Hello, AI Explorer!
Good Saturday. This week: Claude Sonnet 5 is now the default -- Meta's AI reads brain signals and types out your thoughts
Welcome back to your weekly AI digest: human-curated, zero fluff.
Every Wednesday and Saturday, we bring you:
- The week's most important AI developments (no fluff)
- Hand-picked AI tools actually worth your time
- Expert insights on what these changes mean for you
Let's dive into this week's discoveries! ⚡
🔥 This Week in AI
🔥 This Week in AI |
July 4, 2026 |
| 🧠 | Claude Sonnet 5 is here — browses the web, runs terminals, plans and acts at near-Opus quality |
| 🔥 | Palantir's CEO went on CNBC and accused OpenAI and Anthropic of harvesting enterprise data |
| 🇨🇳 | China's Meituan trained a 1.6 trillion-parameter AI on homegrown chips — then open-sourced it |
| 🧬 | Meta's AI reads brain signals and types out your thoughts — no implants, 61% word accuracy |
| 🏛️ | OpenAI offered the US government a 5% stake — worth $42.6 billion at current valuation |
| 🔒 | Apple is now pushing security patches faster — because AI makes exploits faster to build |
| 💻 | Amazon confirmed it builds its own AI chips for Echo and Fire TV — and hinted at more devices |
| 🚀 | Etched exited stealth with $800M raised and $1B in signed contracts for its AI-only inference chip |
| ⚡ | 4 AI Tools This Week: Gamma, Imagine Art, Pika, Final Round AI — all inside. |
ANTHROPIC · MODEL RELEASE
| 🧠 |
Claude Sonnet 5 Is Here — and It Runs Your Computer for You |
Anthropic launched Sonnet 5 on June 30, and the gap between "chatbot" and "AI agent" got a lot smaller. The model plans multi-step tasks, runs browsers and terminals, and executes complex workflows on its own — at performance levels that required Opus just a few months ago. It is now the default for free and Pro plans on Claude.ai.
Pricing is $2 per million input tokens and $10 per million output during an introductory window through August 31. After that it moves to $3 in / $15 out. That is still cheaper than what it cost to get near-Opus quality before. For developers running high-volume agents, the math changed this week.
What changed:
• Agentic by default: Sonnet 5 browses the web, runs terminal commands, calls APIs, and chains tasks together without a human approving each step
• Near-Opus performance at Sonnet price: Scores within striking distance of Opus 4.8 on coding, reasoning, and knowledge — at roughly half the cost
• Better on safety: Lower rates of prompt injection and misuse than Sonnet 4.6, which matters more when the model is taking real actions rather than just replying
Why it matters:
Agents stop being a premium experiment and start being the default. If you use Claude — through the app or the API — what you can ask it to do just expanded significantly. The price drop makes it practical to leave agents running on real tasks without a big API bill at the end of the month.
PALANTIR · INDUSTRY DRAMA
| 🔥 |
Palantir's CEO Went on CNBC and Called OpenAI 'Effing Insane' |
Alex Karp's July 1 CNBC appearance was scheduled to discuss Palantir's new Nvidia partnership. It turned into something else. Karp spent the interview accusing OpenAI and Anthropic of running a "completely insane" model: charging enterprises for tokens that produce no value while quietly harvesting their proprietary data to train future models. His word was "livid." Palantir shares jumped more than 9% that day.
The data harvesting claim is the one that cuts. When a company sends internal documents through an AI API to extract information, summarize contracts, or analyze sales calls — that data has to go somewhere. Karp's claim is that it goes into future training runs without compensation or disclosure. That is not a new concern, but hearing it from a competing CEO at this volume is a different thing.
What Karp said:
• On tokens: He accused OpenAI and Anthropic of billing for "tokens that create no value" — charging for input and output that doesn't move the work forward
• On data: He claimed the labs absorb "alpha" from enterprise clients' proprietary documents, using it to train better models without telling the enterprise or paying them for it
• On national security: He warned against outsourcing battlefield decisions to "the consensus view in Silicon Valley" — the line that went most viral
Why it matters:
Karp is a competitor with clear incentive to attack the other labs — weigh that. But if your company or team runs sensitive documents through an OpenAI or Anthropic API, their data policies are worth reading before the conversation fades. The data training opt-out settings exist. Most enterprise users have not touched them.
| 🇨🇳 |
China Trained the Biggest AI on Local Chips — Then Gave It Away |
Meituan released LongCat-2.0 on June 30 — a 1.6 trillion-parameter model trained from scratch on 50,000 Chinese-made chips, specifically Huawei Atlas-950 hardware. It is the first model of this scale to complete both training and inference without a single NVIDIA GPU. Meituan open-sourced the weights the same day, letting anyone benchmark the claimed performance. The model runs a 1 million token context window.
WHY IT MATTERS:
US export controls were designed to slow China's AI by cutting off NVIDIA access. LongCat-2.0 is direct evidence that the hardware dependency can be built around. A frontier-scale model that runs domestically doesn't just close a benchmark gap — it changes what the export restrictions actually accomplish.
| 🧬 |
Meta's AI Reads Your Brain and Types What You Were Thinking |
Meta published Brain2Qwerty v2 on June 29 — a non-invasive brain-to-text system using an external MEG scanner, not an implant. Participants wear a helmet that records brain activity while they think about typing, and an AI model converts those neural signals into text in real time. Tested on nine people, it reached 61% word accuracy. The previous best non-invasive method was 8%. One participant hit 78%, with more than half their sentences decoded with one word error or less. Meta is releasing the full training code openly.
WHY IT MATTERS:
The primary use is for people who've lost the ability to speak or type because of stroke, ALS, or injury. The 8x accuracy jump over previous non-invasive methods — with no surgery — is the kind of result that moves a research demo closer to an actual clinical tool. No implant required is the part that changes the adoption timeline.
|
⚡ Coming in Hot AI Tools of the Week |
July 4, 2026 |
4 tools worth adding to your workflow right now.
| 🎨 |
AI PRESENTATIONS Gamma Type an outline or paste rough notes, and Gamma builds a complete slide deck in under two minutes — with layout, visual hierarchy, and formatting handled for you. Output goes to slides, a document, or a shareable web page. For anyone who spends time making decks that could be made in a fraction of the time, this is the tool that fixes it. Try Gamma → |
| 🖼️ |
AI IMAGE Imagine Art AI image generator built for commercial use — product shots, illustrations, concept art, and brand visuals from text prompts, with no licensing questions afterward. Outputs are resolution-ready and commercially cleared, which makes it a practical alternative to stock photo subscriptions for teams producing regular visual content. Try Imagine Art → |
| 🎬 |
AI VIDEO Pika Turn text, images, or video clips into short cinematic videos — without touching a timeline. Pika's motion controls let you animate specific objects, swap scenes, or extend clips from a still. Strong for social content, product demos, and quick creative ideation where you want moving visuals without a full video production setup. Try Pika free → |
| 💼 |
AI CAREER Final Round AI Paste in the actual job description and Final Round AI runs a mock interview matched to that specific role — then gives feedback on each answer, flags weak spots, and generates follow-up questions the interviewer is likely to ask. Built for technical and behavioral rounds, useful whether you're preparing for your first interview or your tenth this month. Discover Final Round AI → |
⚡ Quick Hits
🏛️ OpenAI Offered the US Government a 5% Stake Worth $42.6 Billion
Sam Altman reportedly pitched the deal to Trump, Commerce Secretary Lutnick, and Treasury Secretary Bessent — and asked Anthropic, Google, and Meta to match it. Senator Bernie Sanders countered with a proposal for a one-time 50% tax on AI company shares. No deal is signed. Read more →
🔒 Apple Is Shipping Security Patches Faster Because AI Makes Exploits Faster to Build
Apple told Reuters it can no longer rely on the annual update cycle — AI tools now read a patched diff and build a working exploit in the window between public disclosure and user installation. iOS 26.5.2 shipped early with 25+ fixes. No confirmed exploits in the wild. Read more →
🚀 Nvidia Will Give AI Startups Free GPU Access — and Take a Cut of Future Revenue
Instead of cash upfront, Nvidia's new program lets startups access GPU clusters through partners Sharon AI and Firmus Technologies. In return, Nvidia takes a percentage of future cloud and product revenue. It solves the "need to train but can't afford it" problem — at a price paid later, once revenue exists. Read more →
💻 Amazon Confirmed It Builds Its Own AI Chips for Echo and Fire TV
Amazon's devices chief Panos Panay confirmed the AZ3 and AZ3 Pro chips run AI on-device in Echo Show and Fire TV hardware — handling wake words and sensor fusion locally rather than hitting the cloud for every request. He hinted at "on-the-go devices" coming soon, without naming them. Read more →
🔋 Etched Exited Stealth With $800M Raised and $1B in AI Contracts — Before Anyone Knew It Existed
Etched's chip does one thing: run transformer inference. No flexibility, no other workloads — maximum throughput for serving AI models. Built on TSMC's N4P process, first silicon passed validation on the first pass, and $1B in contracts were signed before the public announcement. Shipments planned for summer 2026. Read more →
That's all for this edition! What tools are you most excited to try?
Reply to this email - we read every response and use your feedback to improve future editions.
See you on Wednesday,
The AI Tool Discovery Team
Feedback
What'd you think of today's newsletter?
Vote below to let us know how we're doing.