Presented by
Good morning, AI enthusiasts. Brain-computer interfaces just took a major leap, and Meta didn't need to drill a single hole to do it. Brain2QWERTY v2 reads raw brain waves through a wearable headset while you type, approaching the accuracy of surgical implants with none of the surgery.
If it scales beyond the lab, the next interface layer for human-to-computer interaction may not need a screen, a mic, or a pair of hands. Does this mark the beginning of the end for keyboards?
In today's recap:
Meta's non-invasive brain-to-text system, no surgery required
South Korea's $576B AI chip and humanoid robot push
Apply 12 Factors to Ship Reliable AI Agents
X opens real-time social data to any AI agent
3 new AI tools, prompts, and more
META
Meta decodes brain waves into text without surgery
Meta
Recaply: Meta just unveiled Brain2QWERTY v2, a non-invasive AI system that decodes typed thoughts in real time through a wearable MEG headset, with no surgery or implants required.
Key details:
Brain2QWERTY v2 uses end-to-end deep learning to decode raw magnetoencephalography signals captured while participants typed, bypassing older hand-crafted event-detection pipelines and approaching the accuracy of surgical brain interfaces.
The model trained on approximately 22,000 sentences from 9 volunteer participants, each wearing a MEG device for 10 hours, targeting the millions of people with brain lesions who can't communicate through speech or movement.
Meta is releasing full training code for both v1 and v2 publicly, and partner institution BCBL (Basque Center on Cognition, Brain, and Language) is releasing the full v1 dataset on HuggingFace.
Published June 29, 2026; training code and dataset available now. No commercial deployment timeline has been announced.
Why it matters: The gap between surgical implants and mass-accessible brain interfaces has always been wide. Brain2QWERTY v2 begins closing it, delivering near-surgical decoding accuracy from a wearable device that doesn't require any procedure. If the accuracy holds as the system scales beyond lab conditions, it doesn't just help millions with brain lesions who can't communicate. It signals that the next layer of human-to-computer interaction may not need a keyboard, a screen, or even a voice.
PRESENTED BY AHREFS
AI Marketing Agent With Ahrefs Data Access
AI tools are great at answering questions. But what if they could actually do the work?
Agent A by Ahrefs is an AI marketing agent with access to Ahrefs data and reports. It can analyze competitors, find content gaps, identify technical issues, track brand visibility in AI search, create content strategies, build reports, and much more.
Unlike generic AI assistants, Agent A comes with pre-built marketing skills and can connect to tools like Slack, Notion, HubSpot, WordPress, and Linear to fit directly into your workflow.
See what an AI agent built specifically for marketers can do at ahrefs.com/agent-a.
AI INFRASTRUCTURE
South Korea commits $576B to AI chip dominance
Recaply: South Korea's government just committed $576 billion to AI chip production anchored by Samsung and SK Hynix, targeting doubled DRAM output within five years and commercial humanoid robot deployment by 2028.
Key details:
Samsung and SK Hynix committed 800 trillion won ($518.3B) to build new chip fab clusters in South Korea's southwest, while SK Group, GS Group, and Naver add $357B in AI data center investments across multiple provinces.
The total initiative tops $1 trillion; South Korea controls global HBM supply critical to AI processors, with Samsung and SK Hynix as the two largest memory chipmakers in the world.
SK Hynix chairman noted it took nine years to build a single cluster in Yongjin; Samsung shares closed down 4.86% on the announcement as analysts warned of a potential supply glut if AI demand softens.
DRAM output target doubled within five years; humanoid robots commercialized in 10 major industries by 2028; Hyundai committed $5.8B to a robot manufacturing facility and AI data center in the southwest.
Why it matters: Memory chip prices and GPU availability sit at the center of who can run frontier AI and at what cost. South Korea's $576B commitment from Samsung and SK Hynix is the largest push yet to address that constraint. If new fabs come online on schedule, it could ease the AI compute cost crunch within the decade. The counterweight is a realistic 9-year build timeline for new clusters, meaning today's announcement shapes infrastructure available in the early 2030s, not tomorrow.
GUIDES
Apply 12 factors to ship reliable AI agents
Recaply: In this tutorial, you'll learn to apply the 12-Factor Agents framework, a set of 12 engineering principles for building AI agents that work reliably in production, not just in demos.
Step-by-step:
Go to github.com/humanlayer/12-factor-agents and read Factors 1-3 first. Factor 1: convert natural language to tool calls. Factor 2: treat prompts as code, with version control and tests. Factor 3: own your context window. Decide exactly what the model sees, including your instructions, past history, error logs, and retrieved data.
Apply Factors 4 and 8 as a pair. Tools should be simple structured outputs: text in, action out, result back to context. Own your control flow with a fixed outer loop. Don't let the model decide what step comes next.
Follow Factor 10: build small, focused agents. Cap each one at 3-10 steps with one clear job. Smaller context windows mean sharper model output, fewer errors, and faster debugging when things break.
Apply Factor 12: make each agent a stateless reducer. It should launch, pause, and resume via simple APIs. Combine this with Factor 5: keep execution state and business state in one place so the agent always knows what it's doing without rebuilding context.
Add Factors 7 and 11 last. Factor 7: use tool calls to loop humans in before the agent acts on high-stakes decisions. Factor 11: build triggers from where your users already are. Slack, webhooks, and crons all work. Agents that earn trust get access to bigger tasks.
Pro tip: Don't skip Factor 9. Compact errors into context. A clean error summary helps the model recover. A raw stack trace just causes it to stall.
TOGETHER WITH CATCH
AI Executive Assistant for Normies
Catch is a smart, proactive AI admin agent that handles your scheduling, inbox, emails, follow-ups, and day-to-day admin across Gmail, Outlook, Slack, WhatsApp, and phone.
No setup. No learning curve.
From swamped to sorted, instantly.
XAI
X launches hosted MCP for real-time AI access
Recaply: X just went live with a hosted MCP server, giving any MCP-compatible AI tool direct access to the live X API without extra authentication, setup, or custom code.
Key details:
The hosted X MCP connects agents to the X API through the Model Context Protocol, enabling real-time queries on posts, users, and trending topics without custom integrations. Grok, Cursor, and any MCP-compliant client are supported out of the box.
No authentication setup is required to get started; XDevelopers described the offering as giving agents "access to the best real-time information source in the world," positioning X public data as a live signal stream for AI agents.
The launch was announced via @XDevelopers on June 29, 2026, with documentation at docs.x.com/tools/mcp, marking X's first hosted MCP offering in its push to become an AI infrastructure layer.
Available immediately for any MCP-compatible tool as of June 29, 2026; no pricing, rate limits, or enterprise tier details were disclosed in the initial announcement.
Why it matters: Most agents today reason from documents, pre-indexed data, and cached databases. X's hosted MCP gives them a live public discourse feed for the first time. Agents can now monitor breaking news, track sentiment shifts, or act on trending signals without a custom pipeline. Combined with Open Memory Protocol and Cursor Mobile both landing this week, the tooling layer for always-on ambient agents is assembling faster than most teams can keep up with.
TOOLS
Trending AI Tools
🧠 Open Memory Protocol - An open standard for portable, interoperable AI memory across Claude, ChatGPT, and Cursor through a shared interface, no tool-switching required.
🐱 LongCat-2.0 - A 1.6T parameter open-source MoE model with 1M context, optimized for coding and agentic tasks, free via longcat.chat.
📱 Cursor Mobile - The iOS app for launching and directing Cursor cloud agents from your phone, with voice input, diff review, and remote PR merging.
NEWS
What Matters in AI Right Now?
Cursor shipped a public beta iOS app days after its $60B SpaceX acquisition, letting developers launch and direct cloud coding agents from their phone with voice input, remote control, and live lock-screen notifications when an agent finishes or needs input.
Anthropic deployed Claude on NVIDIA's GB300 Blackwell Ultra GPUs in Microsoft Foundry on Azure, the first time Claude has run on NVIDIA hardware, giving enterprise customers infrastructure-level controls for identity, networking, and runtime policy in agentic workloads.
Salesforce employees expressed concern over Anthropic's growing presence inside Slack, according to The Information, as Claude-powered Slackbot rolls out to enterprise users and Anthropic has hired more than 45 people from Salesforce this year.
The New York Times reported that a Poway, California man lost his life savings to an elaborate AI-generated phone scam, in a case showing how AI-powered voice spoofing is enabling a new class of high-stakes personal fraud.
AI-generated political content is set to swamp the 2026 US midterm elections, the Financial Times reported, with campaign spending on AI video, audio, and targeting reaching a scale that dwarfs prior election cycles.
Meituan open-sourced LongCat-2.0, a 1.6 trillion parameter MoE model with 48 billion active parameters per token, trained entirely on AI ASIC superpods across more than 35 trillion tokens, with 1M-context support and strong performance on coding and agentic benchmarks.
Meta contracted hundreds of workers to pose as minors and probe ChatGPT, Gemini, and Character.AI with over 45,000 prompts about suicide, drugs, and self-harm, in what the company described as routine safety benchmarking, according to Wired.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
