Alex Self published "automated doubt," a workflow that uses specialized subagents to critique artifacts at both the design and implementation stages, front-loading scrutiny. Alibaba open-sourced Open Code Review, a hybrid deterministic-plus-dynamic-agent CLI for line-level code review. A 24-minute primer on "How LLMs Actually Work" walks through the transformer stack end-to-end.

Automated doubt, Open Code Review, and a 24-minute primer on how LLMs actually work

The TLDR Dev digest on June 8 delivered three stories that are all about the same shift in how we build software: the loop is getting more skeptical, more open, and more legible. Alex Self published "My Automated Doubt Development Process," which uses specialized subagents to critique artifacts from multiple technical perspectives throughout the design and implementation phases. Alibaba open-sourced Open Code Review, an AI-powered CLI that automates code reviews with line-level feedback. And a long-form primer, "How LLMs Actually Work," walked through the transformer stack end-to-end.

What You Need to Know: Alex Self published a workflow called "automated doubt" that uses specialized subagents to critique artifacts from multiple technical perspectives, front-loading scrutiny at the design stage and re-running audits at the implementation stage. Alibaba open-sourced Open Code Review, an AI-powered CLI tool that automates code reviews with hybrid deterministic-plus-dynamic-agent architecture. A 24-minute primer on how LLMs actually work walked through the transformer stack, embeddings, positional encoding, attention, and feed-forward networks.

Why It Matters

"Automated doubt" is the AI engineering pattern that replaces code review as the bottleneck. Front-loading scrutiny at the design stage catches hidden assumptions before implementation, and post-development audits catch the rest. This is what production AI engineering looks like in 2026.
Open Code Review from Alibaba is a signal that the big-tech code-review tooling is going open source. If the model is the same, the review tool is the differentiator, and the differentiator is now a commodity.
The "How LLMs Actually Work" primer is the clearest single-document explanation of the transformer stack I have read in 2026. If you have been telling yourself you understand the architecture and you have not read it, read it.
For developers: the review layer is the new IP. The model is the same. The review loop is the difference between shipping and not shipping.

What Actually Happened

Alex Self's automated doubt workflow

My Automated Doubt Development Process describes a workflow that uses specialized subagents to critique artifacts from multiple technical perspectives. The process begins by front-loading scrutiny during the design stage, where agents identify hidden assumptions and architectural gaps within technical specifications. Once implementation is underway, a suite of post-development agents audits the codebase for security vulnerabilities, type safety, and logic errors to make sure the output meets high engineering standards. The pattern is the opposite of the standard "ship code, then review" loop. It treats the design and the implementation as two distinct surfaces, each with their own adversarial agent set. Self's framing is the right one: in a world where the model writes the code, the differentiator is the skepticism applied to it. The doubt is the feature.

Alibaba open-sources Open Code Review

Open Code Review is an AI-powered CLI tool that automates code reviews by analyzing Git diffs for precise, line-level feedback. It uses a hybrid architecture of deterministic logic and dynamic agents, allowing developers to integrate it easily into CI/CD pipelines or local environments. The release matters for two reasons. First, it is from Alibaba, which means the big-tech code-review tooling is going open source at the same time the model layer is going closed-API. Second, the hybrid deterministic-plus-dynamic-agent architecture is the pattern every code review tool is converging on. Pure LLM review is too noisy. Pure deterministic review is too narrow. The combination is what works, and Alibaba is putting it in the public domain.

A 24-minute primer on how LLMs actually work

How LLMs Actually Work is a 24-minute read that walks through the transformer stack end-to-end. LLMs operate by converting text into subword tokens and mapping them to numerical vectors called embeddings that represent semantic meaning. The models use positional encoding to track the order of words and attention mechanisms to allow tokens to weigh the importance of other information in a sequence. Within the transformer stack, feed-forward networks process tokens individually while residual connections and normalization layers maintain mathematical stability across many layers. The piece also addresses a related question in TLDR AI, noting that modern LLMs are mostly built by stacking transformer blocks over and over, and the differences come from what each one was trained on, the scale and configuration choices, and the post-training done on top. The two pieces are the closest thing to a single-document explanation of the architecture that I have read in 2026.

The Take

Here is the pattern this digest is describing: the AI engineering loop is becoming more skeptical, more open, and more legibly documented, and that is the right direction. "Automated doubt" is the workflow that scales the code review function in a world where the model is doing the writing. Open Code Review is the tooling that puts that workflow in every developer's terminal. And "How LLMs Actually Work" is the documentation that lets every developer understand what they are building on top of. None of these are flashy. All of them compound. The next eighteen months of AI engineering will be defined less by model choice and more by the quality of the review and documentation layers around the model. The teams that invest in those layers now are the ones that will be able to ship at the rate the models are improving. The teams that do not invest in those layers are the ones that will be drowning in AI-generated code they do not understand and cannot review. Doubt is the feature. The review loop is the moat. The documentation is the trust contract.

Automated doubt , open code review , how LLMs really work