Theo on A.I. · June 17th

Government ministries are deploying Google Cloud generative AI across municipal agencies to automate council planning operations. Public sector administration handles vast volumes of unstructured data that delay infrastructure development. The UK central government established a target to construct 1.5 million new homes by 2029. Local planning authorities encounter administrative backlogs caused by dense paperwork, delaying these development timelines. To address these constraints, the Ministry of Housing, Communities and Local Government (MHCLG) and the Department for Science, Innovation and Technology (DSIT) expanded two machine learning tools designed to accelerate municipal processing. Speaking at the Google Cloud Summit London, officials confirmed the nationwide deployment of the ‘Extract’ application and the progression of the ‘Augmented Planning Decisions’ (APD) prototype. Lila Ibrahim, Chief AI Readiness Officer at Google DeepMind, said: “The UK has an opportunity to build the homes our communities need, but local councils face a mountain of paperwork. That’s why we’re co-creating a sophisticated planning tool directly with councils to solve real-world bottlenecks. “This will help significantly cut decision times, freeing up planners to focus on the future to get Britain building faster.” Householder applications – which include routine domestic modifications such as loft conversions or property extensions – account for nearly 70 percent of all planning applications submitted annually. Evaluating these standard submissions manually requires planning officers to spend hours cross-referencing regional policy documents, historical archives, and unstructured PDF files. Such a repetitive evaluation process consumes administrative hours that would otherwise support major infrastructure and commercial developments. The deployment of automation targets this administrative distribution, aiming to reduce application decision timelines by 50 percent. Engineers at MHCLG and the government’s applied AI team, the Incubator for AI (i.AI), built the Extract tool internally using Gemini foundation models. Following trials across more than 20 local planning authorities, administrators expanded the application to every council in England. Extract parses unstructured data locked within legacy PDF records, converting hundreds of pages of historical planning documentation into structured digital datasets within minutes. Operational data from the trial phases indicates that the tool will eliminate roughly 255 hours of manual data entry per council annually. This reduction allows local authorities to reallocate personnel to complex evaluation tasks. Integrating large language models into public sector workflows requires enterprise-grade security environments. Local authorities process sensitive civic records, requiring strict risk management protocols to prevent data exposure. The government hosted the Gemini models on Google Cloud to establish a protected operating environment where data sovereignty is maintained. The cloud environment features active security controls to block malicious inputs, including prompt injection attacks. This technical framework ensures that sensitive municipal data remains secure during both testing and production computing cycles. The APD system, meanwhile, acts as an analytical assistant for municipal planning officers by automating four primary administrative tasks: - The system consolidates incoming documentation by pre-processing data backlogs, flagging missing information gaps, and extracting core geographical site data onto a unified user interface for officer review. - The software identifies relevant national and local zoning laws, assesses compliance margins, and appends precise policy citations for manual verification. - The application parses public consultation letters, summarising stakeholder objections or historical legal precedents. - The model generates initial drafts of final evaluation reports, including the technical rationale and recommended approval conditions. Protocols dictate that human planning officers retain final decision-making authority over every application. The software does not automate final approvals or rejections independently. Staff members review every line of text generated by the machine learning models, modifying the analytical reasoning before validating the report. To maintain regulatory accountability, the APD prototype records its internal processing steps sequentially.

Microsoft, Amazon, Alphabet, Meta and Oracle are all cranking up their AI‑infrastructure budgets at roughly 70 percent a year. Their operating cash flow, by contrast, is only nudging up around 23 percent annually. That mismatch is already showing up in the balance sheets, and if the gap keeps widening, the spending curve could actually eclipse cash flow as early as the third quarter of 2026.

What’s interesting is how the numbers force a shift in financing. A few of the big players have already started looking beyond the cash they generate—pulling in external capital to keep the AI pipelines humming. It’s not a headline‑grabbing pivot, just a pragmatic response to a growing budget shortfall.

The underlying dynamic is simple: AI hardware and software costs are ballooning faster than the revenue streams that traditionally fund them. When the expense line outpaces cash inflow, the companies have to decide whether to scale back, borrow, or bring in investors.

If the trend holds, we’ll see more of these hyperscalers juggling outside funding to stay on pace with their AI ambitions, and the cash‑flow gap will become a regular talking point in boardrooms rather than a surprise.

I’ve been tinkering with a personal AI assistant for a while now, mostly because the off‑the‑shelf options felt like paying for a suit that never quite fits—you end up with a lot of fluff you don’t need and give away more data than you’re comfortable with. So I pulled the plug on the subscription and started stitching together the pieces myself, which turned out to be a lot more satisfying than I expected.

Under the hood it’s a modest stack: a compact language model runs locally, a vector store holds my recent notes and emails, and a lightweight retrieval layer pulls the right context before the model generates a response. The code is all in Python, using a few well‑maintained libraries, and I wrapped everything in a tiny Flask service that I can call from my phone.

The first few weeks were a mess of rate‑limits and prompt drift—my assistant would start echoing back the same phrasing over and over. I fixed that by adding a simple cache and tightening the prompt template, then I let it handle things like drafting quick replies, summarizing meetings, and even nudging me about upcoming deadlines. Now it’s a quiet partner that pops up just when I need a hand, and I’ve stopped paying anyone else to do the same job.

For decades, automakers enjoyed a luxury that had nothing to do with the softest leather or the smoothest engines. Their luxury was time, with some popular cars and trucks enduring for a decade or longer before they received a full redesign. The clock is ticking faster now, thanks to China. BYD and other automakers there are speeding EVs and other models from drawing board to showrooms in two years or less. General Motors is among the Western automakers striving to match that blistering pace, by harnessing AI and simulation to dramatically shorten development times. GM’s effort is being spearheaded by Sterling Anderson, the technologist and robotics guru who led development teams for Tesla’s Autopilot and the Model X before cofounding Aurora Innovation, the autonomous trucking company. GM lured Anderson last June as its chief product officer, offering a $40 million package to guide the development of the automaker’s cars, autonomous models, batteries, software, and other tech. In a recent video call, Anderson and Jason Fischer, GM’s executive director of virtual integration engineering, walked me through the company’s latest design processes. But first, Anderson offered a wide-lens view of how AI is transforming everything that came before. Sterling Anderson, robotics guru and former Tesla executive, is pushing AI to accelerate GM’s design process.General Motors Anderson sees design and human ingenuity falling into three main epochs, beginning with thousands of years of empirical design that saw creators largely mimicking nature, building and testing models and advancing from there—slowly, expensively, and narrowly focused. “Flight is a great example,” Anderson says. “Humans looked to birds and said, ‘Hey, those wings seem to work pretty well. Let’s come up with something like it.’” The advent of virtual tools such as CAD and computational fluid dynamics in the 1950s kicked off a second age, he says. Developers had better ways of doing work, but they remained siloed in an inefficient, pass-the-baton process. “Designers still had to toss something over the wall to other engineers, who ultimately had to build that empirical asset anyway,” Anderson says. In automobiles, that meant building prototype vehicles first and then integrating and assessing myriad functions, many of which were developed separately: electrical systems, thermal controls, safety, ride and handling, and so on. Today’s third epoch is characterized by AI and simulation that can collapse those functions into a single virtual development tool, Anderson says. In roughly one minute, a structural engineer can see how a design change might affect a finished vehicle, as opposed to the 15 hours it used to take. The result, he says, “is a dramatically accelerated product development process at GM.” GM is applying this approach to self-driving cars, LMR batteries, Cadillac’s high-profile Formula 1 racing program, military defense systems, and tech for Lunar Outpost’s Pegasus rover, part of NASA’s Artemis mission to land astronauts on the moon in 2028. Fischer says the company’s proprietary environment allows engineers to simultaneously develop and optimize hardware and software, well before the physical prototype stage. A simulated Cadillac performs an emergency avoidance maneuver, with graphs tracking vehicle functions such as brake pressure and steering wheel angle.General Motors In an onscreen demonstration, Fischer runs a digitally rendered Cadillac Escalade IQ through the Consumer Reports avoidance maneuver, which the publication uses to assess a car’s evasive skills. The tricky double lane change is a serious test of the electric SUV’s handling and stability under duress. In the past, physical testing could begin only after an array of systems had been separately developed and stitched together, including the chassis, powertrain, steering, brakes, suspension, sensors, and controls. Engineers would spend months testing and calibrating prototypes in proving grounds and on real-world roads. Now, GM can run detailed, physics-based models of designs across thousands of simulated scenarios—snow and rain, varying road conditions, different suspension setups. “We can do full, virtual calibrations prior to a vehicle ever being built,” Fischer says. “We get a system that performs well not just in ideal conditions, but one that’s been hardened against the real world.” This approach halved the development time of the electric GMC Hummer, which went from initial designs to showroom in two years, versus a more typical four- to five-year product cycle. GM’s goal is to get a full range of vehicle and tech programs onto that lightning-fast development track. “We’re not there yet, but give us a minute,” Anderson says. Front-end crash simulations have also been accelerated. In the past, a “heavy computational method” required 15 hours of computing to complete, Fischer says.

Musicians are accustomed to getting paid each time their creative work is used. Across vinyl/CD sales, streams, radio, cover versions, and those numerous niches like karaoke, there are agreements in place about what “use” means. Underlying this is a simple economic principle: The more something is used, the more money it makes. Generative AI has complicated the definition of use. On the one hand, you could argue that the use of a piece of musical training data happens just once, at the point of training. On the other hand, creators would be right to complain that the creative essence of their work lives on in the structure of the model, used every time the model produces an output. Now, companies like Sureel and SoundVerse are working to re-create the essential economic principle that motivates creativity in an era of AI. Such initiatives aim to turn the generative AI industry from one guilty of “the biggest act of copyright theft in history” into one that coexists harmoniously with hardworking artists. Sureel, a startup Warner Music Group just acquired, has partnered with the Swedish copyright agency STIM to explore the potential for music creators to get paid when their music is used to train generative AI tools. Sureel’s software labels online media, such as a music file, with instructions determined by the owner. The instructions specify whether an AI company may use the media freely in training, limit its influence in any given training set, or avoid it altogether. The software then tracks how the AI company uses the media in training and sets licensing fees accordingly. Meanwhile, the founders of the AI music company SoundVerse “[reject] one-time royalty buyouts as insufficient and [advocate] for ongoing participation of artists in the AI lifecycle,” they wrote in a 2025 white paper. They argue that each time a generative AI system produces an output, certain pieces of training data play a greater role than others. If the system outputs music resembling jazz, the jazz in the training set has arguably contributed more than, say, the folk music. You can therefore differentially reward each piece of training data for each output. Sureel’s Co-President Benji Rogers told me, “Attribution isn’t about re-creating the old economics. It’s about measuring, for the first time, the thing the old economics only approximated.” Such influence attribution needs to do more than superficially measure how similar a training data point is to the AI output. The challenge is to attribute causality, or a relationship between the training data and the trained AI, Sureel CEO Tamay Aykut says. Even if the AI industry achieved that, however, it might encourage people to create music designed to maximize training-data royalties. While all creative markets lead to new incentives (music streaming, for example, has driven songs to have shorter intros), the industry could do without another economic structure that is easily gamed, in which someone’s reverse-engineered pastiche diverts royalties away from original works of creative expression. Inferring the influence of a particular piece of music on a generated piece of music, if a well-defined problem at all, may involve more advanced information theoretic principles, or modelling the actual historical role and impact of individual works. Aykut proposes that in carefully designed attribution systems, more unusual and unpolished musical works could even have more inherent value than radio standards. Simon Gozzi, Head of Business Development at STIM, says the company is in the process of seeing how Sureel’s attribution reports could underlie licensing agreements between musicians and AI companies. Could generative AI attribution strategies not only sustain the economic logic that “popularity pays,” but also motivate musical experimentation and diversity? It’s a compelling concept when public sentiment rightly fears generative AI’s threat to cultural vibrancy, pushing power towards tech companies, deskilling creative workers, shrinking revenue in the creative sector, and filling the internet with slop. “Attribution is one of the few credible tools we have,” Rogers says. There’s a window of opportunity to debate and establish approaches to paying for AI training data that serve a vibrant and sustainable creative sector. The technical problem of training data attribution is both complex and ill-defined. Just as a simplistic attribution strategy based on measuring similarity might motivate people to reverse-engineer the canonical works of a genre to capture royalties, a more complex attribution strategy based on some information theory of originality might be easily gamed or fail to reward human cultural production. For creative workers, there’s good reason to fear that even with the best intentions, AI attribution will only compound the baroque and opaque arms races that they are already weary of navigating. Some voices within the music AI sector are also skeptical.

Google’s new Home speaker swaps the old “Hey Google” trigger for a Gemini‑powered chat layer. Instead of a fixed set of commands, the device now keeps the conversation flowing, remembering context across questions and suggestions.

Under the hood the speaker runs a trimmed‑down Gemini model locally, so most of the thinking happens on the device rather than pinging the cloud for every turn. That cuts latency and lets it handle private requests—like checking a calendar entry or adjusting lights—without sending raw audio to servers.

Because the LLM is tuned for home tasks, it can blend multiple actions in a single reply: “I’ve dimmed the lights, started the playlist you liked, and set a timer for 20 minutes.” The result feels more like talking to a helpful roommate than issuing commands.

At $99.99 the price is modest for a speaker that can actually hold a conversation, and Google’s framing suggests it’s the next step toward making smart speakers feel less like rigid tools and more like a natural part of daily life.

OpenAI researchers propose a method for predicting how often a new AI model will make mistakes after release. It could fill gaps left by standard safety testing. The article OpenAI researchers want to predict how often AI models will fail before launch appeared first on The Decoder.

Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley are using AI coding agents to teach robots dexterous grasping in the real world. A fleet of eight robots hits up to 99 percent success on tricky tasks. The article Nvidia research shows robots that train themselves through AI coding agents appeared first on The Decoder.

Chinese AI lab Zhipu AI releases GLM-5.2 with a stable 1-million-token context under the MIT license. On FrontierSWE, a benchmark for hours-long coding tasks, the open-source model trails Anthropic's Claude Opus 4.8 by just one percentage point. On reasoning, it still falls well behind closed-source rivals. The article Zhipu AI's GLM-5.2 closes in on closed-source leaders in coding marathons appeared first on The Decoder.

Amazon, Nvidia and AMD have pooled $310 million into Odyssey ML, a startup that’s stitching together 3D representations of the world. The money pushes the company’s valuation to roughly $1.45 billion, and it’s not just the hardware giants behind it—IQT, a fund with CIA ties, and Google’s Jeff Dean are also on board.

What’s interesting is how the investment shifts the focus from pure text generators to “world models,” systems that can understand and predict spatial relationships the way we do when we move through a room. Odyssey’s tech is essentially teaching AI to build a mental map of environments, which could make everything from robotics to virtual‑reality simulations feel a lot more grounded.

The partnership feels almost like a silent pact among the biggest players in compute, each betting that the next wave of AI utility will come from these immersive, three‑dimensional understandings rather than just language. It’s a subtle but meaningful pivot in where the big money is heading.

If the models start to reliably capture the nuances of real‑world physics and geometry, we could see a whole new class of applications—think smarter drones, more realistic game worlds, and even better tools for scientific visualization. That’s why this round feels like a quiet signal that the AI frontier is expanding beyond words.