Dev Breakfast · 2026-09-17

Firefox 156 has turned the ad space in the address bar, officially called Firefox Suggest; in the same version, PDF startup is up to 45% faster. My first reaction wasn't "the privacy browser has fallen," but rather to first confirm whether that switch is on or off by default, and if it's turned off, whether it might be quietly turned on again in the next version.

Firefox 156 pushes 'Suggest' ads in the address bar, PDF starts up 45% faster

Firefox 156 has stuffed an ad space into the address bar, officially called Firefox Suggest; the same version also increased the PDF reader's startup speed by up to 45%. Having one selling point and one point of contention in the same update is quite interesting. For people who code, the address bar is something you type hundreds of times a day, so whether the additional recommendations will mix into autocomplete and whether they can be turned off is worth confirming before that 45%. My first reaction wasn't "the privacy browser has fallen," but rather: is this switch on or off by default, and if turned off, might it be quietly turned on again in some version? The business model of privacy products is inherently at odds with their promises, and all users can do is remember the switch's location.

Sources:

• Firefox 156: PDF viewer starts up to 45 percent faster

Karpathy's autoresearch six months later: Shopify uses it to improve 40+ metrics, rekursiv refreshes nanochat record in three days

Karpathy open-sourced autoresearch in March, compressing "AI doing research" into a loop: humans only write program.md, agents only modify train.py, each experiment trains for 5 minutes fixed, looking only at one metric val_bpb, keeping the best and discarding the rest, about 12 experiments per hour, about 100 overnight. The repo has 96k stars to date. In the past six months, it's not only been used for training models: Shopify engineers adapted the same loop into pi's extension pi-autoresearch, generalized to over 40 engineering metrics within the company, and Tobi Lütke used it against the Liquid template engine, with parse+render combined 53% faster and 61% fewer object allocations (he also said "might be a bit overfitting"); on September 14, rekursiv.ai released auto-autoresearch, allowing a group of agents to share each other's experimental records (including failures), running 6,164 experiments in three days, refreshing the nanochat benchmark. For people who code, what can be borrowed is the design principle from the README: one GPU, one file, one metric, plus fixed duration—narrowing the scope of changes and evaluation criteria, so the agent dares to leave it unattended.

Sources:

• karpathy/autoresearch - GitHub
• (auto)²-research - rekursiv.ai
• Autoresearch isn't just for training models - Shopify Engineering

Replacing actions/setup-go: Golang CI scaling actual test

The original post only had a title and one line "Scaling Golang CI by Replacing actions/setup-go," without details on what it was replaced with or the difference before and after scaling. The only thing certain is that someone replaced the step for installing Go in GitHub Actions from the official action, with the reason being that it became a bottleneck as scale increased. If you're running Go CI, this is worth clicking to see the approach—but don't rush to copy it; first measure how much time installing Go actually takes in your pipeline.

Sources:

• Scaling Golang CI by Replacing actions/setup-go

Gemini 3.8 Live launches: adds an "Extended Thinking" mode

Google released Gemini 3.8 Live, along with a new 3.8 Live Extended Thinking mode. The original post only had a title, no parameters, no pricing, no benchmarks, so I can't tell you how much stronger it is than the previous generation. People working on voice and real-time interactions might take note: the Live line has always prioritized low latency, adding "Extended Thinking" means the same model has to trade off between response speed and reasoning depth, and how this trade-off is adjusted and how much latency increases is key to determining whether you can go to production. For model releases, I always first look at what data it was tested on; even if the official numbers are impressive, I put them aside first.

Sources:

• Gemini 3.8 Live and 3.8 Live Extended Thinking

Google copied open-source code and erased engineers' names

On Reddit's reinforcement learning section, someone posted that Google copied their open-source code, removed engineers' names, and gave no credit. The original post only had a title, without posting the specific repository, file comparison, or commit records. For such allegations to hold, two things are needed: which license it is, and whether what was removed was the copyright notice or just author notes—MIT and Apache 2.0 have different attribution requirements. Before evidence is presented, I tend not to jump to conclusions, but the order of "first check the license and diff" cannot be skipped.

Sources:

• Google copied our open-source code, removed engineers' names without credit

Swapping CSS changes the face: the CSS Zen Garden concept finally built

CSS Zen Garden proved one thing back then: without changing HTML, only swapping CSS, the page could be completely transformed. Now someone has truly made this concept usable—the same HTML with different stylesheets results in completely different skins. For front-end developers, the significance of this isn't nostalgia: it shows that when structure, style, and behavior are separated cleanly enough, skinning can be zero-cost. If you're bogged down by coupled components and global styles, it's worth clicking in to see how they made this cut.

Sources:

• The CSS Zen Garden dream, finally shipped

Cloudflare introduces "mixed-use crawlers": choose between search visibility and AI training

Cloudflare has launched a "mixed-use AI crawler" management solution, with the core being to allow sites to maintain search visibility while refusing content to be used for training models. The original post only had a title and one line "Stay discoverable in search while disallowing AI training," without details on how to distinguish or which crawlers it covers.

For people who code, this is worth noting: in the future, robots.txt and UA whitelists might not be a binary choice, but split by purpose. Whether to follow up, first see which UAs it recognizes and how high the probability of accidentally hurting search crawls is.

Sources:

• Accountable Mixed-Use AI Crawlers

Google Play review often exceeds one week, release rhythm needs to change

Google Play app reviews now often take over a week, according to feedback from the developer community. For people who code, this means you can no longer schedule releases based on "submitting and being live the next day"—hotfixes, urgent rollbacks, and such operations need to recalculate buffer time. The original post only had this sentence, without giving specific time ranges or explaining which app types are affected. My suggestion is: treat the review cycle as an uncertain variable in management, and don't set the release window too tight.

Sources:

• hackernews · inputmice

---

The address bar is something you type hundreds of times a day; will you turn off Suggest, or tolerate that 45% PDF speedup? See you tomorrow at 8 AM.

This issue selected 8 out of 59 pieces of information from the past 24 hours on X / Hacker News / GitHub Trending (written hourly throughout the day, fact-checked and compiled in the morning). Content is generated with LLM assistance, each item includes original source links; important decisions should be cross-verified.