标签：#ai

Nearly half a century ago, the free software movement made a demand that was both technical and moral: Users should have the freedom to understand, run, modify, and share the software on which they depended. It was a demand born from practical life with machines. A printer that couldn't be fixed. A program that couldn't be studied. A system that asked its users to accept dependence as the price of progress. That belief shaped modern computing and gave us the tools and norms that made the internet, open infrastructure, and collaborative software development possible. Today, that belief faces its hardest test. The technology has changed, but the warning signs are familiar. In 1980, at the MIT AI Laboratory in Cambridge, Massachusetts, a new Xerox 9700 printer was installed. The previous printer had come with source code that could be modified, inspected, recompiled, and reinstalled. Richard Stallman had changed that software to message users when their print job was done or when there was a jam, a small but meaningful feature since the printer sat several floors away. The new printer arrived with software preloaded and installed, no source code available, no way to modify it. If you needed help or new features, you hoped and prayed Xerox would listen. That loss of agency, alongside other anti-consumer shifts in early software, helped push him toward GNU and the free software movement: the belief that software should be free as in freedom, free to inspect, run, study, modify, understand, and redistribute. AI and inference services today are not too dissimilar. Closed frontier intelligence can make entire companies, governments, developers, and communities dependent on systems they cannot inspect, reproduce, modify, or meaningfully contest. At the dawn of this AI moment, we were promised unfettered intelligence across our products, companies, and codebases. We were told we'd be free to build whatever we wanted. At first, with tab completions. Then whole function blocks.

by Peter Yang, Behind the Craft Today, I want to share 18 hot takes on where I think the AI market is headed. AI is in a weird place right now. The government is restricting access to frontier models, enterprises are becoming conscious of token costs, and everyone’s trying to rebuild their product for agents first instead of humans. I’ve interviewed dozens of AI leaders and spent far too much time following these topics on X/Twitter. Here are 18 hot takes on where I think AI is headed next: The frontier-only AI stack is collapsing The AI super app era is here Traditional software risks becoming a dumb pipe for agents Cloud agents and collaboration are the next wave The Frontier-Only AI Stack Is Collapsing Tokenmaxxing at frontier API prices makes no sense. Uber burned through its entire 2026 AI budget in 4 months, Microsoft moved engineers off Claude Code due to cost, and companies are realizing that running everything on frontier models can get expensive fast. Tokenmaxxing makes sense when you’re on a subsidized $200/month plan but is unsustainable at API rates. Companies will rely on a portfolio of models. Coinbase recently cut its AI spend nearly in half by switching engineers to Chinese open-source models like GLM and Kimi. Airbnb and Pinterest have done the same with Alibaba’s Qwen models. I believe that this will be the default path forward — using frontier for high-stakes work and cheaper models for everything else. China’s open-source strategy is working. Chinese models are taking market share from frontier models at US companies. China is also building the full AI stack — from energy (e.g., solar, nuclear) to data centers to domestic chips. The Chinese government is planning a $295B investment in AI data centers with at least 80% of the chips built domestically. Frontier labs are in a catch-22 situation. If they release great open-source models, they might undercut their own frontier API revenue. If they gate the best models behind a trusted list, companies

What happens when a machine no longer needs to be trained to see something new? That's the quiet question sitting underneath this week's news, buried next to a less invasive brain implant and a handful of robots getting tougher for the real world. Neuralink says it's completed its first "transdural" brain implant, a surgical approach built to reduce trauma during the procedure. As someone who spends a lot of time thinking about how you get sensors close to a human eye without hurting anyone, I find these less-invasive-implant strategies worth watching, because the surgical-risk problem is basically the same one we wrestle with in ophthalmic hardware. Vision is getting less invasive too, in its own way. Roboflow rolled out text-prompt object detection built on SAM3 (Meta's latest segmentation model): you type the class of object you want "forklift," "cracked tile," whatever, and it returns boxes and masks without you collecting a single training image first. That's a real shift. For most of computer vision's history, teaching a model to recognize something new meant labeling hundreds of examples before you could even start; this collapses that step into a sentence. The same week brought several applied builds using the same detect-then-orchestrate pattern: a drone system that patrols for intrusions, a pipeline that inspects transmission lines for damaged cables, and an airport tool that spots foreign debris on the tarmac. The Robot Report's roundup of June's biggest robotics stories leaned heavily on humanoid robots companies going public, new deployments, and production milestones stacking up faster than would have seemed plausible a few years ago. Apptronik unveiled its Apollo 2 humanoid alongside a dedicated data-collection facility built so the robot keeps learning after it's deployed, not just during initial training which quietly answers one of the harder questions in robotics: how do you keep a system improving once it's out of the lab? X Square Robot raised e

For a few days, it seemed like Universal decided that there would be no advanced screenings of Christopher Nolan's The Odyssey for influencers. But on Monday, influencers sat alongside traditional critics and journalists at special showings of The Odyssey specifically for the associated press junket. Despite what it may have looked like, Universal was not […]

This tutorial was written by Néstor Daza . This is the second article in a series about building Claudius , my own Claude-based chatbot ( Github ). The prologue made the case for building it, and for choosing MongoDB as its foundation. Open the conversations collection in Claudius’ database and you find the usual fields of a thread header but nothing else: a userId , a title , some timestamps , and so on, but no array of messages, no messages collection sitting beside it either! The text of every conversation lives somewhere else entirely, in the LangGraph checkpointer, which I wire up later in this series. This absence is a modeling decision, and how I came up with the database schema for my chatbot is the theme of this article. If you come from a relational background, you're used to modeling the data first when designing a database. For a project like this, you would start by finding the entities and normalizing them, and the final schema would come out of the data's structure: a conversations table and a messages table with a foreign key between them, because that is what the data looks like. Document modeling runs the other way. You start from how the application reads and writes, and the shape of the document follows the access patterns. Claudius never reads conversation messages without the agent's full working state wrapped around them, and that state is persisted using the LangGraph checkpointer. A separate messages table would add nothing, since the app would always have to join it back to that state on every read. The access pattern says the messages belong with the agent state, so that is where they go, and conversations are left as the lightweight header the list view actually needs. That inversion, modeling around use rather than around the data, runs through everything below. Schema-flexible is not schemaless This is the misconception lots of people often carry, and it is worth killing on the way in. A document database does not mean no schema; it mea

By Hiral Shah, Senior Director, Product Management, Docusign A major recurring theme among the engineering teams at this week’s AI Engineer World’s Fair in San Francisco is the push to move specialized AI models out of research and directly into high-volume production. At Docusign, that optimization challenge happens at massive scale: we handle millions of transactions daily and have nearly 1.9 million customers in over 180 countries. Organizations have historically lost significant value every year to the friction, delays, and missed obligations that come from treating these agreements as static documents rather than live sources of business data. Much of that trapped value sits inside tables: the pricing schedules, SLA obligations, and contractor rate cards that define enterprise relationships but are often the hardest part of a contract to extract accurately. To solve this, we integrated NVIDIA Nemotron Parse , a vision-language model purpose-built for document understanding, directly into our document processing pipeline. Docusign and NVIDIA took the AI Engineer World’s Fair stage this week to give attendees a look at how the architecture works under the hood. Here’s what that looks like: Why Contract Tables Break General-Purpose AI Contracts routinely contain merged cells, multi-page structures, mixed formatting, and nested layouts that general-purpose vision language models (VLMs) and broad AI models weren't designed to handle. The result is inaccurate extractions that require manual correction, slowing down the workflows they are intended to accelerate. Our teams watch this operational friction play out across real enterprise scenarios every day: System Downtime: When a critical system goes down, operations teams need to know immediately which SLA notification requirements apply and to whom. Resource Tracking: When business stakeholders ask legal what hourly rate was agreed to in a contractor engagement, the answer is often buried deep inside a rate card tabl

by Sergey Karayev, cofounder @ Superconductor Recently, my wife and I sat down to plan an upcoming trip. Naturally, we each asked an AI. Trouble was, I had my chat and she had hers, and they knew nothing about each other. So we served as couriers between chatbots: her idea pasted into my chat, my hotel booking screenshotted into hers, the itinerary reconciled by hand in a Google Doc. I bring this up because your team probably works the same unfortunate way: each person in their own chat or coding agent session, with precious little shared. I've been building software with the same set of people for over a decade. In the past year, we all got a superpower: coding agents that can do extremely impressive things. But each one (Claude Code, Codex, Cursor, etc.) was built for a single player. That's fine and dandy if you're vibe-coding your own little app. It's just you and Claude, and it's absolutely magical. But put that same agent on a team and the magic fades quite a bit. The model is no longer the bottleneck. Coordination is. You don't know who's working on what. You can't see that an agent already tried the approach you're about to attempt, and abandoned it. You spend an hour re-deriving context that a teammate has, because it's trapped in their private chat. Now let me tell you of a better way. On the Superconductor team, every coding agent session is in the cloud, open to anyone else on the team to join. What this enabled was transformative. Code review improved first. My teammate reviews my work by joining the session I built it in. The session holds the full history of decisions, including the dead ends. Instead of Slacking me "why'd you name it this way?" she asks the agent. She gets her answer, and I never waste time answering. She also doesn't have to check out the branch locally — the live app preview in the cloud sandbox does the job. Handoffs became easy. If I have to pass a feature to a teammate, he picks it up with full context: what's done, what's left,

When the enemy is in distress, exploit the opportunity to seize advantage. — The 36 Stratagems, Loot a Burning House Who's Leo — In the last story , he was CoreStack's backend lead — the guy who built the core system alone over five years with zero P0 incidents. Then a new CTO named James showed up, spent $8M on his old employer's product, and laid off Leo's entire team. Thirteen days later, that $8M AI system collapsed — three agents fighting over context, OOM taking down six GPU servers, a 37% order duplication rate, and 2,300 customer complaints. Leo pulled the old system off his laptop, flipped one line of Nginx config, and restored service in thirty seconds. The CEO called him at 3 AM begging him to come back. He came back. Three conditions: kill the paid AI product, AI assists only — never touches the primary pipeline — and engineers decide the architecture, not the guy writing checks. The CEO agreed to all of it. So who's Leo now: CoreStack's CTO. Technically confident to the point of arrogance. Zero talent for upward management. No idea how many people he pissed off on the board with those conditions. Doesn't care. He only knows one thing — the system he built is still running. That's all the proof he needs. Then a Slack message cut him off. The Signal 12:47 AM. CoreStack's CTO gets a Slack notification. The account has no profile picture, no display name, no status. Account creation timestamp at the bottom — 00:43. Four minutes old. Seven characters: Check CodeForge's status page. Leo taps it open. CodeForge's status page is all red. Payment Routing — Major Outage. Investigating. All customers affected. Status has been active for approximately 3 hours. He pulls up CoreStack's CRM. The sales team's prospect list has ShopStream at #2 — a potential whale, with "Current Provider" reading CodeForge. E-commerce platform doing 470,000 transactions a day . An hour of downtime costs them $210,000 . If this drags on until morning? He doesn't want to do the math. Core

OpenAI CEO Sam Altman has reportedly proposed giving 5% of the company’s equity to a U.S. sovereign wealth fund, reviving discussions about letting the public share in the financial gains from the AI boom.

TV Time, the popular TV-tracking app, is shutting down on July 15 as parent company Whip Media pivots toward enterprise AI products.

Insiders say Sam Altman is in active talks with the Trump administration.

FTC urged to reject Elon Musk’s bid to end X monitoring amid AI concerns.

We are so excited to finally announce the winners of the GitHub Finish-Up-A-Thon Challenge, our...

Microsoft follows Amazon, OpenAI, and Anthropic with its new AI deployment group.

HDMI excels for media consumption, while PC gamers prefer DisplayPort.

Originally published on productize.life . Quick answer: pm-skills is a marketplace of around 68 Claude skills for product management across 9 plugins, from strategy and discovery to market research and AI shipping. It is built by Pawel Huryn, author of the Product Compass newsletter. Each skill is not a loose prompt but a named, sourced framework, and one of them audits the gap between documentation and code, a PM lens built for the era of AI-written code. Last week I was reading through a run of repos that pack product work into skills. Some pick one topic and go deep. This one does the opposite: it is the broadest of the bunch. It is called pm-skills, by Pawel Huryn, the author of the Product Compass newsletter. He packs almost the entire product management craft into around 68 skills across 9 plugins, from setting strategy, running discovery, and researching the market, to analyzing data, executing, and shipping software that AI wrote. Usually something this broad ends up shallow. But when I actually opened it, it was not, and one skill in particular made me stop and look for a while , because it covers an angle that only recently became necessary in the era where AI writes code for us. I will tell it in three parts, starting with what it is , then why it is not just a prompt box , and closing with lessons for anyone building products . Terms, gathered once, right here skill a ready-made set of instructions an AI agent (such as Claude Code) can invoke, like a shortcut that wraps one way of doing a task. framework a ready-made way of thinking from the PM world, such as SWOT, JTBD, or RICE, that you once had to read a book to use well. plugin (category) a group of skills that belong to the same topic, such as the discovery category or the go-to-market category. PRD a product spec document that says what will be built, for whom, and how success is measured. Part 1: What pm-skills is It is a marketplace of around 68 Claude skills for PM, organized into 9 plugins, eac

Originally published on productize.life . Quick answer: gstack is an open-source (MIT) skill set that Garry Tan, president of Y Combinator, builds with every day. It turns Claude Code into a team of 23 specialists, CEO, engineers, designers, QA, and a release engineer, forcing every change through a multi-lens review before shipping. The point is not speed; it is taste written into software. Last week I was going through a repo that collects skills for coding, several of them. Most share one theme: helping AI write code in a systematic way, and faster. But one made me stop longer than the rest, called gstack, for two reasons. One: its owner, Garry Tan, president and CEO of Y Combinator, took the stack he actually builds with every day and opened it for free. Two: it does not sell "code faster," it sells "review before you ship." Once I actually opened it, it was not just a toolbox but one of the clearest examples of an idea I have been interested in for a while. On the day AI can write code very fast, the bottleneck of the work is no longer speed. I will tell it in three parts, starting with what it is , then what gstack believes , and closing with lessons for people who build products, not just people who write code . Terms, gathered here in one place agentic coding letting an AI agent run the coding work in its own steps, from planning to writing to review to shipping, not just autocompleting a line at a time. skill a packaged set of instructions an AI agent (like Claude Code) can call, like a shortcut that wraps one way of doing one thing. review lens reviewing one piece of work from several roles, for example as a CEO, an engineer, a designer. taste the sense and judgment of what is good and what is bad, what to build and what not to ship. The part that is still human. Part 1: What gstack is Garry Tan describes gstack in the README plainly, as the way he works. "It turns Claude Code into a virtual engineering team: a CEO who rethinks the product, an eng manager

If you're anything like me, your Chrome toolbar probably looks like a collection of tiny puzzle pieces. Grammar checker. Screenshot tool. Summarizer. Writing assistant. Code explainer. Translator. Email helper. At one point I had more than a dozen extensions installed. Chrome became slower, pages loaded later, and every extension wanted permission to "read and change all your data." Then I started experimenting with AI tools instead. Not everything was better—but some things surprised me. Here's what I learned after replacing most of my browser extensions with AI. 1. Grammar Checkers I used to rely on grammar extensions that constantly underlined my writing. Now I simply paste my draft into an AI assistant and ask: Improve grammar while keeping my writing style. The biggest advantage isn't fixing mistakes—it's preserving tone. Traditional grammar tools often make everything sound the same. AI can make your writing cleaner without removing your personality. 2. Article Summarizers This was probably the easiest replacement. Instead of installing a summarizer extension, I paste the article and ask: Summarize in 5 bullet points Give me the key takeaways Explain it like I'm a beginner What important details are missing? The last prompt is especially useful because summaries sometimes leave out important context. 3. Code Explanation This has become one of my favorite AI use cases. Instead of searching Stack Overflow for every unfamiliar function, I simply paste the code and ask: Explain this line by line Why was this approach chosen? Is there a better alternative? What's the time complexity? The answers aren't always perfect, but they're often enough to understand what's happening before diving into documentation. 4. Writing Commit Messages This is something I didn't expect AI to help with. Instead of writing: fixed stuff I can paste my git diff and ask for a concise commit message. Example: feat: add JWT authentication middleware fix: resolve login redirect loop refactor:

You built a system to surface information so a person could decide. Somewhere it started deciding for them — the output stopped saying "here's what the documents show" and started saying "you should do X." Nobody designed that drift. An LLM, when asked a question, produces an answer-shaped thing, and an answer easily becomes a verdict. What everyone tries A prompt instruction: "Don't make recommendations." "Only state what's in the documents." People add the line and assume the boundary is enforced. Why it doesn't work A prompt instruction is a request, not a guardrail. The model follows it most of the time, then on the input that matters produces a confident recommendation anyway, because nothing structurally prevents it. "Please don't make recommendations" is to a guardrail what a sticky note saying "please don't enter" is to a locked door. And the stakes are higher than they look. When output drifts from evidence to verdict, accountability moves. As long as the system returns evidence and a human decides, the human owns the decision. The moment the system returns a verdict and the human defers, the system is deciding things it was never validated to decide — and when one is wrong, accountability is a blank. High-stakes fields separate evidence extraction from judgment on purpose; most RAG systems erase that line by default. The one shift Decide what the output is and enforce it structurally. An output should declare itself: answer, evidence, missing facts, or out-of-scope. "Return decision material, not a decision" has to live in the output contract and in gates — not in a polite request to the model. The system supplies frames; the human supplies verdicts. This is the output boundary — one of three places production RAG dies. Read the full version on my blog , where this connects to the RAG Failure Diagnosis Kit for teams debugging production RAG.

👋 Hey dev.to community, Last week I shipped a workflow I built for a friend who runs an online shop. He called me again a few days later with a new headache: he's drowning in Purchase Orders. Every single one gets opened by hand, the data typed into a Google Sheet, and that sheet uploaded into his ERP to update his numbers. Hours a week, pure copy-paste. So I built him something to kill that step. He uploads the PO PDFs through a simple n8n form, and a structured Google Sheet comes out the other end. He just downloads it and pushes it to his ERP. How it's set up: The form accepts multiple PDFs at once , so he can batch a whole stack instead of doing them one by one. Each PO loops through on its own so nothing gets jumbled. The extraction runs on the easybits Extractor node ( @easybits/n8n-nodes-extractor ). I set the field structure up in two parts: the header fields that appear once per PO (PO number, PO date, delivery date, mark for, PR number, reference no), plus an articles array for the line items, each holding article name, unit and quantity. That array is the key bit, it gives you one entry per row of the PO table, and I flatten it into one sheet row per article with the header details repeated on each. Two things I added because real documents are messy: Error flagging . If any field comes back empty, the completion screen lists which document and which field didn't extract cleanly, so he knows exactly which PO to double-check instead of trusting it blindly. Document name column . The original filename lands in the sheet next to every row, so if a number looks off he can jump straight back to the source PDF. Workflow JSON is on GitHub: https://github.com/felix-sattler-easybits/n8n-workflows/blob/c38749a68fd6ea4ae6ebff41789d35cceaacdef1/easybits-purchase-order-extractor-workflow/easybits_purchase_order_extractor_workflow.json Anyone else automating document-to-sheet data entry? Curious how you're handling the messy multi-line rows – that was the trickiest par