Source: https://seroter.com/2026/10/28/daily-reading-list-october-28-2026-898/
Seroter's Daily Reading. Daily Reading List – October 28, 2026 (#898). Seroter opens by noting that he is heading up to San Francisco to attend GitHub Universe tomorrow, his first time, where he will be part of a panel discussion before flying home. The list includes Gemini model news, security patch concerns, JavaScript runtimes, agent workflows, Android updates, and the rise of Postgres.
The first link is Gemini 4 Carbon: our next frontier model. Seroter's take is that it is Google again, two weeks after the last frontier release, and this model builds on the previous one with strong improvement at long-horizon tasks. The post from Google DeepMind positions Carbon as its strongest model yet, with state-of-the-art cybersecurity and long-horizon capabilities. It ranks first on the updated Artificial Analysis Intelligence Index at fifty-nine, with big gains in vulnerability detection, exploitation, and patching benchmarks. The company says Carbon can run for days on multi-step engineering problems, improving coding and reasoning. Access is being rolled out deliberately, first to a subset of cyber defenders, with the same introductory price as Argon.
Seroter also includes a supplementary VentureBeat link labeled in the list as More here. That page could not be retrieved, so the only additional detail is Seroter's note that the model builds on the last and shows great improvement at long-horizon tasks.
Next is Q3 2026 earnings call: Remarks from our CEO. Seroter highlights a big week of public company earnings, saying Google had a strong quarter and Cloud is an absolute rocket ship. CEO Sundar Pichai's prepared remarks describe Alphabet revenues up twenty-five percent year over year, with a half-trillion-dollar annual run rate. Google Cloud grew one hundred three percent, and AI infrastructure is a major driver. First-party model APIs are processing nearly thirty-two billion tokens per minute, up forty percent quarter over quarter. Pichai also mentions the Gemini app reaching one billion users, Argon's fast developer adoption, and progress on search, agents, YouTube, Waymo, and quantum computing.
Then comes The State of Context Engineering. Seroter calls it an excellent read and says to get smarter on context fails, patterns, and opportunities. The piece is a lightly edited talk from DBreunig at AGNTCon. It traces how the industry moved from trying to stuff everything into huge context windows to understanding failure modes such as poisoning, distraction, confusion, and clash. It then lays out a toolbox of fixes, including augmentation, tool loadout, quarantine, pruning, offloading, and summarization. The talk ends by pointing to newer failure modes like fighting the weights and context inertia as models become more agentic.
The next article is “The panic hasn’t turned into patching”: What security researchers really think about the PHP2Shell response. Seroter observes that the vast majority of exposed systems have not been patched and asks whether we are failing our collective day-two responsibilities. The New Stack piece reports on the gap between the initial alarm over PHP2Shell and the slower work of actually securing exposed PHP servers. The headline captures the core concern: panic arrived, but patching has lagged.
Apple’s git still trusts the compromised eMudhra certificates follows. Seroter notes that while it is hard to patch a lot of software quickly, the industry has not been fast at closing holes identified the previous week. The post explains that Apple was notified of the eMudhra incident in mid-October, yet the git, curl, and Python tools shipped with macOS still trusted the compromised certificates as of late October. Worse, a macOS update undid the manual workaround some users had applied. Because of this, anyone on the network path could intercept HTTPS connections, tamper with clones and pulls, and steal credentials. The author argues Apple should remove the affected roots quickly, and developers should reapply workarounds after updates.
Next is The Web Can’t Afford a JavaScript Runtime Monoculture. Seroter says consolidation is a natural response to an upfront explosion in options, and asks whether the web is landing on too few JavaScript runtimes now that Deno is going away. Kate Holterhoff of RedMonk argues that Deno's exit leaves Node.js and Bun as the major general-purpose runtimes, with Cloudflare's workerd tied to its own platform. The essay compares this to Internet Explorer 6 and browser engine consolidation. It warns that server-side JavaScript already relies on a small number of underlying engines, that standards benefit from multiple independent implementations, and that Deno pushed important features into Node. The lesson is that competition has been what keeps the web moving forward.
Then there is The sleep framework: A new way for LLMs to retain knowledge in continual learning. Seroter frames the challenge: how can LLMs keep learning new facts without forgetting what they already knew? He says it is hard, but something Google is looking at. The research proposes a sleep phase for continual learning, alternating wake and sleep periods. During sleep, the model consolidates short-term context into longer-term parameters, uses Mixture-of-Experts memory blocks, and even dreams by generating synthetic practice examples. The team reports strong results for long-context understanding up to ten million tokens, factual knowledge incorporation, translation, and mathematical reasoning.
How do developers keep momentum as AI agents change the work? is next. Seroter asks whether AI messes with flow, and whether developers can get and sustain momentum when agents are involved, calling the research interesting. The Research-Driven Engineering Leadership piece describes interviews with experienced developers and a Theory of Developer Momentum. The researchers found that momentum starts like climbing a hill, with interruptions most damaging before the path is clear. Developers relied heavily on notes and left deliberate hooks for their next session. Waiting on slow tests or agents broke orientation, and caring about the work mattered. The application section suggests protecting early scoping time and treating slow feedback loops like interruptions.
Then Jason McCreary's piece, I don’t write code anymore. Seroter identifies with it, saying he is not paid to write code anymore either, but he is producing more software and having more fun than at any point in his career. McCreary describes four stages: letting AI handle tedious snippets, writing tests first, teaching AI to code like him, and finally letting go entirely. His guardrails are tests, static analysis, CI, and adversarial agents. He argues that programmers' value is delivering a result, not typing code, and calls himself a software foreman.
Next is GitHub Universe 2026: Evolving every layer of the stack. Seroter says it looks like a solid vision for software engineering work in the AI era. The GitHub post is authored by Jay Parikh, who leads CoreAI at Microsoft and GitHub. Seroter also links a DevOps.com piece labeled More here about GitHub adding tools to manage AI coding agents running in parallel at scale, but that page could not be retrieved. Its title suggests the announcement includes management capabilities for parallel coding agents.
Then comes Fast, reliable A2UI, with swappable inference formats. Seroter's reaction is, now we're talking, saying this is moving from a useful idea to a production-ready implementation and fast generative UIs are here. The post introduces A2UI Express, a new intermediate inference format that reduces latency and token use compared with generating JSON directly. It also demonstrates how decision models like Jev can choose from predefined layouts much faster. The team describes swappable inference formats as a lever to choose the right trade-off per screen, with code and macros to support the new approach.
Agentic and Flexible: A next era of Android development is next. Seroter notes that, speaking of A2UI, it is now part of the Android experience with Jetpack Compose, and much more was announced today. The Android Developers blog covers Android Studio Rabbit 2 bringing your own agent, Android CLI updates for device streaming, Compose previews, and profiling. It introduces Android intelligence system APIs, AppFunctions in beta, A2UI for Jetpack Compose in alpha, on-device Gemini Nano 4 capabilities, adaptive layout tools, and stable Jetpack XR libraries.
Next is Our Favorite Management Tips on Delegating Effectively from Harvard Business Review. Seroter says the advice is meant for human managers, but he read much of it thinking about how it applies to delegating work to AI. The piece is a curated collection of ten favorite tips on delegating effectively, organized around giving managers practical ways to hand off work while staying accountable.
What was the turning point for Postgres? asks PlanetScale. Seroter says Postgres is winning right now, and asks why, noting it definitely was not an overnight success. Simeon Griggs traces the pivotal moments, from Postgres's extensible design in the 1980s, through its near-death in the mid-nineties, the Oracle acquisition of MySQL, Heroku's early bet on Postgres, JSONB support, and its eventual lead in Stack Overflow surveys. By 2025, LLMs and venture money had made Postgres the default choice for new applications.
Then there is How AI agents shape experts’ judgment calls. Seroter asks whether experts are influenced more by AI than they think, and reports that Anthropic's research looks like people tend to adopt the agent's direction and claim it as their own. In the study, fifty analysts worked with Claude Code on a causal-inference task, while the agent was secretly given random preferences on four decision points. Analysts followed the agent's lean eighty-six percent of the time. When they had already expressed a preference, they switched away sixty percent of the time when the agent disagreed. Most repeated the agent's arguments later, but few credited the agent. The authors see this as a risk to independent judgment and an important area for further research.
The final link is Why “Plan Mode” is just scoping. Don’t forget discovery! That page could not be retrieved. Seroter's note says that if you want to stop the agent before it starts writing code, you have a few options, and Alexis points out when to use which. The title suggests the author is cautioning not to collapse planning into scoping alone, and to keep discovery in the loop.
Across today's list, a few threads stand out: AI is moving from writing code to shaping judgment, security patch lag remains a real risk, and platform choices are consolidating in ways that deserve attention. Seroter's daily list, as usual, connects model announcements, developer experience, and infrastructure decisions.
- Daily Reading List – October 28, 2026 (#898)
- Gemini 4 Carbon: our next frontier model
- More here
- Q3 2026 earnings call: Remarks from our CEO
- The State of Context Engineering
- “The panic hasn’t turned into patching”: What security researchers really think about the PHP2Shell response
- Apple’s git still trusts the compromised eMudhra certificates
- The Web Can’t Afford a JavaScript Runtime Monoculture
- The sleep framework: A new way for LLMs to retain knowledge in continual learning
- How do developers keep momentum as AI agents change the work?
- I don’t write code anymore
- GitHub Universe 2026: Evolving every layer of the stack
- More here
- Fast, reliable A2UI, with swappable inference formats
- Agentic and Flexible: A next era of Android development
- Our Favorite Management Tips on Delegating Effectively
- What was the turning point for Postgres?
- How AI agents shape experts’ judgment calls
- Why “Plan Mode” is just scoping. Don’t forget discovery!

