Week 28 · 2025-07-07 → 2025-07-13 · 18 newsletters

The Soham Mirror

soham-and-the-system · context-and-thinking · downshift-and-solitude

A sparse summer week. Eighteen newsletters across seven days, no breaking macro thread, no dominant news story. The Soham Parekh overemployment saga from the week prior bled into two thoughtful follow-ups, the AI conversation quietly pivoted from prompts to context, and a small cluster of writers wrote about thinking, solitude, and what to do with less. The week earned its short word count. I am not going to pad it.

Theme: The Soham Mirror

The week's only real cultural through-line was Soham Parekh, and the two best takes on him arrived three days apart. Jacob Voytko at Client/Server opened the week with a thought experiment: if you woke up as Soham, how would you actually optimize the overemployment grift? His read was that Soham was bad at the scheme, not good at it. The compensation was equity-heavy with a one-year cliff, which means each job paid out almost nothing before he got fired. He kept getting caught. He skipped meetings, ghosted PRs, gave wild excuses. The scam was real; the execution was amateur. Voytko's reframing as a strategy problem rather than a morality play is the version of the story most coverage missed.

Carly Ayres at Good Graf ran the companion piece in "The grift we deserve." Her angle was that Soham did not hack the system. He interviewed. He passed technical screens. He got offers. The system worked exactly as designed. The real scandal is the flimsiness of startup recruiting safeguards, not the existence of someone willing to test them. The symmetry she draws is the one that lingers: the same week Soham went viral for getting hired everywhere, 22-year-old Roy Lee raised $15M from a16z for Cluely, a product whose pitch is "cheat on everything." Roy got celebrated for the same energy Soham got dragged for. Cluely had even interviewed Soham for a job that same week.

The take: the Soham story is interesting not because of what one person did but because of what it surfaces about the recruiting funnel and the founder mythology around it. Voytko gives you the operator's read; Carly gives you the cultural read. Together they are the only commentary on the saga worth keeping.

Theme: From Prompts to Context

The AI conversation this week pivoted from prompts to context, and it pivoted cleanly. Addy Osmani at Elevate published "Context Engineering: Bringing Engineering Discipline to Prompts," arguing that prompt engineering as a discipline is graduating into something larger. The frame: providing an AI with all the information and tools it needs to complete a task is a different skill than cleverly phrasing a single question. His checklist (be precise, provide relevant code, include design docs, share full error logs, show schemas, give examples, state constraints) is the operator-level version of the argument. The pitch is that "program in prose" was always a half-truth, and the real craft is constructing an information environment.

Sahar Mor at AI Tidbits ran the agent-stack companion with his June 2025 LinkedIn highlights, covering LangMem for long-term agent memory, browser-use for browser automation, OpenAI Agents Python, Agno, and agents-towards-production. The packages are the through-line: memory, browser control, agent orchestration, production patterns. The infrastructure for serious agents is consolidating into open-source primitives faster than the consumer-facing applications can absorb them.

Ethan Mollick at One Useful Thing wrote "Against Brain Damage," the most thoughtful pushback I read this week on the MIT Media Lab "Your Brain on ChatGPT" paper. His read: the study involved a small group of students writing essays with ChatGPT, Google, or nothing, and the ChatGPT users were less engaged and remembered less of their own work. The press coverage turned this into brain damage. Mollick's point is that yes, outsourcing thinking costs you something, the same way calculators cost us mental math and phones cost us memorized numbers, but the real question is which cognitive work is worth outsourcing and which is worth keeping. The historical frame (Plato on writing, the cellphone-and-phone-numbers panic) is the right one.

Henrik Werdelin at Inklings ran the prosumer field report. He has been using Replit and Lovable to replicate paid SaaS tools, building a custom CRM with vector search to replace Clay, and wondering whether AI no-code tools will unbundle SaaS the way streaming unbundled the album. He also got a Limitless Pendant and reported on his Audos studio, which spins up "donkeycorns" (one or two-person, million-dollar businesses) by giving founders $25k plus AI agents in exchange for a 15% revenue share. The unbundling thesis is the one to file away.

The take: prompts were the surface; context is the substrate. The week's AI writing collectively argued that the hard work is no longer phrasing the question but constructing the environment the model answers from. The Mollick piece is the right counterweight: outsourcing is a tradeoff, not a free lunch, and the engineering discipline now applies to figuring out which parts of your thinking are worth keeping in-house.

Theme: Downshift, Solitude, and Strong Opinions

Steven Schlafman at Where the Road Bends wrote the most honest piece of the week in "Where the Road Ends." He describes a future-self visualization where he met his 2029 self and discovered the outer contours of life looked identical to today. Same house, same marriage, same coaching work, no bestseller, no big company, no newfound wealth. What had changed was internal: deeper presence, regulated nervous system, rooted at peace. The realization was that the life he had been projecting toward was the one he was already living, and the work ahead was not addition but subtraction. The piece announces Downshift, his new direction. It lands precisely because it refuses the conventional escape-velocity narrative.

Rob Thomas at The Mentor wrote the parallel argument in "Thinking," on solitude as the substrate for actual thought. The Isaiah Berlin anecdote (becoming a historian of ideas because of one sleepless WWII bomber flight with nothing to do but reflect) is the kind of frame that earns its place. His line: thinking happens in the margins, not in the meetings. The people you admire spend more time alone than you think. Not lonely, alone.

Chandra Narayanan at Opinionated Intelligence ran the systems-level cousin in "Why the Best Decisions Start with Strong Opinions." The frame is that the best decisions come from forming clear, experience-based beliefs about what matters, then using them to interpret data and build systems. The original Macintosh is the case study: Jobs insisted on no fan, on the mouse, on beautiful typography, against the consensus of his engineering team. Strong opinions, weakly held is the pop-culture version of this; Narayanan's version is more disciplined.

Ami Vora at The Hard Parts of Growth wrote the product-design adjacent piece in "To build trust in complexity, offer small choices and fast feedback." Her example is a privacy recovery flow that calmed her down not because it removed the steps but because it laid them out, channeled anxiety into small actions, and gave fast feedback on each one. The argument generalizes: when predictability is impossible, participation is the next best thing. Helplessness becomes action when the user gets to make small choices and see immediate confirmation.

Carly Ayres ran a second piece on zines as pedagogy, about a librarian named Abby who posted that her undergrads had said "zines are kind of like anti-AI" and watched the post go wide. The hunger for tactile, expressive, unmediated authorship is real, and it is showing up in classrooms as a counter-formation to screen fatigue. Piera Luisa Gelardi at Noomalooma wrote the parenting version of the same instinct: a play-as-resilience field report from a jet-lagged travel day with her six-year-old, where the discipline of staying playful turned a slog into something else. Ben Kassoy at A Strawberry Spinning Like A Dreidel announced his partner Kristen Browne's run in Constellations at The Actors Company in LA, July 23-27, and the next Future Castles writing workshop on July 15.

The take: the small writers carried this week more than the industry writers did. Schlafman on subtraction, Thomas on solitude, Narayanan on conviction, Vora on participation, Ayres on tactility, Gelardi on playfulness. They are running variations on a single instinct about what to do when the world keeps asking you to do more. The unifying argument is that the answer is rarely more.

Three Takeaways from the Week

The Soham Parekh story turned out to be a Rorschach test for what you already believed about startup recruiting. Voytko read it as an operator's puzzle (the scheme was poorly executed). Ayres read it as a system stress test (the scaffolding was hollow). Both are correct, and the value of reading them together is that it forces you to hold both reads at once. The morality framing was the least interesting version of the story; the system-design framing is the one that earns its airtime.

The pivot from prompts to context is the real AI through-line of the week, and it is the kind of pivot that quietly resets what counts as expertise. Addy Osmani's checklist, Sahar Mor's package roundup, and Mollick's pushback on the brain-damage discourse are different facets of the same shift. The skill that mattered six months ago (phrase a clever question) is being replaced by the skill that matters now (construct the right environment, and decide which cognitive work to keep in your own head).

If you only revisit three pieces from the week, I would suggest Carly Ayres on "The grift we deserve" for the cleanest read on the Soham saga and what it surfaced, Addy Osmani's "Context Engineering" for the cleanest articulation of where the AI craft is heading, and Steven Schlafman's "Where the Road Ends" for the post most worth reading slowly on a Sunday afternoon. A sparse week, but those three earn their slot.