In partnership with
Hello from The AI Night,
Today in AI:
OpenAI Model Cracks Erdős Unit Distance Conjecture
Anthropic Launches Self-Hosted Sandboxes For Managed Agents
Cohere Launches Command A+ as Open Source
Here's the deal: An internal OpenAI reasoning model autonomously disproved Paul Erdős's 1946 unit distance conjecture, a central open problem in combinatorial geometry. External mathematicians including Tim Gowers verified the proof, which constructs configurations exceeding the long-assumed upper bound.
The Breakdown:
The 1946 problem asks how many pairs among n points in the plane can sit exactly distance 1 apart.
For decades, rescaled square grid constructions producing n^(1+C/loglog(n)) growth were believed essentially optimal.
The model produced configurations with n^(1+δ) unit distance pairs for a fixed δ > 0.
A follow-up refinement by Princeton's Will Sawin shows δ can be at least 0.014.
The construction draws on algebraic number theory, using infinite class field towers and Golod-Shafarevich theory.
The model was general-purpose, not trained for math or scaffolded for proof search.
The bigger picture: Mathematicians assumed AI would verify proofs before discovering them. OpenAI skipped that step entirely. A general-purpose model connected algebraic number theory to discrete geometry in ways specialists never attempted. That is not retrieval. It is original reasoning. The question for every research field just shifted from "can AI help us" to "what has it already seen that we missed.
Here's the deal: Anthropic added self-hosted sandboxes and private MCP tunnels to Claude Managed Agents. The agent loop stays on Anthropic's side, while tool execution moves into the customer's perimeter.
The Breakdown:
Self-hosted sandboxes are in public beta. MCP tunnels are in research preview.
Four partners: Cloudflare microVMs with zero-trust secrets, Daytona stateful sandboxes over SSH, Modal containers with sub-second startup, and Vercel sandboxes with VPC peering.
Customers control compute sizing, runtime images, network policies, and audit logging.
MCP tunnels open one outbound connection from a customer-deployed gateway, no inbound firewall rules or public endpoints.
MCP tunnels work in Managed Agents and the Messages API, configured by org admins in the Claude Console.
Amplitude, Clay's Sculptor, and Rogo are early production users.
The bigger picture: Let me check with security" has stalled more agent deployments than any technical limitation. Anthropic stopped trying to win that conversation and eliminated it instead. Model calls stay with Anthropic. Code and customer data never leave your perimeter. Four sandbox partners already built the infrastructure. The security excuse is gone. Now the only question left is whether the agent delivers.
Here's the deal: Cohere released Command A+, a 218B-parameter mixture-of-experts model, under Apache 2.0. It combines reasoning, multimodal, tool use, and multilingual capabilities from the Command A family into one open-weight model built for enterprise agentic workflows.
The Breakdown:
Sparse MoE design activates only 25B of 218B total parameters, cutting inference cost while preserving capability.
Supports 128K input context, 48 languages, and text plus image inputs with tool use.
Runs on two H100s or one B200 GPU at W4A4 quantization with near-zero quality loss.
Agentic benchmarks jumped sharply. τ²-Bench Telecom rose from 37% to 85% over Command A Reasoning.
Delivers 63% higher output tokens per second, with speculative decoding adding 1.5-1.6x further speedup.
The bigger picture: European banks, Gulf governments, and Asian telecoms all need the same thing. Frontier AI that never leaves their building. Cohere built exactly that. Two H100s, 48 languages, Apache 2.0. Meta and Mistral chase the same buyers but neither jumped from 37% to 85% on telecom benchmarks in one generation. When the first question is "where does our data go," Cohere just became the answer.
2026 State of AEO Report
A year ago, most marketers weren't thinking about AI search. Now it's one of the fastest moving channels in the industry and nobody has a playbook yet.
So we built one. We surveyed hundreds of marketers to find out how they're approaching answer engine optimization, where they're investing, what's actually working, and what isn't.
The result is the 2026 State of AEO Report. Real data. Real strategies. A clear picture of where AI search is headed and how to get ahead of it.
What else you need to know:
Exa raised $250M in Series C funding at a $2.2B valuation led by a16z, building a search engine designed for AI agents now serving 5,000+ companies and 400,000+ developers.
CapCut is partnering with Google's Gemini app, letting users edit images and videos directly inside Gemini using CapCut's creative and editing tools, with the integration rolling out soon.
Google is replacing Gemini CLI with Antigravity CLI, a faster Go-built terminal tool, and will stop serving Gemini CLI requests for free, Pro, and Ultra consumer users on June 18, 2026.
NousResearch's Hermes Agent gained access to hundreds of browser skills through Browserbase's new Browse.sh hub, letting agents perform internet tasks more reliably from a shared catalog users can contribute to.
Alibaba launched Qwen3.7-Max, an agentic model that ran 35 hours autonomously and made over 1,000 tool calls, alongside its Zhenwu M890 chip that triples its predecessor's performance.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
2 ways to support us:
Forward this to your AI-curious friend → https://www.theainight.com
Sponsor The AI Night and reach 500+ AI builders daily → passionfroot.me/theainight