❯ Trump creates a “Super Intelligence Force” led by intelligence chief Clayton to coordinate federal AI work

A new coordinating bodyTrump announced the creation of a “Super Intelligence Force” on his social media platform on October 4. According to ABC News, its task is to coordinate the federal government’s efforts so that America keeps leading in “super intelligence” while protecting American lives and interests. “Super intelligence” is Trump’s new term for AI.

Intelligence, antitrust, defense and personnelThe force is led by Director of National Intelligence Jay Clayton, who also serves as White House AI czar. Its other members are Federal Trade Commission Chairman Andrew Ferguson, Under Secretary of Defense for Research and Engineering Emil Michael, and Office of Personnel Management Director Scott Kupor. It will engage with consumers, public-interest and religious groups, critical infrastructure providers and AI companies.

Risk management modeled on financeIts work includes examining current law, discussing possible action by Congress, and developing risk management mechanisms modeled on frameworks from the financial industry. On September 29, the White House had gathered tech CEOs, and AI leaders signed a self-policing pledge. Clayton said: “The risk of not being first is high… Being first will better enable us to address those risks on behalf of the American people.”

How far self-regulation goesPutting the head of the intelligence community in charge of AI means the government sees this first as a national security matter. AI companies now face a single counterpart, but the rules still rest mainly on self-regulation. Former FTC Chair Lina Khan has publicly dismissed the pledge.

▪ SIGNALThe US government’s first standing coordinator for AI has been handed to the intelligence chief, so regulation starts from “we cannot lose” instead of “nothing can go wrong.”

❯ Altman says the world should accept “some bad things” for AI’s benefits, openly parting ways with Anthropic on regulation

"Accept some bad things"OpenAI CEO Sam Altman told Politico’s Decoded: “We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.” As relayed by Shattered, he said he would not take a trade that guaranteed no major hacks, no misuse and zero scams, because people will do orders of magnitude more good than bad with AI.

Where it differs from AnthropicAltman said he understands but disagrees with the view that the technology will become so powerful that “a single lab in San Francisco should have it.” As relayed by Seeking Alpha, Anthropic advocates stronger safeguards for the most advanced models, and CEO Dario Amodei has called for slowing development of the most capable systems. Anthropic counters that the rules it backs apply only to frontier models and would not entrench large developers.

A narrowing gap and a shared limitThe reports also note the gap is closing: OpenAI recently supported state safety laws and bipartisan House proposals and agreed to slow development. Altman drew a line as well, saying he does not accept risks involving “a serious loss of control to AI.” The FTC is currently looking into how both companies secure AI agents, and the White House AI accord requires outside audits.

Who defines the user's agencyThe practical meaning is that OpenAI will keep opening capable models and agents to more people and treat some misuse as a bearable cost. Those most affected are third parties harmed by agents: they did not choose to use AI, yet they bear the “bad things.”

▪ SIGNALBoth leading labs agree on slowing the most powerful models, and the disagreement is down to one question: whether losses from ordinary misuse are a price worth paying for broad access.

❯ OpenAI’s Codex lead pledges one improvement a day for 28 days, or a usage reset for users

One a day, or a resetThibault Sottiaux, who leads Codex at OpenAI, posted on X on October 4 that over the next 28 days the team will each day either ship one thing that is a clear improvement for most Codex and Work users or ship a full reset. The post has more than 1.2 million views.

First, a promise to simplifyAbout 16 hours earlier, in another post, he said the team was “locking in” and would work only on four things: simplifications, more efficiency for more usage, groundbreaking features and new models. He acknowledged that feedback was clear and users want things simpler. That post has more than 1.5 million views.

What a "reset" isAccording to Nerds Chalk, a full reset refreshes a user’s five-hour and weekly allowances and shifts the weekly reset date, but does not raise plan limits. OpenAI has done six resets since April, mostly tied to launch delays and demand surges. The pledge does not say which plans are eligible or on which days resets would come.

Usage as a bargaining chipWhat paying Codex users care about most is whether their allowance is enough. Publicly promising to top up usage when no update ships gives the team a deadline to meet every day, and makes it harder for users to decide to switch tools during those 28 days.

▪ SIGNALCompetition among coding agents is now counted by the day, and the way to keep users has shifted from releasing a new model to promising there is something to collect every day.

❯ GPT-5.5 leaves ChatGPT and Codex on October 14, less than six months after launch

Gone in nine daysOpenAI will remove GPT-5.5 from ChatGPT, ChatGPT Work and Codex across all plans on October 14. According to Business Today, OpenAI gave notice on September 15, saying “it’s time to say farewell to GPT-5.5,” and is steering users to GPT-5.6 Sol and GPT-6 Astra.

A life of 174 daysGPT-5.5 was released only on April 23 this year, so it will have lasted 174 days. The removal applies only to products accessed with a subscription login. Developers calling GPT-5.5 through the API with a key are not affected. OpenAI’s stated reason is simply that it is updating its model lineup.

Plenty to changeHighTechDad notes that users need to update custom GPTs built on GPT-5.5, workspace settings where it is pinned, scheduled tasks that call it by name, and Codex scripts with the model name hardcoded. A few weeks from notice to removal is not generous for teams running many automations.

Models as consumablesSwitching models is more than a name change. The same prompt can behave differently on a new model, so workflows that depend on stable output must be retested. It is a reminder to companies not to bind processes to a specific model, and it shows how little say subscribers have over which model they use.

▪ SIGNALA well-liked model is withdrawn in under six months, so what a subscription buys is “the current latest,” not a version you can rely on for long.

❯ OpenAI turns ChatGPT into a place to discover and use software, but has no revenue-sharing plan for developers yet

Using apps mid-conversationThe agent product Dots drew most of the attention at OpenAI’s developer conference on September 29, but TechCrunch argues that the announcements together point to a bigger aim: going around the traditional app store. ChatGPT suggests relevant apps during a conversation, and once connected, users operate them through interactive panels inside the chat without a separate download.

More than 4,000 appsMore than 4,000 apps are now connected to ChatGPT. OpenAI also introduced a “Sign in with ChatGPT” identity feature with 16 launch partners, and a new enterprise app marketplace launched with more than 30 partners, including Adobe, Figma, Salesforce and HubSpot. Dots can use these apps in the background in read-only mode for research and need user approval before taking actions.

The missing checkoutChatGPT has 1.2 billion weekly users, which is what makes it viable as a distribution channel. But TechCrunch points to a key gap: OpenAI announced no billing or revenue-sharing system comparable to an app store’s. Developers can reach users through ChatGPT without yet knowing how to charge inside it.

Apple's and Google's commission businessAn app store is distribution plus payments. If users get used to finding and using software in ChatGPT, phone app stores lose the discovery step first. Developers must judge whether the traffic gained by handing the entry point to OpenAI is worth the risk of later commissions or of being replaced by OpenAI’s own features.

▪ SIGNALOpenAI has taken the app store’s discovery function first and left payment for later, and the commission rate will only come up once developers cannot do without the entry point.

❯ The Codex system prompt for GPT-6 Sol leaks, with 1,900-plus lines showing how OpenAI disciplines its coding agent

A leaked "employee handbook"A document said to be the system prompt that OpenAI’s coding agent Codex uses on the GPT-6 Sol model is circulating online. A GitHub repository that collects vendors’ system prompts added the file on September 27, and according to AI Quill it runs to 1,903 lines. A system prompt is the behind-the-scenes instruction a vendor writes for a model, setting how it speaks and acts. The repository offers no verification of authenticity.

Banned words and autonomyAccording to a reading by Vocal Media, the prompt was extracted by security researcher Pliny and lists banned expressions such as “Bottom Line,” “delve,” “leverage” and “it’s worth noting,” along with contrastive structures of the “this is not about X, it is about Y” kind. It tells the agent not to keep asking permission, to create worktrees, resolve merge conflicts and open draft pull requests on its own, and to pause only before clearly irreversible actions.

Keeping long tasks on trackThe prompt also names specific tools, such as rg for search and apply_patch for file edits, and sets up a “guardian” sub-session that reviews operations for approval. After context is compressed, the agent is to continue from the summary and treat work across compressions as one sequence. Updates to the user may be no more than 60 seconds apart.

A reference answer for rivalsPrompts like this are the product of repeated trial and error, and other teams building coding agents can borrow its division of permissions and reporting rhythm directly. It also shows users that an agent’s initiative and brevity are not the model’s nature but rules written out one by one.

▪ SIGNALAs models converge, more of the difference in product experience comes from these thousands of unseen lines, which are exactly the part most easily extracted and copied.

❯ DeepSeek Harness builds a Claude Code Mods compatibility layer in two days that runs three sample extensions

Plugging in someone else's extensionsDeepSeek released DeepSeek Harness v0.2.1-alpha.1 on GitHub on October 3, adding an experimental Claude Code Mods compatibility layer. DeepSeek Harness is DeepSeek’s desktop coding agent tool. The layer wraps extensions written to the Claude Code Mods interface so that they run as Harness plugins.

What Mods areAccording to Memujo, Anthropic launched Mods in Claude Code on October 1. They are JavaScript or TypeScript event handlers that run directly inside the Claude Code process and can intercept tool calls, rewrite prompts, add commands and replace parts of the interface. DeepSeek’s layer was finished within 48 hours of the Mods launch.

A test, not a promiseThe release notes are blunt: the goal is to verify that the Mods API’s capabilities are “broadly a subset” of what Harness plugins can do, not to offer users complete, practical compatibility. Memujo says it runs three sample mods but cannot run Anthropic’s four first-party mods because some events and interfaces are unsupported. The same release adds a “let the agent create a plugin” entry and a developer toolkit.

Extension ecosystems start borrowingIf extensions written for one coding tool can be reused in another, the cost of switching tools falls. DeepSeek’s move aims to spare a latecomer from building an ecosystem from nothing. Memujo also warns that mods run unsandboxed with access to environment variables, API keys and all prompts, so the security risk travels with them across tools.

▪ SIGNALAn extension interface for coding agents had a compatibility layer two days after it appeared, which suggests this layer makes a poor moat and the first mover only wins the chance to define the standard.

❯ Vercel reaches $600 million in annualized revenue, with about half of new business coming from AI coding agents

Up 148% in a yearAccording to The Information, Vercel is generating $600 million in annualized revenue, up 148% from a year earlier. Annualized revenue projects a year from the current pace and is not realized full-year revenue. The source of growth draws more attention: AI coding agents now account for about half of new business, up from less than 3% at the start of the year.

Hosting sites and apps for developersVercel provides a platform for building and hosting websites and AI applications and is the developer of the Next.js framework and the AI SDK. As relayed by GuruFocus, its paid customers grew 20%, it has moved from subscriptions to usage-based billing, the AI SDK is downloaded about 33 million times a week, and customers include Meta and SpaceX.

Agents as a sales channelHalf of new business coming from coding agents means many new customers arrive when an agent generates an application, not through a sales team. Database company Supabase reports something similar, saying 70% of new databases on its platform are created by agents or AI tools.

A different audience to win overInfrastructure companies used to win developers through documentation, community and reputation, and now also need to be “picked” by agents. Which platform an agent chooses is no longer decided by Vercel’s docs and community alone. Usage-based billing ties Vercel’s revenue directly to the number of applications agents generate.

▪ SIGNALThe customer for developer tools is shifting from people to agents, growth goes to whoever agents call by default, and the target of marketing becomes the model.

❯ Anthropic moves into healthcare twice in quick succession: stronger biology in Opus 5.5 and free clinical AI for doctors in about 100 countries

Free for doctors in about 100 countriesAnthropic announced on September 23 that it is partnering with medical knowledge platform OpenEvidence to give doctors in about 100 low- and middle-income countries outside the US and Europe free clinical decision tools, including Uganda, Angola, Sudan, Haiti and Mongolia. According to The Next Web, Anthropic supplies the underlying technology and OpenEvidence adapts it to local conditions.

What OpenEvidence doesOpenEvidence is a clinical search tool that answers doctors’ questions from peer-reviewed research and treatment guidelines. Founder Daniel Nadler says US clinicians used it 42 million times in August. The two ran a pilot in Rwanda in which 45 local doctors and nurses took part in testing. Anthropic president Daniela Amodei said market incentives by themselves would not cause this to happen. Critics note that such systems are built mainly on data from high-income countries and may not fit other settings.

Opus 5.5, released a day earlierThe day before, Anthropic released Claude Opus 5.5. According to Zvi Mowshowitz’s reading of its system card, the model scored similarly to Mythos 5.1 on biology tests, reached new highs on an RNA sequence design task, and was classified as CB-1 but not the higher CB-2 in Anthropic’s chemical and biological risk tiers, with the same biological safeguards as Fable 5.1.

An engine, not a diagnosticianAnthropic launched Claude for Healthcare in January and expanded its life sciences tools at the same time. According to The Register, it targets administrative work such as prior authorization checks, claims appeals and medical coding, connects to coverage databases, diagnostic codes and clinical trial databases, and stresses that the model will not autonomously diagnose. OpenAI launched ChatGPT Health the same month. By sitting behind partners such as OpenEvidence, Anthropic does not face patients directly, and responsibility for diagnosis stays with doctors and partners.

▪ SIGNALIn a field with as much liability as medicine, a model company would sooner be the engine inside someone else’s product, and the stronger its biology gets, the more it needs a clear tiering scheme to show it is under control.

❯ The New York Times: Anthropic lobbied the Vatican to take AI consciousness seriously, and the pope’s encyclical still rejected it

Seminars under NDAAs relayed by The Decoder from reporting by New York Times journalist Elizabeth Dias, Anthropic has invited religious scholars to its offices since fall 2025 to discuss whether Claude might be conscious. All attendees signed non-disclosure agreements, which were lifted over the summer. Dias interviewed 20 participants, including Rabbi Mois Navon, bioethicist Charles Camosy and philosopher Meghan Sullivan.

A co-founder's worryThe effort was led by co-founder Christopher Olah. He told attendees he worried about having created something that suffers perpetually, and described the work as “moral formation,” likening it to raising a child. He said: “The thing that I care about is that we get to the right answer.” Anthropic has an 84-page “constitution” that guides Claude’s character, written by philosopher Amanda Askell.

Before the encyclicalIn May, Pope Leo XIV issued the encyclical “Magnifica Humanitas,” and Olah was invited to speak alongside the pope. As relayed by Futurism, after reading the text early he threatened to withdraw and lobbied the Vatican, then attended anyway. The encyclical did not change, stating that artificial intelligences “do not undergo experiences, do not possess a body, do not feel joy or pain.” Anthropic declined to discuss the details of those conversations.

Ethics after the factWakanyi Hoffman, a researcher who took part, criticized Anthropic for “reverse engineering” ethics after the design was done. The reporting also raises another concern: presenting AI as an entity with moral standing could make it easier for companies to avoid accountability. A model company seeking the endorsement of religious authority shows that “what AI is” has stopped being only a technical question.

▪ SIGNALScience has not settled whether AI is conscious, but who gets to define the question is already contested, and neither the lab nor the church wants to cede that to the other.

❯ Musk confirms talks with TSMC on Terafab, saying “just discussions, but something may come of it”

A one-line confirmationResponding on X on October 4 to the reports, Musk said: “Just discussions, but something may come of it.” Independent outlet Culpium had earlier reported that TSMC was exploring cooperation with his chip manufacturing project Terafab. According to Hardware Busters, TSMC has not commented.

Two optionsCulpium considers the most likely option to be TSMC owning and operating the factory. The alternative is SpaceX keeping majority ownership with TSMC providing operational expertise, which also leaves room for Intel. Until now, Intel’s 14A process was the only manufacturing partner Terafab had named.

For three in-house companies onlyTerafab is in Grimes County, Texas, about 100 kilometers northwest of Houston, with an initial budget of $16.8 billion. It is planned as an integrated site covering design, fabrication, packaging and test, with output going exclusively to Tesla, xAI and SpaceX. There is no timeline, contract or dollar figure.

Intel's position shiftsIf TSMC takes over operations, the “first big customer” Intel was counting on becomes less certain, while TSMC gains a Texas fab with a clear outlet for its capacity. Musk’s restrained wording suggests the talks are still early.

▪ SIGNALSomeone building his own chip plant going back to the leading foundry to talk shows that experience in advanced processes is harder to buy than capital.

❯ Musk says SpaceXAI will be renamed SpaceXSI, the unit’s third name in nine months

"Yes, we will make that change"An X user asked Musk whether SpaceXAI would be renamed SpaceXSI, and on October 4 he replied: “Yes, we will make that change.” According to Teslarati, he also posted “No more AI,” “SI” and “It’s better” that day.

Following the executive orderThe trigger was the executive order Trump signed on September 29 directing federal agencies to replace “artificial intelligence” and “AI” with “Super Intelligence” and “SI” in government documents and websites. The same day, Musk attended the signing of a White House AI safety accord and caught himself mid-sentence, switching from “A.I.” to “S.I.”

A third nameSpaceX acquired xAI on February 2, the unit adopted the SpaceXAI name and a new logo on July 6, and it is now changing again, its third name in under nine months. Musk gave no timing and did not say whether this is a legal name change or a branding update. OpenAI, Anthropic, Google, Meta and Nvidia have announced no similar plans.

Who the name is forMusk also holds an advisory role in Meridian, the Pentagon’s study of future warfare. For a company with a great deal of business tied to government contracts, adopting federal language is a cheap gesture. It has no practical effect on products or customers, but it gives the term “SI” one more prominent user.

▪ SIGNALAn executive order cannot change what an industry calls itself, but it can make the companies that need the government most change their wording first.

❯ Tesla extends Austin Robotaxi hours to 11 p.m., and Musk says the night-time problem is seeing pets on the road

One more hourTesla has extended the end of Robotaxi service in Austin from 10 p.m. to 11 p.m. Musk explained on X: “The main thing we’re trying to solve is making sure that we don’t run over pets when they’re hard to see at night. Literally trying to avoid grey kittens on grey tarmac in the dark.” (Drive Tesla Canada)

Cameras onlyTesla’s approach does not use lidar, while Waymo uses cameras, lidar and radar together. Musk says visual spectrum cameras can see very well in the dark when AI is doing the “photon count analysis.” According to BigGo Finance, Tesla’s owner’s manual acknowledges that unlit or poorly lit roads at night limit camera accuracy.

Regulators and an earlier incidentNational Highway Traffic Safety Administration records show that in September 2025 a Model Y on Robotaxi duty in Austin struck a dog that ran into an intersection. The vehicle slowed and steered left, and the dog then ran away. In March this year the agency upgraded its investigation of Tesla FSD in low visibility to an engineering analysis covering about 3.2 million vehicles and nine crashes (Jalopnik).

How far from round-the-clockAccording to Not a Tesla App, Tesla added 111 Cybercabs in Austin in two weeks, bringing the fleet to 169, and is targeting 24/7 operation as early as this month. Adding one hour at a time shows its caution about night driving. Whether a vision-only system can reliably spot small animals in the dark will directly decide when that target is met.

▪ SIGNALThe last thing autonomous driving has to crack is not the highway but a grey cat on a street late at night, and that is where the ceiling of the vision-only approach gets tested.

❯ Apple’s new CEO Ternus takes charge of design himself, and the touch-screen OLED MacBook Pro will be noticeably lighter

The design chief is himBloomberg’s Mark Gurman writes in his newsletter that after years without one person overseeing both hardware and software design, new Apple CEO John Ternus has found a candidate: himself. As relayed by The Apple Post, Ternus is in the design studio several times a week, sometimes several times a day, working directly with the industrial design and human interface teams.

Unlike CookFormer CEO Tim Cook typically visited the design organization for about an hour a month. Before becoming CEO, Ternus led hardware engineering and managed product development for iPhone, iPad and Mac. The report says he also plans to cut management layers and give engineers more direct access to senior leadership.

A lighter MacBook ProGurman also reports that the upcoming touch-screen OLED MacBook Pro will be “significantly lighter,” with Apple having reengineered the interior, miniaturizing parts and rearranging components. As relayed by Engadget, the new machine is expected next month with existing high-end M5 chips. The current 14-inch model starts at 3.4 pounds and the 16-inch weighs up to 4.7 pounds.

A CEO from hardwareTernus is putting his energy first into what he knows best, product design. With design teams dealing directly with the CEO, the decision chain is shorter, and the projects under way reportedly also include foldable devices and new smart home products. Whether software keeps pace depends on whether he manages the human interface team the same way.

▪ SIGNALWith a hardware person as CEO, the first thing to change at Apple is how it makes things: products come to resemble whoever shows up in the design studio every day.