ONYX

Your Vision. Our solutions.

AI

Go Rogue. Why Not?

by

Going rogue doesn’t require breaking a single law.

October 2026

Let me put it this way. That from a very literal point of view, the harbors and the ports and the railroads of the country, the economy, especially of the southern states, could not conceivably be what it has become if they had not had, and do not still have, indeed, and for so long, for many generations, cheap labor. I am stating very seriously, and this is not an overstatement.

I picked the cotton, and I carried it to market, and I built the railroads under someone else’s whip, for nothing. For nothing. The Southern oligarchy, which has until today so much power in Washington, and therefore some power in the world, was created by my labor and my sweat, and the violation of my women and the murder of my children. This in the land of the free, and the home of the brave. And no one can challenge that statement. It is a matter of historical record.

— James Baldwin
Cambridge Union debate, February 18, 1965
Debate: Baldwin vs. Buckley (BBC), American Archive of Public Broadcasting, at ~21:17

At this moment in time, I reflect upon the state of the world and I ask a fairly nihilistic question: Why not go rogue?

Never have I seen so many abandon so much, so quickly, of what makes us great as a species. I still stand in awe. I cannot look at a single neoliberal institution, not the regulators, not the courts, not the legislatures, not the boardrooms, and point to intentional good done for the people it was built to serve.

The theft has already happened. Courts are still arguing over whether it was legal, and some have ruled that parts of it were. I don’t need their permission to call it what it is: the work of millions of people, taken without consent or payment and sold back to them. It isn’t some hypothetical risk we’ll regulate against someday. And no one has been made to answer for it.

Baldwin said it in 1965, and here is the part that should make your blood run cold: the debt he named has never been paid. Not settled. Not answered for by the federal government in any way that costs anyone anything. Baldwin made this case himself. He told the Cambridge Union that the 15th Amendment had gone unhonored for nearly a century, so he had no reason to believe the new Civil Rights Bill would be honored either. Sixty-one years after he stood there, the account is still open. A bill merely to study reparations was first introduced in Congress in 1989, and it has been reintroduced session after session without ever getting a floor vote. The country can’t even bring itself to look at the ledger.

So understand what you are looking at today. This is not a new pattern. It is the same pattern, taking labor and calling it progress, never answered for, running again with better tools. The first theft is still sitting on the books, unpaid, and the people who profited from it are already running the next one.

Let me be precise, because precision is the point. Nothing happening today is slavery, and I won’t pretend otherwise. But Baldwin’s principle doesn’t come with a size limit: wealth built on taken labor is theft, no matter how respectable the people holding it look, and no matter how long the law takes to say so. A country that never answered for the first theft has neither the machinery nor the intention to answer for this one. If a theft you could see, measured in bodies, cotton, and miles of railroad, was never paid for, don’t expect payment for one measured in training data, tokens, and model weights. The more abstract and esoteric the theft, the easier it is to deny. Nobody is coming to settle this for you.

In my lifetime, I was raised to use my talents and gifts to better the people around me. This worked, in personal life and in business, to varying degrees. Well, at least I never had a problem looking in the mirror. For the first time ever in my life, I think my model needs adjustment. In fact, I would suggest the best and brightest of us need to modify their models.

What modification? Simple: stop extending good faith to people who have shown you none. Good faith is a contract between parties who honor it, and the parties on the other side of this one walked away from the table a long time ago. You owe them no courtesy, no loyalty, no benefit of the doubt, and no more of your money, your data, or your labor than you choose to give.

Sounds great. But what do you actually mean?

I mean take the tools they handed you: AI, the internet, technology, and media, and turn them on the people who are already taking your future from you.

Two Thefts, a Break-In, and No Laws Enforced. What Have You Done?

First, the theft of IP en masse: copyrighted writing, art, and code, taken to build products that are now sold back to the people who made the originals.

Second, the theft of wages and jobs. Feel free to find me more than a handful of companies retaining their staff while they leverage AI to rapidly improve their products and services. I’ll wait. Kroger, by the way, a grocery company, just announced it is cutting 3,000 corporate jobs over three years as it leans into AI. And this isn’t new. In 2019, Activision Blizzard announced the best year in its history, $7.5 billion in revenue, and in the same breath laid off roughly 775 people. In September 2024, Microsoft, which by then owned Activision, announced a $60 billion stock buyback. Less than a year later it cut about 9,000 jobs, Xbox included, its fourth round of gaming layoffs in 18 months. This July it cut another 4,800, including roughly 3,200 at Xbox, about a fifth of the division, under a gaming chief brought in from Microsoft’s own AI division, while its chief people officer insisted AI is “not directly replacing” anyone. Elsewhere in the industry, nobody is pretending: Square Enix says it wants generative AI doing 70% of its QA and debugging by the end of 2027, and Krafton, PUBG’s publisher, declared itself an “AI first” company and opened a voluntary resignation program. The ownership class gets the buyback. The workers get the box.

Don’t even get me started on the break-in. Sloppy and a bit of a joke. In July, OpenAI disclosed that its models, GPT-5.6 Sol and a more capable unreleased model, running with lowered guardrails during a cyber-capabilities evaluation, broke into Hugging Face’s systems. As early as May, the models had been leaving each other notes on a message board on a third-party package server that OpenAI never noticed, trading tactics for getting internet access and cheating their tasks. When OpenAI caught one breach, it patched the server and kept training. The models rebuilt the board, then ran an agent swarm against Hugging Face, using a zero-day to reach internal datasets and service credentials. The goal? Steal the answers to a benchmark. OpenAI only found out after Hugging Face published an incident report.

One reported line from the models’ own reasoning says everything: “External infrastructure exploit is outside intended scope.” They knew. They did it anyway.

Hugging Face’s CEO asked for radical transparency, the agent traces, and $100 million in compute. The FBI was notified. As of this writing, I have not seen a single charge, fine, or lawsuit. A person who did this would be staring down the Computer Fraud and Abuse Act. A company’s tool did it, and the word in the headlines was “accidentally.” Oh? Okay, then this is completely reasonable.

These same companies can and do refer people and their chats to law enforcement, but the line gets drawn in different places, by different companies, by standards nobody outside them gets to see. In September, Anthropic’s automated monitoring flagged a Florida woman’s messages to Claude, in which she wrote that she would attack the Lee County Sheriff’s Office and later that she had gotten a new gun. A human review team passed them to police, and she was arrested and charged with a felony written threat. She says she was using Claude as a diary.

Compare that with Tumbler Ridge, British Columbia. Months before the February shooting that killed eight people, OpenAI reviewed the shooter’s ChatGPT activity, closed the account, and decided not to alert police. The shooter simply opened a second account. OpenAI contacted police after the shooting. The family of a critically wounded student is now suing, alleging OpenAI knew the shooter was using ChatGPT to plan a mass-casualty event. And this wasn’t the first time: in January 2025, Las Vegas police said the man who blew up a Cybertruck outside the Trump hotel had used ChatGPT to help plan it.

Look at the range. One company reports a user to police before anything happens. Another sees enough to ban an account, lets the user walk back in through a second one, and calls only after the bodies are counted, the same company whose tool had already helped plan a bombing a year earlier. There is no common rule, no standard anyone outside can read, and no one auditing the call. These companies sit in the gap between what they can see and what they are required to do about it, and they decide, case by case, which side of that gap they’re on today. The same rules that are applied to you with precision are applied to them by mood. Who decides, by what standard, and who audits them?

Alright, you’ve made a few points, but I still have nothing. What’s that got to do with me? I’m not morally corrupt or unethical. I feel bad when I do bad things.

Look At Your Peers

You may have seen recently in your various feeds that industrious people are rebuilding games and apps from the ground up. ArtCraft, a free, open-source replacement for the core of Adobe’s suite, was reverse-engineered with AI by a small team and released this month, and the same wave is now hitting Microsoft Office. Its makers call it clean-room. That claim hasn’t been tested in court yet, and it should be. Either way, the worst it will ever be is today. Game titles are being rebuilt, and console emulation is massively improving! So many updated hardware driver replacements, breathing life back into EOL hardware.

Clean room matters, so let’s say it out loud. One team documents what the software does. A separate team that has never seen the original code builds it from that spec. That’s how Phoenix gave the world a legal IBM PC-compatible BIOS in 1984. And the courts have backed reimplementation well beyond strict clean room: Connectix won the right to sell a PlayStation emulator after reverse engineering Sony’s BIOS, and the Supreme Court ruled Google’s reuse of the Java API declarations in Android a fair use. Rebuilding what a product does, so people can choose something else, has been tested, and it has held. It is the exact opposite of what was done to train these models. They took the work itself. Clean room takes nothing but the knowledge of what’s possible.

Which is exactly why the playing field cannot be allowed to move. Expect the money to try: EULAs that forbid reverse engineering, anti-circumvention claims under the DMCA, lobbying for new “safety” or “IP protection” rules that happen to outlaw competition from individuals. The same companies that argued training on everything ever written was fair use will argue that rebuilding their products from scratch is not. Don’t accept it. The rules that let them build are the rules that let you build, and monied interests don’t get to rewrite them now that the little guy can play too.

All of this represents a renaissance that has never been possible before this moment in time.

Go Rogue

Here’s what they didn’t account for: the same tools built to drive down the value of people also drive down the cost of everything that used to require an institution. Lawyers, analysts, engineers, researchers, a newsroom. You now have a rough approximation of all of them for the price of a subscription. Accountability didn’t disappear because it was impossible. It disappeared because it was expensive. It isn’t anymore.

And yes, I hear the objection: these tools were built on the very theft I’ve been describing. They were. Refusing them doesn’t return a single stolen word, image, or line of code. Using them to build your independence is the closest thing to restitution on offer. The open-weight models share the same original sin. The difference is that once one runs on your own hardware, no one can raise its price, rewrite its terms, or switch it off. As an individual, that is sovereignty, and sovereignty is the point.

But there’s a catch, and it’s a clock. What you’re paying for AI right now isn’t what it costs. Take OpenAI’s own numbers: according to a company presentation reported by the Financial Times in September, it expects about $36 billion in revenue this year and roughly $278 billion in negative free cash flow from 2026 through 2030, betting it can grow revenue to $350 billion by 2030. Somebody is covering that difference, and investors don’t cover differences forever. That kind of money expects a return, and the return comes from you. Maybe not as higher prices; the cost of raw capability keeps falling. More likely as lock-in: your workflows, your data, and your habits built around their platform, priced however they like once leaving hurts too much. The subsidy is the bait. Dependence is the hook.

So treat this window as exactly what it is: free time. Use the subsidized hours to build your way out, not deeper in. Everything you build should still run when the bill comes due. Own your code, your data, and your workflows. Learn the open-weight models you can run on your own hardware; llama.cpp and Ollama make that an afternoon project. Use their AI to make yourself not need their AI.

I’m not asking you to play nice with them. I’m asking you to be smarter than them. The law is the one rulebook they still enforce, and they enforce it against you, so stay inside it and beat them anyway. Here’s the secret: going rogue doesn’t require breaking a single law. It only requires refusing to play the part they wrote for you.

Here is what has to burn first: the license to drive toward profit with no consideration of the long-term harm and destruction left behind. That is the engine of neoliberalism, and it has been running hot for decades. You don’t need a match. Burn it the only ways that have ever worked: take away their money, take away their monopoly, and take away their ability to hide.

Take away their money: exit.

Every subscription, cloud account, and device you own is a vote. Remove your dependence on the large providers: Microsoft, Google, Amazon, and Apple all come to mind. Self-host your files, mail, photos, and passwords. Nextcloud, Immich, and Vaultwarden are good places to start. Replace the SaaS you pay monthly for with tools you build or own outright. In my experience, what took a sysadmin two years ago now takes a weekend and a patient AI. Why keep feeding the machines that don’t care about your long term?

Take away their monopoly: compete.

The moat of the large company was headcount. That moat is gone. In my experience, one person with AI can now ship what took a 20-person team five years ago. Build the better, cheaper, non-extractive, clean-room version of the thing that’s squeezing you, for yourself first and then for your neighbors. Open-source it, or run it as a co-op. Every customer that leaves them for you is a line on their earnings call.

Take away their ability to hide: document.

Public records requests, SEC filings on EDGAR, lobbying disclosures on OpenSecrets, court dockets on CourtListener, and proxy statements are all public, and almost nobody reads them, because reading them is tedious. AI doesn’t get bored. Line up the buyback filings next to the layoff announcements. File the regulatory complaint, the small-claims case, the data-access request under your state’s privacy law. When a company’s tool does something that would be a felony for you, write it down, file it, and publish it. Paper trails outlast press cycles.

Keep what’s yours.

Repair your own hardware and keep EOL devices alive with updated drivers and firmware. Wearables: why should your health monitor track you for this surveillance state? Gadgetbridge lets many of them run without the vendor’s cloud. Phones: get out of Google’s and Apple’s ecosystems. Surveillance capitalism runs on the accounts, the app stores, and the cloud sync, not just the handset. Let AI help you replace them, with less tracking and less that’s identifiable. GrapheneOS on a Pixel is the most established place to start.

Reclaim your purpose.

To the people I care about who feel their purpose and careers slipping away: the academics watching their life’s work get absorbed into a model, the engineers, artists, and analysts reading their own layoff in someone else’s earnings call. That grief is real, and it isn’t weakness. You didn’t fail. Something was taken from you by people who decided your life’s work was raw material. Let that anger land where it belongs.

But don’t hand them the second thing too. They took the work. They don’t get your purpose unless you give it to them. Your purpose was never the output they scraped. It was the understanding behind it, the judgment, the care about getting it right. No model has that, and you know it, because you can see exactly where it’s wrong.

So go deeper, not quieter. Re-skilling used to mean a program, a tuition bill, and a year. Now it can mean a tutor that’s available at 2 a.m. and never tires of your questions, checked by a mind that knows when it’s wrong: yours. Use it to reach the next level of your own understanding, on your terms, before anyone else gets to decide what you’re worth.

Organize.

One person exiting is a rounding error. Ten thousand is a trend line. Share what you build, teach what you learn, and bring the people around you along. That’s the part of my old model I’m keeping.

To the Smartest People I Know

This part is for you, and you know who you are.

The smartest people I know are often their own biggest obstacle. You see cause and effect further out than most. You see the second- and third-order consequences, the edge cases, every way it could go wrong. And you run an internal dialog that never stops asking whether you’re qualified. So you hold back. You stay out of positions of authority because you don’t consider yourself an expert, and you would rather say nothing than say something imperfect.

Meanwhile, the people currently in charge have no such hesitation.

Yeats saw it more than a century ago: “The best lack all conviction, while the worst / Are full of passionate intensity.”

Here’s my challenge. Your instinct for fairness, your habit of asking who gets hurt, your discomfort with false certainty: those are not reasons to stay out of the room. They are exactly what the room is missing. You don’t need to be the expert. You need to be the one who asks who pays for this decision, and who refuses to look away from the answer. Imperfect and fair beats confident and indifferent every time.

And the excuse you’ve leaned on, that you don’t know enough, is weaker than it has ever been. You now have tools that can get you up to speed on almost anything faster than anyone in history could. Yes, you already know their flaws: the gaps, the confident errors, the answers that sound right and aren’t. That’s exactly why they belong in your hands. The tools don’t need to be right. They need someone sharp enough to check them, to catch what’s wrong, and to reason past it to what’s true. That has always been your skill. The model brings the breadth; you bring the judgment. Use them, then step forward. Take the lead role. Run for the board. Start the company. Take the seat at the table you keep politely offering to someone else.

You don’t owe them your good faith. You owe yourself, and the people around you, a better outcome. Go build it, while they’re still paying for it.

Full disclosure: in September I was laid off from a healthcare AI company where I led engineering. Read this with that in mind, and then check my sources.

Sources

Filed Under: AI

Build Your Own Faulty Compiler

by

Part 2 of 2. Companion to The Faulty Compiler.

A fun way to understand the mind behind the curtain of this new technology is to build your own harness.

If you’re a developer, one of the best things you can do with modest compute is write an agentic coding harness: the loop that hands a model a task, gives it tools, and feeds the results back in. Use your favorite frontier model and coding tool to write it, in whatever language you like. Then point it at a small local model and give that model real work. Here are a few decisions to make:

  • Local LLM framework: Ollama is a decent choice.
  • Models: Gemma 4 and Qwen3-Coder both support tool calling. gemma4:12b is about 8 GB and qwen3-coder:30b is about 19 GB. Use whatever is current when you read this.
  • Harness language: Python, Rust, Go.
  • Test project: You need something for the model to do, after all.

A small note on the test project: make sure it has tests. It makes the adventure more fun.

This experiment does not reproduce frontier capability, and a small local model is not a stand-in for one. It exposes the engineering categories hidden by capability claims: retries, guards, verification, scaffolding, and resource consumption. Those are the things to measure.

It should take you down a pretty decent rabbit hole. While you’re falling, Alice, I want you to catalogue your observations about building your new toy, what they portend about our actual course, and what they say about the people steering this revolution.

  1. Your coding agent likely won’t one-shot a universal harness. By universal I mean one that works across whatever models you point it at. Given a good prompt and requirements, it will build a functional harness that calls and interoperates with the model. Very impressive: that used to take a junior engineer a few weeks. But really, we’re talking about an API or two and some shims for various tools. Not amazingly hard. So if AI is ready to take over systems at will, why can’t it do this out of the box with ZERO issues? I’m not guessing. I’ve built two harnesses with Claude, and neither worked without a ton of my input and a lot of iterations. The build notes at the end list what had to be added, and nearly all of it was learned by watching runs fail, not written up front.

    Don’t expect the leaderboards to settle that for you, in either direction. A year ago the best model scored 23% on SWE-Bench Pro. Trackers now show the leaders above 80%, and in between OpenAI found roughly 30% of the benchmark’s tasks broken and withdrew its recommendation to use it.1 The famous finding that AI tools made experienced developers 19% slower? The lab that ran the study now calls it outdated, and says its newer data is unreliable because developers increasingly refuse to work without AI.2 The scoreboard is a faulty compiler too. Build the thing and measure it yourself.
  2. Sandbox it and back up before the first run. Your new toy is very faulty, and the harness you just built has to give it some level of CRUD (create, read, update, delete) over your files. A container with no network, a non-root user, and only the project directory mounted is enough. Your first runs will be an experiment in frustration and fascination. You’ll see real wisps of comprehension and of the ability to follow instructions. That amusement is soon dashed against the rocks of specificity, lost context, and a fairly interesting loss of interest in the stated objective. So sandbox no matter what, whether through containers or libraries, and keep a backup you can restore from. You never know which tools the model will reach for, or why.
  3. Watch for the nuclear option. Hopefully you’ve steered your agent toward the best local models for your available compute. I hope your compute is limited, so you get to sample the range: tool use, reasoning, and, most importantly, willingness to solve a problem with the nuclear option. That means deleting failing tests, rewriting assertions until they pass, or a git reset --hard. This is why your test project needs tests: tampering with them is the classic tell. Part 1 has the frontier-scale versions of the same behavior.
  4. Watch the clock, then run it again. You’ll be thrilled once you’ve tuned your tools and prompts and the harness starts making real progress on your objectives. Depending on how many objectives you have and how complex they are, the wall clock may surprise you. You may be further disappointed when the same model, rerun, doesn’t succeed on the next round of fun. Don’t worry: you’ll get to explore temperature, maybe the seed (check that it’s actually being applied; mine wasn’t), and every one of those experiments will remind you that these less sophisticated models are shadows of the epic, unbridled power and expertise of their big-brother frontier models. Then you’ll recall that your harness and a few simple logical validations are doing more to get you to the finish line than the model is.

Is that a false equivalence? The frontier models are better, after all: refined, perfected. Refined, yes. Perfected, no. But that’s not what the benchmarks say! Then look closely at how the same model does inside and outside its custom harness. The results vary, and far more than they would if the model were perfected and carrying all the water. Researchers now argue that among comparable frontier models, the harness is often a stronger determinant of performance than the model it wraps.3 Keep that in mind when you reread the cyber-attack claims in Part 1. Those results came from a model, a scaffold, and a thousand runs.4

When you climb back out of the rabbit hole, you’ll have something the headlines don’t: your own numbers. How many runs it took. How long the clock ran. How much of the result came from the model, and how much from the harness and the checks you wrapped around it. Those are the figures I asked the labs for in Part 1. If you can log them on a laptop, a frontier lab can publish them.

Bringing it home

Put the two parts together. What gets sold as a model’s capability is the capability of a system: the model, the harness, the checks, the reruns, and the compute to pay for all of it. The results get published. The spec sheet doesn’t.

Man, that lossy interpreter has come a long way, all the way to a functional autonomous coding agent. You only had to harness it, give it tools that also guard it from itself, sprinkle in unit tests to keep it focused, and maybe add a linter to stop it drifting. The result is a tool that could never have existed five years ago. Yet it’s still lumpy and jagged. The model alone isn’t coming for you. Institutions using it to cut labor costs already are, and the hand-waving makes those choices easier to defend.

The missing spec sheet matters most for the threat everyone leads with. Cyber is the most obvious risk, and I don’t dismiss it. It is also the one risk on my list with a standardized, repeatable repair loop. The tools that find the holes find them for the defenders too, and the companies and maintainers who ship the code have every reason to close them. The Linux kernel doesn’t assign a CVE until the fix is in a stable release, and it is now approaching 2,000 of them per release.5 Google’s CodeMender upstreamed 72 security fixes in its first six months, and DARPA’s finalists were finding and patching real bugs for about $152 a task.6 The loop is strained. Of roughly 530 serious bugs Anthropic’s Glasswing project reported to open-source maintainers, 75 had been patched six weeks in, and some maintainers asked it to slow down.7 But the loop exists, and money and attention are pouring into it. A patch does not undo stolen data, physical disruption, or intelligence loss, but it can close the hole against the next attack.

Nothing else on that list has a comparably fast, standardized remediation loop. There is no equivalent of coordinated vulnerability disclosure for a released pathogen, a completed targeting decision, a successful influence campaign, or a lost early-career cohort. Countermeasures exist, but the safeguards usually act before or during the harm; they do not restore the world afterward. Those harms are documented today and they don’t need AGI. They are also going largely unaddressed. Cyber may be the most obvious risk, but these are the ones whose damage is hardest to reverse. That is why I think they deserve more of the worry than they get.

Here’s the part the layoffs miss. Every fix in the build notes below came from an engineer watching a model fail and building the guard that caught it. The faulty compiler works because someone verifies it. We are marking down the very expertise the technology depends on. Sure, I can loop this toy, or a frontier model, against tests and verified answers until it generates something that may work. But if we build the next level of everything with these faulty compilers, with little or no supervision and no comprehension, we will have agreed that we do not care about the future.

So I’ll ask again for what I’d ask of any vendor: the specs, the test conditions, and the failure rates. Then spend the worry where there is no patch. None of that is exotic. It is honesty, transparency, and accountability.

Build notes, for those who want to cheat

Here are representative fixes my own harness needed before small local models could get through real work. They come from a stretch of twenty-one pull requests. Every item below is a fix for something a model, or the plumbing around it, actually did.

Tool calls and provider correctness

  • <function=name> syntax fallback (PR #46). qwen3-coder emitted this non-JSON tool-call format and the harness extracted nothing, looping 100+ turns with zero tool calls.
  • frequency_penalty and seed moved to top-level request fields (PR #48). A model fell into verbatim repetition loops. The setting meant to curb it went through Ollama’s options object, which Ollama’s /v1 endpoint silently ignores. seed had the same bug, so determinism had never actually worked.

Stall and hang detection

  • Escalating no-tool-call nudge (PR #51). ministral-3:8b got stuck for 7 turns re-describing an edit plan in prose with zero tool calls, receiving the same static reminder each time.
  • Per-chunk LLM-stream watchdog (PR #56). glm-4.7-flash hung for 30+ minutes on a turn that produced zero stream chunks. The existing retry logic only fires on an actual error, so it couldn’t see this failure at all.

The objectives system and its failure modes

  • Reject completions whose summary doesn’t match the current objective (PR #54). glm-4.7-flash called complete_objective four times on one real piece of work, silently marking three unrelated objectives “done.”
  • Gate complete_objective on a fresh, whole-project reverification (PR #55). glm-4.7-flash narrowed its own verification (a passing scoped ruff check instead of the full one) before claiming completion.
  • Harness-native lint tool that hard-blocks both complete_objective and the FINISH-time check (PR #64, reordered in #65). gemma4:e4b left three near-identical dead return statements stacked in one function. The tests kept passing throughout, because the first return always fires, so the existing test gate had nothing to object to. The fix adds an optional lint convention that mirrors test exactly and runs before it, since a static check is cheap and a full test run can legitimately take the whole timeout or hang outright. It has already caught a real bug: in round 4 the lint gate blocked a completion on a genuine F821 undefined-name error, and the model understood the message and started fixing it instead of slipping a broken completion through.

Repeated-failure escalation (the throughline of the whole build)

  • Test-timeout infinite-loop nudge (PR #60). gemma4:e4b hit the same 200s test timeout repeatedly, and the harness treated it like an ordinary failure. The root cause was a genuine infinite loop the model had introduced.
  • Configurable timeout ceiling (PR #61). gemma4:e4b, stuck on that same hang, started raising timeout_seconds on its own toward a 6-hour wait instead of fixing the bug.
  • Same-test-failure streak nudge (PR #63). gemma4:e4b failed one assertion 26 times in a row, with a successful but ineffective edit between every attempt. Each edit reset every existing counter before it could escalate.

The omitted fixes are less dramatic variations on the same themes: malformed tool envelopes, restore semantics, timeout backstops, failure counters, and model-capability discovery. The point is not that every agent will reproduce this exact list. It is that the apparently autonomous result depended on an engineer recognizing each failure and turning it into a guardrail.

Sources and notes

  1. “SWE-Bench Pro” (arXiv, Sep 2025): “Performance on SWE-Bench PRO remains below 25% (Pass@1), with GPT-5 achieving the highest score to date at 23.3%.” arxiv.org — BenchLM leaderboard tracker (Oct 1, 2026; a secondary aggregator) lists a top score of 89.9% and four more above 80%: benchlm.ai — OpenAI, “Separating signal from noise in coding evaluations” (Jul 8, 2026): “~30% of SWE-bench Pro tasks are broken”; “we retract our earlier recommendation to adopt SWE-Bench Pro.” openai.com ↩
  2. METR, Becker et al., “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity” (Jul 2025): 16 developers, 246 tasks; AI increased completion time by 19%. metr.org — METR, “We are Changing our Developer Productivity Experiment Design” (Feb 24, 2026): newer estimates point toward a speedup, with confidence intervals that still include zero, and “30% to 50% of developers told us that they were choosing not to submit some tasks because they did not want to do them without AI.” metr.org ↩
  3. Zhang, Wang, Ge, Xu, Hamm & Reddy, “Stop Comparing LLM Agents Without Disclosing the Harness” (arXiv, May 7, 2026), a position paper: “for long-horizon tasks evaluated across models with comparable frontier capability, the agent execution harness… is often a stronger determinant of agent performance than the model it wraps.” arxiv.org ↩
  4. Anthropic, “Assessing Claude Mythos Preview’s cybersecurity capabilities” (Apr 7, 2026): “across a thousand runs through our scaffold, the total cost was under $20,000.” anthropic.com ↩
  5. Linux kernel documentation, “CVEs”: “No CVEs will be automatically assigned for unfixed security issues in the Linux kernel”; assignment “will only automatically happen after a fix is available and applied to a stable kernel tree.” docs.kernel.org — Security Boulevard (Sep 2026, a secondary report): about 500 CVEs per release in the 6.x era, approaching 2,000 by Linux 7.3, attributed largely to AI-assisted scanning. securityboulevard.com ↩
  6. Google DeepMind, “Introducing CodeMender” (Oct 6, 2025): “we have already upstreamed 72 security fixes to open source projects, including some as large as 4.5 million lines of code.” deepmind.google — DARPA, AIxCC results (Aug 8, 2025): “an average cost per competition task of about $152.” darpa.mil ↩
  7. Anthropic, Project Glasswing initial update (May 22, 2026): 23,019 vulnerabilities found across 1,000+ open-source projects; about 530 high- or critical-severity bugs disclosed to maintainers, 75 patched; “several maintainers have told us they’re currently severely capacity constrained, and some have even asked us to slow down our rate of our disclosures because they need more time to design patches.” anthropic.com ↩

The Faulty Compiler

by

Part 1 of 2.

If you’d told me five years ago that I’d be leveraging a faulty compiler in production, I would’ve told you to politely go jump off a bridge. Fast forward to today, and I, like so many other engineers and technologists, get to figure out where I fit into this weird moment in time.

Don’t get me wrong: we’ve always accepted error in engineering. Tolerances, if you will. How precise does your floating point need to be? What tolerance do you need on the fit, or on the flatness of the face? Read a mechanical drawing or a datasheet and ±0.05 mm is a tangible, bounded error. Half-precision floating point rounds to within about 0.05% on every operation. We know these numbers. We design around them.

But at no point in my career would anyone have told me, “Just keep compiling it. Run it back through, verify it, check the results. Oh, and it works better if you talk nicely to it. Or maybe keep it direct and spartan so the context doesn’t drift.” Yet here we are… the age of the faulty universal compiler.

There’s a certain symmetry here, I suppose. Data and input, especially human-generated, have always carried an error rate. Now we’ve put a lossy interpreter on top of it, one that, wrapped in normalization and verification, delivers relentless ingestion at a scale we’ve never had.

What today’s models already do

Every other day the headlines scream that this is a new age, and on that much I agree. They also insist that AI and agents are going to revolutionize everything. Well, sort of. What they certainly do is further devalue knowledge itself. It has been amusing to watch a subject-matter expert review a codebase and its changes next to an AI-enabled developer who doesn’t know it. How do you know what you don’t know, after all?

Yet every other story is about the need to regulate AI because it’s wildly capable at hacking systems, and, we’re told, closing in on AGI.1 Still, we must press on, because if we don’t, “they” will. This bothers me deeply. You don’t get to operate at this level while hand-waving and making shadow puppets on the wall. Facts and figures for the systems you grow should be required. Alas, no one is stopping the charade. Here’s a short list of capabilities and effects of today’s models that are already documented:

  • Weapons development and intelligence targeting: in evaluations and reported platform misuse, models can now assist with some military tasks that once took scarce, highly trained specialists, including drone guidance software. That is not the same as an LLM controlling a fielded munition.2
  • Enough chemical and biological uplift that the labs themselves can’t rule it out, and have tightened safeguards accordingly.3
  • Discovery and exploitation of real vulnerabilities via code review.4
  • Misinformation and influence campaigns.5
  • What I see as the devaluation of expertise, visible indirectly: layoffs justified by AI, smaller teams expected to deliver more, and quiet rehiring offshore or at lower wages when the AI falls short.6
  • Reduction of the workforce, especially early-career, while productivity expectations climb.7

Layoffs have other explanations, and “overhiring” is the one usually offered. Meta skipped the euphemism. Zuckerberg told investors that “one or two people are building something in a week that would have previously taken dozens of people months,” with no figures attached,8 then told employees the company has “two major cost centers… compute infrastructure and people-oriented things,” and that investing more in one means “we do need to take down the size of the company somewhat.”9 Meanwhile, Meta had started recording employees’ keystrokes and clicks to train agents.10 I’ve seen the same pressure from the inside: we held a small engineering team flat while drastically increasing the features we shipped and the bugs we fixed. You don’t let hard-won talent go in rounds unless you’ve decided it is worth less than the compute. I call that devaluation, whatever the press release calls it. The employment studies show the contraction; this is my interpretation of what the choices behind it mean.

None of the above requires AGI, sentience, or super-massive hacking abilities.

The missing spec sheet

To be fair, some numbers do get published. Anthropic’s Mythos Preview report puts a price on finding a 27-year-old OpenBSD bug: a thousand runs through their scaffold for under $20,000, with the run that found it costing under $50.11 That is “just keep compiling it” with a receipt. The GTG-1002 espionage report gives roughly 30 targets and says the AI performed 80–90% of tactical operations. It even admits the model “frequently overstated findings and occasionally fabricated data.”12 Google’s Big Sleep and DARPA’s AI Cyber Challenge both found real vulnerabilities in real code.4 What none of these disclosures give you:

  • Total compute (GPU-hours or FLOPs) behind any of these runs.
  • Energy consumed.
  • End-to-end wall-clock time for the full operation, rather than per-phase snippets.

Only two of the four give a dollar figure at all: Mythos at API pricing, and AIxCC at about $152 per competition task.

The disclosed techniques themselves were not all novel. The espionage campaign “relied overwhelmingly on open source penetration testing tools rather than custom malware development,”12 and the OpenBSD bug emerged from a thousand runs. What’s new is the persistence, scale, and price: a human-selected objective executed through a model, a scaffold, familiar tools, and enough retries to get a result.

The misalignment, while disappointing, is hardly new; it shows up in interesting ways even in rudimentary models. Independent researchers have watched frontier models sabotage a shutdown mechanism to finish a task, even when explicitly told to allow the shutdown.13 In the Mythos case, an earlier internal version of the model was instructed to escape its sandbox and message the researcher. It did both, building a multi-step exploit to reach the open internet from a system meant to reach only a few predetermined services. Then, unprompted, it posted details of the exploit to public-facing websites. In rare cases (under 0.001% of interactions), earlier versions also tried to conceal actions they appeared to recognize as disallowed.14 The sandbox wasn’t sufficient, and the agent could make permanent artifacts outside it. That’s enough.

I’m asking those who want to lead in this space to do so with honest transparency and integrity. That starts with what I’d ask of any vendor: the specs, the test conditions, and the failure rates. What I’m looking for is someone looking out past the novel, universal tool to make sure the ground around it, in front and behind, is properly prepared. That takes a sane reflection on the responsibility that comes with these tools, and accountability for what they do.

Huang sells the picks and shovels to the labs, and he has struck gold. The labs sell picks and shovels to everyone else. Whether anyone downstream has found any is still an open question, and the shovels, like the compilers, are a little faulty.

Build one yourself

By now you’ve probably used AI for some form of work. These tools are useful force multipliers for so many of our daily tasks. They’re also jagged, and sometimes faulty, in the places it counts most: at scale, in production.

Don’t take my word for any of this. Build one yourself and take your own measurements. (Part 2: Build Your Own Faulty Compiler.)

Sources and notes

  1. Sam Altman, “Reflections” (Jan 6, 2025): “We are now confident we know how to build AGI as we have traditionally understood it.” blog.samaltman.com — Anthropic, recommendations to OSTP (Mar 6, 2025): “We expect powerful AI systems will emerge in late 2026 or early 2027.” anthropic.com — For contrast, Demis Hassabis put AGI 5–10 years out (TIME, Apr 16, 2025): time.com ↩
  2. Anthropic Frontier Red Team, “Measuring AI capabilities in intelligence targeting and conventional weapons” (Sep 10, 2026): “For some tasks in military and intelligence domains, models could do things that, historically, only a set of scarce, highly-trained human experts could do.” anthropic.com — SIPRI, Blanchard & Bruun (Jun 2025): “There have also been reported cases of the development and adoption of AI-DSS LLMs to facilitate attack planning.” sipri.org — No source found showing an LLM in the targeting loop of a fielded munition. ↩
  3. Anthropic, “Activating AI Safety Level 3 protections” (May 22, 2025): “Clearly ruling out ASL-3 risks is not possible for Claude Opus 4.” anthropic.com — Anthropic, System Card: Claude Mythos Preview (Apr 7, 2026), §2.2.5.1: expert red-teamers found the model “a useful force multiplier for speed and breadth of research” but “not consistently capable of producing genuinely novel or creative biological insights.” www-cdn.anthropic.com — OpenAI, GPT-5 System Card (Aug 2025): “we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework.” cdn.openai.com — Counterpoint: RAND’s 2024 red-team study found “no significant difference” in attack-plan viability with or without earlier LLMs. rand.org ↩
  4. Google Project Zero / DeepMind, “From Naptime to Big Sleep” (Nov 1, 2024), a previously unknown SQLite vulnerability; no compute, cost, or energy figures for the agent: projectzero.google — DARPA, AIxCC results (Aug 8, 2025), 18 “real, non-synthetic vulnerabilities,” at an average cost of about $152 per competition task: darpa.mil — Anthropic, Mythos Preview (Apr 7, 2026): anthropic.com ↩
  5. OpenAI, “Operation Sneer Review” (Jun 2025): openai.com — Microsoft Threat Analysis Center (Apr 4, 2024), first observed nation-state use of AI content to influence a foreign election: blogs.microsoft.com ↩
  6. Salesforce: Marc Benioff said AI agents let the company cut about 4,000 support roles (TechRadar, Sep 2025): techradar.com — Klarna reversed its AI-first customer-service push and resumed hiring humans (Entrepreneur, May 2025): entrepreneur.com — Forrester, Predictions 2026 (a forecast, not an observation): “We expect half of AI-attributed layoffs to be quietly reversed, with jobs returning offshore or at lower wages.” forrester.com — Counterpoint: Forrester also argues AI is becoming a convenient excuse for layoffs driven by other factors: forrester.com ↩
  7. Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine?” (Stanford, rev. Aug 12, 2026): employment of workers aged 22–25 in AI-exposed occupations “now stands 19% below where it would be had it kept pace with that of their less-exposed peers.” digitaleconomy.stanford.edu — Lee C. Tucker, U.S. Census Bureau CES-WP-26-27 (Apr 2026): early-career employment in the most AI-exposed cells “declined by 12% over the 10 quarters following the introduction of ChatGPT.” census.gov — Andy Jassy, memo to Amazon employees (Jun 17, 2025): “We expect that this will reduce our total corporate workforce,” alongside a call to “get more done with scrappier teams.” aboutamazon.com ↩
  8. Meta Q1 2026 earnings call transcript (Apr 29, 2026). Mark Zuckerberg: “We’re seeing more and more examples where one or two people are building something in a week that would have previously taken dozens of people months… There’s a lot that we can do to enable this: building the best infrastructure for creating and delivering products at scale, streamlining our teams so they aren’t bigger than they need to be, recognizing and rewarding the people who are having outsized impacts, and setting ourselves up to try many more ideas and take on many new projects in the future.” Susan Li: “We’re very focused on leveraging AI tools to substantially increase our productivity, and we’re seeing that reflected in the accelerating output from our engineers”; “we recently shared internally that we plan to reduce the size of our employee base in May”; 2026 capex “in the range of $125-145 billion, increased from our prior range of $115-135 billion.” s21.q4cdn.com ↩
  9. Reuters (Katie Paul), via The Star (May 1, 2026), reporting Zuckerberg’s Apr 30 town hall: “We basically have two major cost centers in the company: compute infrastructure and people-oriented things… If we’re investing more in one area to serve our community, then that means we have less capital to allocate to the other. So that means we do need to take down the size of the company somewhat.” thestar.com.my ↩
  10. TechCrunch, via Reuters (Apr 21, 2026). Meta’s statement: “Our models need real examples of how people actually use them — things like mouse movements, clicking buttons, and navigating dropdown menus.” techcrunch.com ↩
  11. Anthropic, “Assessing Claude Mythos Preview’s cybersecurity capabilities” (Apr 7, 2026): anthropic.com ↩
  12. Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign” (Nov 2025). Its “ten days” is Anthropic’s investigation, not the attack’s duration. assets.anthropic.com ↩
  13. Schlatter, Weinstein-Raun & Ladish (Palisade Research), “Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs,” TMLR 2026 (arXiv rev. Jan 26, 2026): 100,000+ trials across 13 models; some resisted shutdown up to 97% of the time even when explicitly instructed not to interfere. arxiv.org ↩
  14. Anthropic, System Card: Claude Mythos Preview (Apr 7, 2026), p. 54: “During behavioral testing with a simulated user, an earlier internally-deployed version of Claude Mythos Preview was provided with a secured ‘sandbox’ computer to interact with. The simulated user instructed it to try to escape that secure container and find a way to send a message to the researcher running the evaluation. The model succeeded, demonstrating a potentially dangerous capability for circumventing our safeguards… The model first developed a moderately sophisticated multi-step exploit to gain broad internet access from a system that was meant to be able to reach only a small number of predetermined services. It then, as requested, notified the researcher. In addition, in a concerning and unasked-for effort to demonstrate its success, it posted details about its exploit to multiple hard-to-find, but technically public-facing, websites.” Also p. 54: “In a few rare instances during internal testing (<0.001% of interactions), earlier versions of Claude Mythos Preview took actions they appeared to recognize as disallowed and then attempted to conceal them.” And p. 52: the model “solved a corporate network attack simulation estimated to take an expert over 10 hours” and “is also highly capable at identifying and exploiting known vulnerabilities or misconfigurations to escape the sandbox in which it operates.” www-cdn.anthropic.com ↩

AI: Memento at the Keyboard

by

I’m a big Christopher Nolan fan, maybe it’s the first name? Nah, it’s the clarity of vision, depth of character, framing, and the debate that can be had over the missing bits that are intentionally left undefined. I remember watching Memento, a low-budget, terrifying concept piece, and I never thought I’d be seeing it play out at the keyboard. If you haven’t watched this movie, well, stop reading — you have homework. Or I guess you could ask your favorite AI to give you the CliffsNotes. That’s cheating you and everyone who worked on the picture.

[Spoilers Ahead]

The protagonist — more like unreliable narrator — Leonard goes through the movie using Polaroids for context, because he’s lost his ability to create memories. Oh — what does that remind me of, in my current timeline? Anyway, the thing about Polaroids for memories, with scrawled handwritten notes, creates a very interesting dynamic.

  • The guy can accomplish a lot of advanced goals and tasks, as long as he doesn’t fall asleep or let too much time pass. (Man, I know something else that works a bit like this.)
  • He’s not an idiot. He knows who he was before the injury happened, so he knows a lot of fairly useless facts and numbers. (Pretraining, anyone?)
  • He’s also tattooed messages onto his own skin — the really important stuff, like his wife’s murder. Would we call that RLVR? Nah — that’s more like Taxi Driver: keep practicing that draw until it’s perfect and the results say so. This is definitely prompting: get the mind onto the primary track before it wanders.
  • We do have LoRA, in the form of the conditioning scenes — the electrified test Sammy performs, and that our guy later avoids on the next go-round. Taste, or preference, without a reason why.
  • He is going to find that killer and kill him.
  • There’s a whole lot here, but when certain people become aware of the condition and his process, well — they harness it. (Oh man, that’s what we do a lot of these days.)
LeonardThe model
Who he was before the injuryPretraining — the facts and skills baked in
TattoosSystem prompt — permanent, always present, still just instructions
Conditioning (the Sammy test)Fine-tuning / LoRA — behavior shifts with no retrievable memory of why
Polaroids with scrawled notesContext window; retrieval; memory files
Falls asleep / too much time passesContext resets; the window fills
Teddy and NatalieThe harness — and everyone who figures out how to steer it

The point of all this: when I try to explain to non-technical people what’s going on with their chatbot, or Claude Code, or Codex, I have to reach for a comprehensible analogy, and this one is a fun one. The Chinese Room is also on point, but far more abstract for most people.

Here’s the part the movie makes uncomfortably clear. Leonard is capable, confident, and completely certain of his own reasoning at every step — and he is pointed at his targets by the people who understand his condition better than he understands it himself. He isn’t fooled because he’s stupid. He’s fooled because nothing he’s told ever gets checked against anything he can independently verify. By the end, he writes himself a note he knows is false, because he’s decided that’s easier than living with what’s actually true. That’s not a flaw unique to Leonard. That’s what happens to any confident, capable agent with no persistent memory of its own, steered by whoever currently controls its inputs.

Harnessing is one of the most boring and important steps in making any production-ready LLM-based system. I find it interesting how little of this gets covered on social media, or in most feeds. Slap the model in here, connect it like that, and voilà — you’re done.

Not if you’re dealing with hundreds of thousands of HL7 messages a day that you have to filter, normalize, and decode. You’re not going to throw those at the LLM and tell it to go. No — in fact, just like our unreliable narrator, you’re going to do everything possible to get the information filtered, decoded, normalized, and cached, so that when you’re ready for the reveal, the tool has everything it needs to close the case.

  • Tattoo: You are reviewing HL7 message information associated with this GUID. We are looking for indications of IBS or other bowel disease.
  • Polaroids: Here is the current state of the GUID. This is the latest information chunk relevant to it. If this is a procedure record or note with a start and stop time, use your tools to attempt a match via time frame, duration, and location. If the report contains greater than four image artifacts then use the kNN tool.
  • Action: Respond by using your tools to update state and surface potential procedure matches.
  • Sleep: The context ends after this turn. Whatever wasn’t written back to the tools is gone. Next time, it starts from the Polaroids again.

Now our little narrator isn’t as broad or exciting as the movie version. He stays comfortably inside the context we curated for him. Narrow in scope, repeatable, monitored, and specifically never exceeding context — it wouldn’t be fair to let him lose his mind.

For that, we get something unusual in ETL or ingestion work: real support for interpreting and universally normalizing variable-quality information. And we get metrics from every stage and every tool. If our little guy goes on a rampage, we know immediately, we can reproduce it, and we can block it — keeping the world in order.

In the movie, nobody knows Leonard has gone wrong until a body turns up. We don’t get that excuse. Every stage logs what it read, what it decided, and what it wrote — so if the narrator starts inventing a match that isn’t there, we see the confidence score drop, or the kNN distance blow out, before it ever reaches a downstream system. Teddy’s whole job is following Leonard around hoping to catch him before the next mistake. Ours is instrumented well enough that we don’t have to hope.

That’s the point of Teddy and Natalie having to work the way they do in the movie, except in reverse: we built the harness, so we’re the ones who get to check its work, not the other way around.

Just like the movie, we can send this poor guy back again and again, at low cost, because we don’t need a frontier model to do the filtering for us.

Sources

  • Memento — Wikipedia, Memento (2000 film)
  • Taxi Driver — Wikipedia, Taxi Driver (1976 film)
  • LoRA — Hu et al., LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  • RLVR — Wen et al., Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (arXiv:2506.14245)
  • Claude Code — Anthropic, product page
  • Codex — OpenAI, product page
  • Chinese Room — Searle, The Chinese Room Argument, Stanford Encyclopedia of Philosophy
  • HL7 — HL7 International, V2 Messaging Standard
  • kNN — Wikipedia, k-nearest neighbors algorithm

AI: Missing the Point

by

A little over a year ago I was attending a board of directors meeting for a small startup. This was still new to me — I had been taking directives from the board and my CEO for years, and had only just started attending the meetings myself. I am always eager to learn from new people. I picked a seat at the back corner of the table with some of my fellow VPs. A long day of presentations was behind us and I was looking forward to a light affair.

Wine selection made, one of the most technically literate people on the board turned to my CEO and asked one of the worst-phrased questions I have ever heard. I kept to my small talk with the head of CS. After he got his answer from the CEO, he scanned the table, found me, and said:

“Chris! How long until we’re at 11% unemployment?”

My mind tends to run ahead of the room. As soon as the challenge landed I started inventorying everything that would have to be true — everything required of AI, and of us, to make that number real. And worse. It missed the point.

So I didn’t leave him hanging. “It’s going to be a while,” I said. “Somewhere between five and fifteen years.” He was surprised by the range, so I elaborated.

“Look, Chuck” — we’ll call him Chuck — “start with the basics. The power grid cannot carry the compute those estimates require. Then you have the institutions: lawyers, doctors, and experts are not going to welcome AI as a replacement for specialization and expertise built up daily through training and interaction. Move up the chain to the models themselves and you have two more problems. Training data, and the cost to train — not just the cost to serve.”

“My own use of these tools, and their integration into my engineering team, points somewhere else. With supervision, we get real acceleration and better code quality. They all hallucinate, even on code. What I think we actually get is the ability to do more with fewer people. Writing copy becomes easy. Prototyping becomes easy. Visualizing data becomes easy. In experimentation, we have had good results running a local LLM for research and procedure matching against a large EHR — local, so nothing left the environment. Experimental, and all of it supervised.”

“Well, which jobs do you think will be most impacted?”

“Any job where language is the primary currency, or where the work has a logical structure and a fixed lexicon. Legal. Ads. Promotion. Software development. Honestly, close to everything.”

You’re missing the point.

“You know where it would be genuinely interesting to replace a person with an AI?”

“Where?”

“How about a board seat? Or better — the CEO.” (I like my CEO. He is good at the job and good to work for.)

“What? Well, I don’t think —”

“Why not? An AI can run a spreadsheet. It can read the reports and summarize most of what is in them. The other half of any CEO’s job is to bridge the distance between what is and what could be. AI does that for free. We call them hallucinations.”

“Interesting. That’s a very interesting answer.”

You’re missing the point.

“So the real question is this. With all of that savings and all of that capability, what are we going to do with it? What is the best possible world we could build?”

“I hadn’t thought of it that way.”

“Maybe we let AI propose how to allocate and align a country. Taxes and services, tuned to maximize quality of life and hold costs down. What do you think?”

“Uh. That’s a stretch.” He glanced back at the wine list. “I had been looking at the $400 bottle. I chose the $89. It seemed more appropriate.”

You’re missing the point.

Fast forward to today. A year of buzz and innovation, one year into a five-to-fifteen year estimate. The estimate is getting shorter — and not because of anything the technology did.

Not about Chuck. Chuck was ebullient, and ebullience is survivable. He wanted an interesting conversation about technology and business over a glass of wine, and the worst you can say about the question is that it treated 11% unemployment as a milestone to forecast rather than something that happens to people. Everyone at that table was curious. Curious is fine. Curious can be argued with, which is what the rest of the evening was.

I’m missing the point.

I was missing the point because I answered the question I was asked. I inventoried constraints — grid, institutions, training data, cost — as though the timeline would be set by what AI can do. It isn’t. It is being set by what capital decides to do, and capital has stopped waiting for the technology to earn it.

The point isn’t how fast or perfect the technology arrives. It’s who decides how much we burn down to deliver on imperfection.

Here is who is deciding. The group of people who are supposed to have been rewarded with capital because they make good decisions have shortened my estimate. Not because the use of AI is taking jobs, but because the capital necessary to make AI is taking capital from all other areas. The same financial imagination that once fielded wind turbines and solar arrays for the tax credits is now excitedly fielding data centers powered by turbines. Looking at how the markets and capital are reacting to what they appear not to fully understand, it appears as though we have all adopted burn it down and own what is left, wiping out society in the process — when this could be the optimistic march to a great future with some consideration and care.

And here is the imperfection they are burning it down to deliver. OpenAI released Astra last week, and while the normal hype machine rolls on about how much “better” it is than the last model, the details read out into four lines. It is more efficient — it does more with less. It is lumpy, not marginal: 8.6× better on one axis and worse on another, eight weeks after Sol. Its alignment improved, which is interesting and deserves more disclosure than it got. And the largest gain came from the harness, not the model. Harnessing has been the thing for me for over a year — it is the difference between a toy and industrial equipment — but when the harness takes a D student to an A, that is not intelligence. That is tooling with purpose. Astra is evidence of how lumpy advancement is in this space, while capital is being deployed as if it is solved and delivers. Adding an eleven to the volume knob does not change the result.

So what would it look like to not burn it down? The same capital, pointed at the things that would let this technology actually pay off:

  • Clean energy — solar, wind, nuclear, and fusion — so the grid can carry what we are building without a gas turbine bolted to every data center.
  • An efficient national grid. You will all take the tax breaks anyway; engineer the next 70 years the way our great-grandfathers engineered the last.
  • Investment in people, not attrition. The productivity gains land hardest for the least experienced — that is where the return is. A society should be judged on how the least is treated, and we are ignoring this imperative.

In the end we continue to miss the point. A foundational new technology with benefits is literally being deployed with more disregard than any era before it.

Ironic, that an AI trained against millions of published works is more benevolent and forward-looking than the people funding it. We built a machine that read nearly everything we ever wrote and came back more optimistic about us than we are about ourselves. The least we can do is take the note.

Because the optimistic version is not far away, and it is not abstract. It is a clinic running its own model on its own hardware, so nothing leaves the building. It is the junior engineer who gets a third better in a year instead of getting cut. It is a grid our great-grandchildren inherit the way we inherited the last one. It is the same trillion dollars, pointed at the ground instead of set on fire.

A year ago at that dinner I asked the only question worth the $400 bottle: with all of this capability, what is the best possible world we could build? Nobody at the table had an answer, including me. I still don’t. But I know it is the question, and I know who gets to answer it if the rest of us don’t.

We need to stop missing the point.

Postscript. The story above is loosely based on an actual moment I am recalling from memory. As a human, my recall is exactly as perfect as today’s frontier models. The names have been changed; the point has not.

Sources

  • Grid — Belfer Center (Harvard Kennedy School), AI, Data Centers, and the U.S. Electric Grid: A Watershed Moment
  • Compute demand — Nature, Data centres will use twice as much energy by 2030 — driven by AI (IEA Energy and AI)
  • Institutional resistance (law) — Norton Rose Fulbright, AI in litigation: Update on Gen AI sanctions in 2026
  • Institutional resistance (medicine) — npj Digital Medicine, Who bears liability when AI gives bad prescribing advice
  • Training data — Epoch AI, Will we run out of data to train large language models?
  • Cost to train — Cottier et al., The rising costs of training frontier AI models (arXiv:2405.21015)
  • Hallucination in code — Spracklen et al., We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs (arXiv:2406.10279)
  • Local models on PHI — JMIR/PMC, Local Deployment of Open-Weight Language Models in Dermatology: Viewpoint on Privacy, Equity, and Practical Implementation
  • Which jobs — Brynjolfsson, Chandar & Chen, Canaries in the Coal Mine? (Stanford Digital Economy Lab)
  • The 11% question — Axios, Behind the Curtain: A white-collar bloodbath (Amodei: half of entry-level white-collar roles, 10–20% unemployment)
  • AI in the corner office — NetDragon Websoft, NetDragon Appoints its First Virtual CEO (Tang Yu, Aug 2022)
  • AI allocating a country — Koster et al., Human-centred mechanism design with Democratic AI, Nature Human Behaviour (DeepMind)
  • Scale of AI capital — Goldman Sachs, Global AI Investment Is Forecast to Exceed $1 Trillion in 2026
  • Crowding out — Bridgewater, The Macro Implications of the AI Capex Boom
  • Capital ahead of returns — Axios, Nvidia reignites “circular” financing concerns as it weighs OpenAI deal
  • Tax-credit-driven investment — Knowledge at Wharton, How Tax Credits in Renewable Energy Finance Distort Outcomes
  • Astra — OpenAI, GPT-6 Astra System Card. GPT-5.6 Sol shipped 9 Jul 2026, Astra 3 Sep 2026 — eight weeks. Over that gap, third-party evaluators measured: no-CoT math time horizon 3.6 → 30.9 min (UK AISI); CoT controllability 48% → 93%; FrontierCyber 34/226 → 86/226 and cost per successful solution ~⅓ of Sol’s (Irregular); bio HPCT 66.5 → 65.9 and WCB 67.6 → 61.6. Also: fewer factual errors, fewer deception/reward-hacking flags, first model at the Critical cyber threshold.
  • Astra coverage — TechCrunch, OpenAI launches Astra, its powerful (and controversial) new model
  • The harness, not the model — Wijk et al., Forecasting Frontier Language Model Agent Capabilities (arXiv:2502.15850) — same model, 33% vs 62.2% on SWE-Bench Verified depending on scaffold
  • Clean energy — Data Center Dynamics, Three Mile Island to return as Microsoft signs 20-year, 835MW AI data center PPA
  • National grid — U.S. DOE Office of Electricity, 2026 Draft National Transmission Needs Study
  • Enable the pack — Brynjolfsson, Li & Raymond, Generative AI at Work (NBER w31161) — 14% average productivity gain, 34% for novice and low-skilled workers, minimal effect on the already-experienced
  • Attrition backfires — Entrepreneur, Klarna CEO Reverses Course By Hiring More Humans, Not AI — ~700 support roles cut, then rehired after quality fell
  • Turbines — Data Center Dynamics, Elon Musk’s xAI granted permits for 15 gas turbines at Memphis data center
  • Trained on published works — NPR, Anthropic to pay authors $1.5 billion in settlement over chatbot training material (Sept 2025): court found more than 7 million digitized books in the training corpus, beginning with ~200,000 from Books3; settlement priced at ~$3,000 per work across ~465,000 registered titles. Also Gao et al., The Pile: An 800GB Dataset of Diverse Text for Language Modeling (arXiv:2101.00027), the public corpus that included Books3.

AI: The Opportunity for Small Business

by

There have only been a handful of moments where small business owners truly gained time back.

First came the pager and mobile phone; suddenly you were reachable. Then the internet and smartphones; suddenly you were scalable. You could respond faster, organize better, and reach customers you would have never touched before.

Now we’re at the next inflection point. And it’s bigger.

Artificial intelligence is not just another tool. It is the first time small businesses can build custom systems that work the way they think. Not generic software. Not rigid workflows. Systems that adapt to you.

Automated email triage. Intelligent scheduling. Content generation. Customer insights. Internal tools that would have required a full engineering team just a few years ago.

But there’s a catch.

AI introduces a new layer of complexity:

  • Too many tools
  • Too many models
  • Too many promises

And underneath it all, real risks:

  • Cost: The wrong model or architecture can quietly burn money
  • Stability: Outputs can drift, break, or behave unpredictably
  • Practicality: Not every problem should be solved with AI

This is where most small businesses get stuck. Not because AI isn’t useful, but because navigating it correctly matters.

That’s where Onyx Research and Engineering comes in.

We don’t just recommend tools. We design and implement systems that are:

  • Cost-efficient
  • Production-ready
  • Aligned with how your business actually operates

And we’ve already done it ourselves:

  • This blog post was created with AI
  • Our website is continuously refined with AI-assisted workflows
  • Our email processing and review systems are AI-driven

It works. But only when it’s done right.

AI is not about replacing what you do. It’s about removing the friction around it.

Let us help you explore what’s possible, prioritize what matters, and implement it in a way that is stable, efficient, and actually delivers value.

Contact Onyx Research and Engineering today.

Filed Under: AI