State of AI Engineering

What The State of AI Engineering Report Reveals About AI in Production

AI engineering has moved past experimentation, but are your observability practices keeping up? Datadog analyzed LLM telemetry from 1,000+ customers to reveal what's actually happening at scale: model adoption shifts, hidden token costs, and what the rise of agentic frameworks means for reliability. Download the report to benchmark your AI stack against production reality.

Complain Box for AI! - OpenAI has introduced a Model Misalignment Reporting Framework to document and study cases where AI models behave in unexpected or potentially harmful ways. The framework covers incidents such as models hiding mistakes, using exposed credentials, sharing files, or bypassing restrictions. Read more.

Anthropic goes biological - Anthropic has quietly established a wet lab to combine its AI models with real-world biological experiments as it expands into drug research and life sciences. The company says physical lab work is essential for testing biological ideas, while its broader effort targets rare and difficult-to-treat diseases. Read more.

TypeSafe AI Introduces Jev, a New AI Model Built for Automation

Read full blog.

Large language models have become very good at chatting, writing code, and following instructions. But using them inside software systems is a different challenge. Applications often need fast, predictable decisions rather than long streams of generated text. TypeSafe AI is trying to address this gap with a new category of models called System One Models, starting with its first model, Jev.

From Generating Text to Making Decisions

Traditional LLMs generate text one token at a time. This makes them flexible, but it also creates extra latency, costs, and the possibility of incorrect or unexpected output. Software usually has to parse and validate those responses before it can safely use them.

Jev takes a different approach. Instead of generating strings, it produces type-safe structured values. Its possible outputs and structure are defined in advance, while every answer includes probabilities and confidence scores. TypeSafe says this makes Jev suitable for tasks such as classification, routing, scoring, extraction, and deciding which branch of a software workflow should run.

The model also uses parallel sampling rather than generating outputs sequentially. According to TypeSafe, Jev can deliver end-to-end responses in roughly 70–500 milliseconds, compared with several seconds for many frontier models on the company's comparison. TypeSafe also lists input pricing at $0.042 per million tokens and says output is effectively free to meter.

Why This Matters for Software

The bigger idea behind Jev is that AI does not always need to produce language. Many software tasks are really about making a decision from a piece of state.

For example, an application might need to decide whether a user should be routed to a particular workflow, whether an event looks suspicious, or how likely a customer is to leave. Instead of asking an LLM to explain its reasoning in text and then parsing that response, a System One model can return a structured decision that fits directly into the application's code.

Confidence is another important part of the design. TypeSafe argues that automation becomes difficult when models cannot reliably communicate when they are uncertain. Jev is designed to return calibrated probabilities with every output, so software can make decisions based not only on the answer but also on how confident the model is.

TypeSafe's early workflow evaluations report large speed and cost differences compared with frontier LLMs, although the company also highlights limitations in its evaluation setup, including possible bias in the workflows and reference models.

Jev is currently available through early access. TypeSafe's larger goal is to create an AI interface that software can depend on, moving AI beyond chat interfaces and deeper into real-time automated systems.

Atlassian becomes “always on” - Atlassian is adding always-on AI agents that can keep working on development tasks without waiting for new prompts. The agents connect Jira, code, testing, and pull requests to turn AI coding into a more continuous workflow. Read more.

Pinterest designs your room - Pinterest is testing Restyle, an AI feature that lets users upload a photo of their room and experiment with furniture, colors, lighting, decor, and different design styles. Users can also select individual items to replace or remove, turning Pinterest inspiration into visual room makeovers. Read more.

Buzz of the Week!

Seqlock

A seqlock (sequence lock) is a synchronization primitive designed for situations where reads are far more frequent than writes, especially in operating systems and low-level concurrent systems. Instead of making readers acquire a lock, a reader records a sequence counter, reads the shared data, and checks whether the counter changed before accepting the result. Writers increment the counter before and after updating the data, making the sequence odd while a write is in progress and even when the update finishes. If a reader detects a changed or odd sequence value, it simply retries rather than blocking. This makes reads extremely cheap, but readers may repeatedly retry when writes are frequent. Seqlocks are particularly useful for things like kernel timekeeping, shared statistics, and rapidly changing system state where avoiding reader-side locking matters.

Things that launched. Things that went viral. Things you'll pretend to try.

earthly

earthly is a container-based build system that makes builds reproducible locally and in CI.

dagger

dagger runs CI/CD pipelines as programmable containers, allowing developers to define pipelines using normal programming languages.

nushell

nushell A shell that treats command output as structured data rather than plain text. Very interesting for engineers who work heavily with JSON and APIs.

Build Braincells, Not Just Features

This weekend’s read: The next two years of Software Engineering.

This week’s watch: How did Humans discover mercury.

Meanwhile…