Welcome back to Daily AI Skills.

Here’s what we are covering today:
1. Google’s Ironwood TPU Unveiled
2. Microsoft Debug Gym
3. DeepSite - A free open-source cursor alternative

+ Google’s AI Agents Guidebook

Google Unveils Ironwood TPU at Cloud Next 2025

Google launched Ironwood, its seventh-generation TPU, at Cloud Next 2025, optimized for AI inference with unmatched performance and efficiency for advanced “thinking models” like LLMs and MoEs.

The Details:

• Ironwood scales to 9,216 chips per pod, delivering 42.5 exaflops, with each chip offering 4,614 teraflops and 192 GB HBM at 7.2 TB/s bandwidth.

• Enhanced SparseCore accelerates sparse data tasks, ideal for recommendation systems and financial modeling.

• Inter-Chip Interconnect provides 1.2 Tbps bandwidth for low-latency, pod-scale operations.

• Pathways software enables distributed computing across thousands of TPUs for generative AI.

• Available in 256-chip clusters or 9,216-chip pods, Ironwood boasts 10x performance and 2x energy efficiency over prior TPUs, per X posts.

Read the full article by Google here

AI Still Can’t Debug Like a Human, Microsoft Study Shows

Microsoft Research has released a study showing that AI agents—even those built on top-tier language models—continue to fall short on many software debugging tasks that human developers routinely handle.

The details:

• The study evaluated nine large language models, including Claude 3.7 Sonnet, by running a “single prompt-based agent” through 300 debugging tasks from the SWE-bench Lite dataset.

• Even when powered by top-performing coding models, the agent failed to complete about half of the assigned tasks.

• When debugging tools were introduced, Claude 3.7 Sonnet achieved the highest success rate at 48.4%, followed by OpenAI’s o1 at 30.2% and o3-mini at 22.1%.

• Researchers pointed to the lack of sequential decision-making data—like human debugging traces—in the training data as the primary reason for this performance gap.

Read the full research blog by Microsoft here

The New Browser-Based Cursor Alternative

DeepSite, a free tool hosted on Hugging Face Spaces, uses DeepSeek V3 to let users create websites and games via natural language prompts, making web development accessible to non-coders.

The Details:

• DeepSite, built by Hugging Face’s enzostvs, harnesses DeepSeek V3 to convert text prompts into functional web code, no coding skills needed

• It offers real-time previews in the browser, allowing users to tweak and download projects instantly

• An X post showcased DeepSite creating an anime website with a single prompt, highlighting its ease of use for creative projects

• The tool is open-source, free, and requires no software installation, running entirely in the browser

• Since its early 2025 release, DeepSite has gained traction, with X users noting its potential as a “web version of Cursor”

Try out DeepSite here: https://huggingface.co/spaces/enzostvs/deepsite

Google’s Whitepaper on AI Agents Guide

📩Forward it to people you know who are keeping pace with the changing AI world, and stay tuned for the next edition to stay ahead of the curve!

Maybe you missed out on these:

1. Copilot vs Claude vs Gemini

2. How I Built an AI Influencer in 2 Minutes—and Landed Sponsorships

3. Microsoft Muse, Helix and Google's AI Co-Scientist

4. Bolt vs Lovable

5. Your AI Stack Just Got Superpowers: 4 New Inference Providers Live!