Dev Breakfast · 2026-08-29

openai-python officially begins HTTPX2 migration, with official documentation ready, but custom transports, timeouts, and error handling are major pain points. When replacing the underlying library, the biggest fear is that old workarounds suddenly fail, so run tests first before proceeding.

OpenAI's official Python library begins migrating to HTTPX2: Points to review before upgrading

OpenAI has released migration notes in the openai-python repository, officially launching the upgrade to HTTPX2. For developers using this library to write agents or pipelines, this means the underlying HTTP client will be replaced with a new version, and API call behavior may have subtle changes. It is recommended to run existing tests before upgrading, focusing on checking whether custom transports, timeout settings, and error handling logic are compatible. The official documentation is ready, and it's worth reviewing before proceeding.

💡 Chef's take: Every time the underlying library is upgraded, the biggest fear is not that new features can't be used, but that old workarounds in the code, which were written 'without knowing why at the time,' suddenly fail. Run tests first, then discuss migration.

Sources:

• OpenAI Migration Notes

GLM-5.3 releases weights: doors open for local deployment and fine-tuning

Zhipu AI officially announced the release of GLM-5.3 weights, allowing developers to download the model for local deployment and fine-tuning. This changes the previous decision condition we mentioned of 'waiting for evaluation data to decide'—if you have private deployment or customization needs, you can start testing now; but the actual performance level still depends on third-party benchmark data, so it is recommended to test in your own business scenarios before making a final judgment.

💡 Chef's take: Releasing weights is good, but don't rush to switch online traffic—first see how it performs on your own dataset, then decide whether to say goodbye to your cloud provider's bills.

Sources:

• Hacker News · jeudesprits

10G NIC runs at 300M: Hanselman's troubleshooting log

Scott Hanselman wrote a new blog post, documenting his troubleshooting process for his newly assembled 10 Gigabit Ethernet network that actually runs at only 300 Megabits. The theoretical rate for 10G is 10,000 Megabits, so running at 300M means a 97% performance drop. Such a large gap basically rules out hardware issues and points to configuration, cable, or driver problems. For developers involved in intranet transfers, NAS, or self-hosted data centers, this type of issue is typical—network cards, switches, cables, MTU settings, any incorrect link can cause a cliff-like drop in bandwidth. It's worth clicking to see which pitfall he encountered, so you can avoid them when setting up high-speed networks yourself.

💡 Chef's take: 10G running at 300M, don't blame the ISP first—check if the network cable is bent beyond the radius, or if a patch cable is Cat5e pretending to be Cat6a—hardware doesn't lie, labels do.

Sources:

• Debugging my new network, when 10 Gigabit Ethernet Runs at 300 Megabits

Twelve-Factor App releases 2025 version: major revision to architectural guidelines

The Twelve-Factor App manifesto has released its 2025 revision, marking the largest update in over a decade for this architectural guideline that has influenced a generation of backend developers. The new version has adjusted and expanded upon the original twelve factors, focusing on responding to the new realities of containerization, serverless, and cloud-native eras—such as the process model no longer assuming long-running services, configuration management adapting to dynamic environments, and log processing aligning more closely with modern observability systems. For coders, this means that the 'legacy architecture specifications' in your team may need to be reviewed against the new standards, especially when starting new projects—don't use the old 2011 guidelines for technology choices.

💡 Chef's take: The Twelve-Factor has been updated, but your old system won't automatically become version 12.0—the refactoring bill is still yours to pay, so see which factor hurts the most before proceeding.

Sources:

• Hacker News · jxmorris12

tailcat released: netcat bypassing control plane, directly connecting to Tailscale data plane

Tailscale open-sourced tailcat, a netcat alternative that uses the Tailscale data plane but doesn't depend on its control plane. This means you can use Tailscale's encrypted tunnels for data transfer without logging into an account or configuring node relationships, suitable for quickly transferring files or testing ports in an existing Tailscale network. For scriptwriters and debuggers, one less control plane dependency means one less point of failure. It's already on GitHub, so clone it and run it if you want to try it out.

💡 Chef's take: Skipping the control plane sounds clean, but don't forget—without it, node discovery and key rotation also need to be managed by yourself, so don't turn 'simplification' into 'going naked.'

Sources:

• GitHub - tailscale/tailcat

Claude 48-hour autonomous alignment: safety score does not drop

Anthropic released new research: having Claude autonomously align a small model using 48 hours and 1 GPU, it researched methods, trained, and tested itself, stably improving safety scores without losing capability across 10 types of alignment failure scenarios, and the best method can generalize to unoptimized benchmarks and models 4.7 times larger than the training target. In another test, Sonnet 5 post-trained an early checkpoint of Opus 4.8, achieving safety scores close to the fully aligned trained version. External misuse is a short-term risk, while autonomous alignment is a long-term capability, and the two are not contradictory, but safety assessment must simultaneously monitor external attacks and internal autonomous behaviors. Anthropic has open-sourced the automated alignment research environment, so those who want to reproduce it can start directly.

💡 Chef's take: Letting the model align itself sounds like letting the examinee grade their own paper—fortunately, it didn't give itself a perfect score, but next time it learns to perform well on benchmarks, who will be the proctor?

Sources:

• Anthropic official tweet: Sonnet 5 post-training Opus 4.8 early checkpoint
• Anthropic official tweet: automated alignment research environment open-sourced
• Anthropic official tweet: Claude 48-hour 1 GPU autonomous alignment experiment
• Anthropic official tweet: hill-climbing safety benchmark details
• Anthropic official tweet: 10 types of alignment failure results and generalization performance

Terminal-Bench-Science: a new arena for AI agents in scientific research

Terminal-Bench-Science is now online, a benchmark specifically for evaluating AI agents in scientific research workflows. The project has posted an announcement on its official website, with the title directly pointing to 'Evaluating AI agents on scientific research workflows.' Currently, public information is limited, with specific evaluation tasks, dataset scales, and baseline results not yet released. For coders, the value of this benchmark lies in: it extends the evaluation of agent capabilities from writing code and calling APIs to scientific research scenarios, meaning the dimensions of agent capability assessment are expanding. To follow up, save the official website first and wait for details to be published before judging its value.

💡 Chef's take: Another 'scientific' benchmark—don't get too excited yet—wait until it releases the task set and baselines to see if it's just their own model topping their own leaderboard.

Sources:

• Terminal-Bench-Science official announcement

---

Do you dare to upgrade directly to HTTPX2: Upgrade / Wait and see? See you tomorrow morning at 8.

This issue selected 7 items from a total of 65 pieces of information from X / Hacker News / GitHub Trending over the past 24 hours (written hourly throughout the day, fact-checked, and compiled in the morning). Content is generated with LLM assistance, each item includes original source links, and important decisions should be cross-verified.