H.Headline
Chaos

THE EDITION ARCHIVE

A front page, saved in time.

Making AI Headlines · PM edition

Published · New York time

Permanent edition link →

10 pages · 23 stories

PAGE 1 OF 10

OpenAI Adds a Lab Bench to the Chat Window

The Orchestra Bio acquisition is intended to add people and technology to OpenAI’s Rosalind life sciences tools.

Making AI Headlines newspaper, page 1: OpenAI Adds a Lab Bench to the Chat WindowPage 1 of 10
Page 1 of 10. Read the same stories below.

The Chatbot Would Like to See Your Experiment Plan

OpenAI is acquiring Orchestra Bio, saying it will add the startup’s team and technology to Rosalind, its AI model for life sciences research, and the related Rosalind Workbench. The workbench is intended to help researchers plan experiments and analyze results, moving the company further into tools for scientific work. That is a different assignment from making a paragraph sound more confident. An experiment must contend with materials, measurements and the stubborn refusal of physical reality to accept a revised prompt. The useful question is whether AI can make research planning and interpretation more effective, not whether it can describe a laboratory with impressive enthusiasm. The acquisition also illustrates how AI companies are assembling specialized products around their models. A general assistant can answer questions. A research workbench needs to fit the tasks scientists actually perform, including the parts that do not resemble a chat conversation. The comic promise is a lab partner that never needs coffee. The practical test is whether it helps the people who still need controls, reproducible methods and results that survive scrutiny. Biology remains an unusually demanding customer: it does not award partial credit for a beautifully formatted explanation.

Share articleOpen article permalink →

Your Dashboard Can Now Know You Uncomfortably Well

Google is bringing Gemini’s optional Personal Intelligence feature to Android Auto and Google Built-in. The feature lets Gemini use more personal data to tailor answers, extending a capability that began as a paid offering before becoming available to all Gemini users. For drivers, the appeal is straightforward: an assistant with relevant context may require less explanation. The tradeoff is equally straightforward: personalization depends on information that a less personalized assistant would not use. A car is now another place where a settings review can matter as much as a software update. This is not simply a contest to make the dashboard sound friendly. It is a question of what information belongs in a particular interaction, and whether the convenience justifies connecting it. Optional participation is important because people can reasonably want navigation help without inviting their broader digital lives into the passenger seat. The old car asked whether you wanted to avoid tolls. The new car could aspire to understand why you avoid commitment. Drivers do not need to accept that entire emotional upgrade package. Sometimes a useful assistant is one that knows the destination and leaves the rest of the family business parked.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 2 OF 10

AI Gets Cheaper to Run, Harder to Keep Private

DeepSeek releases its inference framework, while Meta pauses conversation sharing after researchers find public URLs.

Making AI Headlines newspaper, page 2: AI Gets Cheaper to Run, Harder to Keep PrivatePage 1 of 10
Page 2 of 10. Read the same stories below.

DeepSeek Opens the Engine Room

DeepSeek has open-sourced DeepInfer, the inference framework behind its V4.1-Flash model. The vLLM-based system reduces computing costs and energy use when running AI models, adding competitive pressure on Western providers including OpenAI and Anthropic. Inference is the work that happens when a model responds, rather than when it is initially trained. Making that process more efficient matters because a service must pay to produce answers repeatedly. A brilliant model with an expensive operating routine can resemble a restaurant that needs a new oven for every sandwich. An open framework also gives developers something more concrete than a promise of future affordability. They can examine a component of the operating machinery, although using it still requires engineering judgment and an assessment of suitability for their systems. The business stakes are not limited to which chatbot tops a comparison chart. If efficient infrastructure becomes more accessible, providers face pressure to explain what customers receive beyond access to raw computation. The industry has spent years teaching machines to write elaborate answers. Now the less glamorous contest is teaching the bill to use fewer words. That competition may be especially compelling to anyone responsible for approving the bill.

Share articleOpen article permalink →

The Share Button Discovers an Audience

Meta says it paused a Muse feature for sharing AI conversations after researchers found conversation URLs publicly available on the web. Security firm Claroc reported finding more than 6,800 Muse links through public DNS records, bringing attention to the boundary between sharing a conversation and making it broadly discoverable. That boundary matters because an AI chat can feel like a private workspace even when a product offers ways to distribute it. Generating a link does not necessarily give users an intuitive understanding of where that link might appear or who might encounter it. The practical product challenge is therefore larger than adding a friendly icon. Sharing controls need to communicate audience and visibility in terms ordinary people can understand. A button labeled with a cheerful verb cannot carry the entire burden of privacy education. The pause comes while Meta investigates the feature’s privacy. The broader lesson applies across conversational products: a chat window is not automatically a locked room. It may be a room with an export function. Software has spent decades making publication effortless. It is still working on making the consequences equally legible. The ideal sharing interface would offer one additional confirmation: yes, other people are included in the concept of sharing.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 3 OF 10

Washington Checks the Software Family Tree

Federal civilian agencies must remove DeepSeek technology, including covered third-party services, while legacy code gets new AI attention.

Making AI Headlines newspaper, page 3: Washington Checks the Software Family TreePage 1 of 10
Page 3 of 10. Read the same stories below.

GPT-6 Studies the Languages That Never Retired

OpenAI says its latest models perform substantially better on programming languages such as COBOL, Fortran and Visual Basic. On its new MultiLangBench, GPT-6 Astra achieved an 87.4 percent test pass rate across 20 languages, compared with about 38 percent for GPT-5.6 Terra. The improvement targets a practical problem: organizations do not replace important software simply because its language has stopped appearing on fashionable conference merchandise. Older systems can remain central to operations long after the people who first built them have moved on. Better assistance could help developers understand and maintain that code. But a benchmark result is not a guarantee that an entire migration will work. Real systems include dependencies, business rules and accumulated exceptions that rarely fit neatly into a test question. The opportunity is less about making old languages trendy than making essential maintenance less difficult. That is a respectable ambition in an industry that often treats maintenance as the interval between launch announcements. The future of coding may therefore include a machine patiently learning the past. Somewhere, a decades-old program can finally enjoy the satisfaction of watching a cutting-edge model prepare for an exam it has been administering since before the internet became popular.

Share articleOpen article permalink →

Federal IT Gets a Technology Ancestry Assignment

The United States has ordered federal civilian agencies to stop buying and remove DeepSeek products and services from government systems. The Department of Homeland Security cited national security concerns, including the potential accessibility of sensitive government information to Chinese authorities under Chinese law. The order also covers third-party products and services built using DeepSeek technology. Agencies can request exceptions. That wider scope makes the task more complicated than finding an application with a recognizable name and clicking uninstall. For procurement and technology teams, the relevant question becomes what is inside the software they use. A service may arrive through a vendor while depending on technology from another company, creating a chain that is not obvious from its branding. The policy makes software provenance an operational concern rather than an abstract diagram in a presentation. Agencies must connect purchasing decisions with the underlying technology and the security concerns identified by DHS. In the consumer world, a product’s ancestry is often a marketing story about founders and garages. In federal IT, it can become a spreadsheet with consequences. The software family reunion now requires a guest list, a review process and someone willing to explain who brought the embedded model.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 4 OF 10

The Cloud Has a Competitive Weather System

Anthropic reportedly faces compute constraints as cloud partners increasingly see a rival across the negotiating table.

Making AI Headlines newspaper, page 4: The Cloud Has a Competitive Weather SystemPage 1 of 10
Page 4 of 10. Read the same stories below.

Your Supplier Would Also Like Your Customers

Anthropic is reportedly struggling to obtain enough computing capacity from cloud providers that increasingly view it as a competitor. The reported squeeze comes as infrastructure spending grows ahead of a planned November initial public offering. The situation exposes a structural tension in the AI business. A model company may depend on cloud firms for the resources needed to serve customers, while those same firms pursue their own AI ambitions. The relationship must accommodate both cooperation and competition, sometimes during the same procurement conversation. For a growing service, securing capacity is not a decorative expense. It affects the ability to deliver the product. For the supplier, allocating scarce resources involves its own commercial priorities. Neither side can solve that tension by calling the relationship an ecosystem often enough. The useful business analysis concerns dependence: who controls a critical input, how many alternatives exist, and how easily a company can change suppliers. A planned public listing adds another audience interested in those answers. Cloud computing sounds weightless until somebody needs more of it. Then it begins to resemble a crowded apartment building where the landlord also runs a competing business and would like to discuss your electricity usage before renewing the lease.

Share articleOpen article permalink →

SpaceX Finds Considerable Revenue Below Orbit

SpaceX reported $14.6 billion in third-quarter revenue and $4.4 billion in net income, with AI providing a major contribution. Its renamed Super Intelligence business generated $7.5 billion in revenue from operations including Cursor, Grok and its expanding neocloud business, according to its regulatory filing. The compute operation now has 1.7 gigawatts of capacity. Those figures put a distinctly terrestrial business at the center of a company long associated with spaceflight. The servers may support ambitious software, but they still require power and physical infrastructure on Earth. The practical analytical task is to look beyond the corporate label. A company’s familiar identity does not necessarily describe its current revenue mix. Investors and customers assessing the business need to distinguish its different operations rather than assume every dollar arrived wearing a spacesuit. The AI contribution also shows how computing infrastructure can become a substantial business alongside consumer-facing agents and coding tools. Not all of the action takes place in the chat window. There is a useful comedy in a space company finding so much business in rooms full of machines that must never get too hot. Humanity reaches for the stars, then checks the cooling equipment.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 5 OF 10

One Word Leaves the Terms of Service

Meta removes “train” from revised commercial terms after criticism, without adding a commitment to respect robots.txt.

Making AI Headlines newspaper, page 5: One Word Leaves the Terms of ServicePage 1 of 10
Page 5 of 10. Read the same stories below.

The Redline Is Smaller Than the Argument

Meta has removed the word “train” from commercial terms scheduled to take effect October 30 after criticism of the policy. The company has not added language committing to respect robots.txt, the website mechanism used to communicate crawler preferences. Legal experts had warned that the earlier language appeared to grant broad permission to scrape websites linked from business pages. Experts speaking to Nieman Lab suggested the policy could allow Meta to obtain material for AI training without the licensing payments publishers might otherwise seek. Those were assessments of the terms, not established outcomes. The change highlights the gap between editing a document and resolving the commercial questions behind it. Businesses need to understand what permission they grant by maintaining a page, while publishers have an interest in how their linked material may be used. A single removed word can be meaningful, but it is not a substitute for examining the remaining language. The relevant unit of analysis is the policy, not the size of the redline. Terms of service occupy a peculiar literary genre: consequential enough to govern a relationship, yet presented in a form that makes a refrigerator manual feel like a beach read. This revision has given one verb an unusually eventful career.

Share articleOpen article permalink →

AI Learns to Reply Before the Awkward Pause Gets Comfortable

ElevenLabs has released two Eleven Realtime speech-to-speech models for live AI conversations. The company reports that its v1 Max variant scored 98.8 percent on the Big Bench Audio benchmark with a response time of 0.57 seconds. ElevenLabs says no model on the Artificial Analysis leaderboard is both faster and more accurate. That is a competitive claim about a particular comparison, rather than a promise that every spoken exchange will work equally well. Speed matters in voice interfaces because a conversational delay is difficult to ignore. A written response can arrive while someone looks elsewhere. A delayed spoken answer leaves a person holding the floor, wondering whether to repeat the question or apologize to the appliance. Accuracy matters just as much. A fast misunderstanding is not an improvement merely because it reaches the wrong destination efficiently. Real usefulness depends on handling what people actually say and responding appropriately within the flow of a conversation. The industry is therefore competing to remove the pause without removing the meaning. This could make voice assistants more practical. It could also deprive users of their brief, peaceful interval between asking a machine something and hearing it explain how happy it is to help.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 6 OF 10

Developers Get an Exit Ramp

Google joins Amazon in offering tools to move PyTorch applications away from Nvidia GPUs.

Making AI Headlines newspaper, page 6: Developers Get an Exit RampPage 1 of 10
Page 6 of 10. Read the same stories below.

The Hardware Relationship Becomes Less Exclusive

Google has joined Amazon in releasing a developer tool designed to make it easier for PyTorch users to move applications away from Nvidia GPUs. The development challenges Nvidia’s position not just through hardware competition, but through the software developers use to work with it. A computing platform’s strength depends partly on how much effort it takes to leave. Developers build applications, learn tools and establish routines around available systems. Even an appealing alternative can struggle if migration demands a large engineering project before useful work resumes. Tools that reduce that burden can change the assessment. They do not automatically make every alternative preferable, but they can make comparison more practical. Teams still need to evaluate performance, compatibility and operating requirements for their own applications. That is why a developer utility can carry competitive significance beyond its modest appearance. The less dramatic piece of software may help determine which expensive machines a customer can realistically use. The industry likes to celebrate breakthroughs with grand language about intelligence. This contest has a simpler theme: making it easier to take your work elsewhere. In technology, as in apartment hunting, the existence of a usable exit can improve the entire relationship.

Share articleOpen article permalink →

ChatGPT Adds a Door to the Conversation

ChatGPT now lets users lock chats. OpenAI says the feature is intended for situations such as discussing personal matters or planning a surprise, when someone might hand a device to friends or family to show a different conversation. That is a small, recognizable privacy problem. A person can willingly share a screen without intending to share everything accessible from it. The social convention of handing over a phone is not equivalent to opening the entire filing cabinet. The feature addresses that device-sharing situation. Its practical value lies in separating conversations that may otherwise sit close together in the same interface, where an innocent tap can become an unexpectedly comprehensive tour. The broader product lesson is that privacy includes ordinary household interactions, not only distant technical threats. People need controls that fit the way devices move between hands at dinner tables, on couches and during attempts to demonstrate something amusing. A chat assistant may help plan a surprise party, but the surprise traditionally works better if the guest of honor cannot discover the planning thread while inspecting a recipe. After years of expanding what assistants can do, software is still finding useful work in the ancient technology of a closed door.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 7 OF 10

Ownership Stakes Meet Meeting Stakes

A reported Anthropic holding illustrates the scale of AI ownership, while Otter.ai secures new growth capital.

Making AI Headlines newspaper, page 7: Ownership Stakes Meet Meeting StakesPage 1 of 10
Page 7 of 10. Read the same stories below.

One Percent Enters the Conversation

Former Google CEO Eric Schmidt owns about a 1 percent stake in Anthropic, Bloomberg reported. The reported holding is roughly half the size of CEO Dario Amodei’s stake, offering a glimpse into ownership at a prominent AI laboratory. Percentages can sound small without being economically trivial. Their significance depends on the underlying business, the terms attached to ownership and the rest of the capitalization structure. A reported percentage alone does not establish operational control or explain how decisions are made. The useful distinction is between ownership and management. A stake represents an economic interest; the responsibilities and authority of running a company are a separate matter. Collapsing those ideas produces a dramatic story more quickly than it produces a good business assessment. The report also underscores how familiar technology figures can remain connected to the industry through investment, even when their best-known executive roles are elsewhere. The modern corporate map is not a neat row of separate buildings. Financial language has a gift for making large interests sound modest. One percent is the kind of number a phone battery displays just before inducing panic. In a major AI company, it is better understood with considerably more context and considerably less sympathy for the percentage.

Share articleOpen article permalink →

Otter Raises Money to Help Capture the Meeting About Growth

Meeting note-taking company Otter.ai has raised $320 million from General Catalyst’s Customer Value Fund. The fund provides capital without diluting equity and receives returns tied to the results of marketing spending. Otter plans enterprise growth, international expansion, product innovation and go-to-market investment. That structure differs from a conventional equity investment. Rather than exchanging a portion of ownership for cash, the arrangement connects the fund’s return to a defined commercial activity. The distinction matters when assessing both the capital received and the obligations associated with it. For a meeting software business, growth requires more than demonstrating that a machine can produce notes. Customers must decide whether the tool fits their work, and the company must reach those customers at a cost that makes business sense. Marketing performance is therefore part of the operating question, not merely an announcement-day accessory. The broader appeal is understandable: companies can seek resources for expansion without issuing more ownership. That does not make capital consequence-free. It changes the form of the tradeoff. Otter’s category has an especially tidy comic symmetry. Money will support software that records workplace discussions, potentially including discussions about whether workplace software is growing efficiently. At least the follow-up notes should have a clear business model to summarize.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 8 OF 10

The Voice Agent Gets More to Say

Mistral’s Voxtral 3 combines speech tools, multilingual support and access to external tools during live conversations.

Making AI Headlines newspaper, page 8: The Voice Agent Gets More to SayPage 1 of 10
Page 8 of 10. Read the same stories below.

Listening, Thinking and Talking Become a Software Stack

Mistral AI has launched Voxtral 3, a family of models for real-time voice agents. The system combines speech recognition, speech generation and a voice agent software development kit, supports more than 50 languages, and allows agents to use external tools during live conversations. Developers can run it on their own infrastructure or use EU-hosted APIs. Those deployment choices give organizations another factor to assess alongside speech quality: where their conversational systems operate and how they manage data. Putting the components together addresses a practical developer problem. A spoken service needs more than a pleasant voice: it must understand incoming speech, decide what to do and communicate the result without making the conversation feel like a series of disconnected applications. Tool access adds another layer. An agent that can consult outside systems may be more useful than one limited to conversation, but smooth speech cannot compensate for an inappropriate action or a misunderstood request. Multilingual support broadens the potential audience. The real test is whether the product serves the people trying to use it. For developers, the sales pitch is an integrated stack. For callers, the request is more modest: please understand the question before launching into the answer. The future may speak fluently, but it still needs to take turns.

Share articleOpen article permalink →

Muse Joins the Fast-Download Club

Meta’s AI companion Muse surpassed 10 million U.S. downloads in its first 40 days. The milestone places it in a group reached by only 11 other apps, including Threads, while new data also indicates stronger user retention than its AI rivals. Downloads establish the scale of initial interest. Retention addresses a different question: whether people continue using an application after curiosity has done its promotional work. Both matter when assessing a consumer product, but they should not be treated as interchangeable measures. An AI companion faces a particular version of that challenge. Getting someone to try a conversation is one task; becoming a recurring part of that person’s routine is another. The reported retention comparison suggests attention beyond a single visit, without making every download a permanent relationship. For the wider industry, the figures provide evidence that conversational apps can attract large U.S. audiences quickly. They also raise the competitive bar for products trying to turn a novelty into a habit. The app store has become a place where a new companion can arrive with a download button and an onboarding screen. Ten million introductions are impressive. The next commercial question is whether users keep showing up after the digital small talk ends.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 9 OF 10

Safety Plans Get Very Large Organizational Charts

Sam Altman reportedly outlined three AI safety paths, calling the option of one dominant company “terrible.”

Making AI Headlines newspaper, page 9: Safety Plans Get Very Large Organizational ChartsPage 1 of 10
Page 9 of 10. Read the same stories below.

Three Safety Paths, All Requiring a Bigger Diagram

OpenAI CEO Sam Altman reportedly outlined three paths to safe AI in an internal presentation: an international watchdog modeled on the IAEA, a merger of major laboratories, or one company powerful enough to shut down everyone else’s computing. He called the last option “terrible.” The reported alternatives place governance directly beside technical safety. An international body, consolidated laboratories and a dominant company would involve very different arrangements for authority and accountability. Each raises questions about who makes decisions, who checks those decisions and what happens when participants disagree. Concentrating power is not simply another engineering setting to adjust. The organizational chart would certainly be impressive. The difficult box is the one labeled “Who supervises this box?”

Share articleOpen article permalink →

The Houseplant Has Entered the Data Economy

Startup Electric Plant has built an AI-powered translator intended to help humans interpret the subtle electrical signals moving through plant stems. The project treats those signals as information that could make plant activity more understandable. The important distinction is between interpreting biological signals and assuming plants converse like people. A useful system would connect signals with meaningful observations, not merely give a fern an entertaining personality. The idea turns a familiar domestic relationship around. Humans have long spoken to plants without expecting a reply. Now software seeks to make the listening part more informative. If the office plant eventually communicates anything, the strongest comic candidate remains a concise request to stop scheduling meetings beside it and start remembering the water.

Share articleOpen article permalink →

Spotify Clears Some Traffic From the Recommendation Lane

Spotify says it removed more than 50 million tracks considered AI-generated slop or spam from user recommendations over the past year. It also announced recommendation changes to highlight more music made by human artists. The action concerns recommendations, the selections listeners encounter when an algorithm chooses what to present next. That makes filtering a central part of the listening experience, alongside the size of the available catalog. Cheap, abundant production increases the importance of distinguishing worthwhile selections from content that simply exists in enormous quantities. The modern jukebox has an unusual maintenance problem: it does not run out of songs. It risks becoming so full that finding something worth hearing requires a cleanup operation with eight digits in its workload.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑

PAGE 10 OF 10

Useful AI Meets the Small Print

Figma weighs standards and data access as new tools tackle document extraction, photo edits and robotic handling.

Making AI Headlines newspaper, page 10: Useful AI Meets the Small PrintPage 1 of 10
Page 10 of 10. Read the same stories below.

Figma Wants Connections, Not an Open Filing Cabinet

Figma co-founder Dylan Field says MCP is not a perfect protocol. The company plans to support as many standards as possible without exposing all its data. That frames a practical tension in connected AI tools: interoperability can make a product more useful, while broader access creates decisions about what should remain protected. Supporting a standard is not the same as granting unlimited access. The ideal software handshake exchanges what is needed. It does not arrive carrying a moving truck for the entire office.

Share articleOpen article permalink →

The Group Photo Gets Another Chance to Look Awake

Adobe’s mobile Lightroom can now use generative AI to open closed eyes in photos with a tap. The feature is available on iOS, iPadOS and Android. It addresses the familiar group-photo problem in which everyone cooperates except one pair of eyelids. The resulting image is an edit, rather than a recovered record of exactly what the camera captured. That distinction matters even when the purpose is harmless convenience. Family photography can now negotiate with blinking, though it still cannot make everyone agree on which picture looks best.

Share articleOpen article permalink →

A Robot Arm Impresses by Picking Things Up

Eka Robotics has reached a reported valuation of $1.45 billion as its physical AI model draws enthusiasm for handling a wide range of objects. Its robotic arm’s smooth movements have attracted industry attention. Picking things up sounds basic until a machine must cope with different shapes and positions. Breadth of handling is therefore a meaningful engineering challenge, separate from the financial assessment reflected in a valuation. After years of grand promises about intelligence, the industry can still be delighted by a machine competently doing what a toddler attempts before breakfast.

Share articleOpen article permalink →

Cloudflare Gives Document Extraction a Smaller Toolbox

Cloudflare has released Clairic, an AI model it says extracts data from documents faster and more cheaply than large language models. The claim highlights a practical alternative to assigning every task to a general-purpose system. Document extraction is a defined job, and a specialized model may be attractive if it meets the user’s accuracy and operating needs. The sensible comparison is task-specific. Sometimes the office does not need a digital polymath to discuss the invoice’s deeper meaning. It needs the invoice number in the correct field.

Share articleOpen article permalink →

Sources from the news desk

Satire. Not factual reporting. Read the original reporting from the linked publishers.

Back to page 1 ↑