H.Headline
Chaos

THE EDITION ARCHIVE

A front page, saved in time.

Making AI Headlines · AM edition

Published · New York time

Permanent edition link →

10 pages · 23 stories

PAGE 1 OF 10

AI asks to be trusted, then meets the independent examiner

Australia proposes frontier-model audits, while Google DeepMind uses AI to verify Signal cryptography.

Making AI Headlines newspaper, page 1: AI asks to be trusted, then meets the independent examinerPage 1 of 10
Page 1 of 10. Read the same stories below.

Australia tells frontier AI that self-grading is not a safety plan

Australia has proposed mandatory safety rules for frontier AI companies, including independent audits and a new regulator to enforce the requirements. The proposal would bring outside scrutiny to developers of powerful models, with steep penalties for violations. The government's plan matters beyond Australia because leading model makers operate internationally. A company building one system for many markets may have to reconcile its release schedule with different countries' expectations for testing and accountability. Canberra is prioritizing domestic rules before entering international talks, according to the reporting. Independent audits could give regulators a basis for examining safety claims that companies make about their own products. Their practical value will depend on how the requirements are implemented and what the examinations actually cover. A proposed bill is still a proposal, not an operating inspection regime. There is a familiar business tension here: firms want the world to treat their products as consequential, but consequential products tend to attract people with clipboards. The proposed regulator would give those people more than a guest pass. A developer can demonstrate confidence by releasing a model quickly, or by allowing someone outside the company to examine it. Those are not necessarily the same demonstration. Under the proposal, the world's most ambitious software may have to make room on its calendar for an appointment it did not schedule itself.

Share articleOpen article permalink →

Signal's cryptography gets six million lines of very serious homework

Google DeepMind says its AlphaProof Code system mathematically verified the cryptographic core of Signal in four weeks, using Gemini Carbon agents to produce more than six million lines of proof code. The work concerns software that helps protect private communications, making the result relevant far beyond a single engineering team. Formal verification is an unusually demanding kind of checking. Rather than merely seeing whether code appears to work in a selection of tests, a proof aims to establish that specified properties hold. That makes it attractive for security-critical systems, but the work can be painstaking. DeepMind says the process would typically take years and that only about one percent of the proof code was flagged for human review. The company plans to extend the method to critical infrastructure, including power grids and medical devices. The appeal is clear: AI might help specialists tackle more of the mathematical work behind software people depend on. A proof, however, is tied to what is specified and verified, not a universal promise that every aspect of a service is invulnerable. In an industry fond of announcing that AI will do the hard thinking, this is an unusually literal assignment. The machine did not just polish the presentation. It brought six million lines of homework, and some of them still landed on a human desk.

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

Creative software joins the meeting and immediately rearranges the room

Adobe previews experimental AI tools, and Anthropic says AI helped move Claude onto AMD chips.

Making AI Headlines newspaper, page 2: Creative software joins the meeting and immediately rearranges the roomPage 1 of 10
Page 2 of 10. Read the same stories below.

Adobe gives the still photo an itinerary and a personality

Adobe has shown experimental AI features at its Max conference that can turn images into explorable 3D scenes and let creators develop characters through conversation. The demonstrations offer a glimpse of editing tools that treat a picture less like a finished surface and more like a place to wander around. Another experiment turns photo editing into a copy-and-paste process. For artists and designers, such tools could change the early stages of a project. A visual reference might become a space to explore, while a character concept could be refined through dialogue. The convenience also invites a practical assessment of control: creators still need to decide whether the generated scene or personality serves the work they actually want to make. These are experiments, not a guarantee that every showcased feature will become part of a finished product. Adobe's conference has long been a place to display possibilities, and possibilities are especially easy to make look well behaved onstage. The pitch captures a shift in creative software. The tool is no longer content to offer a menu of adjustments. It would like to join the brainstorming session, remodel the setting and ask the protagonist a few questions. Somewhere, a perfectly ordinary photograph is about to be informed that it needs a backstory.

Share articleOpen article permalink →

Anthropic asks AI to help its software leave Nvidia's neighborhood

Anthropic says it used its own AI model to port Claude to AMD graphics processors in three days, a task the company says could otherwise take months. It reports that Claude now runs on four chip families, an effort intended to cut costs and reduce reliance on individual suppliers. Nvidia's position in AI rests partly on CUDA, the software ecosystem developers use to make its chips work efficiently. Moving sophisticated AI workloads to another chip family is not simply a matter of changing the logo on a server. Code and performance tuning can make the transition expensive and slow. Anthropic says its in-house model Mythos handled the port with almost no human intervention. The claim suggests AI coding tools may reduce some switching work. If the approach holds up across more tasks, customers could have more room to compare hardware suppliers. A fast port and a full economic assessment of running a model are different things. Hardware availability, performance and operating costs still matter, as does the work needed to keep a system running on each platform. The comic symmetry is difficult to improve upon: an AI company is asking AI to make AI less dependent on the machinery used to build AI. The industry has found a way to make a technical migration sound like a self-help book, with a very large electricity bill.

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

The AI boom discovers that money and electricity have speed limits

Corporate AI spending rose more slowly in October, while proposed data centers confront local obstacles.

Making AI Headlines newspaper, page 3: The AI boom discovers that money and electricity have speed limitsPage 1 of 10
Page 3 of 10. Read the same stories below.

Corporate AI spending hits a record, then remembers how percentages work

US corporate spending on AI reached a new high in October, but growth slowed to 1.65 percent, according to the Ramp AI Index. The index also found a striking concentration of revenue: one percent of companies account for 80 percent of revenue for OpenAI and Anthropic. The figures offer two ways to look at a market often described as though every office were buying the same amount of automation at the same pace. Total spending can continue to rise even as its rate of growth cools. Meanwhile, a relatively small group of customers can have an outsized influence on the businesses selling the technology. Customer support is emerging as another source of growth, according to the reporting, though it may carry lower margins than coding. That matters because not every popular use of AI necessarily has the same economics. A vendor can add customers and still face questions about the cost of serving them. For corporate buyers, the useful question is less whether the AI boom has arrived than what a particular purchase actually accomplishes. For vendors, the challenge is to turn a wide field of interest into durable business without relying on a tiny group to carry the entire buffet. The record is real. So is the smaller percentage.

Share articleOpen article permalink →

Data center announcements meet the local tradition of needing electricity

The data center boom is encountering local opposition, permitting delays and competition for electricity, raising questions about how many announced projects will actually be built. A proposed facility can look impressively complete in a presentation while still needing approval, power and a viable path to construction. That gap matters to communities considering land use and infrastructure, as well as to companies counting on more computing capacity. A large project can affect local planning long before a server arrives. Yet an announcement alone does not settle when, or whether, the building will open. Electricity is a particularly concrete constraint for an industry that likes to describe its products in terms of clouds. Servers occupy physical space and require power. Local residents and officials may weigh promised economic activity against changes to the landscape and demands on infrastructure. The gap between a plan and a completed facility is not unique to AI, but the industry's appetite for capacity makes it unusually visible. Billions of dollars in proposed infrastructure can encounter a process that runs on different schedules from a product launch. On a slide, a data center is a clean rectangle with a confident label. In the real world, it has neighbors, permits and a power connection. Those features are harder to render away, even with a very persuasive computer.

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 remembers that local work and local privacy still exist

Microsoft previews a command-line tool for Logic Apps, and flight attendants secure protections in a data sale.

Making AI Headlines newspaper, page 4: The cloud remembers that local work and local privacy still existPage 1 of 10
Page 4 of 10. Read the same stories below.

Microsoft gives workflow builders a command line, because the cloud needs a desk

Microsoft has released a public preview of a command-line extension for Azure Logic Apps Standard. The tool supports local project creation, workflow authoring, connector management, run inspection and test scaffolding, bringing development tasks into a local command-line workflow. Logic Apps helps developers build workflows that connect services. The new extension gives teams another way to create and inspect that work, including through tools and routines centered on a terminal. The tasks previously required the VS Code extension, according to the reporting. For developers who prefer repeatable commands, the preview offers another way to incorporate the work into existing practices. Microsoft's product team also presents the terminal interface as a way to make AI-assisted development more deterministic. A preview is still a preview, and teams considering it will need to assess how it fits their projects. The change is not that cloud software has suddenly become less cloudlike. It is that more work on a cloud workflow can start closer to the developer's own machine. The arrangement has a pleasingly familiar shape. An industry spent years explaining that computing was moving away from the local desk, then added a better way to work from the local desk. Progress is sometimes a grand migration. Sometimes it is making sure the person doing the work can type a command without opening another elaborate window.

Share articleOpen article permalink →

Flight attendants remind a data sale that employees are not carry-on luggage

The union representing flight attendants at the defunct Spirit Airlines reached a deal with Google to remove sensitive employee records from a $10 million data sale. The agreement puts employee privacy into a transaction that might otherwise have been discussed mainly in terms of the value of a dataset. Records created in the course of work can outlast the workplace itself. When a company disappears, information associated with it can become an asset in a sale, while the people described in those records still have an interest in how it is handled. The union's agreement shows that those interests can become part of the negotiations. The deal also illustrates a distinction that matters in AI and data businesses: access to a collection of records is not the same as permission to use every sensitive detail about the people inside it. Removing employee information changes what a buyer receives without making the entire transaction vanish. For workers, the stakes are not abstract. A personnel record is not just a row waiting to become somebody else's useful data. It began as information about a person doing a job, and its sensitivity does not disappear when the employer does. The spreadsheet may be ready for its next owner, but the people in it did not necessarily agree to come along.

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

AI infrastructure discovers that moving things is also a job

Google opens a library for transferring model data, while proposed data center power plants raise climate questions.

Making AI Headlines newspaper, page 5: AI infrastructure discovers that moving things is also a jobPage 1 of 10
Page 5 of 10. Read the same stories below.

Google opens a library for the glamorous work of moving model data

Google Cloud has open-sourced TPU Sync, a C++ library for moving model weights and key-value cache blocks among TPU memory, host memory, shared storage and data center networks. It addresses an unglamorous but important part of operating AI systems: getting the right data to the right place. Model weights are central to how a trained system operates, while cached information can help manage ongoing computations. Moving those materials between different kinds of memory and storage is a practical concern when running models at scale. A library dedicated to those transfers gives developers a tool for managing work that sits behind the more visible interface. Open-sourcing the library also lets others inspect and use the code. Whether it helps a particular team depends on its infrastructure and workload, but the release highlights the number of separate engineering problems hidden inside an apparently simple AI response. The library is built around Google's TPU infrastructure, rather than being a promise that every kind of hardware will work the same way. A chatbot can make the result look effortless: a question goes in, an answer appears. The machinery underneath is closer to a warehouse full of carefully labeled boxes being moved between shelves at high speed. Google's contribution is a way to move some of the boxes. The boxes, as usual, did not receive top billing.

Share articleOpen article permalink →

Data centers seek their own gas power, making the cloud bring a generator

Developers of AI data centers have proposed large amounts of on-site gas-fired power capacity in the United States and Europe, according to research examining delays in connecting projects to national grids. The report identifies 181 gigawatts of proposed capacity, putting the energy demands of the computing boom into unusually concrete terms. On-site generation can appear attractive when a data center needs power and a grid connection takes time. But gas-fired plants also carry climate implications, making the proposed response to an infrastructure bottleneck a question for energy and environmental planning. Proposed capacity is not the same as completed power generation, and the eventual scale depends on which projects proceed. The issue reaches beyond a single facility. AI services depend on physical equipment, and physical equipment depends on electricity. Decisions about how to supply that power can affect the communities and energy systems around the buildings, not just the companies operating the servers. The proposed total covers sites in the United States, the United Kingdom and mainland Europe. It is a measure of plans across those regions, not a current output figure for operating plants. The industry's favorite metaphor has finally met its least poetic requirement. A cloud sounds weightless until someone asks where the power comes from. Then the cloud produces a plan for a power station and asks whether it can park it next to the building.

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

AI gets faster at climate scenarios and keeps both sides of the ledger busy

MIT researchers develop a climate-emulator training method, while OpenAI reports rising revenue and losses.

Making AI Headlines newspaper, page 6: AI gets faster at climate scenarios and keeps both sides of the ledger busyPage 1 of 10
Page 6 of 10. Read the same stories below.

MIT speeds up climate modeling, where seconds can hold thousands of futures

MIT researchers have developed a training method intended to produce fast, adaptable AI models for climate prediction. These models, known as climate emulators, can simulate the effects of human activities on future climate conditions and run scenarios in seconds. The speed matters because planning often requires comparing many possible futures rather than relying on a single projection. Researchers can use emulators to explore variations in temperature or sea level, with applications that include planning for power grids, public health and agriculture. Faster methods can make it easier to examine a wider range of scenarios. MIT says its technique streamlines the production of such models. The resulting projections can support decisions about resilience, but the work does not replace the choices communities must make about how to respond. A quick projection still has to be understood, evaluated and applied to real circumstances. There is a useful reversal in the technology's usual sales pitch. Instead of promising to make every task instantaneous, this research makes it faster to consider consequences that unfold over long periods. A model can help researchers compare many combinations of conditions while officials consider what those differences would mean for infrastructure and people. The computer can run through thousands of possible futures while humans do the slower work of agreeing which risks deserve attention. That second task remains stubbornly resistant to a faster processor.

Share articleOpen article permalink →

OpenAI grows revenue and losses, keeping both sides of the ledger busy

OpenAI's quarterly revenue rose 50 percent to $10 billion, while adjusted operating losses reached $10.9 billion, according to the reporting. The figures put rapid sales growth alongside substantial costs, a pairing that makes the economics of the AI business more complicated than a simple race for customers. Enterprise sales helped drive the revenue increase. Selling to businesses can bring large contracts and a broader range of uses, but expanding an AI service also requires substantial spending. The reported loss shows that a growing top line does not by itself settle the question of how efficiently that growth is being operated. The reporting also compares OpenAI with Anthropic, which it says generated $17 billion in revenue and operated more efficiently. Revenue and adjusted operating results are distinct measures and should not be treated as interchangeable. The report also says OpenAI is using lower prices to compete for enterprise customers, adding another consideration to the assessment of future sales and costs. For customers, competition may change what is available and what it costs. For the companies selling the services, the task is to turn that demand into a business whose expenses fit its income. The industry's enthusiasm tends to give every large number its own parade. These figures suggest the parade needs two lanes: one for the money coming in and another for the money going out. Both are moving briskly.

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

AI promises to protect children and deliver the boss some candor

OpenAI commits research funding for children's safety, and Databricks uses an agent to surface information for its CEO.

Making AI Headlines newspaper, page 7: AI promises to protect children and deliver the boss some candorPage 1 of 10
Page 7 of 10. Read the same stories below.

OpenAI funds children's AI safety research, where the easy answer is not enough

OpenAI has committed $2 million to research on AI and children's safety, supporting shared tools to measure risks and opportunities for young people. The commitment puts funding behind a question facing families, researchers and companies as AI products become more accessible. Measuring children's experiences is a different task from simply declaring a product useful or safe. Researchers need ways to examine potential benefits alongside risks, and shared tools can help make that work more consistent. The company's announcement concerns support for research, rather than a finding that those questions have already been resolved. The source of the funding is also relevant to how the work will be understood. AI companies have a direct interest in the technology being studied, while families and researchers need useful evidence about how it affects young users. The practical value of the commitment will lie in what the research tools enable people to learn and compare. That includes the ability to ask questions across different uses of AI, rather than treating every interaction as though it presented the same benefit or risk. Research tools are part of that work, not a substitute for the findings themselves. The technology industry is very good at demonstrating how quickly a system can answer a question. Children's safety requires attention to a less theatrical question: which questions should be asked, and how will anyone know whether the answer is sound? Two million dollars buys research support, not a certificate that the homework is finished.

Share articleOpen article permalink →

Databricks gives its CEO an AI assistant for the things nobody wants to say

Databricks CEO Ali Ghodsi uses the company's revised Genie AI agent to help surface information about the business as the company grows. With 12,000 employees, the challenge is familiar to any large organization: the person at the top cannot know everyone or personally observe every problem. Ghodsi acknowledges that employees may be unlikely to confess what goes wrong directly to him. The agent offers another route to information about the company, using a system designed to retrieve and summarize material rather than relying only on the people in the room. An agent that helps leaders find information may make large collections of company data more accessible. It also changes the route by which a question reaches an answer. A manager can ask a system to look across available information, then use the result as a starting point for further questions. That does not eliminate the need to judge what the system returns or to listen to employees. A tool can make information easier to find without automatically explaining every reason a problem exists. The quality of the underlying information and the questions asked still matter. The attraction is understandable. Executives routinely say they want unvarnished truth, while organizations develop elaborate habits for applying varnish. Databricks is offering software to help locate a less polished view. It is a modern solution to an old management problem: the larger the office gets, the harder it becomes to hear what is happening outside the executive suite.

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 travel agent is back, but the airline wants a human at the keyboard

United restricts personal AI agents from booking tickets, while US brokers sell training environments to Chinese AI labs.

Making AI Headlines newspaper, page 8: The travel agent is back, but the airline wants a human at the keyboardPage 1 of 10
Page 8 of 10. Read the same stories below.

United tells AI travel assistants they can admire the itinerary from outside

United says it will not allow personal AI assistants to book tickets on the airline, citing privacy and regulatory concerns. The company says it has not yet seen many people attempt to use such agents for bookings, according to the reporting. The restriction lands at an early stage for a technology often advertised as a way to delegate everyday online tasks. Booking a flight requires handling personal information and navigating airline rules, which makes it more complicated than asking an assistant to suggest a destination. United's position draws a line between helping a traveler consider options and completing the transaction on the traveler's behalf. For passengers, the practical consequence is that the promised hands-off itinerary still requires human participation when booking with United. For developers, it illustrates that an agent's ability to perform a task does not guarantee that a business will permit it to do so. The airline's concern arrives before widespread use, according to its own account. That gives the policy an unusually anticipatory quality: the company is setting a boundary around a behavior it says it has not yet seen much of. The situation gives a new twist to the old travel routine. A customer can bring a digital assistant along to discuss the trip, but at the point of purchase the airline wants to deal with the traveler. The assistant may have a detailed plan for the entire journey. It just cannot be the one to press the final buttons.

Share articleOpen article permalink →

US training-data brokers find a global market for the practice field

US data brokers are selling roughly half a billion dollars' worth of AI training data to Chinese laboratories, according to an investigation by ChinaTalk. The reporting says the material includes the same reinforcement-learning environments previously built for OpenAI or Anthropic. These environments provide settings in which AI systems can practice tasks and receive feedback. Their sale illustrates that the international AI trade involves more than finished models and physical chips. Training materials and the work of constructing them can also become valuable products across borders. The investigation names brokers including Mercor, Surge AI and AfterQuery, and says at least nine firms are involved. It reports that export controls barely apply to this trade. That gives the business a different policy backdrop from the prominent restrictions on advanced computing hardware. The trade creates a commercial distinction worth noticing. A company may build an environment for one customer while the broader market for training resources extends to other labs and countries. The reporting describes sales by brokers, not a conclusion that every buyer receives the same data or that every arrangement has the same terms. For the AI industry, competition and commerce can occupy the same remarkably crowded room. Companies race to build capable systems, while businesses supplying parts of the training process look for customers. The model may be presented as the star of the show, but the practice field has discovered that it, too, can sell tickets.

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

AI answers a thousand questions while finance asks for payment in advance

Google introduces a decision API, Meta's infrastructure chief is stepping down, and CoreWeave leans on customer prepayments.

Making AI Headlines newspaper, page 9: AI answers a thousand questions while finance asks for payment in advancePage 1 of 10
Page 9 of 10. Read the same stories below.

Google's decision model makes the questionnaire fear for its job

Google DeepMind's Gemini Decisions API can evaluate up to 1,000 closed-ended questions in a single pass across text, images, audio and video. Google says Gemini Decision 1 returns results in 150 to 250 milliseconds at roughly one-fifth the cost of Gemini 3.5 Flash-Lite. The product is aimed at a different task from an open-ended conversation: processing many defined questions quickly. That could be useful when organizations need assessments across large volumes of material. Speed and cost, however, are not the same as correctness, and the quality of the questions remains important. The announcement offers a tidy picture of AI's workplace ambition. A thousand questions can now be processed in less time than a meeting takes to decide who will share the screen. The harder part is making sure those are the thousand questions anyone actually needed answered.

Share articleOpen article permalink →

Meta's infrastructure chief departs, leaving the servers without their longtime guide

Santosh Janardhan is stepping down as Meta's head of infrastructure. He joined the company in 2009, making the change notable for a business whose services depend on extensive computing systems. Infrastructure is the part of a technology company users usually notice only when it fails to behave like invisible magic. It encompasses the capacity that keeps digital products running, even as demand and technical priorities change. A leadership transition in that area draws attention to the people and decisions behind services that otherwise appear to arrive effortlessly on a screen. The departure does not by itself establish a change in Meta's infrastructure plans. It does provide an occasion to remember that the cloud has an organizational chart. The racks and networks may keep humming, but someone still has to be responsible for making the whole arrangement work.

Share articleOpen article permalink →

CoreWeave asks customers to pay early, because AI growth dislikes waiting

CoreWeave is increasingly seeking customer prepayments and borrowing against specific contracts to finance AI growth, according to its third-quarter earnings. The company is also considering shorter-term deals that could lead to higher prices. The approach highlights the financing demands behind AI infrastructure. Building and operating computing capacity requires money, while contracts determine when customers pay for access. Prepayments can change that timing, and borrowing against contracts is another way to fund operations against expected business. Shorter deals involve a different tradeoff, including the possibility of higher prices. Customers and the company must weigh the length of a commitment against its cost. The arrangement is a reminder that the cloud is not merely a collection of powerful machines. It is also a collection of payment schedules, contract terms and financing decisions, all trying to arrive before the next expansion does.

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

AI briefs: the duck gets a job and the translator learns to raise an eyebrow

A robotics experiment, a translation model, coding-agent research and an AI engineering conference round out the desk.

Making AI Headlines newspaper, page 10: AI briefs: the duck gets a job and the translator learns to raise an eyebrowPage 1 of 10
Page 10 of 10. Read the same stories below.

A virtual duck joins robotics, and the cat gets an unsolicited assistant

A robotics experiment used an OpenAI model to help teach a virtual Microduck to waddle and act as a tiny cat assistant. The project shows how AI tools can support experimentation with behavior in a simulated setting. A duck is a less intimidating ambassador for robotics than a factory floor, but the underlying appeal is recognizable: give a model a task and see how it helps shape movement. The cat, meanwhile, has acquired the most modern of workplace problems, an assistant whose enthusiasm may exceed the job description. The duck's employment prospects remain firmly outside the benefits discussion.

Share articleOpen article permalink →

DeepL teaches translation to carry the attitude, not just the words

DeepL says its new Aure model for real-time voice translation carries tone, emotion, emphasis and speaking pace into the target language. The aim is to make translated speech sound more natural, including when a speaker's mood shifts. That is a more complicated assignment than moving words between languages. How something is said can change its meaning, particularly when the words themselves are dry or ironic. DeepL's claim puts delivery alongside vocabulary in the translation job. The software is now being asked to preserve the raised eyebrow as well as the sentence, preferably without needing an actual eyebrow.

Share articleOpen article permalink →

Coding agents learn that finding code is not the same as guessing its meaning

JioHotstar evaluated six approaches to retrieving code for AI coding agents. The assessment found structural search performed well on concrete tasks, while semantic retrieval helped with ambiguous requirements. An agent working on a codebase must first find relevant material, and a precise request may call for a different search approach from a vague one. The assessment also considered accuracy, token use and tool design. The industry's tireless digital programmer still has a familiar problem: before fixing anything, it has to find the useful files. A confident answer is no substitute for looking in the right place.

Share articleOpen article permalink →

AI engineers gather to discuss the features demos politely skip

QCon AI New York has published its program for December 15 and 16, featuring engineers discussing AI systems used in production. Topics include agent permissions, guardrails, evaluation and inference costs. The agenda puts attention on what happens after a demonstration becomes a service people rely on. Permissions and evaluation are less theatrical than a fluent answer, but they shape what a system can do and how its performance is assessed. Speakers include engineers from companies such as LinkedIn, Netflix and DoorDash. The conference promises a rare industry spectacle: people discussing the plumbing while the chatbot waits outside.

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 ↑