Hello, AI Explorer!
Good Wednesday. This week: an Anthropic co-founder putting 60% odds on AI training its own successor by 2028, Peter Thiel backing floating data centers powered by the Pacific Ocean, smart glasses with a built-in AI agent in the lens, and OpenAI and Anthropic both launching $5.5B in PE ventures on the same day.
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 |
May 6, 2026 |
| 🤖 | Anthropic co-founder: 60% odds AI trains its own successor by end of 2028 |
| 🌊 | Peter Thiel backs $140M ocean AI data centers -- 85-meter floating nodes powered by wave energy |
| 🦾 | Mira AI glasses: waveguide display, built-in agent, 60-language translation -- no camera, no cloud |
| 💰 | OpenAI and Anthropic both launched PE-backed deployment ventures the same day -- $5.5B combined |
| 🧠 | GPT-5.5 Instant is ChatGPT's new default -- 52.5% fewer hallucinations on medical and legal prompts |
| 📈 | Sierra raises $950M at $15.8B -- AI customer agents now serve 40% of the Fortune 50 |
| 🏛 | Google, Microsoft, and xAI agree to let the US government review new AI models before release |
| 💉 | Harvard study: OpenAI's o1 correctly diagnosed ER cases 67% of the time vs. 50-55% for physicians |
| ⚡ | 4 AI Tools this week: Be My Eyes, HomeSage.ai, Layla, Tidio -- all inside. |
ANTHROPIC
| 🤖 |
Anthropic Co-Founder Puts 60% Odds on AI Training Its Own Successor by 2028 |
Jack Clark, co-founder of Anthropic, published a new analysis in his Import AI newsletter putting 60% odds on AI systems training their own successors before the end of 2028 -- and 30% odds it happens by 2027. He built the entire case from public benchmark data, not internal Anthropic research.
The numbers are hard to dismiss. On SWE-Bench -- real GitHub coding tasks -- AI scores went from 2% with Claude 2 to 93.9% with Mythos Preview in under three years. METR data shows AI's independent task horizon extended from 30-second runs in 2022 to 12-hour autonomous sessions in 2026, with 100-hour runs projected before year-end. On an Anthropic internal model optimization test, mean speedup jumped from 2.9x in May 2025 to 52x by April 2026.
What Clark's data shows:
• SWE-Bench coding scores: 2% (Claude 2) to 93.9% (Mythos Preview) in under 3 years
• AI independent work horizon: 30 seconds (2022) to 12 hours (2026) -- 100-hour runs projected this year
• Internal model optimization speedup: 2.9x (May 2025) to 52x (April 2026)
Why it matters:
Once AI can reliably train its own successor, model improvement no longer depends on human researcher hours -- it becomes computational. Clark is not predicting superintelligence. He's pointing at a specific, measurable threshold that public benchmarks say we're approaching faster than most people assumed. 2028 is two and a half years away.
AI INFRASTRUCTURE
| 🌊 |
Peter Thiel Bets $140M on Wave-Powered AI Data Centers Floating in the Pacific Ocean |
Oregon startup Panthalassa closed a $140M Series B led by Peter Thiel, valuing the company near $1B. The company builds 85-meter autonomous steel nodes that float in open ocean, convert wave motion into electricity for onboard AI chips, and use seawater for cooling. Inference results travel to land via SpaceX Starlink. No grid connections. No data center zoning battles.
John Doerr, Marc Benioff's TIME Ventures, and Max Levchin's SciFi Ventures also joined the round. The capital finishes a pilot factory near Portland and deploys the first wave-powered compute nodes in the Pacific this year, with commercial rollout in 2027. Thiel told the Financial Times: "Panthalassa has opened the ocean frontier."
How it works:
• Nodes self-steer to remote open-ocean positions using hull shape alone -- no engines needed
• Seawater cooling eliminates the freshwater consumption problem that's made land data centers controversial
• Pacific pilot nodes launch in 2026; commercial deployment begins 2027
Why it matters:
AI data centers face two genuine public crises right now: land opposition and power grid pressure. Space-based compute -- which Musk and Google have both floated -- is still years out. The ocean is available today. Panthalassa's approach doesn't need a new grid, a new power plant, or a zoning approval. That's a meaningful structural advantage in a race where compute availability is the constraint.
| 🦾 |
AI Glasses With an Agent Inside: Mira Ships at $479 With Real-Time Translation and Always-On Memory |
Mira's smart glasses include a dual waveguide display, a built-in AI agent that sends emails and books rides on command, two-way translation across 60+ languages, and real-time transcription. No camera. Audio transcripts store on your phone only -- deleted from servers immediately. At 39 grams, they're roughly half the weight of Meta's Ray-Ban Display. Ask any question and the answer surfaces in your lens, without reaching for your phone.
WHY IT MATTERS:
Smart glasses have been looking for a reason to exist beyond notifications and music. An always-on memory layer you can search after the fact -- combined with live translation during actual conversations -- is a use case that finally justifies wearing them. The no-camera choice also removes the biggest social friction point that killed Google Glass.
| 💰 |
OpenAI and Anthropic Both Launched PE-Backed Deployment Ventures the Same Day |
On the same day, Anthropic announced a $1.5B enterprise services company with Blackstone, Hellman & Friedman, and Goldman Sachs -- pairing its Applied AI engineers with companies building custom Claude workflows. OpenAI separately finalized a $4B "Deployment Company" with 19 investors including TPG, Brookfield, Bain, and SoftBank, at a $10B valuation. Both companies are building what amount to AI-native consulting arms backed by private equity capital.
WHY IT MATTERS:
The bottleneck for enterprise AI is no longer the models -- it's implementation inside legacy, messy businesses. Both frontier labs read the same signal and moved the same day. Whoever lands inside more enterprise systems first builds the stickiest long-term moat. This is the consulting model, rebuilt for the AI era.
⚡ Coming in Hot
AI Tools of the Week
4 tools worth adding to your workflow right now.
| 👁 |
01 Be My Eyes ACCESSIBILITY AI GPT-4 powered visual assistant for blind and low-vision users -- point your camera at anything and ask what it reads, shows, or means. A medication label, a street sign, a restaurant menu. 700,000+ blind users globally. Completely free. If there is one tool on this list that a friend or family member needs to know about, this is it. |
| 🏠 |
02 HomeSage.ai REAL ESTATE AI AI built specifically for real estate agents -- pulls comparable sales, runs market analysis, and writes listing descriptions that used to take 2 hours in about 15 minutes. Real estate agents report cutting their pre-listing prep by more than half. If you work in property, this is the tool your competitors are quietly adopting. |
| ✈️ |
03 Layla AI trip planner that builds your full itinerary, compares hotels, and can book -- all inside one conversation. No switching between tabs, no copying from Google Maps to a notes doc, no cross-referencing reviews manually. Freemium. If you have a summer trip coming up, this saves 2-3 hours of planning time per destination. |
| 💬 |
04 Tidio BUSINESS AI AI chatbot and live chat for small businesses -- handles FAQs, captures leads, and escalates to a human when it matters. Setup takes 15 minutes. 300,000+ websites run it. E-commerce stores report 30% fewer support tickets after installing it. The free tier handles unlimited chats. This is the customer service stack most small teams can't justify hiring for. |
⚡ Quick Hits
🧠 GPT-5.5 Instant is now ChatGPT's default model
OpenAI replaced ChatGPT's default on May 5 with GPT-5.5 Instant, cutting hallucinations by 52.5% on medical and legal prompts. A new "memory sources" panel shows exactly which past chats shaped each reply -- and lets you remove individual entries. Rolled out to all users immediately. See what changed →
📈 Sierra raises $950M at a $15.8B valuation
Bret Taylor's AI customer service platform just raised at $15.8B -- up from $10B just six months ago. Its agents handle mortgage refinancing, insurance claims, and support for 40% of the Fortune 50. AI is replacing call centers faster than most predicted. Read the TechCrunch breakdown →
🏛 Google, Microsoft, and xAI agree to US pre-release AI model reviews
All three companies agreed to submit new AI models to the US Commerce Department's AI standards center for evaluation before public deployment. First voluntary government checkpoint of its kind for frontier models. Read the full agreement →
💉 OpenAI's o1 outdiagnosed ER doctors in a Harvard study
Across 1,000 real emergency room cases, o1 reached the correct diagnosis 67% of the time. Attending physicians scored 50-55%. The gap held across both common and rare presentations. Read the Harvard study →
🏢 Coinbase cuts 14% of staff to go AI-native
CEO Brian Armstrong announced layoffs across roughly 14% of the workforce as part of a strategic shift toward AI-first operations. Coinbase joins a growing list of tech companies restructuring headcount around AI capacity rather than human scaling. See Brian Armstrong's announcement →
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 Saturday,
The AI Tool Discovery Team
Feedback
What'd you think of today's newsletter?
Vote below to let us know how we're doing.