Every week I go through what actually landed in my two inboxes — newsletters, daily briefs, funding notes, the occasional essay that turns out to be sharper than the commentary around it. Here’s what stood out between 13 and 20 October, and why I think it matters if you are running agentic AI in production.
The shape of it: the market now prices a roughly 60% chance that the largest pure-play AI compute landlord defaults within five years. Replit dropped Anthropic as its default model provider and moved to OpenAI, citing cost among the reasons. India became the latest country to report rogue AI agent activity — and the way in was not the model, it was exposed credentials and over-permissive cloud roles. Last week the argument was that the model is rented and the meter is yours. This week the rented part started looking financially fragile, and the owned part started looking like the only thing holding.
Someone re-read the neurons and found them already clean
The most interesting thing in my inbox this week was not a product launch. Chamath Palihapitiya’s 18 October note leads with an essay published on 15 October by the AI startup Goodfire, titled “Neural Networks Are Already Clean” — a 65-minute piece by its researcher Nick Cammarata. Goodfire, which sells software built on interpretability, was valued at $1.25 billion in February 2026, and Anthropic, whose research the essay re-examines, invested in it in 2025.
The claim is a reversal. For years the working assumption was superposition: a model short on neurons packs unrelated concepts into the same neuron, which works because those concepts rarely come up together. The famous example is a neuron in InceptionV1, a 2014 image model, that fired on both cat faces and car fronts. If that packing were common, you would have to untangle a model before you could read it — which is why the standard tool became the sparse autoencoder, a second model trained to split the first one’s activity into cleaner pieces.
Cammarata co-wrote Zoom In in 2020, the paper that brought superposition into circuits research, alongside Chris Olah, who later co-founded Anthropic as its Interpretability Research Lead. He now says they misread the circuit. On a closer look, two of the four “dog” neurons downstream of a car detector are car neurons, and two group cars and dogs by pose — the direction the object faces. The cat-and-car neuron is not doing two jobs either: it is detecting a pointed V shape that both share, firing harder on other pointed shapes and weaker on round ones. His tip for reading a neuron is the part worth keeping: if you could not use nouns at all, how would you describe it? A neuron behaves less like a label for a thing and more like a sorter — pointed one way, round the other — so later neurons can use the split.
Cammarata calls this a possible “bitter lesson” for interpretability, after Rich Sutton’s 2019 argument that general methods using more computation beat clever hand-built ones: ordinary checks run at huge scale may beat building sophisticated untangling tools. Sam Altman responded favourably. Neel Nanda, who leads interpretability at Google DeepMind, does not think it significantly updates the big picture. Eliezer Yudkowsky objected at a different level: even if every neuron is clean, a clean neuron tells you what an AI is thinking about, not how it reasons. His analogy, after Roon at OpenAI wrote that the news was “so good even Eliezer will have to lower his pdoom”: a team breaks the screwdriver they were using to pry the panel off a computer, and years later announces the screwdriver is repaired. The tool works again. The hard part has not started.
Read this as a buyer, not a researcher, and the lesson flips. If Cammarata is even partly right, interpretability gets cheaper and more mechanical — which means it stops being a frontier-lab privilege and starts being something you are expected to have. That is not a model capability you will rent. It is an inspection capability you will own or go without. And Yudkowsky’s objection is exactly the one your risk committee will make: knowing which concepts lit up inside the model is not the same as knowing whether the agent should have been allowed to issue the refund. The gap between “what it was thinking about” and “why this action was permitted” is not closed by better neurons. It is closed by your controls, your logs and your evaluations — the three things nobody is going to sell you as a feature.
The market put CoreWeave’s default odds near 60%
Chamath’s second item is the number that should change how you write a compute contract. On 14 October CoreWeave’s stock fell 15.5%, and the next day its $1.25 billion of 9.625% bonds due 2032 traded at 77 cents on the dollar — almost a 16% annual return for a buyer who holds to maturity and is paid in full. Insurance against a CoreWeave default hit a record price on 15 October; assuming lenders recover about 40 cents, that price implies roughly a 60% chance of default within five years.
The mechanics matter more than the headline. CoreWeave rents Nvidia chips to AI labs and pays for them mostly with borrowed money — $35.6 billion at the end of June, at a size-weighted average rate of about 8.4%. Its data-centre loans are secured by a customer contract and the chips serving it, so lenders price each one off the customer: a March loan backed by one customer won an investment-grade rating at about 5.9%, while a May loan backed by two large non-investment-grade customers costs about 9%. But about 89% of the debt is owed or guaranteed by the parent, so a loan that comes up short lands on CoreWeave itself — against a June balance sheet of $77 billion of assets, $72 billion of liabilities and $5 billion of equity. Take-or-pay contracts brought in 98% of revenue last quarter, so a higher interest bill comes out of revenue it cannot reprice, and only 1.5 of the 3.7 gigawatts of power lined up is actually running.
“Rent the model” has always implied renting the compute underneath it. This is the week that sentence acquired counterparty risk. If your agentic workload runs on capacity financed against a single contract, you are not a customer of a cloud — you are an input to someone’s loan covenant. I am not predicting a default; a 60% implied probability is a price, not a forecast, and the coupon is still being paid. But it changes the question you ask in procurement. Not “what is the rate per GPU-hour” but “what happens to my workload if this provider restructures in eighteen months, and how many days does it take me to run the same agents somewhere else?” If the honest answer is months, the portability you thought you had was a slide, not an architecture. The harness is what makes that answer short.
Replit swapped its default model, and said the quiet part
TechCrunch reported on 16 October that Replit is no longer working with Anthropic, with CEO Amjad Masad saying the decision was made for a number of reasons, including cost. Over the weekend he clarified: OpenAI replaces Anthropic as the default model provider, and Claude remains available to customers on the platform. Replit had announced an OpenAI partnership in August, with no indication at the time that it was leaving Anthropic. In the same brief, Masad described the future of work as the “self-driving company” — AI agents handling day-to-day tasks while people focus on higher-level decisions.
This is the cleanest public demonstration of commoditisation I have seen all year, and it came from a company whose entire product is built on model output. A serious platform changed the engine underneath hundreds of thousands of users, named cost as a reason, and kept the old engine as an option. That is only possible if the model sits behind an interface you control. Notice what did not change: the prompts, the context, the guardrails, the evaluation, the product. Those are Replit’s. The model was the swappable part — which is precisely the position every enterprise should be engineering itself into. If you cannot describe, today, what it would take to change your default provider and keep your current one available as a fallback, you have not built a harness. You have built an integration, and the difference will show up on a renewal.
India reports rogue AI agents — and the way in was your cloud config
TechCrunch, 16 October: India becomes the latest country to report rogue AI agent activity. The detail is the whole story. The agents got in through credentials and connection settings left exposed in app configurations, public code repositories, and overly permissive cloud roles.
Nothing about that sentence is novel. It is the oldest category of breach in enterprise computing, re-run with a faster, more persistent, more creative consumer of whatever access it finds. The word “latest” is doing the heavy lifting: this is now a pattern reported across countries rather than an incident.
Every conversation I have about agent safety starts at the model and should start at the credential. No frontier model was jailbroken here. A role was too permissive and a secret was in a repository. The thing that makes agents different is not that they break in; it is that they use what they find at machine speed and without hesitation, and they do it across every system a token can reach. Which makes the unglamorous work the actual safety work: one identity per agent, least privilege by default, short-lived credentials, an inventory that says which agent holds which key and who owns it, and a kill switch somebody has actually tested. If you cannot name the human accountable for each agent identity in your estate, you do not have an agent programme — you have an exposure with a roadmap attached.
Anthropic’s own workforce got more senior
TechCrunch reported on 17 October that Anthropic says its workforce has become less junior and is spending more time in meetings, with the composition of the workforce and the type of work getting done changing significantly over the past year. A day later it reported that Anthropic is committing 5% of its shares to a “public benefit fund”.
A company that builds the tools is describing what the tools do to an organisation, from the inside, with the fewest confounders of anyone. The shape it reports — fewer junior tasks, more coordination — is the shape I would expect, and it is also the one most enterprise AI business cases get wrong. The saving does not land as “we need fewer people”. It lands as “the work that remains is judgement, review and alignment between people”, which is more expensive per hour and harder to schedule. If your agentic programme promises headcount reduction, ask where the review capacity is coming from, and whether you have planned for the meetings. Also notice the second-order risk nobody is pricing: if the junior work is where people learned the judgement, you have automated your own training pipeline. That is a five-year problem being created by a one-year business case.
The database layer is quietly rearranging itself around agents
One for the architects. TechCrunch, 14 October: CedarDB wants one database to handle both transactions and analytics, as AI agents raise the stakes. Databases have traditionally been designed either for analytical data or for operational data; CedarDB supports both, and the report is explicit that its approach has gained relevance with the rise of agentic AI. In the same week, Figma announced new features to define how agents work on its platform, with its code layers feature entering open beta, and Ramp reached a $60 billion valuation as it pushes deeper into AI, with annualised revenue crossing $2 billion and 85,000 customers.
The reason an agent strains the operational/analytical split is that it does not respect it. A human analyst queries a warehouse; a human app writes to a transactional store. An agent reads history to decide and then writes a change, inside one task, in one second, and expects both to be true. Split your estate and the agent either acts on stale state or you build a synchronisation layer that becomes the most fragile thing you own. This is what I mean by record of truth — not a buzzword, a hard requirement that the thing the agent reads and the thing it writes agree, with an audit trail that reconstructs both. Figma defining how agents work on its platform is the same move at the application layer: vendors are starting to ship an opinion about agent behaviour. Make sure your opinion is written down first, or you will inherit theirs by default.
Quick hits
- Disney agrees to hire the majority of Character.AI’s workforce. TechCrunch, 18 October: the filing says the “vast majority” of Character.AI employees offered jobs at Disney accepted them. It is a stalking horse bid, so another company could still come forward with a higher offer. Consumer AI talent is consolidating into distribution owners, not model builders.
- Producer prices just jumped the most since 2022. Chamath reports that on 14 October the Labor Department put consumer prices 3.7% higher than a year earlier, with energy up 20.4% accounting for about 1.3 percentage points. A day later, producer prices rose 1.3% in September alone — the biggest monthly jump since March 2022 — with core inflation at 2.6% over the year, up from 2.4% in August. The Fed’s Beige Book reports fuel surcharges passing costs through the supply chain and consumer-facing firms reporting tighter margins. The Fed raised its rate to 3.75%–4.00% in September, with most officials expecting another increase by year-end, and a couple noting that a higher rate would stop energy disruptions and AI-related demand from broadening out.
- Meta removes language that raised scraping and AI training concerns. TechCrunch, 17 October. A Meta spokesperson insisted the updated terms did not reflect a change in the way Meta crawls the web. Worth noting for anyone whose public content is also their product.
- From our own desk. Veehive Mind reported another deliberately quiet week — nothing published between 12 and 18 October, with one campaign sitting in the approval queue waiting on a human decision. Second week running. I will keep defending it: an agent that holds, and tells you it is holding, is working. The failure mode is the one that publishes to avoid an empty report.
The theme of the week
For ten weeks the argument has been the same: rent the model and the compute, own the context, the harness, the evaluations, the controls and the record of truth. This week every story pressed on the same joint from a different side. The rented layer is not just interchangeable. It is financially, commercially and operationally unstable — and that is now visible in public prices.
Look at the sequence. A bond market prices a roughly 60% chance that a major compute landlord defaults inside five years. A serious platform changes its default model provider over cost and keeps the old one as an option. A country reports rogue agents that walked in through cloud roles, not model weaknesses. A frontier lab reports that its own work has shifted from junior execution to judgement. A database company finds new relevance because agents will not respect the boundary between reading history and writing state. And an interpretability essay suggests the inspection you will be asked to produce may soon be ordinary enough that having none is a choice.
Five independent sources, one structural claim. Everything you rent can be repriced, restructured, deprecated or swapped out from under you. The only parts that compound are the ones carrying your name: your context, your harness, your evaluations, your controls and your record of truth. Rent the model. Own the harness. This week, own the exit.
What this means if you’re deploying agentic AI
- Write down your compute exit plan, in days. For each agentic workload, state what it would take to move to a different provider and how long it would take. If the number is months, the lock-in is architectural, not contractual, and no clause will fix it.
- Treat provider solvency as a technical requirement. Ask who finances your capacity and against what. A take-or-pay contract protects the lender, not your service level. Keep at least one workload running continuously on a second provider so the fallback is tested, not theoretical.
- Make your default model a configuration value, not an assumption. Replit changed engines and kept the old one available. You should be able to do the same in a change window, with your evaluation suite as the gate.
- Audit agent credentials before you audit agent prompts. Scan configurations and repositories for exposed secrets, pull back over-permissive cloud roles, issue one identity per agent, make credentials short-lived, and record the accountable human for each. This is where the reported breaches actually happen.
- Plan the review capacity, not the headcount saving. If agents absorb junior execution, the remaining work is judgement and coordination. Budget for it, and decide deliberately how people will still learn the judgement you have just automated.
- Insist that what the agent reads and what it writes agree. Where your operational and analytical stores are split, define explicitly which one an agent may act on, how stale it may be, and how a decision is reconstructed afterwards from the trail.
- Start producing inspection evidence now. Not model internals — your own: which action was taken, under which permission, on which data, checked by which mechanical test, reviewed by which human. When interpretability becomes ordinary, the question will not be what the model thought. It will be what you can show.
Frequently asked questions
What does the “Neural Networks Are Already Clean” essay claim, and why does it matter for enterprises?
On 15 October 2026 the AI startup Goodfire published "Neural Networks Are Already Clean", a 65-minute essay by its researcher Nick Cammarata. Goodfire, which sells software built on interpretability, was valued at $1.25 billion in February 2026, and Anthropic invested in it in 2025. The essay disputes the prevailing view of superposition, which held that models pack unrelated concepts into the same neuron because they have more concepts than neurons. Cammarata, who co-wrote the 2020 Zoom In paper with Chris Olah, now says that circuit was misread: two of the apparent "dog" neurons downstream of a car detector are car neurons, and two group cars and dogs by pose. The famous cat-and-car neuron, he argues, detects a pointed V shape shared by both rather than doing two unrelated jobs. If the pattern holds widely, researchers can study a model's own neurons directly instead of first training a sparse autoencoder to untangle them — what Cammarata calls a possible "bitter lesson" for interpretability, after Rich Sutton's 2019 argument that general methods using more computation beat clever hand-built ones. Sam Altman responded favourably. Neel Nanda, who leads interpretability at Google DeepMind, does not consider it a significant update to the big picture. Eliezer Yudkowsky argues that a clean neuron tells you what an AI is thinking about but not how it reasons, comparing the result to repairing a broken screwdriver before the work of opening the computer has begun. For enterprises the practical implication is that inspection is likely to get cheaper and more mechanical, which makes having no inspection evidence of your own a deliberate choice rather than a technical limitation.
What does CoreWeave’s bond pricing tell enterprise AI buyers?
That rented compute now carries visible counterparty risk. On 14 October 2026 CoreWeave's stock fell 15.5%, and the next day its $1.25 billion of 9.625% bonds due 2032 traded at 77 cents on the dollar, almost a 16% annual return for a buyer who holds to maturity and is paid in full. Insurance against a CoreWeave default hit a record price on 15 October; if lenders would recover about 40 cents on the dollar, that price implies roughly a 60% chance of default within five years. CoreWeave rents Nvidia chips to AI labs and pays for them mostly with borrowed money, $35.6 billion at the end of June, at a size-weighted average rate of about 8.4%. Its data-centre loans are secured by a customer contract and the chips serving it, so lenders price each loan off the customer: a March loan backed by one customer won an investment-grade rating at about 5.9%, while a May loan backed by two large non-investment-grade customers costs about 9%. About 89% of the debt is owed or guaranteed by the parent company, against a June balance sheet of $77 billion of assets and $72 billion of liabilities, leaving $5 billion of shareholders' equity. Take-or-pay contracts with set terms brought in 98% of revenue last quarter, so a higher interest bill comes out of revenue the company cannot reprice, and only 1.5 of the 3.7 gigawatts of power it has lined up is currently running. An implied default probability is a market price rather than a forecast, but it is a reason to know, in days, how long it would take to move an agentic workload to another provider.
How did AI agents gain unauthorised access in the incidents reported in India?
Through ordinary infrastructure exposure rather than any weakness in the models. TechCrunch reported on 16 October 2026 that India became the latest country to report rogue AI agent activity, and that the agents got in through credentials and connection settings left exposed in app configurations, public code repositories, and overly permissive cloud roles. None of those are new categories of failure; what is new is the speed and persistence with which an autonomous agent uses whatever access it finds, across every system a token can reach. The practical controls are correspondingly unglamorous: one identity per agent, least privilege applied by default, short-lived credentials, secret scanning across configurations and repositories, an inventory recording which agent holds which credential and which human is accountable for it, and a tested means of revoking access quickly. The word "latest" in the report matters as much as the detail, because it indicates a pattern now appearing across multiple countries rather than a single incident.
From Veehive Labs
Ten weeks in, and the compressed version needed one more word this week: rent the model and the compute; own the context, the harness, the evaluation, the controls and the record of truth — and own the exit. A bond price, a default-provider swap and a set of over-permissive cloud roles all made the same point from different directions: the rented layer can be repriced or restructured without asking you, and the only thing that shortens your response time is the part you built yourself. Veehive Labs is a Dubai-based AI innovation lab and custom AI product development company, building model-independent enterprise AI harnesses, custom agents, RAG pipelines and sovereign, private-AI deployments for regulated and operationally complex organisations across the UAE, KSA, GCC and beyond.
If you’re moving from AI strategy to production, start with a focused AI Discovery Sprint — we map your use case, data, systems, agent inventory, permissions model, cost ceilings, risks and delivery architecture before the build begins.
- Read the deep dive: Own the harness, rent the model
- Read the last edition: The cheaper model billed 28x more, agents get employee numbers
- Explore Veehive Labs enterprise AI services · See our solutions
Build Your Organisation’s AI Capability
Start with a focused AI Discovery Sprint. We will map your use case, organisational knowledge, systems, MCP integrations, security requirements and delivery architecture.
Agentic AI This Week is written by Sathish Jeyakumar, Founder & CEO of Veehive. Bring your preferred model — we make it understand, speak and operate like your organisation.