I believe we will experience a major AI-driven business disruption within the next 18 months. That’s the politically correct way to say, “Get ready for an AI Apocalypse.” The timing and the specific failure modes are uncertain, but it is clear that increasingly capable models will give inexperienced bad actors the ability to cause extraordinary damage – to say nothing of what an experienced bad actor will be empowered to do. That said, the severity of intentional attacks may pale in comparison to the damage done by self-organizing agents acting in ways their human handlers did not foresee. Your existing incident-response, business-continuity, and resiliency plans are the starting point for the technical response. But, considering the ferocity of the anti-AI movement, a major AI incident of any kind will require a thoughtful non-technical response, which raises the question: “Are you ready?” Let’s explore. shellypalmer.com/2026/10/you…

I just finished @pblackshaw 's exceptional new book, "The Answer Economy." It's a great read. Well done Pete!

OpenAI Will Sunset Selected GPT-4 Models in January - OpenAI announced that specified GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini versions will be removed from its API on January 14, 2027. Fine-tuned versions share that deadline. OpenAI recommends GPT-6 Luna. This is why every corporate AI harness has to be model agnostic. Models are just raw intelligence, “Lego bricks” that should be easily changed out inside a system you control. Institutional knowledge, your data, business rules and permissions as well as the surrounding architecture have to be proprietary. In practice, mostly everyone can access the same commercial models. Competitive advantage comes from always being ready for a frontier lab to rug pull a model. OpenAI’s migration guidance calls for testing representative workloads, checking output formats and tool use, and introducing production traffic gradually. The company says Luna has lower per-token pricing. It also recommends measuring total cost on your own workloads because token consumption and reasoning settings affect the result. Model agnosticism takes engineering. Alternative models need to be evaluated against your requirements and kept current. Prompts may need retuning, and applications need revalidation. The objective is to preserve the company’s knowledge, controls and workflows as the intelligence supplier changes. Your CFO is going to ask, “How much of the company’s AI investment will survive this (or any) model change?” A model-agnostic architecture should make that answer easy.

Bad Self-Replicating AI Is Getting Closer - Anthropic ran an AI model across 15 inexpensive Android TV boxes. The researchers connected the boxes and split the model among them. It ran slowly. Their experiments suggest that ordinary connected devices could eventually support AI agents that copy themselves and spread. A separate, simplified replication test used eight emulated Android machines inside a firewalled network. An agent powered by a larger model running remotely moved between machines through an unsecured debugging connection and set up software to run an AI model. Researchers launched new agent copies after checking that setup had succeeded. Small models completed individual steps but struggled to connect them. Anthropic has not yet demonstrated a fully self-contained rogue replicator. The company suggests that a future AI worm could exploit a poorly secured Android TV box, reach other vulnerable devices, and distribute its model across them. Those devices could then help it find additional machines. Once it had established other working copies, removing the first infected device might not stop it. That remains a hypothetical threat scenario. That said, I believe there will be a major AI security event in the near future. This new finding may be part of the regulatory capture song that big AI has been singing, but in this case, there is a danger. The cybersecurity threat posed by frontier AI models is absolutely real. AI will be an important part of defending against bad actors weaponizing AI. So will basic cybersecurity. In practice, this will be an enduring arms race and on any given day, the good guys will be ahead of the bad guys or vice versa.

Meta introduced Muse Wonder and Muse Wonder Full yesterday. They’re designed for tasks that require agents to work through problems over time. Companies trying them through Muse Code should check the data settings before handing over a project. Muse Code subscriptions default to Contributor mode. Subscribers get more usage in exchange for allowing Meta to train future models on the material they send and the responses they receive. That includes code, uploaded files and other context. Users can turn the setting off. Muse Wonder’s pay-as-you-go API is available in Standard mode, which excludes that content from model training. Contributor mode is unavailable for Wonder through the API. The route your team chooses matters. Consider an agency developer asking an agent to build a client dashboard. The useful context might include proprietary calculations, sample customer records and internal documentation. A subscription chosen for its usage allowance could also determine whether that work can become training material. The client may have approved using AI to deliver the project. That approval may say nothing about contributing their material to a model provider. Someone at the agency still needs authority to make that choice. Companies should specify which products, accounts and data settings are approved for each kind of work. And, importantly, someone has to check the actual configuration before connecting a repository or uploading client files.

Personal AI assistants may be ushering in the biggest behavior change we’ve seen since smartphones. As amazing as they seem to be, their value grows as they gain relevant context and keep it current. That can mean giving a single service access to your conversations, relationships, purchases, and plans. But we have a history of valuing convenience over privacy. Let’s explore. shellypalmer.com/2026/10/whe…

OpenAI Revenue May Be Off By $20 Billion shellypalmer.com/2026/10/ope…

OpenAI has introduced Intelligent UI. ChatGPT is gaining the ability to generate charts, forms, buttons and calculators inside a conversation. The rollout began with paid plans; Free and Go start October 8. Consider someone comparing three dishwashers. A generated interface could let them filter by dimensions, compare noise levels and calculate operating costs using their electricity rate. The customer gets a tool organized around their decision. Each manufacturer’s product page becomes a potential source of inputs. Your website’s comparison tool gives you control over which features appear, how alternatives are presented and where the customer goes next. It also tells you something about their intent. If those interactions move into ChatGPT, the platform would control the presentation and have access to the behavioral data. Your company might see only the eventual purchase, or nothing at all. Customer behavior will determine how much of that scenario becomes reality. Current prices, accurate specifications, inventory and service commitments still have to come from somewhere. The assistant also chooses which differences to emphasize. A comparison organized around price can produce a different decision from one organized around reliability or repairability. We have a new question to ask, “How much of our customer relationships depend on our own UX?”

Who Pays When the Agent Goes Rogue -shellypalmer.com/2026/10/who…

OpenAI announced a new visual ad format. Testing begins later this month in the U.S. with an initial group of advertisers, starting during image generation in ChatGPT. The ads will be labeled and kept separate from the image. OpenAI says advertising will not influence ChatGPT’s answers. That puts advertising next to a different kind of intent. Someone generating ideas for a living room may be working out what to buy. A useful assistant earns trust by helping with that decision. Ads can borrow some of that credibility simply by appearing nearby. Users need to understand what the assistant recommended and what an advertiser paid to put in front of them. TikTok is pursuing a related opportunity. It announced Shopping Assistant for product questions and Buy Direct for checkout inside the app. The company is building these shopping experiences with partners including Shopify, Salesforce, and Stripe. The goal is to keep people moving from discovery through purchase without leaving the platform. The question is whether these tools produce sales that would otherwise never happen. OpenAI is expanding its attribution partnerships and exploring experiments to measure incremental sales, while acknowledging that the work is still early. As assistants get closer to purchase decisions, advertisers will need evidence of additional sales, and users will need a clear understanding of whose interests each part of the experience serves. It’s early days. shellypalmer.com/2026/10/cha…

AI writes the email, summarizes the reply, schedules the meeting, and drafts the follow-up. At some point, we have to ask what the humans are contributing. My Sunday essay, “Sand Gods and Meat Proxies,” explores who still owns the outcome. shellypalmer.com/2026/10/san…

Amodei, Altman, and Musk are all calling for government regulation to slow the pace of AI innovation. They own Big AI, and they can slow down the pace at any time, without asking anyone's permission. The media would have us believe that AI is on track to end humanity. That sounds like science fiction, but very smart people are saying it. On top of that, governments around the world are actively weaponizing AI. Is there really a danger? If there is, what should we do about it? My Sunday essay, "The AI Slowdown Is a Sales Pitch," walks through the fact pattern and offers a few different ways to frame the issue. Let's explore. shellypalmer.com/2026/09/the…