasoning without shipping products or chasing revenue.

The details:

SSI has now raised $7 billion total at a $32 billion valuation, with backers including Nvidia, Andreessen Horowitz, Alphabet, Sequoia, and Lightspeed.
Vera Rubin is Nvidia's next-generation GPU architecture; SSI positions itself as an early flagship customer and will collaborate with Nvidia on advancing future compute platforms.
SSI also partners with Google Cloud, meaning the lab now has two of the largest compute suppliers in the industry funding its runway while it pursues foundational research instead of shipping products.

Why it matters: Sutskever is betting that deep research on AI alignment and reasoning, freed from product pressure, will eventually matter more than speed-to-market. The level of compute backing and investor confidence suggests the market agrees — at least for now — that there's value in the long game.

Read the full story →

02 · Models

Claude Opus 4.7 finishes a two-week coding job in 14 hours for $251

Claude Opus 4.7 reimplemented a 61,000-line Apple software program from scratch in 14 hours for $251, doing work that would take a human developer 2-17 weeks. A new benchmark called MirrorCode tests whether AI can rebuild real production software by probing it like a black box—no source code, no internet—and frontier models are now solving it at scale.

The details:

Out of 25 target programs, 17 achieved perfect reimplementation in at least one run and 4 more hit 99%; only 8 programs remain unsolved to 100%.
Both Opus 4.7 and OpenAI's GPT-5.5 reimplemented gotree, a 16,000-line parser, across multiple programming languages for $100–$400 per run.
In a separate robotics test, Opus 4.7 completed a quadruped task suite in 9 minutes autonomously—roughly 20x faster than the 181-minute human record from earlier this year.

Why it matters: Frontier models can now infer the structure of unfamiliar software systems and rebuild them from behavior alone, a capability that mirrors what an AI agent would need to operate inside enterprise systems it has never been shown documentation for. The hardest problems remain those with intricate edge cases and narrow domain requirements, like Python linters and email authentication standards.

Read the full story →

03 · Analysis

Google AI Overviews now appear in 43% of searches, up from 15% a year ago

Google AI Overviews now show up in 43% of all searches, nearly triple the rate from a year ago. This marks a fundamental shift: Google is becoming a destination where it answers your question directly, instead of just pointing you to other websites.

The details:

Google AI Mode visits jumped 121% in eleven months, climbing from 126 million in June 2025 to 279 million by May 2026.
Publishers are losing referral traffic because users read AI-generated answers and never click the links below; cited sources have risen fivefold but citations alone don't drive clicks.
ChatGPT sends far fewer users to external websites: only 6.8% of ChatGPT desktop searches included citations in May 2026, versus Google's higher rate.

Why it matters: Google is keeping users on its own platform longer, which hurts the websites it once funneled traffic to. Publishers are testing new defenses—like requiring payment to let AI systems index their content—but it's unclear if those will work at scale.

Read the full story →

AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

● ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat

● Generate images & video with Sora, Veo, Ideogram

● Compare any two models side by side

● From $8.99/mo · 80+ models, all included

Try AI Box →

aibox.ai

Trusted by 3,000+ teams

04 · Security

Nvidia, Microsoft, IBM launch Open Secure AI Alliance to defend agents

Nvidia, Microsoft, IBM, and 30+ other companies launched the Open Secure AI Alliance on July 27 to build open-source cybersecurity tools for AI agents. The coalition argues that closed AI systems block defenders from investigating breaches — and is lobbying against government restrictions on open AI that could leave only a handful of companies controlling critical security tools.

The details:

Hugging Face used an open-weight GLM 5.2 model to analyze 17,000 actions during its own security breach after closed AI systems refused to run the forensic analysis.
Nvidia is open-sourcing NOOA (NVIDIA Labs Object-Oriented Agent framework) on GitHub to make AI agent behavior easier to test, trace, and audit.
The alliance directly opposes OpenAI and Anthropic's calls for restrictions on open-weight AI, arguing such blanket limits would concentrate power with a few closed providers.

Why it matters: Security teams now have a coalition arguing they need hands-on access to AI tools during live attacks — not APIs they can't inspect or modify. This positions open-source AI as infrastructure, like Linux, rather than a consumer product, and sets up a real policy fight in Washington over who controls the AI defending banks and hospitals.

Read the full story →

05 · Business

Enigma exits stealth with $70M to rethink how humans talk to robots

Enigma, a startup founded by former Microsoft and Israeli intelligence veterans, just raised $70 million to test a radical bet: that robots fail not because AI models are weak, but because the way humans control them is clunky. The company is opening 100+ of its own robots to anyone on the internet to remote-control them and generate data on what actually works.

The details:

Enigma built both the robotic arms and the AI models powering them from scratch — an unusually complete operation for a company less than a year old.
Users can already remotely direct the robots to draw pictures, fence with swords, or run chemistry experiments through the internet from Israel and California hangars.
The $70M round is designed to fund a data-collection experiment on which input methods work best: text, voice, video demonstration, or direct manipulation of what the robot sees.

Why it matters: If Enigma is right that interface design beats raw model power, robotics could move faster by treating this as a UX problem rather than a pure AI problem. The public experiment approach — letting strangers use real robots and watching what they naturally try — is fundamentally different from how competitors train robots, and could uncover design patterns the field has missed.

Read the full story →