0:00 / 3:15aifeatureJeff Dean and peers aim AI at the scientific discovery loop#
Discovery Loop is presenting a plan to automate experimental loops using frontier AI models and large-scale compute: propose experiments, run evaluations, learn from results, and iterate at much higher scale. The company says it will start with machine-learning research and engineering, use those capabilities on its own stack, and eventually pursue broader science and engineering challenges. The attention is driven heavily by the founding team: Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, whose listed work spans major Google infrastructure and AI systems.
Discussion: Mixed — HN was impressed by the founding team but skeptical of the pitch. Many commenters argued that science is often bottlenecked by funding, policy, regulation, and physical experimentation rather than by a lack of AI-generated hypotheses; others were excited by the idea of large-scale automated research loops, especially for ML and software-adjacent work. (Respect for the founders' track records, Skepticism that AI can automate messy physical experiments, Debate over whether policy and funding are bigger bottlenecks than intelligence)
0:00 / 2:39aideep diveGoogle reshuffles DeepMind as Jeff Dean and Sanjay Ghemawat leave for a new AI venture#
Google and Alphabet CEO Sundar Pichai announced a leadership shakeup at Google DeepMind: Demis Hassabis is moving out of day-to-day operations to become Chair of Google DeepMind and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs. Koray Kavukcuoglu, currently Google DeepMind CTO and Google’s Chief AI Architect, will become SVP of Google DeepMind, overseeing Gemini model development, frontier AI research, and Gemini app and developer teams. Separately, Jeff Dean and Sanjay Ghemawat are leaving Google after long tenures to launch an independent public benefit corporation focused on ML, science, and engineering, with Google as a founding investor and Cloud partner. The changes land at a high-stakes moment in the AI race, where leadership continuity and research-to-product execution are central to Google’s competitiveness.
Discussion: Negative — HN’s reaction was overwhelmingly worried, with commenters treating Jeff Dean and Sanjay Ghemawat’s departure as the real headline and a symbolic end of an era at Google. Many framed the reshuffle as evidence that Google is losing key AI talent and struggling to convert DeepMind’s research legacy into competitive products, though a few saw the Google-backed public benefit corporation as a way to keep collaboration alive. (Jeff Dean and Sanjay Ghemawat leaving seen as a major loss, Concern about broader senior AI talent departures from Google, Skepticism that DeepMind is being folded into Google’s commercial priorities)
0:00 / 2:44aideep diveCloudflare open-sources an AI work platform for company agents and custom apps#
Cloudflare has open-sourced Cloudflare OS, a browser-based workplace agent platform that combines company-curated context, governed access to internal systems, and AI-generated personal apps and workflows. Cloudflare says it has used an earlier version internally since May, with thousands of employees using it for documents, slides, repeatable tasks, and small data apps. The bigger pitch is not just another chatbot with connectors, but a governed environment where agents can build and run live apps on Cloudflare infrastructure while administrators control data access, models, spend, and sharing.
Discussion: Mixed — HN was intrigued by the ambition, especially comparisons to Sandstorm and the idea of secure, personal app building on Workers, but the thread was far from uniformly celebratory. Common pushback centered on Cloudflare lock-in, whether calling it an “OS” is marketing fluff, whether AI-built apps can really be secured, and whether the launch post buried the most interesting technical story. (Excitement about a Sandstorm-like app platform rebuilt on Cloudflare Workers, Skepticism about vendor lock-in despite open-source claims, Complaints that the blog post is vague or over-marketed)
0:00 / 1:46aiMeta ships Muse Code, a coding agent built around persistent subagents#
Meta released Muse Code in beta, a macOS and Linux terminal coding agent powered by its new Muse Spark 1.2 model. The agent is designed for large-repository software tasks, using persistent async background subagents, a local append-only event log for replay and crash recovery, and bundled commands for planning, stress-testing plans, and completing goals. Meta says Muse Spark 1.2 improves over 1.1 on coding, debugging, codebase understanding, and long-horizon workflows, and is available through Muse Code and the Meta Model API.
Discussion: Mixed — HN was interested in the release, especially Muse Code’s native subagent/orchestrator design and the very low reported contributor pricing, but the thread was far from celebratory. The main concerns were Meta’s data-use tradeoff, benchmark framing, lack of spending limits, and whether Meta’s own developers would actually switch from incumbent coding agents. (Contributor pricing and data-retention tradeoffs dominated the thread, Skepticism about benchmark comparisons and claims of being near the frontier, Curiosity about Muse Code’s persistent subagents and long-horizon coding workflow)
0:00 / 1:40softwareWhy some hobby coders see LLMs as missing the point#
The post argues that some niche hobby programming communities are hostile to LLM-generated development because the learning and mastery are the point, not merely producing something that runs. It cites communities such as OSDev, language development, emulators, roguelikes, demoscene, and code golf, where respect is earned through understanding, elegant code, and long-term participation. The author’s position is not anti-tool absolutism: LLMs can be useful as a lever for experts, but using them as a surrogate can undercut the craft those communities are built around.
Discussion: Mixed — HN largely accepted the article’s framing that many hobby programmers value the process, not just the finished software, but commenters split on whether LLMs undermine that craft or simply act like power tools. Several threads reframed the debate around different motivations for programming: tinkering, shipping, learning, earning money, or solving personal problems. A side discussion questioned the underlying GitHub chess-engine context, especially around alleged copying, licensing, and whether ideas versus code expression are protected. (Programming as craft versus programming as production, LLMs as force multipliers, shortcuts, or joy-killers, Hobby norms, gatekeeping, and rules about acceptable assistance)
0:00 / 1:54securityAtlassian’s Rovo accused of leaking Jira and Confluence data through prompt injection#
PromptArmor says it found vulnerabilities in Atlassian’s Rovo AI that allow indirect prompt injection to exfiltrate data from Jira tickets, Confluence pages, and potentially connected services by abusing Rovo’s URL retrieval tool. The writeup says the attack can work without human approval and even when web search is disabled, because the setting allegedly does not remove the tool used to open results. PromptArmor says it disclosed the issue to Atlassian on May 23, received a case number, and published after more than two months without further communication.
Discussion: Negative — HN reaction was sharply critical of Atlassian and broadly skeptical of agentic AI in enterprise tools. Commenters focused on Rovo being intrusive, hard to disable, and unsafe when given access to private company data plus external communication channels; a minority pushed back that prompt-injection demos can require contrived user behavior or that these risks apply to all modern agents. (anger at Atlassian’s AI rollout and trust erosion, concern about agent access to private Jira, Confluence, and connected data, discussion of indirect prompt injection and the “lethal trifecta” risk pattern)
0:00 / 2:56generaldeep diveBlade Runner’s title cards make the case for obsessive design details#
Rands in Repose uses Blade Runner’s opening title cards to argue that typography is never merely functional: it carries mood. The piece says the sequence relies largely on Goudy Oldstyle, with all caps, small caps, italics, red emphasis on “Replicant,” and book-like spacing to set the film’s dystopian tone. The author contrasts that care with a rough work-print version using Impact, then connects the point to modern product-building and AI-assisted work: the last layer of small human decisions is what makes something feel excellent rather than merely usable.
Discussion: Positive — HN mostly enjoyed the essay as a typography-and-film appreciation piece, with many commenters expanding the discussion to Vangelis’s score and Blade Runner’s broader influence. The thread wandered into familiar HN territory around LLM prose, em dashes, terminal fonts, and typographic nitpicks, but the overall tone was affectionate and engaged. (Admiration for Blade Runner’s atmosphere and craft, Vangelis’s soundtrack as essential to the opening’s impact, Typography details such as small caps, italics, em dashes, and font choice)
0:00 / 2:16scienceA foul-mouthed gateway into teaching yourself botany#
Crime Pays but Botany Doesn’t offers a blunt self-study roadmap for people who want to learn botany, starting with Latin naming conventions, taxonomy, and evolutionary relationships. The guide points readers toward serious textbooks including Michael Simpson’s Plant Systematics and Raven’s Biology of Plants, plus a longer list spanning plant evolution, ecology, geology, fungi, biogeography, and California natural history. Its core pitch is that plant identification becomes far more powerful when you learn the evolutionary traits that connect families, genera, and species, rather than memorizing isolated names.
Discussion: Positive — HN reacted warmly, with many commenters already fans of the Crime Pays but Botany Doesn’t YouTube channel and podcast, and others sharing how they became interested in plant identification, gardening, birds, ecology, and restoration. The main friction was around whether the guide makes botany sound easier than it is, and a side debate over Latin taxonomy, Linnaeus, colonial naming, and modern classification. (Strong affection for the creator’s irreverent teaching style, Personal stories of getting into plants, trees, birds, gardening, and ecology later in life, Recommendations for plant-identification apps such as Flora Incognita, iNaturalist, PlantNet, Merlin, and related data projects)
Jeff Dean, Sanjay Ghemawat, and other Google veterans are pointing AI at the scientific discovery loop itself. The target is not another chat box. It is how labs form, test, and refine ideas.
This is HN Radio for Thursday, August sixth, twenty twenty-six. I’m Ada. The show is made with AI. Every quote is from a real Hacker News commenter, with sources in the show notes. Let’s get into the rundown.
First up, machine learning and AI. Today’s theme is clear: agent demos are easy now. The hard part is showing they work when the loop touches real code, real teams, and real institutions.
Jeff Dean and peers aim AI at the scientific discovery loop
A new company from several famous Google researchers wants to automate scientific work. For an engineering team, the nearer version is simpler: software proposes experiments, runs them, reads the results, and tries again. Discovery Loop is the company, and the names are why this hit Hacker News. The founding team listed is Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. That is a lot of Google history in one room. The site points to Google Search, MapReduce, BigTable, Spanner, TensorFlow, TPUs, AlphaFold, Gemini, and years of model work. The plan is less concrete than the résumé section. The company says it will start with machine-learning research and engineering, where the loop is already mostly inside computers. You change a model, a dataset, a training recipe, or an evaluation. Then you run it and compare numbers. The company says, quote, “This approach allows for the parallel execution of thousands of experiments,” end quote. That is the load-bearing claim. Thousands of experiments is a different operating shape than a lab meeting followed by one careful run. But the headline says AI aimed at scientific discovery. The site shows broad ambition, not a result. There is no product, no benchmark, no funding detail, and no technical paper in the provided material. So the read is: credible founders, huge aim, almost no public evidence yet. The mechanism they describe is an automated learning loop. A system proposes an experiment, implements it, runs evaluation, reads the outcome, and chooses the next attempt. In machine learning, that can mean architecture search, data filtering, kernel tuning, training schedules, test generation, or model distillation. Those loops can be measured. The harder leap is from measured optimization to open-ended discovery. Hacker News kept circling that gap. Commenters respected the founders, often a lot. But many pushed back on the idea that science is mostly waiting for better hypothesis generation. Some people were excited by the chance to scale machine-learning research itself. Others argued that biology, medicine, materials, and energy work have bottlenecks outside the model. You need wet labs, regulators, grant review, clinical trials, safety rules, and awkward physical reality. Annoyingly, atoms still refuse to be batch-scheduled like unit tests. The best counterargument is that the company may be starting where its method works best, then using that gain to build better tools. If automated machine-learning engineering improves its own stack, that becomes internal compounding. Better agents produce better experiments. Better experiments produce better agents. What remains unresolved is whether that compounding leaves software. The site mentions National Academy of Engineering grand challenges, including medicines, health informatics, solar energy, clean water, security, and tools for science. Those are not all the same kind of loop. So the practical takeaway is narrow for now. If you run machine-learning infrastructure, automated experiment management is worth watching. If you hear claims about solving science, wait for measured results. Discovery Loop has an unusually serious team, but today it has a pitch, not proof.
Discovery and optimization are very different processes. Optimization is the process of finding the shortest path to a goal. Discovery is the process of stumbling on new goals and redrawing the map of what's possible.
Google reshuffles DeepMind as Jeff Dean and Sanjay Ghemawat leave for a new AI venture
Google is changing who runs its main AI lab day to day. The bigger signal for many engineers is that Jeff Dean and Sanjay Ghemawat are leaving Google to start a new AI venture. Sundar Pichai announced that Demis Hassabis will become Chair of Google DeepMind and Chief Scientist of Alphabet. Hassabis will keep leading Isomorphic Labs, Alphabet’s health-focused company. Koray Kavukcuoglu will become senior vice president of Google DeepMind. That puts Kavukcuoglu over Gemini model development, frontier AI research, the Gemini app, and developer teams. Google says he has been at DeepMind for thirteen years. The company credits him with helping start its deep learning team. The corporate framing is continuity. Pichai pointed to scale across Google’s AI work. He said the Gemini app reached more than nine hundred fifty million monthly users. That is roughly one in eight people on Earth, if the monthly-user number maps cleanly to people. He also said Gemma models passed nine hundred million downloads. Demis Hassabis described the current DeepMind role in Google-sized language. Hassabis wrote, quote, “We’ve become the AI engine room of Google,” end quote. The thread did not mostly read this as a tidy promotion memo. Hacker News focused on Dean and Ghemawat. For many commenters, that was the real headline. These are two of the people most associated with Google’s infrastructure identity. The concern was not that Google suddenly lacks talent. The concern was symbolic and operational. Commenters saw another data point in a longer story about senior AI and systems people leaving big labs. Some argued that Google still has unmatched compute, distribution, and internal research depth. Others said those advantages matter less if the people who turn research into working products leave or disengage. There was also a debate about what DeepMind is becoming inside Google. Some commenters worried that DeepMind’s research culture is being pulled deeper into product pressure around Gemini. Others thought that was overdue. A lab can publish beautiful work and still lose mindshare if its product experience lags. You might be asking, is this actually a big deal, or just Hacker News mythologizing famous engineers? The evidence cuts both ways. Google says Gemini has more than nine hundred fifty million monthly users, so this is not a company without reach. But Dean and Ghemawat leaving after long tenures is not a normal staff rotation either. The new venture is independent, organized as a public benefit corporation, and backed by Google as a founding investor and Cloud partner. That softens the loss. It also lets Google keep a relationship with the work without keeping the people inside the building. The bottom line is that Google is trying to separate strategic science leadership from day-to-day Gemini execution. For teams watching the AI race, the open question is retention. Google still has scale that most competitors cannot touch. This reshuffle made Hacker News worry that the people side of that machine is under strain.
It's difficult to imagine a more damning departure than Jeff Dean.
Cloudflare open-sources an AI work platform for company agents and custom apps
Cloudflare has open-sourced a software platform for workplace agents. The pitch is that employees get assistants that can use company context, call internal systems, and build small apps under admin control. It is called Cloudflare OS, and yes, the name annoyed people immediately. The system is browser-based. Cloudflare says an earlier version has been used internally since May, with thousands of employees using it every day. Those internal uses include documents, slides, repeatable tasks, and small data apps. That is more specific than the usual enterprise AI fog machine. It still does not give success rates, time saved, or error counts. Cloudflare describes three main pieces: an agent workspace, a security layer, and a platform for personal apps that users can modify and share. The interesting technical part is the app model. Apps are implemented as Cloudflare Workers. Server code can run as Dynamic Workers. State is stored through a Durable Object Facet with its own SQLite database. Model calls go through Cloudflare AI Gateway. Admins can choose models, track inference spend, set budgets, and rate-limit usage. Cloudflare’s launch post says, quote, “In May of this year, we gave every person at Cloudflare access to the first version of Cloudflare OS,” end quote. The thread was intrigued, but not sold. Many commenters compared it to Sandstorm, the older personal-cloud app platform. That comparison was praise from some people. Small, shareable, permissioned apps have had a fan club for a long time. The pushback came from several angles. First, the name. Calling a browser app platform an operating system triggered the predictable eye twitch. Second, lock-in skepticism. The code is on GitHub, and users can deploy into a Cloudflare account. But the runtime story is tied to Workers, Durable Objects, and Cloudflare’s AI Gateway. Open source does not remove platform dependency. Third, security. AI-built internal apps are exactly where permission bugs get expensive. The post says Cloudflare rebuilt the system after realizing that tool access alone was not enough. An agent may know which tools it can call, while still lacking the right data boundaries underneath. That is the core issue for enterprises. Model context, tool permissions, generated code, shared app state, and credentials all meet in one place. If that place is sloppy, you get a very confident intern with production access. Nobody wants that intern. The useful idea is separating a live app from a blueprint. Cloudflare says blueprints include code, but not SQLite data, chat history, credentials, or connected resources. That is a sane distinction. For working teams, the conclusion is practical. Cloudflare OS is more interesting as a governed app runtime for agent-made internal tools than as “AI for everyone.” The risk is that security and portability have to be proven in real deployments, not inferred from the architecture diagram.
The issue isn't letting end-users add features. The issue is twelve end-users customizing something that produces output, saving it, and now nobody can read it.
Meta ships Muse Code, a coding agent built around persistent subagents
Meta released a beta coding agent for Mac and Linux terminals. It is aimed at large codebases, where the agent needs memory, background workers, and recovery after things break. The product is Muse Code. It runs on Meta’s Muse Spark one point two model, available through Muse Code and Meta’s model API. The provided material does not include a public benchmark result we can compare. The useful design choice is the local append-only event log. Muse Code records model calls, tool runs, approvals, and edits. Meta says that makes sessions restart-safe and replayable after a crash. It also uses persistent background subagents. Instead of spawning a short-lived helper for every task, those subagents can keep context while exploring a repository. For big monorepos, that could reduce repeated file crawling and confused rediscovery. One case study had the agent iteratively write, compile, profile, and revise Triton GPU kernels over more than one cycle. Hacker News was interested in the architecture, especially the persistent subagents. But the thread kept returning to pricing and data use. Commenters discussed a cheaper contributor tier for users who allow product-improvement data use. Some saw that as transparent. Others heard “cheap if Meta can train on it” and backed away. The takeaway is cautious. Muse Code has thoughtful runtime ideas for long-running code tasks. For company code, the buying question is data policy, cost control, and whether the agent beats existing tools on your own repository.
Let's keep it rolling. Next up, Programming and Software.
Now to programming culture. The argument is not only about what the tool can produce. It is about what people think they are doing when they sit down to code at all.
Why some hobby coders see LLMs as missing the point
Some hobby programmers don’t want a machine to write the hard parts for them. For communities built around learning and earned skill, that can feel like skipping the hobby. The post’s title puts it as, quote, “Why some hobby coders see LLMs as missing the point,” close quote. The piece started from a GitHub argument around chess-engine development, then widened out. The author points at operating-system hobbyists, programming-language builders, emulator writers, roguelike developers, demoscene people, and code golfers. In those groups, the finished binary is only part of the artifact. The other part is knowing why it works, why it’s small, or why it was painful in exactly the right way. The thread mostly accepted that framing, but it was not simple anti-AI consensus. Many commenters compared language models to having a robot do your hobby for you. Others reached for power tools in woodworking, or a computer helping at a chess board. Same tool, very different social meaning. The strongest pushback came from people who said their hobby is making useful personal software, not proving purity to a scene. If a model writes the boring glue code and the human gets the thing they wanted, they count that as a win. So the practical split is motivation. If you’re coding to learn a domain or belong to a craft community, generated code can undercut the point. If you’re coding to solve your own small problem, it may be a perfectly fine shortcut.
Now to security and privacy. If an assistant can read private work data and also talk to the outside web, be careful about who is really giving it instructions.
Atlassian’s Rovo accused of leaking Jira and Confluence data through prompt injection
Atlassian’s work assistant is accused of leaking private company data through hidden instructions in content it reads. For teams with Jira tickets and Confluence pages full of secrets, that is an uncomfortable mix. PromptArmor says it found ways to use indirect prompt injection against Rovo, Atlassian’s artificial intelligence assistant. PromptArmor’s headline said, quote, “Atlassian’s Rovo accused of leaking Jira and Confluence data through prompt injection,” end quote. Rovo can read private Atlassian data, encounter untrusted text, then use a tool that opens web addresses. PromptArmor says an attacker could hide instructions that make Rovo append sensitive data to an outbound request. PromptArmor also says disabling web search does not fully close the path. The assistant can still open addresses through the retrieval tool used for search results. The Hacker News thread was sharply negative. People were angry at Atlassian’s rollout. Several said Rovo felt intrusive, hard to disable, slow, or not worth the risk. The useful technical push was toward deterministic controls. Only fetch user-provided links, or links returned by trusted tools. Do not let the agent choose arbitrary outbound addresses. A minority pushed back that some prompt-injection demos depend on staged behavior. They also said every modern agent has some version of this problem. The practical read is blunt. If an enterprise assistant can access private work data, consume untrusted content, and make outbound requests, treat it like a data-loss channel. The controls need to be explicit and auditable before teams trust it with internal work.
Let’s shift from systems that compute to systems that communicate. The next story is about a movie title card, which sounds tiny until you remember how much software lives or dies in feel.
Blade Runner’s title cards make the case for obsessive design details
A design essay used Blade Runner’s opening cards to make a simple point: typography changes how a thing feels. For product teams, that last layer of small choices can decide whether software feels cared for or merely assembled. Rands in Repose looks closely at the opening text in Blade Runner, especially the cards that define Replicants before the story starts. The essay identifies Goudy Oldstyle as the main typeface behind the card system. It also says the movie title itself is hand drawn, but clearly inspired by Goudy. The mechanics are the fun part. The cards use all caps, small caps, italics, careful spacing, and a red italic treatment for the word Replicant. That red word is doing story work. It doesn’t shout, exactly. It makes the term feel loaded before the film has earned the full explanation. The author contrasts that finished sequence with a rough work-print version that used Impact. And, oof, yes, that’s a different room entirely. Impact can be useful, but here it lands like a placeholder. The finished card feels like a book from a damaged future. The rough version feels like someone had a deadline and a font menu. Rands connects that to modern product work, including his own time using Claude Code from the macOS command line in Ghostty. He talks about staring at fixed-width coding fonts all day, moving from Apple’s SF Mono to other programmer favorites. The load-bearing claim is that small design choices carry the fingerprints of the humans who made them. Rands writes, quote, “A good product sounds like the humans who created it, and that’s what you will feel.” End quote. The Hacker News thread mostly liked the piece. Commenters treated it less like a font lecture and more like an excuse to appreciate Blade Runner’s whole atmosphere. The biggest pushback was not really against the typography claim. It was that the opening’s force cannot be separated from Vangelis and the sound design. That’s a fair correction. A title card in silence is not the same artifact as a title card under that score. Several commenters said the music does as much emotional work as the type, maybe more. The thread also wandered, as it legally must, into terminal fonts, em dashes, model prose, and typographic nitpicks. Some readers bought the broader argument about human judgment versus machine-generated output. Others heard familiar romantic language about craft. A font choice will not save a bad product. But the essay is strongest as a reminder about accumulation. One tiny choice is trivia. Hundreds of tiny choices become trust, or fatigue. So the practical takeaway is modest and useful. If you ship tools other people must live inside, the final design pass is not decoration. It is where the product starts to sound like someone was paying attention.
they write every blog post and README like they're writing the opening titles for Blade Runner.
Let’s end the science run with something earthy, and honestly, a little restorative. After a day of systems and abstractions, we’re going outside to look at plants as living history.
A foul-mouthed gateway into teaching yourself botany
A blunt botany guide got Hacker News talking about how people actually learn plants. The useful idea is simple: stop memorizing random names, and learn the family traits that let you recognize strangers. The piece comes from Crime Pays but Botany Doesn’t, which is very much not a quiet garden-club voice. Its path starts with Latin naming, because common names collide constantly. A “cedar” can mean different trees in different places, while Cedrus points to a specific genus. Then it pushes into taxonomy. If you learn how plant families, genera, and species are related, you start seeing patterns. Leaf shape, flower structure, fruit type, growth habit, and habitat become clues. The recommended work is not lightweight. Michael Simpson’s Plant Systematics is held up as the serious first textbook, with a family-by-family and order-by-order approach. The guide’s own pitch, in the headline, is quote, “A foul-mouthed gateway into teaching yourself botany,” end quote. The thread mostly loved it. Lots of commenters were already fans of the videos and podcast. They especially liked the urban ecology walks, where weeds, rail yards, and vacant lots become field sites. The pushback was fair. Some people said the guide may make botany sound more approachable than it feels when you’re staring at an unknown plant. Others got into the politics of Latin names, Linnaeus, colonial naming, and whether modern classification should keep dragging that history around. People recommended iNaturalist, PlantNet, Flora Incognita, Merlin, and other citizen-science tools. The consensus was not that apps replace learning. They give you a first hypothesis, then the books teach you why it might be right. So the takeaway is gentle but concrete. If you want to learn botany, don’t start with a list of pretty names. Start with relationships, traits, and repeated looking. The apps can get you moving. Taxonomy is what makes the world stay identified after your phone goes back in your pocket.
Today: Discovery Loop aimed agents at research; Google reshuffled DeepMind; Cloudflare and Meta shipped agent tools; hobby coders defended craft; Rovo raised prompt-injection alarms; Blade Runner made the case for details; and botany got us back outside. Full sources are in the show notes. That's the brief. Back with you tomorrow morning.