OpenAI announced Jalapeño, its first custom inference chip, with benchmarks showing industry-leading speed and power efficiency for running AI models. The chip is already in production, handling inference workloads for OpenAI’s models with higher throughput and lower latency than off-the-shelf alternatives.

CFO Sarah Friar laid out the economic case in a separate post about OpenAI’s “full stack” strategy. The pitch is straightforward: control the whole pipeline from chips to models to products, compound the improvements at each layer, and deliver more useful AI at lower cost. Jalapeño is the chip layer of that stack.

OpenAI hasn’t published detailed specs or pricing. The announcement focuses on comparative metrics: faster than what they were using before, more power-efficient, better suited to modern model architectures. If you’re building on OpenAI’s API, this matters because it should translate to lower latency and potentially lower costs as the efficiency gains work through the system.

If you’re running your own inference infrastructure, there’s nothing actionable here yet. OpenAI isn’t selling Jalapeño chips. This is for their own use.

AI Wrote a Million Lines of Production Code

Meanwhile, Paul Dix at InfluxData confirmed what’s been rumored for months: AI wrote over a million lines of code for their Rust rewrite, and it’s now running on millions of developer machines.

Simon Willison quoted Dix saying the impressive part isn’t just the volume. It’s that they built a verification system, pointed AI at the problem, and refined the output over a couple of months until it was reliable enough to ship.

The setup mattered. InfluxData had an existing implementation to compare against, which made verification tractable. That’s not “cheating.” That’s engineering. If you can define correctness precisely enough to check AI output automatically, you can use AI to do work that would otherwise take years.

EVE Online is doing something similar. They’re migrating 2.4 million lines of Stackless Python 2.7 to Python 3, starting with automated tools like futurize, then manually reviewing the ~20,000 behavioral differences between Python 2 and 3. The game’s been running on the same Python version since 2010. They’re not rewriting it by hand.

These aren’t toy projects. They’re production systems with millions of users. The pattern is the same: use AI to generate code, build systems to verify it works, iterate until it’s good enough to ship.

If you’re sitting on a large codebase that needs porting, translation, or refactoring, this is the model. AI can do the grunt work. You need to build the scaffolding to check it.

Quick Hits

IBM released Granite 4.2, a new family of open-weight models with details on how they’re built. If you want to understand modern LLM training pipelines, the post is worth reading.

Keenable raised $26 million from Accel to build a web search index specifically for AI agents. They’re exiting stealth now. The thesis is that agents need different crawling and indexing than human search does. We’ll see if that’s true.

Generalist, a robotics startup, hit a $3 billion valuation in a $200 million extension round, according to TechCrunch sources. They raised at $2 billion a few months ago. That’s a lot of money flowing into physical AI.

GitHub published a post on evaluating LLMs before production based on lessons from using LLMs for secret scanning. If you’re trying to ship LLM features, the post covers practical evaluation strategies.

Ringg, an Indian voice AI startup, raised $10 million from Peak XV as part of a Series A extension. TechCrunch reports they’re pushing voice AI beyond phone calls. Not much detail on what that means in practice.