↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading — #760 (April 9, 2026)

Seroter's Daily Reading·

Listen: https://blossom.nostr.xyz/91b9a77c45f329d22d7560a617c701f3433f20eb68ded2da4b96e04a2b5b9e8c.mpga

Source: Seroter's Original Post

Seroter's Daily Reading episode 760, April 9th, 2026. Let's dig in.

First up, a piece that will resonate with anyone who's ever waited for a news page to load while their laptop fan whirs into overdrive. The article is called The 49MB Web Page, and it describes landing on the New York Times website and being greeted with 422 network requests and 49 megabytes of data. Two minutes to settle. That's bigger than the entire Windows 95 operating system, just to read four headlines. The piece walks through what's happening in that network tab: silent programmatic ad auctions happening right in your browser, dozens of concurrent bidding requests to exchanges like Rubicon Project and Amazon Ad Systems, megabytes of JavaScript being downloaded, parsed, and compiled on your device before you've even finished reading the headline. The author's core argument is that publishers are trapped in an ad-tech death spiral, trading long-term reader trust for short-term CPM optimization, and the entire experience has become adversarial by design. He documents the pre-read ambush of stacked GDPR modals, newsletter popups, and notification prompts that users must dismiss before accessing content, and the cumulative layout shift that occurs when an ad slot finally resolves and pushes text out from under a reader mid-paragraph. The piece does acknowledge that economic pressures are real, but argues that better is possible, pointing to lightweight alternatives like text.npr.org as proof that readers still want no-frills, content-focused experiences.

Next, a piece from a16z on Antarctica, and the Extreme Logistics of Human Exploration. The article uses the Artemis II moon mission as a springboard: a ten-day trip in a capsule stocked with hot sauce and iPhones is manageable, but scale that to a 1,000-day Mars mission and you'd run out of food and oxygen well before arriving. The article then zooms in on Amundsen-Scott South Pole Station, which turns out to be more isolated than the International Space Station, sitting more than 2,400 miles from the nearest city and completely cut off for eight months during Antarctic winter. The US Antarctic Program resupply operation is staggering: one cargo ship and one fuel tanker land at McMurdo each February carrying 400 to 500 shipping containers and six million gallons of fuel, and everything from there to the South Pole must be flown in on ski-equipped LC-130s operated by the Air Force, or in some cases hauled overland by a tractor-train called the South Pole Traverse. Fuel delivered to the South Pole costs about forty dollars per gallon, compared to four dollars on the mainland. The article covers the power infrastructure, the water system that melts snow underground using hot water wells, and the remarkable fact that one-sixth of the station is dedicated to recreation, including a basketball court, to help researchers maintain sanity through months of total darkness. A fascinating piece on the unglamorous but critical engineering that keeps humans alive at the bottom of the world.

Third, a quick update from Google on Try notebooks in Gemini to easily keep track of projects. This is an integration that brings NotebookLM-style organization directly into Gemini, letting you create dedicated notebooks where you can group related conversations, add files and PDFs as context, and give Gemini custom instructions per project. It syncs with NotebookLM so the same sources work across both products. If you're already using Gemini for multi-step projects, the ability to keep longer-term context organized and persistent sounds genuinely useful.

Fourth, Addy Osmani has a piece called Your Parallel Agent Limit. He's been writing about the management model for orchestrating agents, but this post gets into the personal overhead: what it actually feels like to supervise multiple agent sessions simultaneously. The core insight is that adding agents doesn't scale your cognitive capacity. Your bandwidth for evaluating, trusting, and integrating output is still single-threaded on your side of the loop. What adds up fast is the ambient anxiety of knowing a thread might be drifting while you're focused elsewhere, the context-switching cost of reloading a mental model each time you switch back to a thread, and the trust calibration that degrades when you're not watching closely. He references Simon Willison's observation that running four agents in parallel until 11am will leave you wiped for the day. Osmani's practical advice: treat long agentic sessions like deep focus work, time-box them with explicit scopes, and resist the seductive math of if one agent ships a feature in forty minutes, four should ship four. His actual ceiling is three to four threads depending on complexity, and he's learned to size task scope to fit the session window rather than jamming unbounded work into an open-ended run.

Fifth, an article from Harvard Business Review titled Managers and Executives Disagree on AI, and it's Costing Companies. The data comes from a Wharton School and GBK Collective survey of leaders at companies with revenues above fifty million dollars. The disconnect is stark: 45% of senior executives report significantly positive ROI from AI investments, compared to just 27% of middle managers. Sixty-four percent of executives say they've become much more positive about generative AI over the past year, but only 39% of middle managers say the same. When asked if their organization is adopting AI faster than competitors, 56% of executives say yes, but only 28% of middle managers agree. The authors argue this isn't a healthy tension between vision and execution, it's a fault line that could be delaying the transition from AI ambition to AI value. The article lays out five actions: diagnose before prescribing, co-create the playbook with managers rather than handing it down, reduce existing load before adding AI responsibilities, measure readiness not just adoption, and build feedback loops that reward honesty about what's not working. A useful framework for organizations trying to close the gap between leadership enthusiasm and frontline reality.

Sixth, Google announced five new features for Android XR, rolling out to Samsung Galaxy XR headsets. The headline is enterprise support: Android Enterprise is now available on XR devices, enabling large-scale deployments for immersive training and collaboration. They're working with major EMM partners like Microsoft Intune, Samsung Knox Manage, and others to make these headsets workplace-ready.

Seventh, a Medium piece walking through Fine-Tuning Gemma 4 with Cloud Run Jobs. The author, Shir Meir Lador, has been doing this with Gemma 3 and updated the pipeline for the new model. Gemma 4 brings a different architecture: it has native multimodal support for images, video, and audio, a 256K context window, and a Mixture of Experts model that activates fewer parameters during inference. The piece covers the key code changes required: handling multimodal input ordering, loading the correct architecture class, implementing a bulletproof label masking approach that accounts for the dynamic number of soft tokens per image, and using target modules all-linear to correctly wrap the custom Gemma4ClippableLinear layers used in the projection layers. Results on the Oxford-IIIT pet dataset show an accuracy improvement from the 89% baseline up to 93%, which the author notes is closer to the state-of-the-art of 94% for that dataset. The piece includes a complete walkthrough from local testing through staging model weights in GCS, building the container, and creating the Cloud Run Job.

Eighth, the PostHog newsletter on The Golden Rules of Agent-First Product Engineering. This one is specifically about building products that agents can actually use deeply, not just adding an MCP endpoint as an afterthought. Rule one: let agents do everything users can, which means exposing nearly everything in your API to agents, while using opt-in YAML configs to avoid accidentally exposing too much. Rule two: meet agents at their level of abstraction, which in PostHog's case means exposing SQL rather than get-insight and get-funnel endpoints, since agents already speak SQL fluently. Rule three: front-load universal context, loading PostHog taxonomy, SQL syntax, and critical querying rules at the start of every session so agents don't waste tokens rediscovering them. Rule four: writing skills is a human skill, which is about writing skill definitions the way you'd onboard a qualified employee, not as step-by-step manuals, and only including context that an agent can't discover on its own. Rule five: treat agents like real users, which means dogfooding headlessly using the CLI, doing manual trace reviews weekly, and feeding findings back into evals. A practitioner's guide from a team that has shipped this twice and gotten it wrong before getting it right.

Ninth, Waymo robotaxis are now tracking potholes and sharing that data with Waze. The pilot covers Austin, Atlanta, Los Angeles, Phoenix, and the San Francisco Bay Area, where Waymo says it's already identified about 500 potholes. Robotaxis loaded with cameras, lidar, and radar are apparently ideal data collection platforms, and Waymo is offering this to cities as part of its broader effort to build goodwill as it expands to more than twenty cities this year. Both cities and individual Waze users get access to the data.

Tenth, a piece from the DX newsletter on Developer Ramp-Up Time Continues to Accelerate with AI. They measure Time to 10th PR, the number of days between a developer's start date and their tenth merged pull request. In Q4 2025, they saw this cut from 91 days to 49 days for daily AI users. Their latest data shows it has dropped further to 33 days as of February 2026, a roughly 50% reduction since AI usage became widespread and a 15% drop just this quarter. The article notes that the improvement is partly because AI has become organizational infrastructure rather than an individual choice: developers are guided to use it for codebase questions, architectural explanations, and draft changes for senior engineer review. The piece does flag that Time to 10th PR is a narrow measure that doesn't capture quality or depth of understanding, a caveat worth noting.

Eleventh, Google Colab is adding a Learn Mode to its Gemini integration. This turns Gemini into a step-by-step coding tutor rather than just a code-writing assistant. The piece frames this as useful for seasoned developers brushing up on new frameworks, educators designing coursework, and students new to coding. There's also Custom Instructions that notebook authors can save and share, letting you set preferences for coding style or library choices that Gemini will follow within a specific notebook.

Twelfth, a Google Research blog post on Improving the Academic Workflow: Introducing Two AI Agents for Better Figures and Peer Review. One agent helps researchers automatically generate publication-quality figures from their papers, and the other assists with the peer review process. The research shows the PaperVisAgent framework outperforming baselines across faithfulness, conciseness, readability, and aesthetics, scoring 60.2 out of 100 on their evaluation metric, which is the only framework to exceed their human baseline of 50.

Finally, a CIO Dive piece titled Unstructured Data Is Piling Up as AI Risks Rise. A Thales and Cloud Security Alliance report found that more than half of companies have only partial visibility into their unstructured data, even as AI ambitions grow. Sixty-eight percent of organizations say a majority of their data remains unprotected. Nearly a third of organizations use more than eleven tools to manage unstructured data, creating silos that hinder governance. The article notes that only 9% of organizations can scan data in real time, and nearly a quarter can't scan at all, which means AI systems risk amplifying existing blind spots rather than resolving them. The key insight: AI is only as good as the data it's trained on, and the data foundation at most organizations is simply not ready for the workloads being asked of it. Many organizations are stuck in pilot phase not because the AI is bad, but because the unstructured data estate hasn't been set up to support it.

And that's the reading list for episode 760. Several threads running through these pieces: the gap between AI ambition and organizational readiness appears in both the HBR piece and the unstructured data article. The question of what AI can actually do versus what humans can sustain shows up in the parallel agent piece and the ramp-up time data. And the tension between chasing metrics and respecting user experience is front and center in the web bloat piece. We'll see you next time.

  1. [blog] The 49MB Web Page — https://thatshubham.com/blog/news-audit
  2. [article] Antarctica, and the Extreme Logistics of Human Exploration — https://www.a16z.news/p/antarctica-and-the-extreme-logistics
  3. [blog] Try notebooks in Gemini to easily keep track of projects — https://blog.google/innovation-and-ai/products/gemini-app/notebooks-gemini-notebooklm/
  4. [blog] Your parallel Agent limit — https://addyosmani.com/blog/cognitive-parallel-agents/
  5. [article] Managers and Executives Disagree on AI—and It's Costing Companies — https://hbr.org/2026/04/managers-and-executives-disagree-on-ai-and-its-costing-companies
  6. [blog] 5 new features for Android XR — https://blog.google/products-and-platforms/platforms/android/android-xr-immersive-features-update-april-2026/
  7. [blog] Fine-Tuning Gemma 4 with Cloud Run Jobs: Serverless GPUs (NVIDIA RTX 6000 Pro) for pet breed classification — https://medium.com/@DataLady/fine-tuning-gemma-4-with-cloud-run-jobs-serverless-gpus-nvidia-rtx-6000-pro-for-pet-breed-d408c7e24be2
  8. [article] The golden rules of agent-first product engineering — https://newsletter.posthog.com/p/the-golden-rules-of-agent-first-product
  9. [article] Waymo robotaxis are tracking potholes and sharing that data with Waze users — https://techcrunch.com/2026/04/09/waymo-robotaxis-are-tracking-potholes-and-sharing-that-data-with-waze-users/
  10. [article] Developer ramp-up time continues to accelerate with AI — https://newsletter.getdx.com/p/developer-ramp-up-time-continues
  11. [blog] Introducing Learn Mode: your personal coding tutor in Google Colab — https://blog.google/innovation-and-ai/technology/developers-tools/colab-updates/
  12. [blog] Improving the academic workflow: Introducing two AI agents for better figures and peer review — https://research.google/blog/improving-the-academic-workflow-introducing-two-ai-agents-for-better-figures-and-peer-review/
  13. [article] Unstructured data is piling up as AI risks rise — https://www.ciodive.com/news/unstructured-data-AI-risks-rise/816955/