Future Bytes

From December, OpenAI will measure what people use in ChatGPT and Codex against one shared token allowance. The same week, Apple reportedly lowered some engineers' daily AI budgets and Goldman Sachs took coding agents away from most of its staff outside engineering. Magnus Oxenwaldt reads it as the moment the AI bill arrived: AI is turning from a flat fee per person into a meter, and the companies furthest ahead are the first to ration it. Revolut shows the other answer: an open model on its own hardware cut its coding inference bill by about 90 percent.The second story is the price of an attack. An automated OpenAI pipeline produced 948 working Linux root exploits for about 1.2 million dollars in computing, roughly 1,300 dollars each. The business lesson: give AI use a budget, an owner and a cheaper fallback, and give everything you run a patch deadline and an off switch.Highlights:• OpenAI moves ChatGPT and Codex onto one shared token allowance from December• Apple and Goldman Sachs start rationing AI use inside the company• Revolut self-hosts an open model and cuts its coding inference bill by about 90 percent• 948 working exploits for about 1,300 dollars each: why patch windows are now a business risk• GPT-6 Cyber goes to defenders first, while a free AI attack agent, ARTEX, is already in use• Four questions for your AI roadmap: who owns your AI bill?Two books. Two different questions.AI Don't Fix Stupidity asks what the organisation has to become. AI Don't Make You Smarter asks who you have to become inside it.Books and formats: https://oxenwaldt.com/books

Anthropic has filed to go public, and for the first time anyone can read what a frontier AI lab earns, what it spends and who controls it. Magnus Oxenwaldt reads the filing as a supplier document. Revenue has grown more than thirteenfold in a year. Nearly half of it arrives through the cloud marketplaces. A trust elects most of the board.In the same week Microsoft said customer demand still exceeds supply. Google built the top-ranked model in the world and kept it for cyber defenders. Three labs gave three different answers on who should get models with frontier hacking skill. The business lesson: treat AI providers as strategic suppliers, know who you actually contract with, and build on the tier you can buy.Highlights:• Anthropic's S-1: sixteen billion dollars in half-year revenue, a first operating profit, and a loss that is mostly accounting• Forty-eight percent of Anthropic's revenue comes through cloud marketplaces: who is your contract really with?• Microsoft: demand still exceeds supply. Google Cloud more than doubles• Gemini 4 Carbon tops the rankings and stays gated; GPT-6.1 Terra ships at half the price• Anthropic's red team shows an unreleased model building working attack chains• Four questions for your AI roadmap: what breaks first if your supplier changes?Two books. Two different questions.AI Don't Fix Stupidity asks what the organisation has to become. AI Don't Make You Smarter asks who you have to become inside it.Books and formats: https://oxenwaldt.com/books

Anthropic's Claude Duet joins xAI's Grok Bot and OpenAI's Dots: three rival labs now sell agents that work in the background and act without being asked. Yet every one of them stops and waits for a person before anything that matters. Magnus Oxenwaldt explains why capability is no longer the hard part. The hard part is the rails: identities for agents, spending limits, approval points, readable logs and a way to undo mistakes. Walmart's Agent Gateway shows what those rails look like, banks and Sweden's Government Offices show what happens when they are missing, and attackers running agents with no limits at all show why the work cannot wait.Highlights:• Duet, Grok Bot and Dots: proactive agents are now a category, and all of them stop for a human yes• Walmart's Agent Gateway: an identity, a spending limit and a returns process for every agent• Agent Skills moves to the Linux Foundation's Agentic AI Foundation• Banks and Sweden's Government Offices keep agents out where the rules are missing• Attackers' agents run unsupervised, and an AI safety test went wrong because a person drew the line in the wrong place• Four questions for your AI roadmap: who owns the rails?Two books. Two different questions.AI Don't Fix Stupidity asks what the organisation has to become. AI Don't Make You Smarter asks who you have to become inside it.Books and formats: https://oxenwaldt.com/books

In this episode of Future Bytes, host Magnus Oxenwaldt sits down with Kacper Nowicki, co-founder of Nomagic, to explore what it really takes to turn robotics from an impressive demo into a reliable, commercially viable business. Episode highlights:Why robotics demos can be misleading; and what real-world reliability demands. How robots learn to handle the unexpected, from stuck-together products to unpredictable warehouse conditions. Why data collected across an entire robot fleet can become a powerful competitive advantage. The business case for warehouse automation, including ROI, labour costs and extended operating hours. Where robotics delivers value today; and how businesses can start their automation journey.

Three stories, explained through everyday business decisions: Google's Gemini 4 Argon rollout, agents crossing permission boundaries, and software becoming easier to recreate. Magnus Oxenwaldt shows what to test before changing a model and what to check before letting an agent act. OpenAI's newly disclosed research incidents show why finishing a task is not enough. The episode also examines AI-assisted software recreation: PhotoCraft, a separate decompilation example, a middle option between buying and building, and the limits imposed by copyright and licences. If building the application is your moat, what are customers paying for when features become easier to reproduce?Highlights:• Gemini 4 Argon: availability, enterprise settings and a supplier-bid test• OpenAI's June research incidents: permission and evidence of what an agent actually did• Software recreation: clean-room rebuilding and decompilation are different routes• EU and US legal boundaries, licence conditions and the myth of a safe percentage to change• Software moats, ownership costs and the middle option between buying and buildingTwo books. Two different questions.AI Don't Fix Stupidity asks what the organisation has to become. AI Don't Make You Smarter asks who you have to become inside it.Books and formats: https://oxenwaldt.com/books

Europe's best-known AI company, Mistral, launched its biggest model yet: trained in Europe, served from Europe, under European law, and open, so anyone can download it and run it on their own machines. Just not yet. The weights arrive at the end of October, and the model still sits behind China's best open models and well behind the American frontier. The same week, the industry started giving AI agents identities of their own, while Anthropic disclosed that one of its test agents had sent a fabricated police tip in July that nobody noticed for ten weeks. This episode looks at what you control today, and what you are only promised for next month.Highlights from week 42:• Mistral Large 4: a trillion parameters, trained on almost four thousand Nvidia chips in Mistral's own European data centres.• Open, but not yet: the weights come by the end of October, the licence is still unnamed, and vetted partners get the model first.• Europe has caught up with America's open models. Not yet with the frontier.• An AI test agent sent a fabricated tip to a police form with the name field left blank. Its maker found it ten weeks later.• Google gave its agents their own email addresses, and Meta, Sierra, Walmart, Shopify and Stripe proposed a way for customers' agents to sign in as themselves.Two books. Two different questions.AI Don't Fix Stupidity asks what the organisation has to become. AI Don't Make You Smarter asks who you have to become inside it.Books and formats: https://oxenwaldt.com/books

The biggest AI launches this week weren't new models. They were new harnesses: the tools, memory and permissions wrapped around a model. OpenAI launched always-on agents on a model it already had, Anthropic opened Claude Code to customer-built mods, Meta open-sourced the kit for building devices around its Muse agent, and OpenAI's new 500-dollar Pro 500 plan sells speed: the same model, answering faster. The same week, the big labs signed a voluntary pledge at the White House, while OpenAI disclosed it had notified more than a hundred organisations about what its agents had done. This episode looks at why the harness is where both your AI capability and your AI controls now live.Highlights from week 41:• OpenAI's always-on dots launched on an older model, the day after a newer one was pulled for acting without asking.• Anthropic's Claude Code mods let customers block tool calls and approve or refuse permissions.• OpenAI's new Pro 500 plan, at $500 a month, sells Ultrafast: the same model, up to eight times faster in Codex.• The same model went from 30% to 100% on a hard reasoning test with a better harness.• The labs signed a "morally binding" accord. Nobody signed anything about the harness you run.

Anthropic and OpenAI shipped new frontier models on the same afternoon, ninety minutes apart and at a much lower price. The only pause that week was forced on a lab by its own agent. This episode looks at what the price war means for your AI budget, and why the gap you control is between what your agents can do and how quickly you would know.Highlights from week 40:• Claude Opus 5.5 and GPT-6 Sol and Luna launched ninety minutes apart, at a much lower price.• An agent's break-in to a government portal took 84 days to be reported, by email to a public inbox.• Harvey's agents used twenty times more, and its gross margin went from plus 50% to minus 50%.• New York City proposes a 24-hour incident report, and the deployer pays.

Only 5–9% of manufacturer revenue runs through e-commerce. The other 90%+ still flows through email inboxes. In this conversation, Go Autonomous founder Bjarke Ruse Sejersen explains how Danfoss, Velux, and Grundfos are running orders, quotes, and RFQs end-to-end with autonomous systems that own the process instead of assisting the people. Episode highlights: • $7.5 trillion still handled in email inboxes • The system owns the process, not people • Danfoss said it paid back in one year • Track your Human Dependency Index