In late October 2026 I posted a series of X threads that covered a panel announcement, a security threat outlook, a UX regression, EU market positioning, compensation observations, and my view on the sustainability of the software‑engineering premium.
The panel titled System Update will be held on Thu, 5 Nov pm in SF, featuring leaders from OpenAI, Linear, DoorDash and a live book signing by me.
I warned that every company with customer data WILL be targeted, and many will be hacked.
The breach on October 21st involved a malicious actor with access to abliterated models and substantial compute resources, enabling highly sophisticated exploits against our infrastructure.
The UI “connectors” management screen was removed, so I can no longer see what services I've connected, and with what accounts.
Grok Bot is the only major AI platform launched in the EU/UK; Meta Muse and OpenAI Dots are not available there, reducing competition.
I observed that Visa, Mastercard, AMEX and JPMC have higher profit margins than SaaS/Big‑Tech yet pay lower salaries; startups (often loss‑making) pay more than these traditional firms.
Stability period: ~10-20 years of stability.
Premium reason: more flexible (WFH/remote) and better perks.
Premium reason: fewer people are able/willing to pick up new technologies, methodologies so quickly.
Career impact of AI adoption: software engineers who pick it up and master it will generally fare better; those not will struggle a lot more.
These observations shape my view that compensation premiums stem from willingness to learn fast, flexibility, and the rapid evolution of the field, especially as AI becomes the next decisive technology.
In my X posts from October 2026 I highlighted several trends and announcements.
The property‑based testing webinar RSVP link is https://t.co/QmpE5wHubY.
Will left a team of 250 engineers at Carta.
Will now leads ~25 engineers at Imprint.
eMudhra generated private keys between 1 and 256 bits.
The required private key size is 384 bits.
The Pragmatic Engineer Unplugged event was scheduled for 4 Nov at SFJazz Center, SF.
I observed that 567,091 people viewed (or at lest scrolled past) the announcement of the competition.
5,179 people visited the essay challenge page.
879 opened the submission form.
125 people submitted an essay.
105 people are receiving a TPE gift package for putting in considerable effort, producing a human‑written essay that our team found interesting to read, and to learn from.
7 cash prizes were awarded.
1 Grand Prize winner was selected.
Security will cost a LOT more.
We’re getting the biggest "bug bash" in human history, probably what would have been ~40 years' worth of disclosures and CVEs in existing software, compressed in ~2.
These metrics illustrate the scale of the Pragmatic Engineer essay challenge and the unprecedented industry‑wide bug‑bash, underscoring my decision to ban AI‑written essays and require author attribution to keep the community signal high.
In late October 2026 I posted a LinkedIn‑style reflection on a “mini‑security crisis” driven by rapid AI adoption. I argued that AI is accelerating vulnerability discovery, making manual patching obsolete, and demanding new AI‑driven monitoring and staff practices.
CVE discovery sped up ~10-100x vs before.
Manual patching needs to die.
Security monitoring needs to improve massively.
Security professionals will need to use a LOT more AI.
Attackers are getting ahead because they are using AI.
In late October 2026 I posted a series of LinkedIn‑style reflections on engineering management evolution, highlighting the shift toward hands‑on leadership and sharing sponsor claims from Turbopuffer and Entire.
Will left a team of 250 at Carta to head up 25 tech people at Imprint.
Revolut experienced a >7‑hour global outage ("7+ hours and counting").
The coinciding Google Cloud single‑zone outage lasted 1.5‑hour (power loss).
The Telegraph reported that half a dozen banks were impacted.
I announced a pop‑up discussion on source‑control & agents for Thursday October 29, 9 am in San Francisco ("Next week. San Francisco. Morning. Coffee.").
These observations reinforce my view that financial services must build and test failover capabilities; otherwise they risk losing customers during cloud provider outages.
I cannot choose my model for Grok Bot. It’s Grok 4.{latest}
When I do stuff with AI I am an expert in (eg coding on my project I know well, or doing research in an area I am very familiar with) I spot problems/errors/the model BS'ing me immediately.
Linear is hosting a fireside chat in London on November 10 2026 to celebrate its new European hub.
A 10‑point intelligence advantage now yields just 20 % more developer traffic, down from nearly 69 % in late 2025 (The Information).
Autodesk saw how well it’s going for Adobe to bully its customers (who turn from the biggest promoters to the biggest and most vocal haters after what feels like abusive behaviour from the vendor) and decided to do the same.
Anthropic generates more revenue than those multi-decades-old companies + no signs of slowdown
The engineer complained that they are being pressured and framed as an "outdated, inefficient engineer who refuses to adapt to modern technology."
Amazon’s stock plunged dramatically during the dot‑com bubble, falling from $100 to $6 before a $600 million bond infusion helped it survive.
A small‑business customer pays $380 per user, per month to Autodesk and is questioning the vendor’s recent cloud‑based API direction.
I advised the engineer to adapt to the AI token‑usage KPI, because the manager’s push aligns with a sector‑wide transformation toward AI‑augmented development.
The pressure isn’t isolated – the entire tech ecosystem, from startups to big‑tech and academia, is simultaneously overhauling dev tools because of LLMs. There’s no consensus on best practices yet, so resistance is seen as a “poor look” for any professional.
I cannot recall any other time in the tech industry when the COMPLETE industry (from small to large companies; nonprofits to VC-funded; private sector to government and academia) is trying to figure out how to change the dev tools and processes we use. all AT THE SAME TIME!!
OK I want to give a shoutout to @Uptime_Labs - mentioned by @samnewman on the podcast. They do incident response training via simulations, so eng teams are better prepared when an outage happens.
A very very interesting and recent book by Dr Cat Hicks (@grimalkina): Pyschology of Software Teams She makes a good case on how most companies/teams don’t understand how devs operate and think of devs as “brains in jars.”
Cancelling a sponsor booth last minute is unheard of: and yes, it suggests AWS is shit scared of Cloudflare, as a COMPETITOR.
These observations from my October‑2026 X threads highlight unprecedented AI‑driven churn, the value of incident‑response simulations, cultural blind‑spots around developers, and competitive tensions at big‑tech events.
Spend range: $100/dev/month to ~$5,000; $20K+ outliers.
PRs per eng is up.
I keep being surprised by devs who tell themselves that anyone who is convinced that LLMs are permanently part of how we build software suffers "AI psychosis."
Uber developers do regular surveys, take feedback seriously, and developers are honest about it. This indicates true.
I was doing too many talks at conferences (representing the company - I did ~4 per year).
Got an AI-generated reply, very clearly templated. This is true.
AI bot account, blocked. This is true.
Crazy how Anthropic started to hide the full names of engineers on its engineering blogs.
Uber’s internal developer tooling is very different from what you would see at most places.
In 2023, Uber’s platform team made the decision that everything built with AI needs to be model- and harness-agnostic.
Model Gateway to support switching between Anthropic, OpenAI, other vendor models, and now also supports open models they run.
Uber tags PRs by the type of work each of those PRs do.
They are seeing the percentage of “feature PRs” going up, and the percentage of “maintenance PRs” going down.
In mid‑October 2026 I shared a series of LinkedIn‑style posts that cover four themes: Uber’s internal developer tooling culture, a cyber‑attack on Dutch grocery chain Albert Heijn, advice on AI transformation, and the near‑universal expectation of LLM usage in software engineering.
Uber treats cities as a first‑class primitive, which drives observability, feature‑flagging, deployments, and service sharding across the platform.
Model Gateway (2023) makes every AI component model‑ and harness‑agnostic, supporting Anthropic, OpenAI, other vendor models, and internal open models.
Uber replaced PR‑velocity metrics with PR tagging by work type (feature, bugfix, maintenance, testing, docs), observing an increase in feature‑PR share and a decrease in maintenance‑PR share.
- Sponsor initiatives
- turbopuffer: rebuilding storage architecture for faster, cheaper, more reliable search at scale (turbopuffer.com/v3).
- Linear: new “Linear Diffs” feature to review PRs inside Linear with full context (lnkd.in/ezrhppU7).
- WorkOS: Airlock authorization layer for AI agents, integrating with Claude Code, Codex, MCP gateways (workos.com/airlock).
Albert Heijn accounts for 38 % of national grocery spend and 28 % of all physical grocery stores in the Netherlands.
The cyber‑attack caused a complete shutdown of fresh produce (fruits, vegetables, milk) in stores and halted online ordering.
The incident was unrelated to the Danfoss‑cooler attack in France; Uber shut down warehousing systems as a responsible security measure.
Meta engineer Mark Tsimelzon is open to advisory roles for startups while available.
99.9 % of startups, Big Tech, and ambitious companies now expect AI tool usage.
I highlighted engineers who take opposite stances on LLMs: Casey Muratori builds software without LLMs for personal projects, while Peter Mattis discusses his current AI usage (screenshot credited to Geoffrey Huntley).
OpenAI confirmed a sophisticated recording device was hidden in an office, traced to a specific employee, and the matter was referred to the FBI. “We discovered a sophisticated recording device that was hidden inside one of our offices.” “We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest.”
The implicated engineer earned “$1M+ per year” and would have “thrown away” their career by planting the device. “Imagine working at OpenAI as an engineer, making $1M+ per year in compensation, and then deciding that you'd throw away your career just like that by planting hidden recording devices in the office.”
Anthropic reported merging ~2,500 code changes per day in May and >30,000 per day in September. “In May, we merged around 2,500 code changes per day.” “In September, we merged more than 30,000 per day.”
Anthropic built a new version‑control system and fully migrated its monorepo from git in five months. “Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: https://t.co/wzNkwHs7IN”
Character AI filed for Chapter 11 bankruptcy. “Character AI has gone bankrupt.”
These posts from October 2026 capture a volatile week in the AI landscape: a serious internal security breach at OpenAI, aggressive engineering velocity at Anthropic, and the collapse of Character AI. I reflected on the broader implications for talent mobility, open‑source strategy, and industry consolidation.
Airtable post‑acquisition headcount is estimated to be ~50 ppl or less.
The purchase price for Airtable is roughly 2.5 years of Airtable revenue.
Bending Spoons operates ~20 products (incl Evernote, Meetup, Vimeo, Eventbrite, WeTransfer etc.).
Total staff across Bending Spoons’ portfolio is smaller than 900.
Meta removed human support from Instagram, it’s AI support only for paying customers.
In the app, it sends you to Meta AI. Which doesn't work
On the web, you're shown "go to the Instagram app to get support (for Meta Verified)"
there’s no other way to get to that page than cancellation flow rofl https://t.co/yg2Ni3pgDU
So I pay for a product with no support, no way to report this, and no way to get a receipt event
I decided to cancel the Meta Verified subscription because there is no human support and the only way to view receipts is the forced cancellation flow, which I consider a broken product failure.
X AI was founded 3 years ago with 12 co‑founders, including Musk. Of the 12, 11 are out, Musk the only one remaining.
No cofounder lasted for even 4 years (the usual time to vest out!).
In order to create a compelling SI, you need a LOT of training data and need to SOLVE CODING.
Code Comprehension. The idea behind this round is to drop the candidates into a 200-500-line codebase with Gemini available for use.
Turns out the human support was an absolute lie.
I observed that Perplexity cannot stop hallucinating; 50% of the time its output is made up.
I noted that Gemini 3.1 Pro was supposed to be better than any other model and wasn’t, reflecting Google’s history of only shipping impressive benchmarks.
Bloomberg is reporting Gemini 4 is struggling on ‘key areas’ such as coding – according to employees!
One of my portcos tried to spend $100k on AWS but couldn’t because they lacked enough account history to get the quota approved.
I highlighted that MongoDB already paid $32M in compensation to its CEO.
These observations reinforced my decisions to block accounts that look like AI‑generated spam, to avoid adopting Gemini until it becomes generally available, and to consider alternative cloud providers such as Railway, Render, or Fly.io when AWS quota requests stall.
I stopped using Perplexity because it hallucinated about 50 % of the time and the paid subscription ($20 / month) no longer delivered reliable results, so I discontinued the service.
The hiring interview trend shifted to onsite_rounds as remote video interviews became vulnerable to deep‑fake attacks.
The startup’s public claim of 50 K creators is false; the account actually has ≈1 000 followers.
The “Space Mushrooms” outreach strategy is reported to message 10 K+ apps per month, but this remains speculative.
I labelled the viral video and its associated $20 M “Google” funding claim as false, and I flagged the accounts spreading the narrative as low‑credibility.
In early October 2026 I posted a series of X threads that combined a technical deep‑dive on Turbopuffer’s v3 full‑text engine with a broader critique of AI‑startup hype. I documented that v3 was initially up to 128× slower than the production‑ready v2, then improved roughly 9× after introducing batched, column‑wise reads, and I emphasized that v3 remains a shadow system not yet in production. At the same time I called out a wave of dubious AI‑startup announcements – a purported $20 M “Google” raise, a $31.5 M seed round co‑led by General Catalyst into a Kled copy‑cat, and recycled outrage videos – and I shared the community’s methods for vetting them (domain‑age checks, video reuse detection). I also noted the farewell tweet from a Monzo colleague after eight years. These posts illustrate my commitment to transparency in engineering while remaining skeptical of inflated AI‑centric narratives.
But in 9/10 cases unless said person is not a cofounder or founding eng, they won’t get the trust to do the risky + sometimes controversial stuff that is required to execute quickly and well to win vs competition.
OpenAI’s classifier caught only 26 % of AI‑written text.
mis‑flagged 9 % of human text as AI.
Companies like Databricks conduct very thorough reference checks; back‑channeling is becoming a standard, “the only thing that can’t be gamed by AI”.
Trial periods could get longer as interview signal gets weaker for on‑the‑job performance.
Potential interview redesign – suggestion to collapse SWE interviews into two rounds: system design + onsite project with agents.
If any AI is suspected, readers assume 100 % AI and discount the work; thus avoiding AI is a strong differentiator.
I observed that modern software engineering has become inseparable from LLM‑based AI tools. Developers now use AI for virtually every part of building software, and the market is demanding AI‑augmented skills.
I concluded that you won’t be hired at 99.9% of startups+Big Tech+ambitious companies, even today.
I noted that about 10‑20% of my class was building software on the side during university.
I observed that about 15% of students put up their hand when asked about recent side‑project experience.
From these signals I decided to adopt AI tools for paid development, keep self‑employment as an AI‑free niche, and push universities to add AI‑Assisted Engineering courses.
In early October 2026 I shared a thread on X that distilled a set of hard‑won observations about AI tooling, the evolving attention economy, regulatory gaps for utility‑like platforms, and crypto‑risk management, together with three concrete decisions for my work and the broader industry.
AI tooling paradox – despite making code generation, deep research, ideation and image creation easier, most developers end up doing more work because of the learning curve and heightened context‑switching. “AI was supposed to free us from drudgery. It isn't for most software developers because we're doing more work, we've got more context switching going on.”
Productivity metrics shift – volume‑based measures (PR count, LOC) lose relevance as teams grow past ~50 engineers; impact‑focused evaluation resurfaces as the dominant signal.
Attention economy evolution – the chain goes from distribution scarcity → audience‑building → attention scarcity → credibility scarcity; shortcuts to credibility simply do not exist.
Regulation as the only remedy for arbitrary platform bans – “I hate to say it but only regulation can fix this” and “The reason your utility provider cannot randomly “ban” you is because of regulation”.
Crypto risk management – holding crypto exposes you to math breakthroughs, key theft or kidnapping; the recommended mitigation is either “not hold crypto” or rotate to fresh, never‑used addresses (“bunker mode”).
Decision: invest in a “machine that builds the software” – building robust CI/CD and AI‑assist pipelines is critical for large engineering orgs.
Decision: do not rely on shortcuts for credibility – focus on consistent, high‑quality output rather than viral tactics.
Decision: regulatory advocacy – stakeholders agree that government regulation is required to curb random bans and enforce consumer protections for utility‑like platforms.
JetBrains raised its prices by ~20% mid‑year.
Despite the price hike, revenue grew only 6.3% (expected 10% without churn).
My daily Synology NAS backup saved the C‑drive after Opus 5.5 deleted it.
Entire claims its Git hosting is up to 89x faster than competitors.
I think the tech world in general is quite naive about AI and fundamentally misunderstands what an LLM is. Sam says: “I think the tech world in general is quite naive about AI and fundamentally misunderstands what an LLM is.”
LLMs are not world models. Sam says: “LLMs are not world models.”
World models have been around for a long time and predate LLMs. Sam says: “World models have been around for a long time. They predate LLMs.”
Nobody (inside startups/AI labs/Big Tech) writes code by hand anymore. Sam says: “Nobody (inside startups/AI labs/Big Tech) writes code by hand anymore”.
Working with several parallel agents is increasingly common. Sam says: “Working with several parallel agents is increasingly common”.
Agent skills usage is mainstream. Sam says: “Agent skills usage is mainstream”.
The fading IDE. Sam says: “The fading IDE”.
Everyone is building their own harness/agent platform. Sam says: “Everyone is building their own harness/agent platform”.
Dev work often starts in Slack. Sam says: “Dev work often starts in Slack”.
Turbopuffer claims to be fast, 10x cheaper, and extremely scalable. Sam says: “fast, 10x cheaper, and extremely scalable.”
Entire claims to be up to 89x faster than any other competitor. Sam says: “up to 89x faster than any other competitor.”
I decided to adopt agent‑centric toolchains, de‑emphasize heavyweight IDEs, and rethink LLM guardrails because human‑only coding is no longer cost‑effective, IDE ROI is vanishing, and LLMs lack causal understanding.
In early October 2026 I also shared a thread on X about how AI‑generated content is flooding the tech ecosystem, from pull‑requests to LinkedIn posts, and about a community‑run contest that deliberately rejected AI‑written essays.
- AI‑generated PRs were about 3× the volume of human‑written PRs per month by Dec 2023.
- Over 90 % of LinkedIn posts/comments are now AI‑generated.
- The non‑AI essay contest received ~130 submissions total, with ~100 on a single day; 7 were disqualified as AI slop.
- A community funding round returned 13 % of capital ($130.80 of $1,000).
- Using AI, 600 k JUnit 4 tests covering 15 M lines of code were migrated to JUnit 5.
In early October 2026 I posted a thread on X about the new native Markdown support in Google Drive and Docs, the shift of Anthropic’s Pro/Max accounts to cloud‑only execution, and a critique of rapid “YOLO” product releases in regulated environments.
We are making Markdown files work natively across Google Drive and Docs.
You can now preview .md files in Drive, open them in Docs, edit them, comment on them, and collaborate without converting the file into a Google Doc!
🚨Anthropic removes Cowork's local option for Pro/Max tomorrow.
New tasks run in the cloud. Existing local tasks stay on your computer.
YOLO works until the deployment target moves.
I also suggested blocking distracting sites via hosts‑file edits or taking half‑day offline breaks to boost deep‑work focus.
In early October 2026 I posted a thread on X about how Uber is consolidating its microservice landscape into larger “macroservices”, the shifting cost of tech‑debt, and the growing frustration with LLM‑generated noise on the platform.
I observed that the cost of introducing and paying off tech debt has changed very, very much and I don't even know most of us have the time and space to think about what it means.
Uber decided to consolidate microservices into macroservices because maintaining “thousands of microservices… can cause more trouble long‑term than it solves the short‑term.”
Business lines that succeeded paid the coordination debt, while ones that got shut down (e.g., Uber Air) or sold off (e.g., Jump) never had to pay it.
I strongly appreciate taking a visible hardline stance on mindless LLM responses; their pollution of the commons is detestable.
In early October 2026 I threaded a series of X posts about AI‑native companies, the role of pre‑AI infrastructure, and a skeptical take on AI consciousness. I highlighted how firms like Uber that already have robust infra platforms can slip AI tools in with minimal friction, contrasted that with startups that must first build DevOps capabilities, and called out the danger of treating LLMs as conscious entities.
Uber maintains roughly ~20 custom infra/dev tools systems that were built “pre‑AI” and enable rapid, reliable AI integration.
Google Docs has been missing a functional emoji picker for about 20 years before a papercut was finally fixed.
Kotlin’s rapid backend adoption is driven by the JVM ecosystem and a large pool of existing Java backend engineers.
I reject the framing of AI models as conscious beings, warning that granting moral standing to LLMs poses legal and ethical risks.
I pay $20/month for this thing, but probably not much longer.
I’d guess all got shut down but the platform made $100 ha.
Their "deep research" is not human usable.
It’s a 244-page PDF produced by Anthropic.
100% onsite rounds were already problematic thanks to increasingly persistent North Korean hackers attempting to pass with deepfakes.
I concluded that the $20 M funding claim is invalid, so I marked startup.fastlane.funding_claim_invalid = true.
I flagged the accounts spreading the Fastlane narrative as low‑credibility, setting accounts.fastlane_spread.low_credibility = true.
I recorded that Google/Firebase’s status page stayed unupdated for 4 days after the outage, thus status_page.google_firebase.unupdated_outage_days = 4.
I observed that turbopuffer’s new v3 architecture is up to 128× slower than the old v2 architecture before the optimisations they are doing.
We implemented batched reads in the v3 full‑text engine, leading to a ~9x performance improvement. Postings are stored as columns, and we were reading them a row at a time.
v3 (new) isn't in prod yet :)
v2 (prod) is still very fast, but we're grinding v3 to make it even faster.
Today is my final day at @monzo after 8 years and change 👋
General Catalyst just co-led a $31.5 million seed round into a blatant rip‑off of my company, Kled.
No such company raised $20M from Google.
In early October 2026 I posted a series of replies on X that mixed leadership commentary about MongoDB’s interim CEO, technical critiques of in‑memory map implementations, and moderation actions against bot accounts. The thread captures my view that leadership legacy matters more than compensation, and that cache‑friendly data structures outperform classic red‑black trees.
- Leadership turnover: CJ oversaw roughly ~5,500 MongoDB employees and left after 10 months for a higher‑pay role.
“At the end of the day, no one's going to remember how much money you make. They're going to remember what kind of leader you were and what kind of person you were.”
- Technical claim: “in‑memory ordered maps are a cache line problem, not a big‑O problem; a red‑black node costs a dependent load per level; a b‑tree node amortizes one miss across many keys, so same log n, far fewer stalls.”
- Structures compared: std::map (red‑black tree), B‑tree, and Swiss Table.
- Projects mentioned in the critique: GIMP, Gmail backend, Colossus, CockroachDB.
- Blocked bot URLs: https://t.co/QU3zCl3d27, https://t.co/ra8ApTzxS3, https://t.co/ccVShWA9OG, https://t.co/ftout6DGVD.