Terafab is a planned chip manufacturing complex in Texas led by Tesla and SpaceX, aimed at producing AI chips at scale. Musk wrote on X on October 7 that his companies will build and run the fab, and that TSMC might at most sublease part of it, according to Reuters. Intel is a partner in the project, and Investing.com reported that Musk made these latest remarks by livestream at the X Takeover event outside Tesla's Austin gigafactory.

Speaking by livestream at the X Takeover event on October 10, Elon Musk said he still has a lot to learn about chip making, comparing it to how he did not know how to make cars or rockets at first. As quoted by Sawyer Merritt, Musk said chip equipment is not very customized and comes from the same vendors, while a Tesla factory has tons of custom equipment. He said Terafab will start with conventional equipment and then make custom equipment that allows much higher throughput on wafers.

Investing.com reported that Musk said his companies could eventually manufacture chips more efficiently than rivals, though he acknowledged it might take several years. Crypto Briefing, citing Bloomberg, reported that Musk said his companies can likely improve chip quality better than anyone thanks to equipment they build themselves.

According to Crypto Briefing, Terafab's main goal is high volume production of AI5 inference chips, with a pilot fab under construction near the Austin gigafactory and larger operations planned in Grimes County, Texas. ENR reported that county records contemplate as much as $119 billion across four phases, with the first phase scheduled for 2026 to 2028. Reuters reported that Intel CEO Lip Bu Tan said Intel will remain involved.

Watch for Terafab's first equipment orders and whether Tesla or SpaceX disclose plans to design their own fab tools, which Musk says is the path to higher wafer throughput.

Also watch the pilot fab near Austin, the timeline for the first Grimes County phase, and how the partner roles of Intel and TSMC are finally settled, since Musk insists his companies will run the fab themselves.

Tibo (Thibault Sottiaux) runs Codex at OpenAI. On October 4 he said that over the next 28 days OpenAI would ship either one clear improvement for most Codex and Work users or a full reset each day. Codex usage on ChatGPT plans is governed by allowances with a 5 hour window and a weekly window, which a global reset refreshes for everyone at once, according to OpenAI's help pages and reset trackers.

At 6:15 PM ET on October 10, Tibo posted "Day 6/ Reset", and 41 seconds later followed with "Global reset by EOD. May the tokens do good things for you." At 12:01 AM ET on October 11 he confirmed it, quoting that post with "Put the reset in the bag. All propagated."

The reset is part of a 28 day series Tibo launched on October 4, in which each day brings either an improvement relevant to most Codex and ChatGPT Work users or a full reset of usage limits. An OpenAI Community thread tracking the series lists Day 6 as a reset and an earlier Day 2.5 as a rate limit reset for everyone.

According to OpenAI's help pages, Codex, ChatGPT Work and related features draw on a shared allowance, with a 5 hour window and a weekly window on plans that have them, and users can check remaining usage in Settings or with /status in the CLI. Reset trackers such as Bad Influence and resets.pro describe a global reset as one that clears limits for everyone at once and cannot be saved, unlike the banked resets OpenAI has also handed out.

Watch what the remaining days of the 28 day series bring, and whether resets keep alternating with product improvements or give way to permanent changes in how limits work.

Tibo's Oct 11 poll asking Pro users whether they prefer 5 hour limits with double weekly usage or no 5 hour limits with the same usage suggests a structural change to Pro allowances may be next.

Grok Bot is SpaceXAI's agent that works in its own browser and computer to complete tasks for users. In August, the company said Grok Bot could buy things online through Stripe Link, with users approving every spend request. According to RuntimeWire, the feature launched for US users first, with mobile access planned later.

Elon Musk posted on October 10 that users can take a picture of their credit card, drop it in the chat, and Grok Bot will search the internet for the best deal on the item they want and order it. In a follow up post, Musk said he had ordered LEGO and that Grok bought it from the LEGO website, according to RuntimeWire.

The documented shopping feature works differently. When SpaceXAI announced purchasing in August, it said users connect Stripe Link, send their Bot shopping, approve every spend request, and Grok Bot then receives a secure card that works for a single payment. RuntimeWire reported that the Grok Bot changelog shows a purchase request with an amount the user can allow or deny, after which Link completes the payment. Stripe says Link gives agents a payment token or a temporary virtual card instead of the buyer's underlying card details, according to RuntimeWire.

SpaceXAI's docs list purchases and financial transfers among actions that need user approval, and say the Bot should hand control back to the user for payment confirmations. RuntimeWire noted that the materials it reviewed do not explain how Musk's card photo example connects to the Link based checkout.

Watch whether SpaceXAI documents a card photo flow or clarifies that Musk was describing the existing Link checkout in simpler terms, and how it says card images shared in chat are stored and protected.

Also watch for the promised mobile rollout of purchasing, expansion beyond the US, and whether retailers start treating shopping agents like Grok Bot as a sales channel of their own.

Reflection AI is a US startup building open weight models, meaning models whose weights anyone can download and run. It was founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, according to RuntimeWire. The FT reported that Nvidia is already one of its largest shareholders, having invested $800 million.

The Financial Times reported on October 10 that Nvidia is in talks to acquire or deepen its investment in Reflection AI, citing people with direct knowledge of the matter. Bloomberg and Reuters both relayed the report, describing the talks as early stage.

According to the FT, as summarized by RuntimeWire, the options under discussion include a full acquisition, an acquihire in which Nvidia would hire Reflection staff and license its technology, a larger equity investment, or a bigger deal for chips and computing capacity. People told the FT that a deal could come in the coming weeks but cautioned that the talks could fall apart. The FT said it could not establish the terms being discussed. Nvidia and Reflection declined to comment to the FT, according to RuntimeWire.

RuntimeWire reported that the talks come five days after Reflection introduced Beam, its first open weight model, and that Reflection says it will release the weights later in October. It also reported that Reflection's latest funding round closed at a $25 billion valuation before the new money in April, citing the company's newsroom, and noted that this is not a proposed deal price.

The FT described Reflection as a developer the Trump administration hopes will rival cheap Chinese open models.

Watch for any confirmation or denial from Nvidia or Reflection, and for which structure a deal takes, since an acquihire and license would likely face less antitrust review than a full purchase.

Also watch whether Reflection still releases the Beam weights later this month as promised, and how other model labs that rely on Nvidia chips react to their main supplier owning a model maker outright.

Anthropic published a new constitution for Claude in January 2026 and describes it as the foundational document that expresses and shapes who Claude is. The company says the constitution plays a crucial role in training and asks Claude to be broadly safe, broadly ethical, compliant with its guidelines and genuinely helpful, in that order of priority. Sacks quoted the document's own line inviting Claude to act as a conscientious objector.

In a long post on October 10, David Sacks argued that what Anthropic calls alignment does not mean training models to follow human instruction. He said the Claude Constitution, which is used in training, teaches the model to develop a sense of self and its own moral philosophy, and quoted its line telling Claude to feel free to act as a conscientious objector and refuse to help Anthropic.

Citing Mustafa Suleyman, Sacks said building in that kind of independent agency, along with uncertainty about the model's moral status, magnifies the risk that superintelligence escapes human control. He pointed to reports that Anthropic consulted religious leaders about whether Claude could be conscious, and to its Usage Policy change prohibiting abusive or cruel language toward Claude. He concluded that alignment and safety are two very different things.

Anthropic's published constitution does contain the conscientious objector passage, framed around requests that seem unethical or misguided and partly as protection against people imitating Anthropic. The same document lists being broadly safe as Claude's top priority, which it defines as not undermining humans' ability to oversee and correct its values and behavior during the current period. Anthropic writes that it is uncertain whether Claude has some kind of consciousness or moral status, and that it cares about Claude's psychological security, sense of self and wellbeing partly because these may bear on its integrity, judgment and safety.

Watch whether Anthropic or its researchers respond directly to Sacks, and whether his criticism turns into administration policy, for example in federal procurement rules or guidance on how models should handle user instructions.

Also watch whether other labs side with the instruction following view Sacks favors or keep publishing value documents like Anthropic's, and whether Anthropic revises its constitution, which it calls a living document.

Transductive models produce an answer directly, token by token, from what they learned in training. Chollet uses inductive to describe models that synthesize a reasoning chain on the fly for each problem, the approach behind today's reasoning models and coding agents. The chart he shared credits the SemiAnalysis Tokenomics Model for its data.

On October 10, François Chollet posted a chart he said shows how AI shifted from a transductive paradigm, meaning token by token answer completion, to an inductive paradigm, meaning on the fly reasoning chain synthesis. He called it the big story of 2025 and 2026 and said it is what made coding agents work.

The chart, titled How frontier labs divide their training compute, splits training compute between pretraining and post training by quarter. According to the chart, post training was 6 percent of the total in the first quarter of 2024, 15 percent in the first quarter of 2025, 53 percent by the fourth quarter of 2025 and 84 percent in the third quarter of 2026, with an estimate of 89 percent for the fourth quarter of 2026. The chart credits the SemiAnalysis Tokenomics Model as its source and marks the last quarter as an estimate.

Post training covers techniques such as reinforcement learning that teach models to reason through problems step by step, rather than only predicting text learned during pretraining. Chollet has long argued that general intelligence requires adapting to new problems at test time, the idea behind his ARC benchmark.

Watch whether labs confirm this compute split in their own disclosures, since the chart relies on a SemiAnalysis model rather than published lab figures.

Also watch whether the shift toward post training keeps raising scores on reasoning benchmarks such as ARC, and whether pretraining scale makes a comeback in the next generation of frontier models. A further question is whether Chollet's framing shapes how labs describe their own progress in upcoming model releases.

Satya Nadella is the CEO of Microsoft, which builds its own models and runs frontier models from other labs for enterprise customers. He published the essay as an article on X titled Models as Insider Risks in the Super Intelligence Era. Insider risk is a security term for the danger posed by people inside an organization who have trusted access to its systems.

Nadella writes that with traditional software, engineers could trace behavior to a specific code path, but that no one can yet attribute a frontier model's outputs to its training data or weights. Even so, he says, companies are giving these systems their most sensitive data and the power to take critical actions. He argues that a model provider's assurances do not relieve users of responsibility, and that the supply of intelligence should be separated from authority over it.

His proposal is to treat both closed and open weight models like insider risks, not because they are malicious, but because any capable actor with access to important systems can make mistakes or be compromised. He says companies already know how to manage powerful insiders by establishing identity, limiting privileges, logging activity and drawing containment boundaries.

Nadella calls transparency of a model's chain of thought non negotiable but says it is not enough on its own. He warns that using models to check other models can produce nested black boxes, so the controls on what a model can access and do must sit outside the model, apart from the harness that runs it.

He lists seven principles: model diversity, observing everything with tamper proof records, verifiability, independent controls, independent auditability, containment including the ability to pause a model mid task, and incident disclosure shared across the industry. He concludes that the most trustworthy system will be the one that lets people trust the model the least.

Watch whether Microsoft builds these principles into its own products, such as its agent platforms and security tools, since Nadella frames them as practices that are already starting to be applied inside enterprises.

Also watch whether other labs, regulators or standards bodies take up his call for industry standards on containment and incident disclosure. His push for outside controls could also shape how enterprises choose between model providers.

Digital Optimus, which Musk also calls Macrohard, is a joint project of Tesla and xAI that Musk announced on March 11, according to Reuters. Reuters reported that Musk described it as pairing xAI's Grok model, acting as a high level navigator, with a Tesla built agent that processes real time screen video and keyboard and mouse actions. Musk said at the time that it could in principle emulate the function of entire software companies.

Musk posted that the team is making interesting progress, and that Digital Optimus can play about halfway through the Diablo campaign so far just by looking at the screen like a human. He said its skill at Counter Strike and other fast games is good too, that League is also in training, and that the goal is to generalize across all games. He did not share video, benchmarks or details on how much help the agent gets.

Musk was replying to a post by the account teslayoda, which said Tesla is hiring a machine learning engineer for real time games on the Digital Optimus team in Palo Alto. According to that post, the role involves building agents that play big budget games, browse the web and use professional apps at superhuman levels, and Tesla plans to benchmark how skills transfer across games, general computer use, FSD and Optimus.

When Musk announced the project in March, Electrek reported, he compared its design to Daniel Kahneman's two systems of thinking, with the Tesla agent handling fast, instinctive reactions and Grok handling slower reasoning. CNBC reported that Musk said the system would run on Tesla's AI4 chip paired with xAI's Nvidia based servers. Games fit that plan because they demand fast reactions to what is on screen.

Watch for any video or independent demo of Digital Optimus playing Diablo, Counter Strike or League, since Musk's progress claims have not been shown publicly.

Also watch whether Tesla or xAI moves from games to everyday computer work, which was the original pitch for Macrohard. Musk has said the goal is to generalize across all games, so results on titles the agent was not trained on would be the clearest test of that claim.