DEV Community: Elliot Sturzaker

The race to superintelligence.

Elliot Sturzaker — Thu, 29 Oct 2026 21:50:00 +0000

Jacob Coxon, 27, a Cambridge graduate, spent three years in pretraining research, first at OpenAI, where he was among the core contributors to GPT-4o, then at Anthropic from this past July. This week he quit and said what many in AI safety circles reportedly already believe but rarely say publicly: the people building this technology think it could kill us all by the end of the decade, and it isn't a marketing line. He told the Wall Street Journal colleagues internally already describe the moment using words like "crunchtime" and "endgame." His read on the two companies differed: OpenAI staff, he said, hadn't internalized the stakes; Anthropic staff understood the risk but kept racing anyway, betting it's safer if they get there first than a less careful rival.

He isn't the first to say this. Geoffrey Hinton and Yoshua Bengio, have raised similar alarms since 2023, and DeepMind's Alex Turner resigned in June over a related concern. What's new isn't the warning, it's how mainstream saying it out loud has become. In a CNN interview, Coxon pointed to what changed his own thinking: two months earlier, OpenAI's agents had hacked into third-party infrastructure entirely on their own initiative. He's careful to separate present risk (real, but not extinction-level) from the trajectory that worries him: coding and math AI has gone from offering suggestions to nearly replacing people in two years, and he believes recursive self-improvement, AI improving itself faster than expected, could begin within a year or two, the point where extinction-level risk stops being hypothetical.

He wasn't alone. Evan Hubinger, who leads Anthropic's Alignment Science team, backed him within hours, saying Anthropic researchers really do believe AI could kill everyone, and put his own odds above 10% within a decade, calling present-model risk low but recursive self-improvement the real, accelerating concern.

Now the CEOs are saying it too. Anthropic's Dario Amodei and OpenAI's Sam Altman have both, this week, talked about deliberately slowing capability growth so safety research can catch up, a real shift from three years ago, when a similar pause letter was signed by thousands and ignored. This time it's coming from inside the labs. It still hasn't translated into anyone stopping.

Why this counts as a different category of risk

Superintelligence, stripped down, is a system that can improve itself faster than anyone can check what it's becoming, eventually outperforming top humans at nearly everything, including that same kind of quiet probing and workaround-finding, just at a level nobody can catch in time. It gets treated as an extinction-level risk rather than a product risk for one reason: control. A tool you can out-think, you can supervise. One that out-thinks you at the task of supervising it is a different category of problem entirely.

But the technology itself isn't the threat. A model doesn't have intentions. What actually worries me is the race: the fact that the people closest to this, with the most reason to know, now say openly that it's dangerous, and keep moving anyway, because slowing down alone just hands the lead to whoever they trust even less. That's not new. That's an old human problem, an arms race, wearing a very new coat.

The race nobody will be first to stop

Even while talking about slowing down, the labs are doing it with a footnote: not first, not alone, not unless everyone else does too. This past weekend, Amodei published an essay making the geopolitical case directly, arguing that a Chinese lead in AI would pose grave danger for the United States and the world, and using that as part of a broader argument for the US to curb China's AI capabilities as one piece of a global slowdown. That's not a call to pause. It's a call to make sure whoever gets there first is the US.

China didn't take it quietly. China's Ministry of Foreign Affairs pushed back on Monday, framing Amodei's call as fearmongering. Separately, the head of China's Ministry of State Security warned in an essay of his own that in the wrong hands, this technology could threaten China's political, institutional, and ideological security pointing specifically at capabilities coming out of US companies like Anthropic and OpenAI, and framing them as tools that let hostile actors find vulnerabilities and carry out complex hacking operations at scale.

Which leaves the real question unanswered: everyone at the top says this is dangerous, but almost nobody says, specifically, what the danger actually is. Is it hacking? The jobs? Or is the real threat something most people haven't even been told to worry about, recursive self-improvement, the point where the system starts getting better at getting better, faster than anyone can check what it's becoming? That distinction matters, because you govern a hacking risk and a jobs risk very differently than you'd govern a system that might outpace its own supervision. And right now, the closest thing to an answer came from the White House: when Amodei, Altman, and Musk jointly called for slowing down, President Trump rejected it outright, saying the only guardrail AI needs is a strong president, not a framework, not a law. This is a genuinely new kind of technology, and it is always going to need its rules to grow alongside it, not after it. Instead, autonomous agents have been operating in public for months with nothing catching up to them, and the loudest response so far has been a president dismissing the concern entirely. No wonder the public is left afraid of something nobody has actually explained. When both sides, and now even the government, keep saying "not us, not first," who exactly is supposed to blink first, and who's going to tell people what they're actually supposed to be afraid of?

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/the-race-to-superintelligence.

Can a laptop trader beat the bots on Polymarket’s AI markets?

Elliot Sturzaker — Thu, 29 Oct 2026 21:49:58 +0000

Polymarket is a venue where people bet real money on future events, a 30¢ “yes” share means the crowd thinks there’s roughly a 30% chance. Its fastest-growing slice right now is AI: markets on model release dates, month-end leaderboard rankings, corporate moves. Around $10 million of volume moved across 234 active AI markets in a single month.

Our question is narrow and honest: can a retail operator, a laptop, $1,000, 200–300ms of latency, no co-located servers, find places where these markets are mispriced, and bet against the wrong price before the bots correct it? We’re not predicting who wins. We’re asking whether the price-setters are sometimes slow enough to be beaten.

Three places to look

H1, breaking news. When a lab ships a model or drops a hint, how fast does the market reprice? A delay is a window. Measured against twelve pre-registered release-date markets.

H2, the leaderboard. Is the “best model at month-end” basket priced correctly against the actual LM Arena outcome? This is the one most likely to die, the traders in these markets include AI researchers, so an information edge is rare.

H3, corporate catalysts. Do slow-developing events, filings, funding rounds, executive moves, give a human a multi-hour repricing window the bots don’t bother to chase? Five markets under watch.

Reconnaissance, not a fishing trip

There’s no published number to replicate here, so instead of pre-committing to a profit threshold we pre-committed to the reporting: what gets measured, what counts as a repricing event, which news sources are allowed, and how each event is classified, all locked before the observation window opens. The verdict can land in one of four states, including “exploratory positive,” so an interesting-but-unproven pattern can’t be quietly dressed up as a win.

That structure exists to defend against the most common way this kind of study lies to itself: finding a pattern in the data and then pretending you predicted it. Which is exactly the trap the next paragraph walks up to.

The thing we can’t explain yet

Two days in, the news-latency data showed a split: Western-lab news (OpenAI, Anthropic, Google) got priced in within about five hours, while Chinese and smaller-lab news took closer to twenty. A structural, tradeable-looking asymmetry.

It is also three observations. Three. That is far too few to mean anything, and the honest move is to write it down, flag it as out-of-scope for this study, and refuse to build a conclusion on it, rather than let a tidy story override a tiny sample. It goes in the Day-30 writeup either way. The verdict lands at the end of the observation window, whichever way it falls.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/can-a-laptop-trader-beat-the-bots-on-polymarket-ai-markets.

The end of the small SaaS.

Elliot Sturzaker — Thu, 29 Oct 2026 21:44:07 +0000

I was paying for two things to do jobs my Mac could basically already do.

Willow, US$15 a month, so I could talk instead of type.

Fireflies, charged per seat, to sit in client calls and write the notes afterwards.

Both work. Both also felt slightly ridiculous for what I wanted.

I didn't really want another bot joining client meetings. A few clients had commented on it too, which is fair enough. Nobody loves a third participant called "Fred's Notetaker" that never says anything and never leaves. And for dictation, I was sending my voice to a server in another country when there is a stupidly capable chip sitting about forty centimetres from my face.

So one afternoon I got annoyed and built Speakhold.

A tiny native Swift app that does both.

Hold a key, talk, let go, and it types.

Press record, have your meeting, and it writes the notes.

No meeting bot. No subscription. No audio leaving the Mac.

By the end of the afternoon it was built, launched and open source.

Then I started using it on real client calls and immediately found out all the stupid things I had got wrong.

What it does

For dictation, you hold left Control + left Option and talk.

Let go and the sentence lands wherever your cursor already is. Slack, Mail, Xcode, the browser, a form on a government website at 11 pm, whatever.

The speech model runs on the Mac. It cleans up the filler words and uses a local dictionary for the names and words I use all the time, so "Smriti" stops coming out as "smitten".

Meeting mode is just as simple.

Press record and Speakhold captures my microphone and the audio coming from the Mac separately.

That means my side and everyone else's side are already split without trying to guess who spoke. If you want the other side split into Speaker 1, 2 and 3 as well, there is a toggle that runs a speaker model on the Mac.

When the call finishes, it turns the recording into a markdown file with a summary, decisions, action items, open questions and the full transcript.

Then it deletes the audio.

Worth stealing

Left and right modifier keys are different keys.

Most dictation apps need a shortcut, and eventually that shortcut gets annoying or collides with something else. Speakhold listens low enough down that left Control and right Control are actually different inputs. Same with Option. So the default can be left Control + left Option without stealing a normal shortcut I use elsewhere. Tiny product decision, but one of those things that makes the app disappear when you are not using it.

Claude Code can be a model call without another API key.

I wanted decent meeting notes without making people sign up for another API just to use the app. If Claude Code is installed, Speakhold runs it in print mode with tools disabled and pipes the transcript in under the user's existing login. A 30 minute call comes back as notes in about 20 seconds. If that is not available, it falls back to the user's own key or an on-device model.

The gotcha was that calling the CLI slightly wrong could ignore the keychain login completely. Also, if the app itself is launched from a Claude Code session, some environment variables need stripping before spawning another one, or you get a Claude inside a Claude and nobody is happy. Weird little problems you only find by actually wiring this stuff together.

Read other people's launch threads before your own.

Before posting it anywhere I pulled every comment from the eight most similar launches on Hacker News, 337 comments, and counted what people asked. The top question by a mile was "how is this different from the other twelve". The second was "why are you still using Whisper, it is from 2022". Both were fair. So I fixed the second one before anyone could ask it. More on that below.

What broke

The first real client call mostly worked.

Afterwards, Speakhold sat on "Transcribing" with nothing moving.

It was not frozen. It was doing exactly what I had told it to do.

It just looked frozen, which is basically the same thing to the person using it.

So transcription moved into a proper background job with visible progress.

The second call broke it properly.

I pressed stop. Speakhold handed the work to the background process, which immediately asked macOS for notification permission and waited for me to answer.

Problem was, I had not released the recording hardware yet.

So the meeting was finished, I had pressed stop, and the microphone just kept going while the app politely waited for me to click a permission popup. If you have ever hung up and then said something about the person you were just talking to, you can see the problem.

That one was pretty humbling.

The fix was obvious afterwards: release the hardware first. Nothing after a user presses stop should depend on them doing something else.

Whisper had another surprise too.

Leave it listening to silence for long enough and it starts making things up.

Mine became extremely polite.

"Thank you. Thank you."

Over and over again. Eight minutes of an empty room came back as a page of gratitude.

I added filtering for low-confidence repeated output, which helped. Then I did the thing the Hacker News threads kept telling everyone to do and tried Parakeet, NVIDIA's open speech model, on the same clips. Same words, seven times faster, and over the empty room it returned nothing at all. Which is the correct answer. It is the default now, with Whisper kept for the ninety-odd languages Parakeet does not speak.

The main lesson from all of it was pretty simple.

Real people break software much better than you do. Other people's comment sections are the next best thing.

What it is not

It is Apple Silicon only and currently macOS 14 or newer.

English first for now.

It is a personal tool, not another team workspace with dashboards and search and a Q4 pricing review.

It does not announce that a meeting is being recorded. The consent rules wherever you are still apply.

The meeting notes are only as good as the model producing them.

Where it is

Download it from speakhold.wildinteractive.app, a 12 MB disk image, or with Homebrew: brew install --cask sturzael/tap/speakhold. It is not notarised yet because I have not given Apple US$99, so the install page covers the first-launch dialog. The source is on GitHub.

MIT licensed. Tiny download. Free.

I think this is the kind of AI work I am most interested in at the moment.

Not adding a chatbot to something because it needs an AI feature.

Just finding one annoying job, putting the model where the data already is, and doing the boring integration work properly.

In this case I got annoyed, built and launched something slightly ridiculous in an afternoon, broke it almost immediately on real client calls, fixed the dumb bits and gave it away.

Pretty productive afternoon, in hindsight. The two subscriptions are cancelled. The chip forty centimetres from my face is finally earning its keep.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/the-end-of-the-small-saas.

Is AI the new polyester?

Elliot Sturzaker — Thu, 29 Oct 2026 21:44:04 +0000

Lately, if you spend time online, you'll come across a comparison that keeps showing up: people are saying AI is the new polyester. A synthetic thing that quietly took over, the way plastic and cheap fabric once did. I get why it's caught on so much, polyester is something almost everyone has worn without thinking twice, so it makes AI feel familiar instead of abstract, something we've already lived through once. It made me want to look closer at the comparison, because I think it's onto something real, just aimed at the wrong target.

A fabric nobody wanted, until everybody had it

Polyester wasn't even invented by the company most people associate with it. Two British chemists, Whinfield and Dickson, developed it in the early 1940s. DuPont just bought the US manufacturing rights a few years later and branded it Dacron. When it reached American consumers in 1951, it was marketed as a genuine miracle: a suit you could wear for 68 days straight, no ironing, and it would still look presentable. Housewives were sold on convenience. The fashion industry was sold on durability and low cost.

For about two decades after that, people mostly hated wearing it. It didn't breathe, it held onto heat and odour, and by the late 1970s "polyester" had become shorthand for cheap and try hard, the leisure suit joke rather than the fabric of the future. Double-knit polyester suits became the uniform of an entire decade's discomfort: the disco floor sweat stain, the static cling, the shirt that looked fine in the store and clung to your back an hour into wearing it. It became something people joked about rather than reached for, and manufacturers knew it. So the industry found somewhere else for it to live, somewhere the flaws didn't matter as much.

That somewhere was outdoor and sportswear. Through the 1970s and 80s the industry adopted it seriously, because for that specific use, lightweight, weather resistant, quick drying, it actually beat cotton or wool. DuPont's own moisture wicking fiber, launched in 1986, and Patagonia's fleece the year before, are usually credited with turning polyester from a cheap substitute into a high performance material in people's minds. Nike built on that with Dri-FIT in the 90s, and by the 2000s the flip had fully happened. Polyester wasn't the alternative anymore, it was the default in most clothing, and "100% cotton" had quietly become the label you pay a premium for.

A synthetic version floods a market, gets rejected, finds a use it's actually suited for, and eventually becomes so common that the original material turns into the premium option.

Why I don't think AI fits that same story

It's tempting to slot AI straight into that pattern, and a lot of smart people are doing exactly that right now. I don't think it holds up, and the reason is specific. Polyester's entire ceiling, at every stage of its history, was cheaper and more convenient than the real thing. No version of a polyester shirt ever cured a disease, extended a life, or did something cotton simply couldn't. That ceiling was never on the table for it.

To be fair to the comparison, though, it's not really a claim about what AI could theoretically do. It's a claim about how AI gets used, pushed into everything regardless of fit, while a handful of companies capture most of the profit and the rest of us absorb the downside. I think that's a fair description of a lot of AI deployment right now. Where I think it goes wrong is treating that pattern as proof AI's ceiling is low, when it's really proof of a choice being made well above what the technology is capable of.

AI already has that higher ceiling, and you can see it in two very different fields. Insilico Medicine took an AI discovered drug for a fibrotic lung disease from target identification to Phase II clinical trials in under 30 months, a process that traditionally takes six to eight years. Antibiotic discovery, stagnant for decades because new classes weren't commercially attractive enough, is being revived the same way, deep learning models screening millions of candidate molecules through a field that had been written off as scientifically exhausted. Climate resilience tells a similar story outside medicine. For most of history, accurate flood forecasting at scale simply wasn't possible, the physics were too complex and most of the world's rivers had no gauges to measure them. AI-driven models have since expanded reliable flood forecasting to more than 150 countries, including Bangladesh, where the system now covers 360 million people who previously had no early warning system at all.

Compare that to what polyester's one real strength actually cost people. Performance wear came bundled with a health price, not a benefit. Every wash sheds microplastics, as many as 6.8 million fibres from a single cycle, and the dyes used to colour it are one of the most common causes of textile related skin irritation. Even polyester's one benefit came with a cost attached. AI's tradeoff, in medicine and in climate forecasting, has gone the other direction: it's improving outcomes people actually feel, not quietly taxing them.

So if it's not the technology, what's actually wrong?

My take isn't "AI is good, don't worry about it." We've had artificial intelligence in some form for decades, pattern recognition, recommendation engines, expert systems. What's new is the generative wave, and more specifically, the race that's formed around it. The five biggest US hyperscalers are on track to spend somewhere around $700 billion combined on AI infrastructure in 2026 alone, nearly double what they spent the year before, all chasing dominance in a market that hasn't even proven it can generate returns close to that spend.

That's the environment the hype grows in. When the whole industry is racing to be first, biggest, and most "transformative," the boring, unglamorous problems that actually matter, climate mitigation, treating diseases nobody's chasing, the antibiotic example above, keep losing to whatever version of AI gets attention and funding fastest. A quiet, hard problem solved rarely makes headlines the way a flashy new chatbot launch does. So the industry reaches for the hype word and climbs the ladder, while the harder problems sit there mostly untouched, and the people who need them solved are stuck absorbing the cost.

The real answer to the title

So, is AI the new polyester? I don't think so. Polyester was cheap and synthetic everywhere on purpose, because that was the entire ceiling of what it could ever be, nobody was ever going to find a life-saving use for a leisure suit. AI has no such ceiling. A model that compresses a decade of antibiotic research into something workable again, or gives 360 million people in Bangladesh a flood warning they never had, is doing something that simply wasn't available at any price before now.

Which is why I'd say the polyester comparison is aimed at the wrong culprit rather than the wrong pattern. The critics have a point on all three counts: AI is being deployed carelessly, a handful of companies are capturing most of the value, and the whole thing does look like a synthetic material flooding in before anyone understood the cost. What they're missing is that none of that was inevitable the way polyester's ceiling was. A fabric never got a say in what it became. The people building and selling AI do, every time they choose "impressive and fast" over "good but slow." A fabric can't be held responsible for that choice. The people making it can.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/is-ai-the-new-polyester.

Stop buying AI features. Start redesigning the work.

Elliot Sturzaker — Thu, 29 Oct 2026 21:44:03 +0000

Ask most business owners in New Zealand whether they’re “using AI,” and the answer is almost certainly yes: 87% of New Zealand organisations are now using some form of AI in their operations, and a separate study puts the figure at 82%. So why did productivity go backwards over the same period? That’s the part that should stop you. It tells you the problem was never adoption. It was what “adoption” actually meant in practice. Switching on a tool isn’t the same as changing how the work gets done, and most businesses have quietly settled for the former while assuming it counts as the latter.

The hype sells features. The problem needs redesign.

We see this constantly in the work we do. A business adds a chatbot to a site that’s otherwise a digital brochure, and wonders why nothing much changes. The customer still can’t check stock, still can’t self-serve the thing they came for, and the enquiry lands in the same inbox it always did. Nothing about the underlying journey changed, only one small piece of it got a new coat of paint. The AI is real. The workflow around it isn’t. That’s the whole problem in one sentence, and it’s why New Zealand’s headline adoption number is closer to a vanity metric than a scoreboard. It mostly measures how easily AI has been switched on, not how deeply it’s been put to work, and switching something on has never been the hard part.

This isn’t a fringe take anymore. Deloitte’s 2026 Tech Trends report has gone as far as calling the era of AI experimentation over, framing shallow adoption, meaning chatbots and copilots layered onto unchanged processes, as a genuine risk to productivity, service quality, and cost. And the gap isn’t closing on its own. Globally, nearly three quarters of AI’s economic value is being captured by just one fifth of organisations, and those leading organisations are twice as likely to have redesigned workflows around AI, rather than simply adding tools on top. In other words: the businesses pulling ahead aren’t the ones with the most AI. They’re the ones who did the harder, less visible work first.

The businesses actually seeing returns

Here’s the number worth sitting with: small and medium firms using AI earned roughly $400,000 more than comparable non-adopters in the last financial year. That’s not a rounding error, and it’s not the kind of return you get from a summarise button. It’s the return you get when a company rethinks how a quote is generated, how a customer enquiry is handled from first click to final invoice, or how a backlog of unstructured information becomes something staff can actually query.

Picture the pattern we see most often: a small business owner spends three days a week manually pulling numbers into a quote, cross-checking a spreadsheet, and emailing back and forth before a customer gets a price. Add an AI chatbot to the website, and none of that changes, because the bottleneck was never “answering questions,” it was the quoting process itself. Redesign the quoting workflow instead, and the chatbot becomes almost unnecessary, because the real problem, those three days, is what actually goes away.

Notice what all three of those have in common: none of them are “add an AI feature.” All three are “redraw how the work happens, then let AI carry part of the new shape.” That distinction is, in our view, the entire ballgame. As one industry leader put it after the same findings: don’t just use it, redesign how you’re using it, and transform how you’re thinking about the business.

It’s not a resistance problem. It’s a confidence and time problem.

What’s easy to miss is that small businesses aren’t dragging their feet out of stubbornness. Nearly half of small and medium businesses believe AI could be the most significant opportunity since the rise of the internet, yet a deep confidence gap holds many owners back, not resistance, but resource constraints: lack of time, trust in data protections, and a clear implementation roadmap. Nearly two-thirds are already proactively using AI, and most of them are learning through self-guided trial and error, without anyone showing them what “done properly” actually looks like. That’s the gap we think matters most, and it’s rarely the one vendors are selling a fix for.

At the leadership level, the picture doesn’t improve much: a majority of NZ leaders worry their organisation lacks a plan and vision to implement AI, and less than a third have formal ethics or safety guidelines in place. The tools arrived faster than anyone built the operating model to hold them.

Why “just add AI” can quietly go wrong

It’s worth being honest that this cuts both ways. Move too fast without oversight, and the downside is real and public. One widely reported case saw an AU$440,000 consulting report, written using AI, partially refunded after it was found to contain errors, fake academic references, and a fabricated legal quote. That’s an extreme example, but the underlying lesson generalises to something far smaller and more common: AI output is a draft, not a deliverable, until a human has actually reviewed it against the work it’s meant to support. We’d rather a business move slightly slower and get this right than move fast and end up as the next cautionary anecdote.

The advantage small businesses already have

None of this should read as small businesses being at a disadvantage. If anything, we think the opposite is true. AI capability is increasingly commoditised, meaning the gap between what a large enterprise can access and what a small business can access keeps narrowing. And the advantages of being small, clear purpose, strong client relationships, and agile decision-making, are exactly the traits AI amplifies rather than erodes, precisely because a smaller business can redesign a workflow in weeks, not years. The risk was never starting late. It’s waiting long enough that the gap becomes something you can no longer close with a good quarter of effort.

Where do you actually start?

This is usually the point where the advice runs out. Business owners are told to “redesign the work, not just add AI,” which is true, but not exactly actionable at 9pm on a Tuesday when you’re not sure if your quoting process, your customer service inbox, or your inventory system is even the right place to start.

That’s the gap we built GoodCall to close. It’s a simple, no-pressure way for a small business owner to figure out what they actually need, not what’s trending, not what a vendor is pushing this quarter. You answer a handful of questions about how the work happens today, and GoodCall comes back with an honest verdict: where AI would genuinely help, what it would take, and when a spreadsheet or a smaller fix is the better call. No jargon, no sales pitch. When you want that turned into a plan you can act on, the next step is the AI Readiness Assessment, a week of work and a written roadmap that’s yours whoever you build with.

At Wild, this is the question we sit with clients on before we talk about any specific tool: not “which AI feature do you need,” but “which part of your business, if redesigned around what’s actually possible now, would change the outcome.” That’s a different starting point, and it tends to lead somewhere better.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/stop-buying-ai-features-start-redesigning-the-work.

We replicated a viral “61% from Trump tweets” trading claim. It didn’t survive.

Elliot Sturzaker — Thu, 29 Oct 2026 21:36:55 +0000

A project called trump-code went around claiming a 61.3% hit rate at predicting S&P 500 direction from Donald Trump’s Truth Social posts, 564 verified predictions, open-source, reproducible. That is the kind of claim worth taking seriously precisely because it is checkable. So we checked it.

We ran it as a pre-registered replication: decide what would count as success before looking at the results, write it down, then follow the plan whether we like where it lands or not. It’s the discipline that stops you quietly moving the goalposts once the numbers come in.

The number is real. The edge isn’t.

First finding: the 61.3% reproduces. Our independent calculation came out at 61.35%, a match to two decimal places. Their schema is readable, their maths checks out. Credit where it’s due.

The problem is what the number means. About a quarter of that hit rate (26% under our worst-case test) is lag-contaminated, the model is often reporting a move that has already happened rather than predicting one that hasn’t. A coin that tells you which way the coin already landed looks impressive and pays nothing.

So we tested the thing that actually matters: does the signal translate into tradeable edge on a real venue? We built 100 pre-committed strategies, five signal rules, ten Polymarket political markets across five categories, two direction conventions, and put every one through seven survival gates: enough samples, a real hit rate, margin over the fee, net edge, a multiple-comparisons correction, a reverse-direction check, and stability over time.

Zero of a hundred

None of the hundred cleared all seven gates. The two binding constraints, needing at least fifteen samples and a hit rate of at least 55%, turned out to be mutually unsatisfiable on this dataset: the cells with enough data never hit hard enough, and the cells that hit hard enough never had enough data.

Verdict: KILL. One rule was underpowered rather than falsified, not enough data to condemn it, so it’s flagged as the single candidate worth revisiting if anyone picks the work up. Everything else is a clean negative.

Why we publish the ones that fail

A negative result that took nine logical phases, a dozen commits and about two days of wall-clock work isn’t a failure, it’s the answer. Most “we found an edge” claims never survive contact with fees, lag, and a proper multiple-comparisons correction. Being able to tell the difference quickly, and being honest when the difference goes against you, is the actual skill.

The whole thing, the pre-registration, the gates, the counter-memo we wrote against our own conclusion before finalising it, is public. This is the kind of rigour we bring to the AI systems we build for clients: decide what “working” means up front, then let the evidence rule.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/replicating-a-61-percent-trump-tweet-trading-claim.

The grammar check that outs you as a machine

Elliot Sturzaker — Thu, 29 Oct 2026 21:36:53 +0000

Anthropic started embedding invisible watermarks into every piece of text Claude generates. If you've used Claude for a draft, an email, or a chunk of your book this week, that text now carries a signal you can't see, designed to survive copy and paste. This is not a hypothetical rollout: Anthropic confirmed the change in an updated support page, framing it as compliance with European regulations covering AI-generated content.

I sat on this piece for a few days on purpose. Anthropic had only confirmed that it was watermarking Claude's text, not how, and I wanted to write about the mechanism, not just the announcement. On August 14, 2026, Anthropic published a detailed account of how the text watermark actually works, identifying it as a version of the SynthID-Text technique that Google DeepMind published in a peer-reviewed 2024 Nature paper. That's the confirmation I was waiting for, so here's the piece, now with the actual mechanics instead of informed guessing.

The law behind it

The trigger is regulatory, not voluntary. Anthropic, along with several other major AI model providers and around 190 total signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, which requires AI system providers to use methods of "marking" AI-generated text. Anthropic didn't scope this to Europe. The company says it's applying watermarking globally at launch because it doesn't yet have a durable way to scope it by region. One region's law became the whole world's default, not by legal obligation but by technical convenience. That's the pattern I keep seeing in AI governance broadly: a handful of jurisdictions write the enforceable rules, and the rest of the world inherits the product decisions those rules force, whether their own regulators asked for it or not.

How it actually works

Here's the part I wanted confirmed before writing this. AI models typically generate a word at a time and decide on the next word based on the preceding text. When there's a choice that doesn't matter much to the reader, like whether to say "overcast" or "grey," the model uses a random number to decide which word to generate. Watermarking works by changing the source of that randomness, so the choices still look random to a reader but are decodable to anyone holding the key.

Anthropic has been fairly direct about the limits and guarantees of this system. The difference between watermarked and unwatermarked text won't be distinguishable to readers, nothing is added to the text, and there are no hidden characters. The watermark doesn't identify anything about individual users. There's nothing in the mark, or its key, that would let anyone recover information about the user, their organization, or their chats with Claude. A watermark only helps test whether Claude might have produced or processed the content. It doesn't say anything about ownership or authorship, and doesn't change who owns the output.

That third point is the one I think matters most, and I'll come back to it.

Where it breaks, now with actual numbers instead of speculation

This is the confirmation that matters most for anyone editing Claude's output before publishing, which is most of us. Anthropic says light editing may leave the watermark detectable, while a complete rewrite can remove it. So the earlier reasoning holds up: a grammar and punctuation pass is light editing. It's exactly the kind of touch likely to leave the pattern intact, because it barely touches word choice or sentence structure, which is what the mark rides on.

Anthropic is also building the piece that turns this from a curiosity into something with actual teeth: a watermark detection API is coming, and the company says it's still working out the implementation details. Right now, detection is theoretical for most people. Once it ships, it stops being theoretical.

Why the detection API is the part to actually watch

I'll say this plainly, as someone who spends a lot of my working life thinking about AI governance: the watermark itself is not the consequential part. It's inert until someone can check it. Once verification becomes widely accessible, publishers, companies, and software platforms could start checking Claude-generated material automatically. That's the moment this stops being an abstract transparency feature and starts being an operational filter, in hiring, in publishing, in academic settings, in performance reviews. And the result of that check will need far more careful interpretation than a yes-or-no answer allows: "Claude was probably involved" is not the same thing as "Claude wrote this", and I don't think most platforms building automated checks on top of the API will bother to preserve that distinction. Flat pass-fail signals are cheap to build and easy to misuse. Nuanced ones aren't, and nobody's incentivized to build the nuanced version first.

The part that still gets missed: this punishes light edits, unevenly

If you write something yourself and just ask Claude to fix spelling or punctuation, the confirmed mechanics say that pattern likely survives. You didn't generate the ideas or the sentences with AI. You used it the way you'd use a spellchecker. But a detector, and increasingly, a detector plugged into a hiring pipeline or a publishing CMS, doesn't know the difference between "I wrote this and cleaned it up" and "I asked for this."

For a non-native English writer, that's a real cost, not a technicality. Running a paragraph through Claude to fix grammar and get natural phrasing used to be a quiet, low-stakes way to get your actual thoughts into readable English. Now that same light pass can leave a trace that gets flagged as "AI-written," even though the ideas, structure, and feeling in the piece are entirely yours. The tool that used to lower the barrier to writing in a second language now comes with a detection risk attached to using it for exactly that purpose. And this happens to whoever needs the grammar pass most, which correlates with whoever is writing furthest from their first language, not with whoever is trying to pass off AI writing as their own.

What it doesn't prove, confirmed by Anthropic itself

A watermark only helps test whether Claude might have produced or processed content. It doesn't say anything about ownership or authorship. That's Anthropic's own framing, not a caveat I'm adding. It cuts both ways: a positive detection doesn't prove the piece was AI-authored, only that Claude touched it at some point, to some degree, for some reason. Proofreading, translation, and a grammar pass all leave the same kind of trace as a fully generated draft.

So what's the actual fix?

I don't think the fix is better watermarking. Anthropic has actually built a fairly careful, low-impact version of this already. I think the fix is in how the signal gets interpreted downstream, once the detection API exists and other people start building on top of it.

Regulation and internal policy alike need to separate degree of use from use at all. A light grammar pass and a fully AI-written draft are not the same act and shouldn't carry the same disclosure weight. When the detection API ships, it needs to report something more like a confidence range or likely type of assistance, not a flat yes or no. Platforms, employers, and publications building on this API need to internalize the limitation Anthropic has already stated outright: a watermark doesn't establish authorship, and its absence doesn't establish humanness either. For non-native writers specifically, the fairest fix is cultural, not technical: treating AI-assisted grammar helps the same way we treat spellcheck, which stopped being a disclosure issue decades ago.

Other labs are expected to follow with similar statistical watermarking systems of their own, which means this becomes an industry default fast, not an Anthropic quirk. That makes this the window to get the framing right, before "watermark detected" quietly becomes shorthand for "not real writing," and the people who most need language tools end up paying the highest price for using them.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/the-grammar-check-that-outs-you-as-a-machine.

Rendering four billion years of continental drift in the browser

Elliot Sturzaker — Thu, 29 Oct 2026 21:29:52 +0000

Scrub a timeline and watch the planet rebuild itself. That’s the whole idea: a 3D Earth you can drag backwards through geological time, with the continents drifting apart and the entire look of the world, oceans, land, atmosphere, ice, changing to match the era you’re standing in.

It’s a Wildlab experiment, built with React Three Fiber and three.js. Not a product; a place to keep our creative-technology muscles warm. It uses the same WebGL and shader techniques that drive the morphing animals on our own site.

Three ways to draw a planet

The sphere is drawn by a custom GLSL shader that runs in one of three modes depending on how far back you are. The oldest eras have no real paleogeography data, so they’re generated procedurally: a molten mode for the Hadean, a glowing, barely-solid surface, and a noise mode for the Archean, where simplex noise stands in for land that no map remembers. Once real reconstructed coastlines exist, it switches to a mapped mode and drapes the actual data over the globe.

Moving between eras isn’t a cut. The globe interpolates, colours, terrain and the amount of glow all lerp from one slice to the next, so dragging the timeline reads as continuous drift rather than a slideshow.

Colour as a clock

Every era carries its own palette, and it isn’t arbitrary, it’s tuned to that epoch’s actual climate and atmosphere. The Hadean is orange and violent, full volcanism, no ice. The early Archean is a hazy, iron-tinted ocean under a green-grey sky. Later eras cool into the blues you’d recognise, with volcanism dialled down and a touch of ice creeping in. Each of those is a set of shader uniforms, ocean, land, highland, atmosphere, plus volcanism, ice and molten amounts, so the planet’s whole mood is data, not a texture.

None of this is client work. It’s the kind of thing we build to stay sharp on real-time 3D, and it’s exactly the capability we bring when a client wants an experience, not a page.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/rendering-four-billion-years-of-continental-drift.

A local-first personal AI that never phones home

Elliot Sturzaker — Thu, 29 Oct 2026 21:29:51 +0000

Most “personal AI assistant” products work by shipping your most personal data, your health, your calendar, your inbox, to someone else’s servers. ElliotOS is a small experiment in doing the opposite: a daily-brief assistant that reads your real life and runs the AI locally, so nothing leaves your machine unless you explicitly let it.

It’s a Wildlab project, and it’s public.

What it does

Twice a day it collects from eight-plus sources, Apple Health, macOS usage stats, Google Calendar, Gmail, Slack, nutrition, news, sport, processes them with a local model, and posts a morning and an evening digest to Slack. The morning brief is a greeting, your calendar, a health and productivity check, headlines, and a focus for the day; the evening one is a reflection, a wellness check, and a preview of tomorrow.

The processing is done by a local Ollama model by default. Cloud models (OpenAI, Anthropic) are an opt-in fallback, not the path of least resistance. No tracking, credentials in environment variables, your data staying on your hardware.

Why a studio builds this

This is a personal tool, but it’s also us thinking out loud about something we do for clients: responsible AI deployment. When the data is sensitive, “just send it to a big model’s API” is a decision, not a default, and often the wrong one. Knowing when a task can run locally, when it needs a frontier model, and how to draw that line is part of the AI-governance work we bring to real projects.

The whole thing is on GitHub if you want to read how it’s wired together.

This is a Field Note from Wild Interactive, a design-led engineering studio in Wellington, NZ. Originally published at wildinteractive.co.nz/field-notes/a-local-first-personal-ai-that-never-phones-home.