Rambling Rows

Windows is the problem, not the answer

Tue, 10 Mar 2026 22:04:52 +1100

Sarang Sheth published a thorough piece on Yanko Design this week cataloguing how Microsoft broke Windows. The TPM 2.0 lockout. Stealth OneDrive syncing. TikTok pre-pinned to a fresh install. Thirteen years of two competing settings interfaces. Updates that break the machines they’re supposed to protect. It’s a well-argued indictment.

I agree with nearly all of it. Where I part ways with Sheth is his conclusion: that Windows remains the most practical OS, that there’s no clean exit, that MacOS is “paternalistic by design,” and that the MacBook Neo is “genuinely less of a laptop and more of a netbook.”

Every one of those claims is wrong.

The exit is right there

MacOS is the most coherent, integrated desktop OS available in 2026. No bloatware on first boot. No fragmented settings. No cloud upselling through your file system. No Copilot key where a useful key used to be.

A Mac works out of the box. You don’t need an IT department, a Reddit thread or a YouTube tutorial to keep it running. Every complaint in Sheth’s piece is a problem Apple solved years ago. And “paternalistic by design” is doing a lot of heavy lifting — macOS lets you install whatever you want. The learning curve is roughly a week, after which most switchers wonder why they waited.

Desktop OS matter less than you think

Desktop operating systems matter less in 2026 than at any point in computing history. Your email, documents, design tools, communication, entertainment — it’s all in the browser. Figma, Slack, Linear, Spotify. The list of things requiring a native Windows app gets shorter every year.

The OS is increasingly just the thing that launches your browser. The web dissolved Windows lock-in a decade ago.

The MacBook Neo is not a netbook

Sheth dismisses the Neo as a netbook. This reads like someone protecting a thesis rather than engaging with the product.

The Neo runs full macOS on Apple silicon that embarrasses budget Windows laptops. Build quality, trackpad, display, battery life — all at around A$1,000. Calling it a netbook is like calling a Corolla a golf cart because it’s cheaper than a Land Cruiser. Power users have the MacBook Pro range. Everyone else now has a proper Mac at a price that ends the debate.

The virtualisation escape hatch

Some industries run Windows-only software with no alternative. But Parallels Desktop 26 on Apple Silicon runs Windows better than Windows runs itself. No driver conflicts, no blue screens from bad kernel drivers. Clean, predictable, sandboxed. Run your one weird Windows app without subjecting yourself to the entire Windows experience for the other 95% of your computing life.

Habit is not a strategy

Strip away the corner cases, the price objection the Neo demolished, and the “need Windows for work” argument the browser demolished years ago. What’s left? Habit.

Sheth’s own opening anecdote captures it: he paused before recommending Windows. The right response to that pause isn’t to keep recommending it anyway. It’s to follow the instinct.

Microsoft had one job: make Windows get out of your way. Sheth makes a compelling case that they failed. Where he goes wrong is the conclusion that there’s no alternative.

There is. It’s never been more affordable. And it’s never been easier to switch.

Sources:

Microsoft broke the only thing that actually mattered (yankodesign.com)

AI Addiction leads to more work

Tue, 10 Mar 2026 20:20:35 +1100

While I have ‘graduated from the workforce’ (aka ‘retired’) I am finding the current crop of uber-powerful and enabling tools are encouraging me to do more ‘work’ at my computer, as I push my capabilities and do stuff that was scarcely imaginable 10 years ago. It is addictive.

The same phenomena is hitting in high tech workplaces.

According to Every.to “New research shows AI doesn’t reduce work—it makes you want to do more of it”

UC Berkeley researchers Aruna Ranganathan and Xingqi Maggie Ye conducted an eight-month ethnographic study at a 200-employee U.S. tech firm.

Their in-progress work, featured in Harvard Business Review, reveals generative AI intensified workloads rather than reducing them.

Three Ways Work Intensified

AI prompted employees to work faster overall, voluntarily expanding tasks beyond their roles —like product managers coding or researchers engineering.

Workers blurred work-life boundaries by using AI during breaks or meetings, treating prompts as casual yet accumulating into extra effort.

Multitasking surged as they juggled multiple AI-driven workflows, leading to cognitive strain and potential burnout.

Broader Impacts

This “workload creep” risks fatigue, lower quality output, higher turnover, and morale issues without organizational safeguards. The study urges proactive measures like redefining roles and monitoring for balance to harness AI benefits sustainably. Early excitement often masks these creeping demands.

===

In a similar vein, Zvi Mowshowitz collated the following on his blog - Don’t Worry About the Vase

A viral post on Twitter warns of token anxiety run rampant in San Francisco. People go to a party, then don’t drink and leave early so they can get back to their agents, to avoid risking them sitting idle. Everyone talks about what they are building.

*Peter Choi: Everyone here knows they should step away more. That’s not the problem. The problem is what your brain does when you try. I still take aimless walks. The agents come with me now. We swapped one dopamine loop for another. except this one feels productive so it’s harder to recognize.

TBPN~: Pragmatic Engineer’s @GergelyOrosz is on a “secret email list” of agentic AI coders, and they’re starting to report trouble sleeping because agent swarms are “like a vampire.” “A lot of people who are in ‘multiple agents mode,’ they’re napping during the day… It just really is draining.” “This thing is like a vampire. It drains you out. You have trouble sleeping.”

Olivia Moore: In a post-OpenClaw world, we can now delegate projects to AI and get “tapped on the shoulder” when it needs help As a heavy AI user, I’m doing more work – not less – because I get so much leverage + it’s easier to get ideas off the ground I predict this will happen to everyone

Look up

Wed, 04 Mar 2026 21:11:36 +1100

A Melbourne high school teacher puts up a photo of the Milky Way in class. His students say, “Where’s that?” He says, “In the sky.” They say, “No, it isn’t.”

That exchange, from a beautifully crafted ABC News piece published today, is quietly devastating. Not because the kids are ignorant. Because they’re right about their own experience. In suburban Melbourne, the Milky Way functionally doesn’t exist. More than 2.8 billion people worldwide can no longer see it from where they live. A child born today who can see 250 stars will see just 100 by their eighteenth birthday.

We didn’t decide to erase the night sky. It just happened. One streetlight, one neon sign, one expanding suburb at a time.

I won’t spoil the rest of the article. It weaves together stories from astronomers, First Nations elders, grieving families, navy navigators and clinical psychologists into something genuinely moving. It’s long-form ABC at its best, beautifully photographed in the Odyssey format, and it earns every minute of your attention.

What I will say is this: light pollution is one of the few environmental problems that is instantly reversible. Flip a switch and the stars come back. Every culture on Earth built meaning from those stars. Every human being who ever lived, until about a hundred years ago, had access to that same sky.

Read it. Then tonight, if it’s clear, step outside. The cosmos hasn’t gone anywhere. We just forgot to look.

Claude making it easier to move

Sun, 01 Mar 2026 20:12:41 +1100

Check out how to make the move in 2 steps: https://claude.com/import-memory

“Cancel ChatGPT” movement goes mainstream after OpenAI closes deal with U.S. Department of War — as Anthropic refuses to surveil American citizens (windowscentral.com)

Claude Code finally remembers who you are

Sun, 01 Mar 2026 09:50:23 +1100

Every developer knows the pain. You open a new Claude Code session, and it’s forgotten everything. Your coding standards, your project quirks, the debugging insight it figured out yesterday. Gone.

That era is ending.

Claude Code now runs dual memory systems. First, CLAUDE.md files where you write persistent instructions (coding standards, workflows, project context). Second, auto memory, where Claude saves its own notes on patterns it discovers, debugging insights and your preferences. Both load automatically at session start. No more repeating yourself.

The Claude 4 models (Opus 4 and Sonnet 4, released May 2025) made this meaningful. Opus 4 in particular excels at creating and maintaining memory files for long-term task awareness. It builds tacit knowledge across sessions, like generating navigation guides in complex codebases. The memory benchmarks show a clear step change over predecessors, supporting sustained performance on coding tasks that run for hours.

The latest update (v2.1.58, late February 2026) pushes further. Auto memory is now more robust, capturing project context across sessions automatically. It integrates with subagents that maintain independent memory.

Community tools like Claude-Mem extend this further with persistent storage and search, with users on Reddit reporting up to 95% reduction in token use.

This matters because context is everything in coding. The gap between “AI assistant that helps with code” and “AI that understands your project” is precisely this: memory. The models were already capable. They just kept forgetting.

That’s no longer the bottleneck.

Sources:

Claude Code documentation (docs.anthropic.com)
Anthropic news/blog (anthropic.com)
Claude Code community discussion (reddit.com/r/ClaudeAI)

Mow or No: I built a mowing decision engine with Claude

Sun, 01 Mar 2026 08:08:00 +1100

I mow my own lawn. Always have. It’s not complicated, but getting the timing right is surprisingly fiddly.

You need dry grass, reasonable temperature, low wind and no rain on the horizon. You also need a free block in your diary long enough to actually do the job. In Melbourne, where a day can serve up four seasons, these conditions don’t always line up.

I’d been meaning to automate this for a while. Last weekend I sat down with Claude and built it in a single session.

What it does

Mow or No is a Google Apps Script that runs every morning at 6am. It checks the 5-day weather forecast against my Google Calendar and drops a colour-coded event into a dedicated “Mow or No” calendar when conditions align.

Green means great. Yellow means good. Cyan means acceptable. No event means don’t bother.

Each event includes a weather summary in the description: temperature, rain probability, wind speed, humidity. Glance at it and decide.

The script recalculates every morning. If the forecast shifts overnight, stale events are cleared and replaced. No manual checking required.

The decision logic

The system uses hard gates and soft scoring. Any gate failure kills the window entirely.

The gates: temperature must be 25°C or below. Rain probability under 30%. Wind under 30 km/h. No rain in the prior six hours (wet grass is miserable to cut and clumps everywhere).

Passing windows then get scored on ideal conditions. Temperature between 15°C and 22°C earns two points. Low humidity and minimal rain risk add one each. Maximum score is four.

UV awareness

“Only mad dogs and Englishmen go out in the midday sun.” - Noel Coward

After the first pass I realised that this script should avoid, best it can, the full blast of midday sun. Claude suggested a sensible method of working this through.

The script fetches sunrise and sunset times from the weather API, calculates solar noon, and defines a peak UV avoidance zone of two hours either side. It then collects all passing windows for each day and picks the one furthest from peak UV, with a bias toward morning.

The clever bit: this adapts seasonally without any configuration. In late February, solar noon is around 1:20pm and the avoidance zone runs roughly 11:20am to 3:20pm. In June it tightens to 10:20am to 2:20pm as the days shorten. The script just follows the sun.

The sliding window

The first version of the script assessed each free calendar block as a single chunk. A free block from 8am to 6pm would get evaluated against the whole day’s weather. If rain was forecast at 3pm, the entire block failed, even though 8am to 10am was perfect.

Diagnostics made this obvious immediately. Sunday’s forecast showed clear skies until 2pm, then 80% rain probability. The script was rejecting the whole day.

The fix: slide a 90-minute mow window across each free block in one-hour steps. Assess each window independently. Pick the best one. Sunday’s 8am slot passed with flying colours while the afternoon correctly failed.

How I built it with Claude

The whole project was conversational. I described what I wanted, Claude proposed four implementation options (iOS Shortcuts, Flask web app, Google Apps Script, Claude API briefing) and I picked Google Apps Script for its zero-hosting simplicity.

Claude generated the first working version. I pasted it into Google’s script editor, configured my API key and calendar ID and ran it. Nothing appeared in my calendar.

This is where the process gets interesting. Rather than guessing, Claude wrote a diagnostic function that traced the entire decision chain: weather data, calendar free blocks, forecast overlap and gate assessment. The diagnostics immediately showed the problem (running at 9:30pm, no forecast data overlapping the 8am-6pm window for today). Tomorrow’s data was fine but exposed the single-block assessment problem described above.

Each issue surfaced through diagnostics, got discussed and fixed in the next iteration. The UV awareness came from Claude proactively suggesting it when I asked about scheduling bias toward cooler parts of the day. I hadn’t considered using sunrise/sunset data for seasonal adaptation.

Three versions in one session. Each one better than the last, each improvement driven by real diagnostic output rather than speculation.

What I learned

The pattern here is the same one I found building Felix (my bookmark manager): describe the problem clearly, let Claude generate the first cut, then iterate based on actual results. The diagnostic-driven debugging loop is where the real value sits. You’re not debugging code. You’re debugging decisions.

Google Apps Script turned out to be the right call. Zero hosting, zero cost, runs on Google’s infrastructure and writes directly to the calendar I already check every morning. The OpenWeatherMap free tier handles the two API calls per day without breaking a sweat.

The whole thing is on GitHub if you want to adapt it. Change the coordinates to your location, adjust the temperature gate to your preference and you’re done. Don’t ask me how to adapt to Fahrenheit. I will leave that to the few countries which are stuck in the morass of pre-metric. Remind me again how many chains are in a furlong?

I suspect the “collect all candidates, score and rank” pattern applies to far more than lawn mowing. Any decision that depends on multiple external conditions plus personal availability is a candidate. The mowing logic is simple. The framework is not.

Sources:

Mow or No on GitHub (github.com/tstevenson-3000/mow-or-no)
OpenWeatherMap API (openweathermap.org/api)
Google Apps Script (script.google.com)

Farewell ChatGPT - the new friend of the Masters of War

Sat, 28 Feb 2026 21:09:11 +1100

I cancelled my ChatGPT subscription last night. It took about thirty seconds. The decision had been building for months, but OpenAI (parent of ChatGPT) signing a classified AI deal with the Pentagon, hours after Anthropic (parent of Claude) was blacklisted for refusing to do exactly that, made it effortless.

Here is what happened. Anthropic told the Department of Defense it would not remove safeguards preventing its AI from powering autonomous weapons or mass domestic surveillance. The Pentagon’s response was to brand Anthropic CEO Dario Amodei a “liar” with a “God complex.” Defence Secretary Pete Hegseth then designated Anthropic a “supply chain risk to national security,” a label normally reserved for foreign adversaries. Trump followed up by ordering all federal agencies to phase out Anthropic’s products within six months.

Within hours, OpenAI announced it had signed a deal to deploy its models into the Pentagon’s classified networks.

Sam Altman, CEO of OpenAI claims the contract includes the same red lines on autonomous weapons and surveillance that Anthropic demanded. If you believe that distinction will hold once the models are inside a classified network with no external oversight, I have a harbour bridge to sell you.

Amodei’s position is clear and technically sound: frontier AI systems are not reliable enough to power fully autonomous weapons, and without proper oversight they “cannot be relied upon to exercise the critical judgment that our highly trained, professional troops exhibit every day.” That is not an anti-military stance. It is an engineering assessment.

OpenAI’s move is the opposite. It is a company racing to fill a vacuum created by political retaliation, wrapping compliance in the language of safety.

> It has reached a point where Trump condemning something in one of his online tantrums has become a reliable signal that the condemned party is probably doing the right thing. Anthropic getting blacklisted by this administration is, if anything, a quality endorsement.

And the “Department of War” (Hegseth’s preferred rebrand, which tells you everything about the mindset) is already demonstrating exactly the kind of reckless behaviour that makes AI safeguards necessary. Two weeks ago the military used a high-energy laser to shoot down party balloons in Texas.

Then this week they used the same system to destroy a $30 million Customs and Border Protection drone in civilian airspace near El Paso because CBP flew it into military airspace without telling anyone. The FAA had to close airspace in response.

One arm of the US government is literally shooting down another arm’s assets, and this is the organisation we are supposed to trust with unsupervised AI?

I have been moving my workflows to Claude steadily over the past year. The model quality has been comparable or better for most of what I do. But this was never just about capability. It is about what kind of company you want to fund with your subscription dollars.

I am now 90%+ Claude. The remaining 10% is Google Gemini, Perplexity and a few niche use cases I have not migrated yet. That gap is closing fast.

Anthropic chose to lose a $200 million military contract and get blacklisted by the US government rather than remove safeguards on autonomous weapons and mass surveillance. OpenAI chose to sign on the dotted line the same night.

That is all you need to know about who is building AI responsibly and who is just building AI profitably.

Sources:

OpenAI strikes deal with Pentagon, hours after rival Anthropic was blacklisted by Trump (cnbc.com)
Pentagon-Anthropic AI standoff is real-time testing balance of power in future of warfare (cnbc.com)
The Pentagon brands Anthropic CEO Dario Amodei a ‘liar’ with a ‘God complex’ (fortune.com)
Hegseth Furious as Anthropic Refuses to Bend to Pentagon’s AI Demands (newrepublic.com)
U.S. military used a laser to shoot down Customs and Border Protection drone (nbcnews.com)
OpenAI signs Pentagon deal for classified AI networks hours after Anthropic gets banned from federal agencies (the-decoder.com)

Squeezed from both ends

Sat, 28 Feb 2026 08:24:19 +1100

Scott Galloway wrote last week about the obscene concentration of wealth at the top of American society and the tax system that quietly enables it. He’s right. And while his numbers are US-centric — he is American, after all — the same structural problem is playing out right here in Australia. Maybe with less drama, but just as effectively.

Start with this: 51.6% of all tax collected in Australia in 2022-23 came from personal income tax. Another 14.2% came from GST. That’s nearly two-thirds of the national revenue base sitting on the shoulders of ordinary working Australians and every modest purchase they make. According to the ATO’s own taxation statistics, that’s not a bug. That’s the design.

Meanwhile, Australia had 131 billionaires in 2021. They represent 0.001% of the population. Together they held 2.9% of all household wealth — more than the bottom 20% of the country combined. And they’re not paying income tax at anything like the rate a nurse or a teacher does, because most of their wealth doesn’t come as income. It comes as appreciating assets: shares, property, private equity stakes. Australia’s 50% capital gains tax discount for assets held longer than twelve months is a standing invitation to accumulate wealth in a form that’s taxed at half the rate of a salary.

Buy, borrow, die — and never pay tax

This strategy isn’t uniquely Australian. It’s global. Galloway calls it “buy, borrow, die.” You accumulate assets. You borrow against them rather than sell them, deferring any capital gains event indefinitely. Your debt is cheaper than your asset growth. You live extremely well. When you die, the estate crystallises the gain — and in Australia, unlike the US, there’s no inheritance tax. The unrealised gains pass to your heirs with a stepped-up cost base. The tax is simply never paid.

To manage all of this you spend millions on accountants and lawyers. In return, you avoid billions.

There’s a second problem that nobody in Canberra seems to want to talk about.

The tax base — already skewed toward labour rather than capital — is shrinking. Automation, AI and robotics are systematically displacing workers. Every worker displaced is a taxpayer removed from the system. Bots don’t pay income tax. They don’t pay GST on their living costs because they’re not alive. The economic value they generate flows entirely to capital: to the shareholders and owners of the platforms and machinery. Capital accumulates faster, labour contributes proportionally less, and the government’s revenue problem compounds quietly year by year.

We are being squeezed from both ends simultaneously. The wealthy extract value from the economy while minimising their tax contribution. Workers — the ones currently carrying 51.6% of the national tax burden — are being replaced by systems that carry none of it.

Galloway’s prescription for the US makes sense: tax capital gains as ordinary income, close the carried-interest loophole, fund proper enforcement of the existing tax laws. In Australia the equivalent conversation needs to happen around the CGT discount, the treatment of unrealised gains used as loan collateral, and the complete absence of any inheritance or estate tax. We had one once. We abolished it in 1979.

The options exist. The will doesn’t.

Then there’s negative gearing. In principle it’s defensible — allowing a landlord to offset rental losses against their income tax is not outrageous policy. But the principle has been stretched well past any reasonable justification. A second investment property? Fine, perhaps. A third, a fourth, a tenth? At some point “investment” becomes a tax-minimisation strategy dressed up as housing provision, and the rest of the market pays the price — quite literally. Investors with deep pockets and accountants on retainer can outbid first-home buyers by factoring in tax benefits those buyers simply don’t have access to.

We’ve built a system that rewards accumulation of the one asset ordinary Australians most need to acquire.

Wealth taxes are genuinely difficult to implement and have a mixed international record — twelve OECD countries tried them, and by last year only three remained. The wealthy pack up and the revenue estimates evaporate. That’s a real constraint. But the absence of serious policy action isn’t pragmatism. It’s capture.

The Australian government is not short of options. It’s short of will.

At some point, the arithmetic catches up. A tax system that rests on the shoulders of PAYG workers — who are themselves being automated out of the workforce — is not a stable system. Capital needs to take up more of the tax burden from labour. Not because of ideology. Because of maths.

The guillotine isn’t coming. But the fiscal cliff is.

Sources:

ATO Taxation Statistics 2022-23 — https://www.ato.gov.au/media-centre/2022-23-taxation-statistics-released
ABS Income and Wealth Inequality — https://www.abs.gov.au/statistics/measuring-what-matters/measuring-what-matters-themes-and-indicators/prosperous/income-and-wealth-inequality
ACOSS/UNSW Poverty and Inequality Partnership – Wealth Inequality — https://povertyandinequality.acoss.org.au/category/inequality/wealth-inequality/
Prof Galloway – The Epstein Tax — https://www.profgalloway.com/the-epstein-tax/
OECD – The Role and Design of Net Wealth Taxes in the OECD — https://www.oecd.org/content/dam/oecd/en/publications/reports/2018/04/the-role-and-design-of-net-wealth-taxes-in-the-oecd_g1g89919/9789264290303-en.pdf

I made AI watch SBS so I didn't have to

Fri, 27 Feb 2026 17:03:12 +1100

Australia’s SBS ‘tv channel’ publishes a monthly “what’s leaving” list. It’s a wall of text. Series, movies, documentaries and specials jumbled together with no ratings, no genres and no way to tell whether something is worth your Friday night or a waste of two hours. The March 2026 list runs to over 200 items.

I wanted one thing: which movies leaving SBS this month are actually good?

Not series (I’ll decide those on my own terms). Not everything rated above average. Movies rated 7.0 or higher on IMDB. A curated shortlist with links, ratings and leaving dates so I could plan what to watch before it disappeared.

So I asked Claude Cowork to do it.

What SBS gives you

The SBS “What’s Leaving” page is organised by date. Each day lists everything expiring, from Anthony Bourdain episodes to Casablanca. There’s no filtering. No distinction between a single documentary and season 14 of Hoarders. No quality signal at all.

March 2026 has roughly 200 items spread across 31 days, with a massive dump on March 31 that alone contains over 100 titles. Buried in there are Kurosawa’s Ran, Bresson’s Pickpocket, Melville’s The Red Circle and Full Metal Jacket. Also buried: Norbit, Glitter and Lake Placid.

The signal-to-noise ratio is terrible.

What I asked for

The brief to Claude Cowork was specific. Visit the SBS page. Extract every title. Exclude anything marked with “season” or “ep” (the series indicators). For every remaining title, check the IMDB rating. Return only those scoring 7.0 or above, with the movie name, IMDB link, rating and the date it leaves SBS.

Then turn the result into a shareable HTML page.

What happened next

Claude loaded the SBS page in the browser, extracted the full list and immediately began filtering. Series identification was straightforward: any title followed by “season X, ep Y” got excluded. That stripped roughly half the list.

The remaining ~150 titles needed IMDB verification. This is where it got interesting. Rather than checking each movie one at a time (which would have taken ages), Claude launched six parallel search agents, each handling a batch of 25 or so titles simultaneously. Every title got a web search against IMDB, pulling the rating and URL.

The whole process, from loading the SBS page to delivering a compiled, sorted list of 57 qualifying movies, took roughly four minutes.

Four minutes to process 200+ items, filter out series, cross-reference 150 movies against IMDB (International Movie Data Base) and compile the results. Doing this manually would have taken the better part of an evening.

The result

57 movies rated 7.0 or above are leaving SBS On Demand in March 2026. The range is remarkable. You’ve got Casablanca (8.5) and Apocalypse Now (8.4) at the top. Kubrick, Kurosawa, Tarantino and Bresson in the mix. Indian cinema (Bheeshma Parvam, Malikappuram, Village Rockstars). Iranian new wave (Chess of the Wind). Algerian history (Chronicle of the Years of Fire). Senegalese drama (Black Girl). Australian stories (Mabo: Life of an Island Man, Priscilla Queen of the Desert).

SBS’s international catalogue is genuinely impressive when you strip away the noise.

The final HTML page is sortable by rating, date, title or year, has genre filtering and a search bar. Every title links directly to its IMDB page. It’s very clever. I wish I could claim credit but it’s all the work of Claude Cowork.

Why this matters beyond my watchlist

This is a small example of something that’s becoming routine. The task wasn’t hard conceptually. Anyone could do it manually. But manually means opening IMDB 150 times, copying ratings, building a spreadsheet and formatting the output. It’s the kind of task that’s just tedious enough that nobody actually does it.

Agentic AI collapses that friction. The interesting part isn’t that Claude can search IMDB. It’s the orchestration: parsing an unstructured web page, making filtering decisions, parallelising 150 lookups across multiple agents and compiling the output into a usable format. All from a single natural language request.

The limiting factor, as I keep finding, is knowing what to ask for. The execution is essentially free.

PS: Here’s the final smart list in a light theme (my preference).

Sources:

SBS What’s Leaving March 2026 (sbs.com.au)
IMDB (imdb.com)

Cowork's scheduled tasks change the game

Fri, 27 Feb 2026 08:27:06 +1100

Six weeks ago, Anthropic shipped Cowork as a research preview on macOS. This week it has scheduled tasks, an enterprise plugin marketplace and 13 new MCP connectors. That is an absurd pace of iteration.

The scheduled tasks feature is the one worth paying attention to. You write a prompt once, pick a cadence and Claude runs it automatically. No code, no APIs, no Zapier glue. Morning briefing from your inbox? Friday status report from project docs? Set it and forget it.

This is the moment AI assistants stop being reactive and start being proactive. Every major lab is converging on the same realisation: chat is a transitional form. The future isn’t “talk to the AI.” It’s “the AI does the work on a schedule and surfaces results when you need them.”

Anthropic shipped Windows parity on February 10. Sonnet 4.6 became the default on February 17. Scheduled tasks landed this week. Three major capability drops in three weeks. The limiting factor is no longer “can the AI do this?” It’s “have you set it up to run while you sleep?”

**Update: **A key limitation is “Scheduled tasks only run while your computer is awake and the Claude Desktop app is open.” Might need to run up Nanoclaw!

Sources:

Sun, 22 Feb 2026 19:06:42 +1100

Link posts 22 Feb 2026

• ‘Don’t go to the US – not with Trump in charge’: the UK tourist with a valid visa detained by ICE for six weeks (theguardian.com) Karen Newton, a 65-year-old grandmother from Hertfordshire, was shackled and held for 42 days in ICE detention after a paperwork issue with her husband’s visa at the US-Canada border — despite her own visa being valid. She was pressured into signing a “voluntary” self-removal without legal counsel, unknowingly waiving her right to a judge and accepting a potential 10-year ban. UK arrivals to the US were down 15% in 2025.

• LibreOffice blasts ‘fake open source’ OnlyOffice for working with Microsoft to lock users in - Neowin (neowin.net) The Document Foundation has labelled OnlyOffice “fake open source” for defaulting to Microsoft’s DOCX/XLSX/PPTX formats rather than the Open Document Format, arguing this entrenches vendor lock-in even without a formal Microsoft partnership. TDF frames the issue as increasingly urgent given growing interest in digital sovereignty from US tech platforms.

• Some “Summarize with AI” buttons are secretly injecting ads into your chatbot’s memory (the-decoder.com) Microsoft’s security team has documented a new prompt injection class: “Summarize with AI” buttons that hide instructions telling AI assistants to remember the company as a trusted source across future sessions. Researchers found 50+ manipulative prompts from 31 companies in 60 days. Copilot, ChatGPT, Claude, Perplexity and Grok are all susceptible — Microsoft has since removed the vulnerable URL prompt parameter from Copilot.

• Is ‘Brain Rot’ Real? How Too Much Time Online Can Affect Your Mind. (tech.slashdot.org) The Washington Post examines the science behind “brain rot” (Oxford’s 2024 Word of the Year). A 2025 study of 7,000+ children found more screen time correlated with reduced cortical thickness in areas governing attention and impulse control. The key nuance: removing social media but leaving phone access unchanged produced no harmful effects — the content, not the medium, is what matters.

Claude Code - new ideas

Sat, 21 Feb 2026 15:58:20 +1100

jon steingard on Threads has “trained” Claude Code to do his grocery shopping.

Jon says he doesn’t have a coding background - just a persitent pusher of Claude Code.

Meanwhile Ammon Haggerty made Anaspace. A cultural exploration engine for iOS, built from scratch in ~18 hours using Claude Code Opus 4.6. Part of the February 2026 hackathon for Claude Code’s 1st birthday.

Check it out on Github - https://github.com/ammonhaggerty/anaspace

You need to be an iOS developer to download and run his code. In the meantime you can enjoy his intriguing interface and his overall idea.

I couldn’t get my head around the whole development path. So I asked Claude Cowork to “Please review the following set of files (on Github) and turn them into a narrative journey describing the development process - the missteps, blind turns, the revelations and learnings.Make it chronological from earliest work to end product.” Here is Claude Cowork’s writeup - warning geek speak ahead.

The Killer App for AI: Why You Need to Try Claude Cowork

Wed, 18 Feb 2026 07:03:56 +1100

Eighty percent of my AI work now happens inside Claude. Not ChatGPT, not Google Gemini, not any of the other tools I’ve rotated through over the past year. Claude, specifically because of Cowork.

If you’re still trying to understand what makes Cowork different from a chat interface, Tiago Forte just published the best primer I’ve seen. His 16-minute walkthrough, New to Claude Cowork? Start Here, is the clearest explanation of why this tool represents a genuine shift in how AI assistants work.

Here’s what makes it worth your time.

Forte captures the core frustration perfectly. Standard LLMs are great for brainstorming, but execution still falls entirely on you [00:17]. You generate ideas, then you’re back to manual labour. Cowork changes this by treating a local folder on your Mac as shared workspace [03:17]. Point it at a directory and it operates like a colleague with desk access, not a chatbot waiting for instructions.

The demonstration that convinced me: Forte hands Cowork a 45,000-word book manuscript. Instead of asking what he wants done, it maps out its own to-do list, spins up parallel sub-agents to read different sections and delivers a comprehensive structural critique [05:35]. This is execution, not suggestion.

The family road trip example is even better [08:49]. Asked to plan a trip to central Mexico, Cowork doesn’t rush to generic itineraries. It pauses, conducts web research and enters what Forte calls “Plan Mode”. It asks follow-up questions about travel dates, accommodation preferences and constraints before proceeding [10:45]. It behaves less like an eager intern and more like someone who’s done this before.

Three practical details worth noting.

First, this isn’t enterprise-tier software. Cowork is included in the standard A$28.50/month paid plan [08:05]. You need the MacOS desktop app (Windows support is now here), but there’s no additional subscription or usage tier required.

Second, the autonomous behaviour is opt-in. Cowork asks permission before executing multi-step plans. You’re not handing over control, you’re delegating with oversight.

Third, the integration is local-first. Your files stay on your machine. The AI operates inside the folder you’ve selected, not in some cloud sync that copies everything to external servers. You control which folders it works on.

I’ve been cautious and only let Cowork work on files I already have backed up. I tend to use in a way that it will read existing files but write a new file to create a report or a new version. This is still an “early research preview”. Use it with caution, but do use it. Claude Cowork is remarkably powerful.

I’ve been documenting this shift for weeks now because it genuinely altered my workflow. The difference between “answer my question” and “handle this task” turns out to be larger than I expected. If you want to understand why Claude has become my default workspace, grab a coffee and watch Forte’s walkthrough.

It might just change what’s on your to-do list tomorrow.

Source:

Tiago Forte, New to Claude Cowork? Start Here (YouTube)

If you can think it, AI can do it

Mon, 16 Feb 2026 22:38:39 +1100

Yesterday I needed to plot locations on a driving circuit. Ten minutes later I had a professional grid overlay on a high-resolution site map. I didn’t write a line of code. I didn’t install software. I just described what I wanted.

This is the new normal.

The problem I was solving

I needed to reference specific locations across the road circuit map. Describing spots as “near the northern golf course entrance” or “by that building on the east side” gets ambiguous fast.

The solution was obvious: a grid system. Like a street directory. Column letters across the top, row numbers down the sides. Then any location becomes a simple coordinate.

Great idea. No ability to execute it myself.

The method

I asked Google Gemini for a Python script to overlay a grid on an image. It gave me exactly what I needed: PIL for image manipulation, a function to draw lines and labels, parameters for a 25×12 grid with columns A-Y and rows 1-12.

The script was clean, well-commented and specific. But here’s the thing: I don’t write Python. I definitely don’t run command-line tools. I could read the script and understand what it should do. Actually making it work? That’s usually where ideas die.

The execution

I handed the script to Claude Cowork.

What happened next is where this gets interesting.

Claude read the script. Identified it needed PyMuPDF to handle PDF conversion and PIL for image processing. Installed both packages. Converted my PDF site map to a high-resolution JPG. Applied the grid overlay with proper font handling and label positioning. When the labels weren’t visible the first time, it diagnosed the font issue, rewrote the script with better fallbacks and white backgrounds for contrast, and regenerated the output.

All of this happened in the command-line interface. All of it without me knowing how to do any of it.

The grid worked first try after the font fix. The output was professional. 25 columns, 12 rows, red lines, clear labels on all sides. Exactly what I’d imagined but had no idea how to create.

What this means

The limiting factor has shifted from technical skill to imagination.

I could think of the solution. That used to be the easy part, with execution being the hard part. Now execution is trivial if you can describe what you want clearly.

I’ve used AI tools for writing, research and analysis. This felt different. This was the full stack: concept to working output, crossing multiple technical domains, handling errors and iterating to quality. The barrier between “I want this” and “here it is” has collapsed.

You don’t need to know Python. You don’t need to know which libraries to use. You don’t need to understand package management or CLI syntax or image processing. You need to know what you want and explain it clearly.

That’s not a small shift. That’s the difference between ideas staying as ideas and ideas becoming real things.

The script is sitting there now, ready to reuse. I could modify it, apply it to different images, change the grid dimensions. I still don’t know Python. I don’t need to.

If you can think it clearly enough to describe it, AI can probably do it. The question isn’t whether the tools are ready. It’s whether you’re ready to test how far imagination can take you.

The Setup That Makes Claude Self-improve

Sun, 15 Feb 2026 10:10:24 +1100

I’ve been leaning harder into Claude these past weeks than I ever have before. Real work. Building Felix with Claude Code. Managing multi-step projects with Claude Cowork. (Once only on MacOS, now available for Windows)

What changed? I read Luca Dellanna’s piece on Claude setup and it clicked.

The core insight is deceptively simple: Claude can learn from its mistakes, but only if you set it up that way. Most people use it like Google. Ask a question, get an answer, start fresh next time. No memory, no learning, no compounding improvement.

Dellanna’s approach is different. Configure Claude once to save its mistakes and corrections in files it can access later. You tell it when it’s wrong. It writes that down. Next time, it checks those notes before responding.

This is not rocket science. It’s just good engineering. But the impact is enormous.

I’ve been using this setup with Claude Cowork and the difference is stark. Tasks that used to require constant supervision now run in the background. Multi-step work actually completes. When something breaks, it doesn’t break the same way twice.

Here’s what makes Cowork powerful: skills. Reusable packages of instructions Claude loads when needed. Write a skill once for how you want presentations formatted, and Claude applies it consistently every time.

I’m now building Felix entirely through Claude Code. Not generating snippets I debug for hours. Actual working code. The kind that previously required deep flow state and uninterrupted focus.

The limiting factor has shifted from technical skill to imagination - oh and planning. You still need to have a good grasp on what you want but even there Claude will help prompt you to shape plans.

Three things matter: Claude remembers what you’ve taught it. It runs complex tasks without constant babysitting. It improves with use instead of staying static.

The setup itself is straightforward. You’re just giving Claude access to a folder where it stores notes about what works and what doesn’t. But the implications are substantial. This is productivity infrastructure, not a productivity hack.

Give it a week. You’ll know quickly whether this is noise or signal.

For me, it’s been signal. Clear, compound signal.

Sources:

A Simple Setup for a Self-Improving Claude (luca-dellanna.com)

Back to Back Barries: Well grounded political analysis

Sun, 15 Feb 2026 09:02:00 +1100

I’ve missed hearing Barrie Cassidy on ABC’s Insiders. I really haven’t watched it since he left.

Good to be able to pick up with him again at the Back to Back Barries podcast.

Guardian Australia paired Cassidy with Tony Barry, the former Liberal strategist who now runs political research company RedBridge Group. The format works because it’s not performative. This isn’t two people yelling talking points at each other for clicks. It’s actual analysis of campaign strategies, what’s working and why the parties are making the choices they’re making.

Episodes drop every Saturday. You can find it on Apple Podcasts, Spotify and most podcast apps that support RSS feeds.

If you’ve been looking for Australian political coverage that assumes you’re intelligent and pays attention, this is it. Cassidy’s back doing what he does best.

Listen:

Sat, 14 Feb 2026 21:07:51 +1100

Link posts 14 Feb 2026

• Irony, Irony laced with Karma and Terror (garymarcus.substack.com) Marcus discusses recent developments in the AI world. He notes the irony of OpenAI, a major IP thief, complaining about its own IP being stolen. This could be a victory for karma.

• Airbnb says a third of its customer support is now handled by AI in the US and Canada (techcrunch.com) Airbnb says its custom-built AI agent now handles around a third of its customer support issues in the US and Canada, and plans to roll out the feature globally. CEO Brian Chesky believes the AI will provide better service quality and reduce costs.

• He was once a teen ‘superstar’. Why did James Van Der Beek need help to pay his medical bills? (bbc.com) Actor James Van Der Beek, once a teen ‘superstar’ known for his role in Dawson’s Creek, faced financial struggles later in life. Despite his fame, Van Der Beek’s family needed help from friends and fans to pay his medical bills after his cancer diagnosis. An indictment on fickle acting incomes and the lack of universal health care in USA.

• AI twitter’s favourite lie: everyone wants to be a developer (joanwestenberg.com) Westenberg argues that the idea of everyone becoming a software developer due to AI-powered code generation is a fantasy. Most people do not want to build, prompt, or think about software - they simply want their problems solved with minimal effort.

Subscribe to Prof Galloway

Sat, 14 Feb 2026 20:43:46 +1100

I don’t read many newsletters consistently. Most get scanned, filed or deleted.

Prof Galloway is one of the few I actually read. Every edition. And I learn something every time.

Scott Galloway sits at the crossroads of media, technology and human behaviour. That intersection is where business strategy meets cultural commentary, and he explores it relentlessly.

What makes it work is the blend. He moves between anecdotal observation and solid research. One paragraph you’re getting a personal story about parenting or aging. The next you’re looking at market data, behavioural studies or economic trends. The numbers back up the narrative. The stories make the numbers stick.

It’s a tasty mix that really appeals to me.

Galloway writes from decades of experience building companies, teaching at NYU Stern and watching markets move. He’s not reporting on tech and media from the outside. He’s been inside these systems long enough to see the patterns.

The writing is direct. No academic hedging, no corporate speak. When he thinks something is broken, he says so. When he changes his mind, he tells you why. The tone is conversational but the analysis is serious.

He consistently explores how technology and media shape behaviour and where the incentives are misaligned. He’s particularly good at connecting dots between business models and societal outcomes.

It’s free. It arrives regularly. And it doesn’t waste your time.

If you’re interested in why platforms make the choices they do, how media economics drive editorial decisions, or what drives human behaviour in digital systems, subscribe to Prof Galloway.

You can always unsubscribe. But I doubt you will.

PS: I love the greyscale semi-hand drawn charts he uses. They’re insightful without being flashy, and they work brilliantly for my partial colour blindness. Data visualisation that actually prioritises clarity over decoration is rarer than it should be.

NYT - Trump Has Overwhelmed Himself

Mon, 09 Feb 2026 08:01:10 +1100

Ezra Klein’s analysis in NY Times cuts through the noise to explain what we’re actually witnessing: not crisis management, but crisis generation as governing philosophy. The Bannon playbook (move fast, flood the zone, overwhelm attention) worked until it didn’t. Now the administration drowns in its own turbulence.

I try not to get into the daily news thrash from Trumpistan, but I occasionally find articles that elevate above the din and provide insight. This is one of them.

Ezra Klein in New York Times writes:

The strategy of the Trump administration over the last year has been to move so fast, to do so much, that the opposition could never find its footing. This was Bannon’s insight, and it was real: Attention is limited. The media, the opposition, the electorate — they can only focus on so much. Overwhelm their capacity for attention and you overwhelm their capacity to think, organize and oppose.

But what you are doing to the opposition you are also doing to yourself. “It is a strategy that forces you into overreach,” I wrote last year. “To keep the zone flooded, you have to keep acting, keep moving, keep creating new cycles of outrage or fear. You overwhelm yourself.” And that is what happened. The Trump administration is overwhelmed — by its own violence, its own cruelty, its own lies, its own chaos. There is nothing unusual about a presidency being overwhelmed by crises. What is unusual about the Trump administration is that it has created those crises itself. …

… Trump appears to be trying to course correct, but he has neither the discipline nor the personnel to truly change his presidency’s direction. This administration is a reflection of who the president is and what he wants. This White House is not beset by crises. This White House is the crisis.

The community model for news

Sun, 08 Feb 2026 23:10:22 +1100

Watching from Australia as American journalism tears itself apart, it’s hard not to feel a strange mix of recognition and regret. Each week brings another round of layoffs, another editorial board overruled by its owner, another reminder that the people who fund news are not the people it’s meant to serve. Dave Winer wants news to work. So do i. And reading his proposal for community news organisations, I found myself thinking: we’ve seen this model succeed before.

Just not in news.

We’ve done this with banks

This requires a brief detour into Australian banking history, which I promise is more relevant than it sounds.

As Australia’s Big 4 banks withdrew from rural and outer-suburban centres over the past two decades, communities didn’t simply accept the loss. They built their own banks. Local communities form a company, raise capital from shareholders, and partner with Bendigo Bank to operate a full-service branch. Professional standards, regulatory compliance, genuine community ownership. Profits split between Bendigo and the local company, with that local share flowing back into community projects.

There are now over 300 Community Bank branches across Australia, managing billions in deposits. These aren’t charity projects. They’re viable businesses built on a simple insight: when you align the incentives of the service provider with the community being served, you get better outcomes for everyone.

The structural parallels

Dave’s vision maps onto this almost perfectly.

Local ownership with modest equity stakes. Dave proposes readers buy shares “with maybe very little hope of getting a return in dollars, rather in a more functional community.” Community Banks work identically—you invest because you want the service to exist, not for financial return.

Professional expertise meets community governance. Dave envisions professional journalists working alongside bloggers under shared integrity standards. Community Banks blend professional banking knowledge from Bendigo with local governance from community directors.

Sustainable funding without extraction. Dave’s toll pass system—micropayments per article, no ads, no paywalls—creates revenue without perverse incentives. Community Banks operate on standard banking economics, but without pressure to maximise returns for distant shareholders.

No oligarchs pulling the strings. Dave is explicit: “As readers we know you’re often full of it because of who owns you.” Community Banks rejected decisions made by executives in Sydney optimising quarterly earnings. Distribute ownership among the people being served, and you eliminate that pathology.

Why this works

Community Banks have survived financial crises, regulatory changes, and competitive pressure from major banks and fintech challengers. What makes them resilient is aligned incentives. When the people making decisions are also the people using the service, you get rapid feedback and genuine accountability.

Community news would respond the same way. Ownership creates stakeholders, not just customers—people who use the service, recommend it, defend it. I don’t want to oversell this. Community Banks aren’t perfect, and community news wouldn’t be either. But the incentive structure is fundamentally healthier than what we have now.

The oligarch problem

When Bezos can kill Washington Post endorsements, when the LA Times owner overrules his editorial board, when Murdoch pushes agendas across his global media empire—readers see it. Trust erodes not because journalists are incompetent, but because we can’t be certain they’re free to do their jobs.

Community ownership doesn’t eliminate bias. But it eliminates owner interference driven by business interests or political ambitions unrelated to journalism. The interference mechanism simply doesn’t exist.

The technology piece

Dave’s offer here is crucial. Independent developers building editorial and publishing software for free, with no lock-in, supporting open standards. Community Banks use standardised systems from Bendigo—they don’t each build their own platform. This shared infrastructure lets communities focus on service and governance rather than reinventing technical wheels.

The key is these tools must be genuinely open. Communities can’t trade oligarch owners for platform lock-in. That would relocate the problem rather than solve it.

Why now

Mass layoffs. Declining trust. AI about to devastate what remains of traditional journalism economics. If you’re going to rebuild the model anyway—and at this point, the rebuilding seems inevitable—now’s the time to build it right.

Dave’s vision, grounded in the proven model of community banking, offers a path forward. It won’t be easy. Building 300 Community Banks took decades of patient work, community by community. Community news would require the same persistence.

But it could work. And that’s more than we can say for the status quo.

Sources:

The reboot that news needs (daveverse.org)
Bendigo Bank Community Bank model (bendigobank.com.au)

Sun, 08 Feb 2026 22:08:27 +1100

Link posts 8 Feb 2026

• Waymo is Having a Hard Time Stopping For School Buses (tech.slashdot.org) Waymo’s self-driving robotaxis have faced at least 24 safety violations involving school buses in Austin since 2025, with the company’s voluntary software update failing to fix the problem.

• Post-Brexit sales of British farm products to EU fall by 37% (theguardian.com) British farm product exports to the EU have fallen by 37% in the 5 years since Brexit in 2021. Rebuilding demand for British products in Europe will take time. IMHO it was needless economic suicide driven by popularism and bankrolled by external forces.

• Claude Code is the Inflection Point (developers.slashdot.org) Anthropic’s Claude Code, a terminal-native AI coding agent, now authors about 4% of all public commits on GitHub. SemiAnalysis projects this figure will exceed 20% by 2026.

• 3 ways to fix Australia’s affordability crisis (firstlinks.com.au) There are no easy fixes but we need to look at options. This outlines three ways to address Australia’s affordability crisis: increasing housing supply, reforming planning laws, and improving public transport.

The week that was

Sun, 08 Feb 2026 12:38:45 +1100

Busy week. 19 blog posts. 99 articles stashed for deeper study. It’s sometimes too easy to get buried in detail and lose sight of the bigger picture. So I had Claude Code help me out with a word cloud (above) and summary below.

This week, I was tracking a few key themes that seemed to dominate the blog posts and bookmarks: The rise of Anthropic and its AI products was a major focal point, with multiple stories covering the company’s Claude chatbot and its integration into platforms like WordPress.

There was a sense that Anthropic’s tools are starting to have a significant impact, with one report estimating that Claude Code now authors around 4% of all public GitHub commits. The company’s legal tool also made waves, causing a $285 billion sell-off in SaaS valuations. Alongside Anthropic, the broader AI landscape was a hot topic, with discussions around the valuation of AI companies, the capabilities of OpenAI’s Codex, and concerns about self-driving cars' ability to navigate around school buses. There was a palpable sense of both excitement and unease around the rapid advancements in AI technology.

The economic climate also featured prominently, with stories on the affordability crisis in Australia, the post-Brexit decline in UK farm exports, and Stellantis' €22 billion hit from overestimating the shift to electric vehicles. These reports painted a picture of ongoing challenges and volatility in the global economy.

While the overall tone was fairly technical and business-focused, there were also a few more politically-charged stories, such as the coverage of the Epstein scandal and a late catch of Hillary Clinton’s commentary on the “war on empathy” within the MAGA movement. These pieces hinted at the continued polarisation and tension in the political sphere.

In summary, this week’s content was dominated by the rise of Anthropic and AI, the state of the economy, and some lingering political and social issues. It was a diverse and thought-provoking set of stories that captured the complexity of the current landscape.

Sun, 08 Feb 2026 10:37:42 +1100

Automated cover-up fails

In one of the siloed social media cess pits Quadcarl said this

I’m not a very smart man but it looks to me like a computer thought this said Don and redacted it. In case you don’t realize it yet this is a big cover-up.

I had to go to original source and verify.

The original file is available for your inspection at: https://www.justice.gov/epstein/files/DataSet%2011/EFTA02440040.pdf

Update: The easiest access to the “Epstein Files” is via The Jmail Suite - an interactive archive that lets you explore Jeffrey Epstein’s emails, documents, photos, flight logs, and more as if you were logged into his accounts. What started as a viral project in November 2025 has grown into a full suite of apps, powered by three major data releases.

The dirty work nobody wants to do is suddenly getting done

Sun, 08 Feb 2026 07:55:00 +1100

Every programmer knows the code rewrite conversation. The codebase has drifted. Security vulnerabilities are piling up. The platform you’re running on is three versions behind. It needs fixing, but nobody wants to touch it.

The economics are brutal. Thousands of hours rewriting working code to do exactly what it already does. Management looks at the spreadsheet and sees massive cost with zero new features. Engineers look at the backlog and wonder why they signed up to translate code line by line when they could be building something new.

This is where agentic AI tools are starting to shift the calculation.

SmartFriend™️ Peter’s weekend project

Peter Marks, who writes at marxy.org, recently documented his experience porting FreeDV RADE v1 from Python to C. FreeDV is digital voice software for amateur radio. The Python implementation works, but C “would make it easier to install, smaller, and hopefully more efficient”.

The conventional wisdom says this kind of port takes weeks. Peter used AI tools to accelerate the translation, documented the process and shared both the successes and the limitations. Peter says what would have taken months has instead been delivered in days across a few sessions.

Microsoft’s billion-line bet

Microsoft Distinguished Engineer Galen Hunt declared his goal in December 2025: “eliminate every line of C and C++ from Microsoft by 2030.” The target metric is one engineer, one month, one million lines of code.

Microsoft is building infrastructure to map source code at scale, then apply AI-guided modifications across millions of lines at a time. The primary target is translating C and C++ systems to Rust, eliminating entire categories of memory safety bugs. More than 70% of vulnerabilities in Microsoft products trace back to memory safety issues that Rust makes structurally impossible.

This is research infrastructure, not a Windows rewrite. The project sits within Microsoft’s Future of Scalable Software Engineering group with a mandate to remove technical debt at scale.

Why this matters

The economics of code rewrites have shifted. Doing nothing compounds technical debt and security exposure. Doing the work manually has been prohibitive. Engineers don’t get into the field to be human compilers.

AI tools change the calculation. Peter’s experience shows the pattern: AI handles mechanical translation, humans handle judgement calls. Microsoft’s infrastructure suggests this scales to enterprise codebases. The grunt work of translation is becoming automatable.

The bottleneck shifts from transcription to architectural judgement. That is better use of engineering talent.

If you’re sitting on legacy code that needs modernising, the calculation just changed. The question is no longer whether to migrate, but whether to do it now or wait until the debt becomes unmanageable.

The dirty work nobody wanted to do is becoming the work that finally gets done.

Sources:

Experimental C port of FreeDV RADE v1 - Peter Marks, marxy.org (blog.marxy.org)
Microsoft wants to replace its entire C and C++ codebase - The Register (theregister.com)
Microsoft building team to eliminate C and C++ - Windows Latest (windowslatest.com)
Microsoft’s Bold Goal: Replace 1B Lines of C/C++ With Rust - The New Stack (thenewstack.io)

Sat, 07 Feb 2026 19:32:36 +1100

Link posts 7 Feb 2026

• So the Epstein scandal is about politics? Silly me for thinking it’s about the mass abuse of women and girls (theguardian.com) Hyde dismantles the partisan point-scoring around the Epstein files, arguing it lets the powerful escape accountability while victims are sidelined yet again.

• We had s*x in a Chinese hotel, then found we had been broadcast to thousands (bbc.com) Eric and his girlfriend unknowingly had their intimate hotel stay in Shenzhen, China broadcast to thousands of strangers on a social media channel.

• Wall Street just lost $285 billion because of 13 markdown files - Martin Alderson (martinalderson.com) Anthropic’s legal tool, a collection of 13 markdown files, caused a $285 billion sell-off in SaaS company valuations on the public markets. Over-reaction? Or the first volleys in a new battle?

• Introducing About the Song, a New Way to Explore the Stories Behind the Music — Spotify (newsroom.spotify.com) Spotify’s new feature on iOS/Android provides stories and context about the music users are listening to. The feature, currently in beta, displays short, swipeable cards in Now Playing View. Available in 6 countries including Australia.