Your agent needs more than 2 projects

You prompt. The agent builds. Then it asks for a database.

Ghost is postgres made for this. Spin one up in seconds. Fork it like a branch. Delete it when you're done. Pay nothing when it's idle.

Your agent gets full sql, mcp support, as many databases. No dashboards. No provisioning. No forgotten dev databases draining your card at month end.

Build a weekend app. Fork the schema three different ways. Throw two of them out. Ghost doesn't care. The next prompt can spin up a fresh one.

You're already vibe-coding the app. Stop wiring up the backend.

Unlimited databases. Unlimited forms. 100 compute/hrs a month. 1tb of storage. Free.

Download Ghost

Today:

• Google Unveils TPU 8t and TPU 8i for the Agentic AI Era

• Alibaba Launches Qwen3.7-Max for Long-Horizon AI Agents

• DeepMind Uses AI and Lean to Solve 9 Open Erdős Problems

• Anthropic’s Project Glasswing Finds 10,000+ Serious Software Vulnerabilities

• ChatGPT Can Now Create and Edit PowerPoint Slides Directly

Google Unveils 8th-Gen TPUs: Hardware Purpose-Built for the Agentic Era

Anticipating the intense computational demands of continuous, multi-step AI agents, Google has introduced its eighth-generation Tensor Processing Units (TPUs). For the first time, Google is splitting its architecture into two specialized chips: the TPU 8t for training and the TPU 8i for inference.

Key Details:

• TPU 8t (The Training Powerhouse): Designed to reduce frontier model development cycles from months to weeks. A single superpod scales to 9,600 chips, delivering 121 ExaFlops of compute and 2 petabytes of shared high-bandwidth memory. It targets massive compute throughput and maximum uptime.

• TPU 8i (The Reasoning Engine): Tailored for the latency-sensitive environment of AI agents "swarming" in complex workflows. To prevent idle time, it breaks the "memory wall" by packing 288 GB of high-bandwidth memory and 384 MB of on-chip SRAM (3x more than the previous generation). It delivers 80% better performance-per-dollar over its predecessor.

• System-Wide Efficiency: Both chips run on Google’s custom Axion Arm-based CPUs and utilize fourth-generation liquid cooling. Thanks to these full-stack integrations, the 8th-gen TPUs deliver up to 2x better performance-per-watt than the previous Ironwood generation. They will be generally available later this year.

Qwen3.7-Max: A Relentless Frontier Model for Autonomous Execution

Alibaba’s Qwen team has officially launched Qwen3.7-Max, a proprietary model heavily optimized to serve as a versatile foundation for agent workflows. Rather than just excelling at single-turn chats, Qwen3.7-Max is designed to sustain autonomous execution across hundreds or thousands of steps.

Key Details:

• Marathon Autonomous Execution: To prove its long-horizon capabilities, Qwen3.7-Max was tasked with optimizing a GPU kernel on an unseen hardware architecture. Over a 35-hour period, the model autonomously executed 1,158 tool calls—writing, compiling, profiling, diagnosing bugs, and redesigning the kernel—ultimately achieving a 10x speedup with zero human intervention.

• Top-Tier Benchmarks: It rivals or beats current heavyweights like Claude 4.6 Opus Max and DeepSeek V4 Pro on major reasoning and coding benchmarks, scoring an impressive 92.4 on GPQA Diamond and 60.6 on SWE-Pro.

• Cross-Scaffold & Physical Versatility: Qwen3.7-Max is designed to generalize seamlessly across different agent frameworks (like Claude Code and OpenClaw). Impressively, it was also showcased operating in the physical world, using tool calls to successfully navigate and control a robotic dog.

• Availability: The model is currently accessible via the Alibaba Cloud Model Studio API.

AlphaProof Nexus: AI is Now Solving Open Mathematics Problems

AI has notoriously struggled with mathematical hallucinations, but a newly published paper from Google DeepMind and collaborating researchers reveals a massive leap forward. Their new framework, AlphaProof Nexus, addresses unreliability by pairing an LLM’s generative intuition with the strict, logical verification of the Lean formal proof assistant.

Key Details:

• Solving the Unsolved: In a large-scale evaluation, AlphaProof Nexus autonomously resolved 9 out of 353 open Erdős problems—notoriously difficult math problems, some of which have been unsolved for decades. It also proved 44 out of 492 open conjectures from the On-Line Encyclopedia of Integer Sequences (OEIS).

• Agentic Verification Loops: The system works by deploying subagents that propose proofs and iteratively refine them based on compiler error feedback from Lean. If a proof cannot be formally verified, it is rejected, virtually eliminating hallucinations.

• High Impact, Low Cost: Researchers noted that the autonomous system achieved these breakthroughs at a compute cost of only a few hundred dollars per problem.

• Real-World Deployment: AlphaProof Nexus is no longer just a prototype; it is currently being actively deployed by researchers across combinatorics, optimization, algebraic geometry, and quantum optics to accelerate mathematical discovery.

🧠RESEARCH

DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards

"DelTA" is a new training method that makes AI models smarter at reasoning. By acting as a filter, it helps the AI recognize exactly which words in its answers earn high scores, rather than rewarding generic formatting. This approach significantly boosts AI performance in complex math and coding tasks. 

Sensor2Sensor: Cross-Embodiment Sensor Conversion for Autonomous Driving

Self-driving cars require massive amounts of 3D data to learn safely. "Sensor2Sensor" is an AI tool that transforms ordinary, flat real-world dashcam videos into rich, multi-view 3D sensor data, including laser depth scans. This breakthrough unlocks endless online driving footage for developers to train smarter and safer autonomous vehicles.

Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention

AI models often struggle to remember long texts efficiently. "Gated DeltaNet-2" solves this by giving the AI separate controls for "erasing" outdated information and "writing" new facts into its compressed memory. This prevents the AI from scrambling its thoughts, dramatically improving its ability to scan and recall long documents. 

📲SOCIAL MEDIA

🗞️MORE NEWS

Anthropic’s Project Glasswing Update Anthropic shared an update on Project Glasswing, an initiative that uses its newest AI model to find and fix hidden security flaws in essential computer systems. The AI acts like an automated security guard, discovering complex coding bugs so partner organizations can patch their systems before hackers have a chance to break in.

ChatGPT Inside Microsoft PowerPoint OpenAI has released a built-in ChatGPT tool for Microsoft PowerPoint that helps users instantly create, edit, and summarize presentation slides. By reading your raw notes or documents, the AI automatically builds professional slide layouts, though users should double-check the results since the tool can occasionally delete or change content by mistake.

Perplexity Shares Its "Bumblebee" Security Tool The AI search company Perplexity is sharing its internal security tool, Bumblebee, for free to help other businesses protect the computers their programmers use. Bumblebee safely scans a programmer's laptop for harmful, hidden software without accidentally triggering the very computer viruses it is trying to find.

Anthropic Explores Microsoft's Custom AI Chips To keep up with massive user demand, Anthropic is reportedly in early talks to rent custom-made computer chips directly from Microsoft. This move would give the AI company another way to run its powerful chatbots while reducing its heavy reliance on popular hardware manufacturers like Nvidia.

Google Expands "Pomelli" for Small Businesses Google has updated its AI marketing tool, Pomelli, to help small businesses create professional advertising materials without needing a big design budget. The tool can now study a company's existing website to learn its unique visual style, and then use that knowledge to automatically generate custom branding guides and full websites.

Anthropic Nears Massive $30 Billion Funding Round Anthropic is preparing to secure over $30 billion in new funding as soon as next week, driven by the company's exploding sales growth. If successful, this massive cash injection could push the company's total value past $900 billion, placing it ahead of OpenAI as the most valuable artificial intelligence startup in the world.