article

AI Dev Essentials #16: GitHub Spark, ChatGPT Agent & Gemini's flash-lite

Explore the latest in AI dev: GitHub Spark, ChatGPT Agent, Gemini 2.5, open-source models, major tool updates, and expert workflows for automation success.

Hey Everyone 👋,

John Lindquist here with the 16th issue of AI Dev Essentials!

Due to popular demand, I'm hosting my first Claude Code Power User workshop on August 1st. After posting a single tweet about on Friday, we've almost already sold out 🤯! I'll have a schedule for future dates out soon, but if you want to grab a slot while they're still available then grab a ticket here.

I posted my first 10 lessons of "AI Dev Essentials" of a course for "Building an AI-Driven Markdown Memory System with MCP" over on egghead.io. That course will have many, many more lessons to come and will focus on "learning from mistakes" and "AI wisdom" which is at the heart of AI-driven development. I'm taking the approach of "publishing as I go" (kinda similar to how LitRPG authors use patreon so readers can don't have to wait for the entire book). Everything is moving so fast right now that I'm convincing this is the only viable model of education.

Otherwise, I'm spending all of my time chaining together AI CLIs (Claude Code, Gemini, etc) to build up pipelines for automating work. It's absolutely bonkers what you're able to do with enough AIs linked together and as an automation enthusiast (maybe "zealot" is a better word?) I'm having an absolute blast building these and I'm extremely excited to share them both in the Claude Code Power User workshop and in future lessons on egghead.io.

Now, onto the news!

🎓 New egghead.io Lessons This Week

Create an AI-Automated MCP-Driven Markdown Knowledge Base

As mentioned above, the initial 10 lessons of this course dig into capturing knowledge from past conversations, daily notes, and across projects while in the flow of using AI tools to help establish a living, breathing knowledge base. Many, many more lessons to come, but laying down the groundwork here:

https://egghead.io/courses/create-an-ai-automated-mcp-driven-markdown-knowledge-base~km3h6

🚀 Major Announcements

GitHub Spark Enters Public Preview for Copilot Pro+ Users

Thomas Dohmke, GitHub's CEO, announced that GitHub Spark is now in public preview for Copilot Pro+ subscribers, marking a major milestone in natural language app development.

Key features:

GitHub Spark represents the evolution from GitHub Universe 2024's technical preview to a production-ready tool, combining GitHub Next's experimental approach with Copilot Pro+'s premium capabilities.

(GitHub Spark)

When the sleeping giant wakes, the startup quakes (I made that up myself 😇). The big dog GitHub is coming for the bolt/lovable/replits of the world. My guess is they'll annihilate an entire product space simply by bundling it with your current subscription and making it "good enough". I'm cheering for the little guys, but I think we've all seen this story play out before...

ChatGPT Agent: AI With Its Own Computer

OpenAI launched ChatGPT Agent on July 17, 2025, giving AI access to its own virtual computer environment to autonomously complete complex, multi-step tasks while maintaining user oversight.

Technical capabilities:

Pricing:

(OpenAI Blog)

I've run about 10 queries through this. It's fine, manus.ai is better. I expect it to rapidly improve. ChatGPT's agent just doesn't have enough access to enough tools to compete with manus. But again, same story as the Github Spark announcement. It just has to be "good enough" and bundled in your subscription.

Qwen3-Coder-480B: Open-Source "Matches" Claude Sonnet 4

Alibaba released Qwen3-Coder-480B on July 23, 2025, delivering state-of-the-art coding performance in a fully open-source model that rivals proprietary solutions.

Technical specifications:

Performance & availability:

Qwen Code CLI:

(Official Blog)

I've heard extremely mixed reviews of this model. From my current understanding, it's great, but not as good as the hype and gamified benchmarks would have you believe. I hate to downplay it, because Qwen and Kimi K2 are both amazing open-source models, but being "close" to Anthropic's models just isn't good enough. I'm know I'm lucky/privelege to be able to afford a $200 a month subscription, but all the little paper cuts from worse models add up quickly. I'm still trying to figure out the best flow for Groq + K2, because raw speed is definitely a huge factor in focused development, but all of my agentic work is done with Claude Code.

🛠️ Developer Tooling Updates

Gemini 2.5 Flash-Lite GA: Speed Meets Efficiency

Google launched Gemini 2.5 Flash-Lite as generally available on July 22, 2025, delivering 1.5x faster performance than its predecessors at the lowest cost in the Gemini 2.5 family.

Key specifications:

Real-world impact:

(Google Developers Blog)

Speedy, cheap, multi-modal = yes please! If you only use it for transcribing video, then it's still a massive win. Someday this speedy cheap models will be able to perfectly follow plans generated by the slow smart models and then we'll be absolutely cookin'!

💼 AI Business & Ecosystem

Stargate Expands: 4.5 Gigawatts for AI's Future

OpenAI and Oracle signed a massive expansion deal on July 22, 2025, adding 4.5 gigawatts of data center capacity to the Stargate project—enough power for 4 million homes.

Scale and impact:

Infrastructure details:

(OpenAI Announcement)

Those are big numbers. Might as well say a bajillion dollarie-doos for a zillion power plants. It's this weird combination of all the AI providers trying to out-brag each other to convince investors to keep dumping in money, but they're also serious about it and actually building it. I think a lot about if today's models just had more power for more speed/context, it would be a massive multiplier on our work even if models never improved.

Lee Robinson Joins Cursor to Shape AI Education

After 5 years helping build Vercel from $1M to $200M+ ARR, Lee Robinson transitioned to Cursor in July 2025 with a mission to teach developers about AI-assisted coding.

Background and impact:

Cursor's education initiatives:

(Lee Robinson Profile)

Lee is a legend. Awesome guy, great teacher, and I'm super happy for him. Make sure to follow him everywhere because he's a huge addition to the Cursor space and I can't wait to see what he produces.

Claude for Financial Services: AI Meets Wall Street

Anthropic launched Claude for Financial Services on July 15, 2025, featuring deep integrations with S&P Global, FactSet, Morningstar, and other major financial data providers.

Performance highlights:

(Anthropic Announcement | S&P Global Partnership)

I would love to be in the room for these meetings where AI nerds sell their models to Wall Street. We need a few episodes of "Silicon Valley" to capture the absurdity of it all. I wonder what the AI Overlords will think of the concept of "money" and the value of Wall Street once they finally take over 🤔

🤖 Model & Research Updates

Gemini Deep Think Achieves IMO Gold Medal Standard

Google DeepMind's enhanced Gemini with "Deep Think" reasoning achieved gold medal performance at the 2025 International Mathematical Olympiad, scoring 35/42 points.

Technical achievements:

(Google DeepMind Blog)

AI solving problems that challenge the world's brightest mathematical minds... And we'll supposedly get access to this in the public later this year. I'm not nearly qualified enough to speak to this subject (I majored in technical writing then went to law school...), but I know that math powers the computer I'm typing this on and can only guess at the implications.

OpenAI o3 Alpha Spotted on Web Arena

A mysterious "o3-alpha-responses-2025-07-17" model appeared on Web Arena benchmarks, showing exceptional performance including generating complex SVG applications in one shot.

(TestingCatalog Report)

OpenAI continues to test new capabilities in public. The SVG generation capability suggests improved spatial reasoning and code generation working together.

🔧 Tools & Integrations

Code Agent Orchestrators Comparison

Addy Osmani highlighted the growing ecosystem of code agent orchestrators:

(Tool Comparison)

We're seeing the emergence of meta-tools for managing AI agents. This is a natural evolution as workflows become more complex.

VibeTunnel for Remote Claude Code Management

Ian Nuttall showcased VibeTunnel for managing Claude Code agents remotely via phone or any browser, enabling progress monitoring and task assignment from anywhere.

(VibeTunnel Demo)

We're really going to have to establish boundaries between work and home. While I'm happy we're starting to see "agents on phones", I also need to disconnect for my own sanity.

⚡ Quick Updates

✨ Live Workshop: Transform into a Claude Code Power User ✨

Full details: https://egghead.io/workshop/claude-code

Learn the latest power user workflows to maximize your productivity. Join me live as I walk through building CLI chains, video prompting, triggering workflows on hooks, and much more:

➡️ Register Now

Limited spots available. Secure yours today!

Thanks for reading! If you have any questions or feedback, feel free to reply directly to this email.

Read this far? Share "AI Dev Essentials" with a friend! - https://egghead.io/newsletters/ai-dev-essentials

egghead.io